draft, wip, but that's the idea

* Squash for refactor: Replace monolithic cache classes with modular LayeredCache (#38077)

- Introduces CacheLayer and Cache base classes
- Ports Static, Dynamic, Offloaded, Quantized, Hybrid, etc. to use layers
- Implements method/attr dispatch across layers to reduce boilerplate
- Adds CacheProcessor hooks for offloading, quantization, etc.
- Updates and passes tests

* fix quantized, add tests

* remove CacheProcessorList

* raushan review, arthur review

* joao review: minor things

* remove cache configs, make CacheLayer a mixin (joaos review)

* back to storage inside Cache()

* remove cachebase for decorator

* no more __getattr__

* fix tests

* joaos review except docs

* fix ast deprecations for python 3.14: replace node.n by node.value and use `ast.Constant`

More verbose exceptions in `fix_docstring` on docstring formatting issues.

* Revert "back to storage inside Cache()"

This reverts commit 27916bc2737806bf849ce2148cb1e66d59573913.

* cyril review

* simplify cache export

* fix lfm2 cache

* HybridChunked to layer

* BC proxy object for cache.key_cache[i]=...

* reorder classes

* bfff come on LFM2

* better tests for hybrid and hybridChunked

* complete coverage for hybrid chunked caches (prefill chunking)

* reimplementing HybridChunked

* cyril review

* fix ci

* docs for cache refactor

* docs

* oopsie

* oopsie

* fix after merge

* cyril review

* arthur review

* opsie

* fix lfm2

* opsie2

.circleci/create_circleci_config.py

@@ -303,7 +303,7 @@ non_model_job = CircleCIJob(
     docker_image=[{"image": "huggingface/transformers-torch-light"}],
     # networkx==3.3 (after #36957) cause some issues
     # TODO: remove this once it works directly
-    install_steps=["uv venv && uv pip install ."],
+    install_steps=["uv venv && uv pip install .[serving]"],
     marker="not generate",
     parallelism=6,
 )

.github/workflows/build_documentation.yml

@@ -18,6 +18,10 @@ jobs:
       notebook_folder: transformers_doc
       languages: ar de en es fr hi it ko pt tr zh ja te
       custom_container: huggingface/transformers-doc-builder
+      # Temporary pin to work around datasets exception in the docbuilder.Remove after docker images and main have
+      # the right dependencies (which **should** be the case by 2025-07-20). See
+      # https://github.com/huggingface/transformers/actions/runs/16365952006/job/46243081358?pr=38545
+      pre_command: uv pip install datasets>=2.15.0
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -15,3 +15,7 @@ jobs:
       pr_number: ${{ github.event.number }}
       package: transformers
       languages: en
+      # Temporary pin to work around datasets exception in the docbuilder. Remove after docker images and main have
+      # the right dependencies (which **should** be the case by 2025-07-20). See
+      # https://github.com/huggingface/transformers/actions/runs/16365952006/job/46243081358?pr=38545
+      pre_command: uv pip install datasets>=2.15.0

.github/workflows/get-pr-info.yml

@@ -0,0 +1,157 @@
+name: Get PR commit SHA
+on:
+  workflow_call:
+    inputs:
+      pr_number:
+        required: true
+        type: string
+    outputs:
+      PR_HEAD_REPO_FULL_NAME:
+        description: "The full name of the repository from which the pull request is created"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REPO_FULL_NAME }}
+      PR_BASE_REPO_FULL_NAME:
+        description: "The full name of the repository to which the pull request is created"
+        value: ${{ jobs.get-pr-info.outputs.PR_BASE_REPO_FULL_NAME }}
+      PR_HEAD_REPO_OWNER:
+        description: "The owner of the repository from which the pull request is created"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}
+      PR_BASE_REPO_OWNER:
+        description: "The owner of the repository to which the pull request is created"
+        value: ${{ jobs.get-pr-info.outputs.PR_BASE_REPO_OWNER }}
+      PR_HEAD_REPO_NAME:
+        description: "The name of the repository from which the pull request is created"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}
+      PR_BASE_REPO_NAME:
+        description: "The name of the repository to which the pull request is created"
+        value: ${{ jobs.get-pr-info.outputs.PR_BASE_REPO_NAME }}
+      PR_HEAD_REF:
+        description: "The branch name of the pull request in the head repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REF }}
+      PR_BASE_REF:
+        description: "The branch name in the base repository (to merge into)"
+        value: ${{ jobs.get-pr-info.outputs.PR_BASE_REF }}
+      PR_HEAD_SHA:
+        description: "The head sha of the pull request branch in the head repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_SHA }}
+      PR_BASE_SHA:
+        description: "The head sha of the target branch in the base repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_BASE_SHA }}
+      PR_MERGE_COMMIT_SHA:
+        description: "The sha of the merge commit for the pull request (created by GitHub) in the base repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_SHA }}
+      PR_HEAD_COMMIT_DATE:
+        description: "The date of the head sha of the pull request branch in the head repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_COMMIT_DATE }}
+      PR_MERGE_COMMIT_DATE:
+        description: "The date of the merge commit for the pull request (created by GitHub) in the base repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_DATE }}
+      PR_HEAD_COMMIT_TIMESTAMP:
+        description: "The timestamp of the head sha of the pull request branch in the head repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_HEAD_COMMIT_TIMESTAMP }}
+      PR_MERGE_COMMIT_TIMESTAMP:
+        description: "The timestamp of the merge commit for the pull request (created by GitHub) in the base repository"
+        value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_TIMESTAMP }}
+      PR:
+        description: "The PR"
+        value: ${{ jobs.get-pr-info.outputs.PR }}
+      PR_FILES:
+        description: "The files touched in the PR"
+        value: ${{ jobs.get-pr-info.outputs.PR_FILES }}
+
+
+jobs:
+  get-pr-info:
+    runs-on: ubuntu-22.04
+    name: Get PR commit SHA better
+    outputs:
+      PR_HEAD_REPO_FULL_NAME: ${{ steps.pr_info.outputs.head_repo_full_name }}
+      PR_BASE_REPO_FULL_NAME: ${{ steps.pr_info.outputs.base_repo_full_name }}
+      PR_HEAD_REPO_OWNER: ${{ steps.pr_info.outputs.head_repo_owner }}
+      PR_BASE_REPO_OWNER: ${{ steps.pr_info.outputs.base_repo_owner }}
+      PR_HEAD_REPO_NAME: ${{ steps.pr_info.outputs.head_repo_name }}
+      PR_BASE_REPO_NAME: ${{ steps.pr_info.outputs.base_repo_name }}
+      PR_HEAD_REF: ${{ steps.pr_info.outputs.head_ref }}
+      PR_BASE_REF: ${{ steps.pr_info.outputs.base_ref }}
+      PR_HEAD_SHA: ${{ steps.pr_info.outputs.head_sha }}
+      PR_BASE_SHA: ${{ steps.pr_info.outputs.base_sha }}
+      PR_MERGE_COMMIT_SHA: ${{ steps.pr_info.outputs.merge_commit_sha }}
+      PR_HEAD_COMMIT_DATE: ${{ steps.pr_info.outputs.head_commit_date }}
+      PR_MERGE_COMMIT_DATE: ${{ steps.pr_info.outputs.merge_commit_date }}
+      PR_HEAD_COMMIT_TIMESTAMP: ${{ steps.get_timestamps.outputs.head_commit_timestamp }}
+      PR_MERGE_COMMIT_TIMESTAMP: ${{ steps.get_timestamps.outputs.merge_commit_timestamp }}
+      PR: ${{ steps.pr_info.outputs.pr }}
+      PR_FILES: ${{ steps.pr_info.outputs.files }}
+    if: ${{ inputs.pr_number != '' }}
+    steps:
+      - name: Extract PR details
+        id: pr_info
+        uses: actions/github-script@v6
+        with:
+          script: |            
+            const { data: pr } = await github.rest.pulls.get({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              pull_number: ${{ inputs.pr_number }}
+            });
+
+            const { data: head_commit }  = await github.rest.repos.getCommit({
+              owner: pr.head.repo.owner.login,
+              repo: pr.head.repo.name,
+              ref: pr.head.ref
+            });
+
+            const { data: merge_commit }  = await github.rest.repos.getCommit({
+              owner: pr.base.repo.owner.login,
+              repo: pr.base.repo.name,
+              ref: pr.merge_commit_sha,
+            });
+
+            const { data: files } = await github.rest.pulls.listFiles({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              pull_number: ${{ inputs.pr_number }}
+            });
+
+            core.setOutput('head_repo_full_name', pr.head.repo.full_name);
+            core.setOutput('base_repo_full_name', pr.base.repo.full_name);
+            core.setOutput('head_repo_owner', pr.head.repo.owner.login);
+            core.setOutput('base_repo_owner', pr.base.repo.owner.login);
+            core.setOutput('head_repo_name', pr.head.repo.name);
+            core.setOutput('base_repo_name', pr.base.repo.name);
+            core.setOutput('head_ref', pr.head.ref);
+            core.setOutput('base_ref', pr.base.ref);
+            core.setOutput('head_sha', pr.head.sha);
+            core.setOutput('base_sha', pr.base.sha);
+            core.setOutput('merge_commit_sha', pr.merge_commit_sha);
+            core.setOutput('pr', pr);
+
+            core.setOutput('head_commit_date', head_commit.commit.committer.date);
+            core.setOutput('merge_commit_date', merge_commit.commit.committer.date);
+            
+            core.setOutput('files', files);            
+            
+            console.log('PR head commit:', {
+              head_commit: head_commit,
+              commit: head_commit.commit,
+              date: head_commit.commit.committer.date
+            });
+
+            console.log('PR merge commit:', {
+              merge_commit: merge_commit,
+              commit: merge_commit.commit,
+              date: merge_commit.commit.committer.date
+            });
+
+      - name: Convert dates to timestamps
+        id: get_timestamps
+        run: |
+          head_commit_date=${{ steps.pr_info.outputs.head_commit_date }}
+          merge_commit_date=${{ steps.pr_info.outputs.merge_commit_date }}
+          echo $head_commit_date
+          echo $merge_commit_date
+          head_commit_timestamp=$(date -d "$head_commit_date" +%s)
+          merge_commit_timestamp=$(date -d "$merge_commit_date" +%s)
+          echo $head_commit_timestamp
+          echo $merge_commit_timestamp
+          echo "head_commit_timestamp=$head_commit_timestamp" >> $GITHUB_OUTPUT
+          echo "merge_commit_timestamp=$merge_commit_timestamp" >> $GITHUB_OUTPUT

.github/workflows/get-pr-number.yml

@@ -0,0 +1,36 @@
+name: Get PR number
+on:
+  workflow_call:
+    outputs:
+      PR_NUMBER:
+        description: "The extracted PR number"
+        value: ${{ jobs.get-pr-number.outputs.PR_NUMBER }}
+
+jobs:
+  get-pr-number:
+    runs-on: ubuntu-22.04
+    name: Get PR number
+    outputs:
+      PR_NUMBER: ${{ steps.set_pr_number.outputs.PR_NUMBER }}
+    steps:
+      - name: Get PR number
+        shell: bash
+        run: |
+          if [[ "${{ github.event.issue.number }}" != "" && "${{ github.event.issue.pull_request }}" != "" ]]; then
+            echo "PR_NUMBER=${{ github.event.issue.number }}" >> $GITHUB_ENV
+          elif [[ "${{ github.event.pull_request.number }}" != "" ]]; then
+            echo "PR_NUMBER=${{ github.event.pull_request.number }}" >> $GITHUB_ENV
+          elif [[ "${{ github.event.pull_request }}" != "" ]]; then
+            echo "PR_NUMBER=${{ github.event.number }}" >> $GITHUB_ENV
+          else
+            echo "PR_NUMBER=" >> $GITHUB_ENV
+          fi
+
+      - name: Check PR number
+        shell: bash
+        run: |
+          echo "${{ env.PR_NUMBER }}"
+
+      - name: Set PR number
+        id: set_pr_number
+        run: echo "PR_NUMBER=${{ env.PR_NUMBER }}" >> "$GITHUB_OUTPUT"

.github/workflows/model_jobs.yml

@@ -97,14 +97,6 @@ jobs:
         run: |
           python3 utils/print_env.py
 
-      - name: Install datasets main
-        working-directory: /transformers
-        run: python3 -m pip install --no-cache-dir git+https://github.com/huggingface/datasets.git@main
-
-      - name: Install torchcodec
-        working-directory: /transformers
-        run: python3 -m pip install --no-cache-dir torch torchvision torchaudio torchcodec --index-url https://download.pytorch.org/whl/cu126
-
       - name: Show installed libraries and their versions
         working-directory: /transformers
         run: pip freeze

.github/workflows/pr_run_slow_ci.yml

@@ -0,0 +1,199 @@
+name: PR slow CI
+on:
+  pull_request_target:
+    types: [opened, synchronize, reopened]
+
+jobs:
+  get-pr-number:
+    name: Get PR number
+    uses: ./.github/workflows/get-pr-number.yml
+
+  get-pr-info:
+    name: Get PR commit SHA
+    needs: get-pr-number
+    if: ${{ needs.get-pr-number.outputs.PR_NUMBER != ''}}
+    uses: ./.github/workflows/get-pr-info.yml
+    with:
+      pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
+
+  # We only need to verify the timestamp if the workflow is triggered by `issue_comment`.
+  verity_pr_commit:
+    name: Verity PR commit corresponds to a specific event by comparing timestamps
+    if: ${{ github.event.comment.created_at != '' }}
+    runs-on: ubuntu-22.04
+    needs: get-pr-info
+    env:
+      COMMENT_DATE: ${{ github.event.comment.created_at }}
+      PR_MERGE_COMMIT_DATE: ${{ needs.get-pr-info.outputs.PR_MERGE_COMMIT_DATE }}
+      PR_MERGE_COMMIT_TIMESTAMP: ${{ needs.get-pr-info.outputs.PR_MERGE_COMMIT_TIMESTAMP }}
+    steps:
+      - run: |
+          COMMENT_TIMESTAMP=$(date -d "${COMMENT_DATE}" +"%s")
+          echo "COMMENT_DATE: $COMMENT_DATE"
+          echo "PR_MERGE_COMMIT_DATE: $PR_MERGE_COMMIT_DATE"
+          echo "COMMENT_TIMESTAMP: $COMMENT_TIMESTAMP"
+          echo "PR_MERGE_COMMIT_TIMESTAMP: $PR_MERGE_COMMIT_TIMESTAMP"
+          if [ $COMMENT_TIMESTAMP -le $PR_MERGE_COMMIT_TIMESTAMP ]; then
+            echo "Last commit on the pull request is newer than the issue comment triggering this run! Abort!";
+            exit -1;
+          fi
+
+  get-jobs:
+    name: Get test files to run
+    runs-on: ubuntu-22.04
+    needs: [get-pr-number, get-pr-info]
+    outputs:
+      jobs: ${{ steps.get_jobs.outputs.jobs_to_run }}
+    steps:
+      - name: Get repository content
+        id: repo_content
+        uses: actions/github-script@v6
+        with:
+          script: |
+            const { data: tests_dir } = await github.rest.repos.getContent({
+              owner: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}',
+              repo: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}',
+              path: 'tests',
+              ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
+            });
+
+            const { data: tests_models_dir } = await github.rest.repos.getContent({
+              owner: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}',
+              repo: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}',
+              path: 'tests/models',
+              ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
+            });
+
+            const { data: tests_quantization_dir } = await github.rest.repos.getContent({
+              owner: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}',
+              repo: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}',
+              path: 'tests/quantization',
+              ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
+            });
+
+            core.setOutput('tests_dir', tests_dir);
+            core.setOutput('tests_models_dir', tests_models_dir);
+            core.setOutput('tests_quantization_dir', tests_quantization_dir);
+
+      # This checkout to the main branch
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: "0"
+
+      - name: Write pr_files file
+        run: |
+          cat > pr_files.txt << 'EOF'
+          ${{ needs.get-pr-info.outputs.PR_FILES }}
+          EOF
+
+      - name: Write tests_dir file
+        run: |
+          cat > tests_dir.txt << 'EOF'
+          ${{ steps.repo_content.outputs.tests_dir }}
+          EOF
+
+      - name: Write tests_models_dir file
+        run: |
+          cat > tests_models_dir.txt << 'EOF'
+          ${{ steps.repo_content.outputs.tests_models_dir }}
+          EOF
+
+      - name: Write tests_quantization_dir file
+        run: |
+          cat > tests_quantization_dir.txt << 'EOF'
+          ${{ steps.repo_content.outputs.tests_quantization_dir }}
+          EOF
+
+      - name: Run script to get jobs to run
+        id: get_jobs
+        run: |
+          python utils/get_pr_run_slow_jobs.py | tee output.txt
+          echo "jobs_to_run: $(tail -n 1 output.txt)"
+          echo "jobs_to_run=$(tail -n 1 output.txt)" >> $GITHUB_OUTPUT
+
+  send_comment:
+    # Will delete the previous comment and send a new one if:
+    #   - either the content is changed
+    #   - or the previous comment is 30 minutes or more old
+    name: Send a comment to suggest jobs to run
+    if: ${{ needs.get-jobs.outputs.jobs != '' }}
+    needs: [get-pr-number, get-jobs]
+    permissions:
+      pull-requests: write
+    runs-on: ubuntu-22.04
+    steps:
+      - name: Check and update comment if needed
+        uses: actions/github-script@v7
+        env:
+          BODY: "\n\nrun-slow: ${{ needs.get-jobs.outputs.jobs }}"
+        with:
+          script: |
+            const prNumber = ${{ needs.get-pr-number.outputs.PR_NUMBER }};
+            const commentPrefix = "**[For maintainers]** Suggested jobs to run (before merge)";
+            const thirtyMinutesAgo = new Date(Date.now() - 30 * 60 * 1000); // 30 minutes ago
+            const newBody = `${commentPrefix}${process.env.BODY}`;
+            
+            // Get all comments on the PR
+            const { data: comments } = await github.rest.issues.listComments({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: prNumber
+            });
+            
+            // Find existing comments that start with our prefix
+            const existingComments = comments.filter(comment => 
+              comment.user.login === 'github-actions[bot]' && 
+              comment.body.startsWith(commentPrefix)
+            );
+            
+            let shouldCreateNewComment = true;
+            let commentsToDelete = [];
+            
+            if (existingComments.length > 0) {
+              // Get the most recent comment
+              const mostRecentComment = existingComments
+                .sort((a, b) => new Date(b.created_at) - new Date(a.created_at))[0];
+              
+              const commentDate = new Date(mostRecentComment.created_at);
+              const isOld = commentDate < thirtyMinutesAgo;
+              const isDifferentContent = mostRecentComment.body !== newBody;
+              
+              console.log(`Most recent comment created: ${mostRecentComment.created_at}`);
+              console.log(`Is older than 30 minutes: ${isOld}`);
+              console.log(`Has different content: ${isDifferentContent}`);
+              
+              if (isOld || isDifferentContent) {
+                // Delete all existing comments and create new one
+                commentsToDelete = existingComments;
+                console.log(`Will delete ${commentsToDelete.length} existing comment(s) and create new one`);
+              } else {
+                // Content is same and comment is recent, skip
+                shouldCreateNewComment = false;
+                console.log('Comment is recent and content unchanged, skipping update');
+              }
+            } else {
+              console.log('No existing comments found, will create new one');
+            }
+            
+            // Delete old comments if needed
+            for (const comment of commentsToDelete) {
+              console.log(`Deleting comment #${comment.id} (created: ${comment.created_at})`);
+              await github.rest.issues.deleteComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                comment_id: comment.id
+              });
+            }
+            
+            // Create new comment if needed
+            if (shouldCreateNewComment) {
+              await github.rest.issues.createComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: prNumber,
+                body: newBody
+              });
+              console.log('✅ New comment created');
+            } else {
+              console.log('ℹ️ No comment update needed');
+            }

.github/workflows/self-comment-ci.yml

@@ -29,7 +29,7 @@ jobs:
     runs-on: ubuntu-22.04
     name: Get PR number
     # For security: only allow team members to run
-    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu", "ivarflakstad"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
+    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu", "ivarflakstad", "stevhliu"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
     outputs:
       PR_NUMBER: ${{ steps.set_pr_number.outputs.PR_NUMBER }}
     steps:

.github/workflows/self-scheduled-caller.yml

@@ -7,7 +7,7 @@ on:
     - cron: "17 2 * * *"
   push:
     branches:
-      - fix-dataset-run_object_detection-and-add-torchcodec-trigger-ci
+      - run_scheduled_ci*
   workflow_dispatch:
     inputs:
       prev_workflow_run_id:
@@ -24,7 +24,7 @@ on:
 
 # Used for `push` to easily modify the target workflow runs to compare against
 env:
-    prev_workflow_run_id: "15988665799"
+    prev_workflow_run_id: ""
     other_workflow_run_id: ""
 
 
@@ -50,8 +50,64 @@ jobs:
     uses: ./.github/workflows/self-scheduled.yml
     with:
       job: run_models_gpu
-      slack_report_channel: "#transformers-dummy"
+      slack_report_channel: "#transformers-ci-daily-models"
       docker: huggingface/transformers-all-latest-gpu
       ci_event: Daily CI
       report_repo_id: hf-internal-testing/transformers_daily_ci
     secrets: inherit
+
+  torch-pipeline:
+    name: Torch pipeline CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_pipelines_torch_gpu
+      slack_report_channel: "#transformers-ci-daily-pipeline-torch"
+      docker: huggingface/transformers-pytorch-gpu
+      ci_event: Daily CI
+      report_repo_id: hf-internal-testing/transformers_daily_ci
+    secrets: inherit
+
+  example-ci:
+    name: Example CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_examples_gpu
+      slack_report_channel: "#transformers-ci-daily-examples"
+      docker: huggingface/transformers-all-latest-gpu
+      ci_event: Daily CI
+      report_repo_id: hf-internal-testing/transformers_daily_ci
+    secrets: inherit
+
+  trainer-fsdp-ci:
+    name: Trainer/FSDP CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_trainer_and_fsdp_gpu
+      slack_report_channel: "#transformers-ci-daily-training"
+      docker: huggingface/transformers-all-latest-gpu
+      ci_event: Daily CI
+      report_repo_id: hf-internal-testing/transformers_daily_ci
+    secrets: inherit
+
+  deepspeed-ci:
+    name: DeepSpeed CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_torch_cuda_extensions_gpu
+      slack_report_channel: "#transformers-ci-daily-training"
+      docker: huggingface/transformers-pytorch-deepspeed-latest-gpu
+      ci_event: Daily CI
+      working-directory-prefix: /workspace
+      report_repo_id: hf-internal-testing/transformers_daily_ci
+    secrets: inherit
+
+  quantization-ci:
+    name: Quantization CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_quantization_torch_gpu
+      slack_report_channel: "#transformers-ci-daily-quantization"
+      docker: huggingface/transformers-quantization-latest-gpu
+      ci_event: Daily CI
+      report_repo_id: hf-internal-testing/transformers_daily_ci
+    secrets: inherit

.github/workflows/self-scheduled-intel-gaudi.yml

@@ -84,8 +84,6 @@ jobs:
       machine_type: ${{ matrix.machine_type }}
       folder_slices: ${{ needs.setup.outputs.folder_slices }}
       runner: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
-      report_name_prefix: run_models_gpu
-
     secrets: inherit
 
   run_trainer_and_fsdp_gpu:
@@ -104,11 +102,10 @@ jobs:
       folder_slices: ${{ needs.setup.outputs.folder_slices }}
       runner: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
       report_name_prefix: run_trainer_and_fsdp_gpu
-
     secrets: inherit
 
-  run_pipelines_gpu:
-    if: ${{ inputs.job == 'run_pipelines_gpu' }}
+  run_pipelines_torch_gpu:
+    if: ${{ inputs.job == 'run_pipelines_torch_gpu' }}
     name: Pipelines
     strategy:
       fail-fast: false
@@ -161,20 +158,20 @@ jobs:
 
       - name: Run all pipeline tests on Intel Gaudi
         run: |
-          python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_pipelines_gpu_test_reports tests/pipelines -m "not not_device_test"
+          python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports tests/pipelines -m "not not_device_test"
 
       - name: Failure short reports
         if: ${{ failure() }}
         continue-on-error: true
         run: |
-          cat reports/${{ env.machine_type }}_run_pipelines_gpu_test_reports/failures_short.txt
+          cat reports/${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports/failures_short.txt
 
-      - name: "Test suite reports artifacts: ${{ env.machine_type }}_run_pipelines_gpu_test_reports"
+      - name: "Test suite reports artifacts: ${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports"
         if: ${{ always() }}
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ env.machine_type }}_run_pipelines_gpu_test_reports
-          path: reports/${{ env.machine_type }}_run_pipelines_gpu_test_reports
+          name: ${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports
+          path: reports/${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports
 
   run_examples_gpu:
     if: ${{ inputs.job == 'run_examples_gpu' }}
@@ -248,8 +245,8 @@ jobs:
           name: ${{ env.machine_type }}_run_examples_gpu_test_reports
           path: reports/${{ env.machine_type }}_run_examples_gpu_test_reports
 
-  run_deepspeed_gpu:
-    if: ${{ inputs.job == 'run_deepspeed_gpu' }}
+  run_torch_cuda_extensions_gpu:
+    if: ${{ inputs.job == 'run_torch_cuda_extensions_gpu' }}
     name: Intel Gaudi deepspeed tests
     strategy:
       fail-fast: false
@@ -305,20 +302,20 @@ jobs:
 
       - name: Run all deepspeed tests on intel Gaudi
         run: |
-          python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_deepspeed_gpu_test_reports tests/deepspeed -m "not not_device_test"
+          python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports tests/deepspeed -m "not not_device_test"
 
       - name: Failure short reports
         if: ${{ failure() }}
         continue-on-error: true
         run: |
-          cat reports/${{ env.machine_type }}_run_deepspeed_gpu_test_reports/failures_short.txt
+          cat reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports/failures_short.txt
 
-      - name: "Test suite reports artifacts: ${{ env.machine_type }}_run_deepspeed_gpu_test_reports"
+      - name: "Test suite reports artifacts: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports"
         if: ${{ always() }}
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ env.machine_type }}_run_deepspeed_gpu_test_reports
-          path: reports/${{ env.machine_type }}_run_deepspeed_gpu_test_reports
+          name: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
+          path: reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
 
   send_results:
     name: Slack Report
@@ -327,8 +324,8 @@ jobs:
         setup,
         run_models_gpu,
         run_examples_gpu,
-        run_pipelines_gpu,
-        run_deepspeed_gpu,
+        run_torch_cuda_extensions_gpu,
+        run_pipelines_torch_gpu,
         run_trainer_and_fsdp_gpu,
       ]
     if: ${{ always() }}

.github/workflows/self-scheduled-intel-gaudi3-caller.yml

@@ -23,7 +23,7 @@ jobs:
     name: Pipeline CI
     uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
     with:
-      job: run_pipelines_gpu
+      job: run_pipelines_torch_gpu
       ci_event: Scheduled CI (Intel) - Gaudi3
       runner_scale_set: itac-bm-emr-gaudi3-dell
       slack_report_channel: "#transformers-ci-daily-intel-gaudi3"
@@ -47,7 +47,7 @@ jobs:
     name: DeepSpeed CI
     uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
     with:
-      job: run_deepspeed_gpu
+      job: run_torch_cuda_extensions_gpu
       ci_event: Scheduled CI (Intel) - Gaudi3
       runner_scale_set: itac-bm-emr-gaudi3-dell
       slack_report_channel: "#transformers-ci-daily-intel-gaudi3"

.github/workflows/self-scheduled.yml

@@ -135,6 +135,7 @@ jobs:
       folder_slices: ${{ needs.setup.outputs.folder_slices }}
       machine_type: ${{ matrix.machine_type }}
       slice_id: ${{ matrix.slice_id }}
+      runner_map: ${{ needs.setup.outputs.runner_map }}
       docker: ${{ inputs.docker }}
       report_name_prefix: run_trainer_and_fsdp_gpu
     secrets: inherit

.gitignore

@@ -167,3 +167,6 @@ tags
 
 # ruff
 .ruff_cache
+
+# modular conversion
+*.modular_backup

conftest.py

@@ -28,6 +28,7 @@ from transformers.testing_utils import HfDoctestModule, HfDocTestParser
 
 NOT_DEVICE_TESTS = {
     "test_tokenization",
+    "test_tokenization_mistral_common",
     "test_processor",
     "test_processing",
     "test_beam_constraints",

docker/transformers-all-latest-gpu/Dockerfile

@@ -26,10 +26,12 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
 # 1. Put several commands in a single `RUN` to avoid image/layer exporting issue. Could be revised in the future.
 # 2. Regarding `torch` part, We might need to specify proper versions for `torchvision` and `torchaudio`.
 #    Currently, let's not bother to specify their versions explicitly (so installed with their latest release versions).
-RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA && python3 -m pip uninstall -y tensorflow tensorflow_text tensorflow_probability
+RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA && python3 -m pip uninstall -y tensorflow tensorflow_text tensorflow_probability
 
 RUN python3 -m pip uninstall -y flax jax
 
+RUN python3 -m pip install --no-cache-dir -U timm
+
 RUN python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git pytesseract
 RUN python3 -m pip install -U "itsdangerous<2.1.0"
 

docker/transformers-pytorch-amd-gpu/Dockerfile

@@ -1,10 +1,10 @@
-FROM rocm/pytorch:rocm6.4_ubuntu22.04_py3.10_pytorch_release_2.6.0
+FROM rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.7.1
 LABEL maintainer="Hugging Face"
 
 ARG DEBIAN_FRONTEND=noninteractive
 
-ARG TORCH_VISION='0.21.0'
-ARG TORCH_AUDIO='2.6.0'
+ARG TORCH_VISION='0.22.0'
+ARG TORCH_AUDIO='2.7.0'
 
 RUN apt update && \
     apt install -y --no-install-recommends git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-dev python3-pip python3-dev ffmpeg git-lfs && \

docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile

@@ -21,7 +21,7 @@ RUN python3 -m pip install --no-cache-dir './transformers[deepspeed-testing]' 'p
 # Install latest release PyTorch
 # (PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops.)
 # (https://www.deepspeed.ai/tutorials/advanced-install/#pre-install-deepspeed-ops)
-RUN python3 -m pip uninstall -y torch torchvision torchaudio && python3 -m pip install --no-cache-dir -U torch==$PYTORCH torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA
+RUN python3 -m pip uninstall -y torch torchvision torchaudio && python3 -m pip install --no-cache-dir -U torch==$PYTORCH torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA
 
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
 

docker/transformers-pytorch-deepspeed-nightly-gpu/Dockerfile

@@ -19,7 +19,7 @@ RUN python3 -m pip uninstall -y torch torchvision torchaudio
 # Install **nightly** release PyTorch (flag `--pre`)
 # (PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops.)
 # (https://www.deepspeed.ai/tutorials/advanced-install/#pre-install-deepspeed-ops)
-RUN python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
+RUN python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
 
 # `datasets` requires pandas, pandas has some modules compiled with numpy=1.x causing errors
 RUN python3 -m pip install --no-cache-dir './transformers[deepspeed-testing]' 'pandas<2' 'numpy<2'

docker/transformers-quantization-latest-gpu/Dockerfile

@@ -26,7 +26,7 @@ RUN [ ${#PYTORCH} -gt 0 ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch';
 RUN echo torch=$VERSION
 # `torchvision` and `torchaudio` should be installed along with `torch`, especially for nightly build.
 # Currently, let's just use their latest releases (when `torch` is installed with a release version)
-RUN python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA
+RUN python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA
 
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
 

docs/source/en/_toctree.yml

@@ -72,8 +72,6 @@
       title: Caching
     - local: kv_cache
       title: KV cache strategies
-    - local: serving
-      title: Serving
     - local: llm_tutorial_optimization
       title: Getting the most out of LLMs
     - local: perplexity
@@ -100,13 +98,15 @@
       title: Distributed inference
     - local: perf_infer_cpu
       title: CPU
-    - local: tf_xla
-      title: XLA
     title: Optimization
   - local: agents
     title: Agents
   - local: tools
     title: Tools
+  - local: serving
+    title: Serving
+  - local: transformers_as_backend
+    title: Inference server backends
   title: Inference
 - isExpanded: false
   sections:
@@ -141,8 +141,6 @@
       title: GPU
     - local: perf_train_cpu
       title: CPU
-    - local: perf_train_tpu_tf
-      title: TPU
     - local: perf_train_special
       title: Apple Silicon
     - local: perf_train_gaudi
@@ -433,6 +431,8 @@
         title: DiffLlama
       - local: model_doc/distilbert
         title: DistilBERT
+      - local: model_doc/doge
+        title: Doge
       - local: model_doc/dots1
         title: dots1
       - local: model_doc/dpr
@@ -443,6 +443,10 @@
         title: Encoder Decoder Models
       - local: model_doc/ernie
         title: ERNIE
+      - local: model_doc/ernie4_5
+        title: Ernie4_5
+      - local: model_doc/ernie4_5_moe
+        title: Ernie4_5_MoE
       - local: model_doc/ernie_m
         title: ErnieM
       - local: model_doc/esm
@@ -477,6 +481,8 @@
         title: GLM
       - local: model_doc/glm4
         title: glm4
+      - local: model_doc/glm4_moe
+        title: glm4_moe
       - local: model_doc/openai-gpt
         title: GPT
       - local: model_doc/gpt_neo
@@ -519,6 +525,8 @@
         title: Jukebox
       - local: model_doc/led
         title: LED
+      - local: model_doc/lfm2
+        title: LFM2
       - local: model_doc/llama
         title: LLaMA
       - local: model_doc/llama2
@@ -563,6 +571,8 @@
         title: MobileBERT
       - local: model_doc/modernbert
         title: ModernBert
+      - local: model_doc/modernbert-decoder
+        title: ModernBERTDecoder
       - local: model_doc/mpnet
         title: MPNet
       - local: model_doc/mpt
@@ -693,6 +703,8 @@
         title: Zamba2
       title: Text models
     - sections:
+      - local: model_doc/aimv2
+        title: Aimv2
       - local: model_doc/beit
         title: BEiT
       - local: model_doc/bit
@@ -709,6 +721,8 @@
         title: D-FINE
       - local: model_doc/dab-detr
         title: DAB-DETR
+      - local: model_doc/deepseek_v2
+        title: DeepSeek-V2
       - local: model_doc/deformable_detr
         title: Deformable DETR
       - local: model_doc/deit
@@ -735,6 +749,8 @@
         title: DPT
       - local: model_doc/efficientformer
         title: EfficientFormer
+      - local: model_doc/efficientloftr
+        title: EfficientLoFTR
       - local: model_doc/efficientnet
         title: EfficientNet
       - local: model_doc/eomt
@@ -1035,6 +1051,8 @@
         title: PaliGemma
       - local: model_doc/perceiver
         title: Perceiver
+      - local: model_doc/perception_lm
+        title: PerceptionLM
       - local: model_doc/phi4_multimodal
         title: Phi4 Multimodal
       - local: model_doc/pix2struct
@@ -1087,6 +1105,8 @@
         title: Vision Text Dual Encoder
       - local: model_doc/visual_bert
         title: VisualBERT
+      - local: model_doc/voxtral
+        title: Voxtral
       - local: model_doc/xclip
         title: X-CLIP
       title: Multimodal models
@@ -1144,4 +1164,3 @@
       title: Environment Variables
     title: Reference
   title: API
-

docs/source/en/agents.md

@@ -14,5 +14,9 @@ rendered properly in your Markdown viewer.
 
 -->
 
+# Agents
+
+(deprecated)
+
 > [!WARNING]
 > Agents and tools were spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. They were removed from `transformers` in v4.52.

docs/source/en/attention_interface.md

@@ -60,11 +60,11 @@ You will see it prints "I just entered the attention computation" as many times
 
 ## Dynamically switching attention function
 
-You could dynamically change the model's attention function as well, by overriding the `config._attn_implementation` field:
+You could dynamically change the model's attention function as well:
 
 ```python
 # Back to use original sdpa implementation
-model.config._attn_implementation = "sdpa"
+model.set_attn_implementation("sdpa")
 
 model(torch.ones(1, 5, dtype=int))
 ```
@@ -72,6 +72,34 @@ model(torch.ones(1, 5, dtype=int))
 and it will stop printing the statements, as it now uses the `sdpa` attention.  
 This allows to quickly change an attention function, without needing to reload the model!
 
+## Different attention per backbone in multimodal models
+
+For multimodal models different attention functions may work better for each backbone module. For example, some vision backbones perform better in fp32, but are incompatible with FlashAttention. To continue using FlashAttention while keeping the vision encoder in fp32, create a dict and map each config to an attention implementation as shown below.
+
+```python
+from transformers import AutoModelForImageTextToText
+
+model_id = "facebook/chameleon-7b"
+
+attention_implementation_per_backbone = {"vision_config": "sdpa", "text_config": "flash_attention_2"}
+model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation=attention_implementation_per_backbone)
+
+# NOTE: keys in the attention implementation have to be the same as the sub-config names
+for key in attention_implementation_per_backbone:
+    assert key in model.config.sub_configs, f"Invalid key in `attention_implementation`"
+
+# You can omit certain backbones - the default attention function (SDPA) will be used
+# This is equivalent to the previous example
+model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation={"text_config": "flash_attention_2"})
+
+
+# Set the same attention implementation for all backbones with single string, same as in non-multimodal models
+model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation="eager")
+
+# Alternatively use a dict with an empty key for global configuration
+model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation={"": "eager"})
+```
+
 ## What about new args needed in my custom attention function?
 
 But indeed, what if the new function requires a new arg to be properly used? It's no issue! Models supporting the

docs/source/en/auto_docstring.md

@@ -64,9 +64,9 @@ Arguments can also be passed directly to `@auto_docstring` for more control. Use
     It builds upon the standard Transformer architecture with unique modifications.""",
     custom_args="""
     custom_parameter (`type`, *optional*, defaults to `default_value`):
-        A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
+        A concise description for custom_parameter if not defined or overriding the description in `auto_docstring.py`.
     internal_helper_arg (`type`, *optional*, defaults to `default_value`):
-        A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
+        A concise description for internal_helper_arg if not defined or overriding the description in `auto_docstring.py`.
     """
 )
 class MySpecialModel(PreTrainedModel):
@@ -85,13 +85,40 @@ class MySpecialModel(PreTrainedModel):
     def __init__(self, config: ConfigType, custom_parameter: "type" = "default_value", internal_helper_arg=None):
         r"""
         custom_parameter (`type`, *optional*, defaults to `default_value`):
-            A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
+            A concise description for custom_parameter if not defined or overriding the description in `auto_docstring.py`.
         internal_helper_arg (`type`, *optional*, defaults to `default_value`):
-            A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
+            A concise description for internal_helper_arg if not defined or overriding the description in `auto_docstring.py`.
         """
         # ...
 ```
 
+You should also use the `@auto_docstring` decorator for classes that inherit from [`~utils.ModelOutput`].
+
+```python
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Custom model outputs with additional fields.
+    """
+)
+class MyModelOutput(ImageClassifierOutput):
+    r"""
+    loss (`torch.FloatTensor`, *optional*):
+        The loss of the model.
+    custom_field (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*):
+        A custom output field specific to this model.
+    """
+
+    # Standard fields like hidden_states, logits, attentions etc. can be automatically documented if the description is the same as the standard arguments.
+    # However, given that the loss docstring is often different per model, you should document it in the docstring above.
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+    # Custom fields need to be documented in the docstring above
+    custom_field: Optional[torch.FloatTensor] = None
+```
+
 </hfoption>
 <hfoption id="functions">
 
@@ -171,7 +198,7 @@ class MyModel(PreTrainedModel):
 
 There are some rules for documenting different types of arguments and they're listed below.
 
-- Standard arguments (`input_ids`, `attention_mask`, `pixel_values`, etc.) are defined and retrieved from `args_doc.py`. It is the single source of truth for standard arguments and should not be redefined locally if an argument's description and shape is the same as an argument in `args_doc.py`.
+- Standard arguments (`input_ids`, `attention_mask`, `pixel_values`, etc.) are defined and retrieved from `auto_docstring.py`. It is the single source of truth for standard arguments and should not be redefined locally if an argument's description and shape is the same as an argument in `auto_docstring.py`.
 
     If a standard argument behaves differently in your model, then you can override it locally in a `r""" """` block. This local definition has a higher priority. For example, the `labels` argument is often customized per model and typically requires overriding.
 
@@ -245,7 +272,7 @@ When working with modular files (`modular_model.py`), follow the guidelines belo
 The `@auto_docstring` decorator automatically generates docstrings by:
 
 1. Inspecting the signature (arguments, types, defaults) of the decorated class' `__init__` method or the decorated function.
-2. Retrieving the predefined docstrings for common arguments (`input_ids`, `attention_mask`, etc.) from internal library sources like [`ModelArgs`], [`ImageProcessorArgs`], and the `args_doc.py` file.
+2. Retrieving the predefined docstrings for common arguments (`input_ids`, `attention_mask`, etc.) from internal library sources like [`ModelArgs`], [`ImageProcessorArgs`], and the `auto_docstring.py` file.
 3. Adding argument descriptions in one of two ways as shown below.
 
     | method | description | usage |
@@ -253,7 +280,7 @@ The `@auto_docstring` decorator automatically generates docstrings by:
     | `r""" """` | add custom docstring content directly to a method signature or within the `__init__` docstring | document new arguments or override standard descriptions |
     | `custom_args` | add custom docstrings for specific arguments directly in `@auto_docstring` | define docstring for new arguments once if they're repeated in multiple places in the modeling file |
 
-4. Adding class and function descriptions. For model classes with standard naming patterns, like `ModelForCausalLM`, or if it belongs to a pipeline, `@auto_docstring` automatically generates the appropriate descriptions with `ClassDocstring` from `args_doc.py`.
+4. Adding class and function descriptions. For model classes with standard naming patterns, like `ModelForCausalLM`, or if it belongs to a pipeline, `@auto_docstring` automatically generates the appropriate descriptions with `ClassDocstring` from `auto_docstring.py`.
 
     `@auto_docstring` also accepts the `custom_intro` argument to describe a class or function.
 

docs/source/en/cache_explanation.md

@@ -82,24 +82,18 @@ When you use Transformers' [`Cache`] class, the self-attention module performs s
 
 ## Cache storage implementation
 
-The actual storage of key-value pairs varies between cache implementations. As an example, consider the [`DynamicCache`].
+Caches are structured as a list of layers, where each layer contains a key and value cache. The key and value caches are tensors with the shape `[batch_size, num_heads, seq_len, head_dim]`.
 
+Layers can be of different types (e.g. `DynamicLayer`, `StaticLayer`, `SlidingWindowLayer`), which mostly changes how sequence length is handled and how the cache is updated.
 
-In [`DynamicCache`], the key-value pairs are stored as two lists of tensors. Each tensor in the lists have the shape `[batch_size, num_heads, seq_len, head_dim]`.
-- `key_cache`: A list of tensors, one for each layer.
-- `value_cache`: A list of tensors, one for each layer.
+The simplest is a `DynamicLayer` that grows as more tokens are processed. The sequence length dimension (`seq_len`) increases with each new token:
 
-When new tokens are processed:
-
-1. For each layer, the new key and value states are concatenated with the existing cache.
 ```py
-self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
-self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
+cache.layers[idx].keys = torch.cat([cache.layers[idx].keys, key_states], dim=-2)
+cache.layers[idx].values = torch.cat([cache.layers[idx].values, value_states], dim=-2)
 ```
 
-2. The cache grows dynamically as more tokens are processed. The sequence length dimension (`seq_len`) increases with each new token.
-
-3. The cache maintains a count of seen tokens through `self._seen_tokens`. This is updated when the first layer processes a new token.
+Other layer types like `StaticLayer` and `SlidingWindowLayer` have a fixed sequence length that is set when the cache is created. This makes them compatible with `torch.compile`. In the case of `SlidingWindowLayer`, existing tokens are shifted out of the cache when a new token is added.
 
 The example below demonstrates how to create a generation loop with [`DynamicCache`]. As discussed, the attention mask is a concatenation of past and current token values and `1` is added to the cache position for the next token.
 
@@ -134,6 +128,34 @@ for _ in range(max_new_tokens):
 print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0])
 "[INST] Hello, what's your name. [/INST]  Hello! My name is LLaMA,"
 ```
+
+## Cache position
+
+The cache position tracks where to insert new tokens in the attention cache. It represents the *absolute* position of each token in the context, independent of padding or batch structure. Suppose you already cached `N` tokens and are now processing `K` new tokens. The cache position for the new tokens will range from `N` to `N + K - 1`. In other words, you're processing tokens at positions - `[N, N + 1, N + 2, ..., N + K - 1]`.
+
+Cache position is used internally for two purposes:
+
+1. Selecting new tokens to process in the input sequence and ensuring only tokens that haven’t been cached yet are passed to the model's `forward`.
+2. Storing key/value pairs at the correct positions in the cache. This is especially important for fixed-size caches, like [`StaticCache`], that pre-allocates a specific cache length.
+
+The generation loop usually takes care of the cache position, but if you're writing a custom generation method, it is important that cache positions are accurate since they are used to write and read key/value states into fixed slots.
+
+
+```py
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
+
+model_id = "meta-llama/Llama-2-7b-chat-hf"
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="cuda:0")
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+messages = [{"role": "user", "content": "You are a helpful assistant."}]
+inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda:0")
+generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=10)
+
+```
+
+
 ## Legacy cache format
 
 Before the [`Cache`] class, the cache used to be stored as a tuple of tuples of tensors. This format is dynamic because it grows as text is generated, similar to [`DynamicCache`].
@@ -143,7 +165,7 @@ The legacy format is essentially the same data structure but organized different
 - The tensors have the same shape `[batch_size, num_heads, seq_len, head_dim]`.
 - The format is less flexible and doesn't support features like quantization or offloading.
 
-If your project depends on this legacy format, you can convert between [`DynamicCache`] and a tuple of tuples as shown below with the [`~DynamicCache.from_legacy_cache`] and [`DynamicCache.to_legacy_cache`] functions. This is helpful if you have custom logic for manipulating a cache in a specific format.
+If your project depends on this legacy format, we recommend to convert to [`DynamicCache`] with [`~DynamicCache.from_legacy_cache`]. Note that legacy cache format is deprecated and not used anymore in `Transformers`. You can convert back to tuple format with [`DynamicCache.to_legacy_cache`] functions, which is helpful if you have custom logic for manipulating a cache in a specific format.
 
 ```py
 import torch
@@ -159,4 +181,4 @@ generation_outputs = model.generate(**inputs, return_dict_in_generate=True, retu
 
 cache = DynamicCache.from_legacy_cache(generation_outputs.past_key_values)
 legacy_format_cache = cache.to_legacy_cache()
-```
+```

docs/source/en/conversations.md

@@ -25,10 +25,7 @@ Check model leaderboards like [OpenLLM](https://hf.co/spaces/HuggingFaceH4/open_
 
 This guide shows you how to quickly start chatting with Transformers from the command line, how build and format a conversation, and how to chat using the [`TextGenerationPipeline`].
 
-## transformers CLI
-
-
-### Interactive chat session
+## chat CLI
 
 After you've [installed Transformers](./installation.md), chat with a model directly from the command line as shown below. It launches an interactive session with a model, with a few base commands listed at the start of the session.
 
@@ -52,68 +49,7 @@ For a full list of options, run the command below.
 transformers chat -h
 ```
 
-The chat is implemented on top of the [AutoClass](./model_doc/auto), using tooling from [text generation](./llm_tutorial) and [chat](./chat_templating).
-
-
-### Serving a model and using MCP tools
-
-> [!WARNING]
-> This section is experimental and subject to changes in future versions
-
-Powering the `chat` interface, we have a server that takes user messages and returns completions. The server has a chat completion API compatible with the OpenAI SDK, so you can also quickly experiment with `transformers` models on existing aplications. To launch a server separately, use the `transformers serve` CLI:
-
-```bash
-transformers serve Menlo/Jan-nano
-```
-
-Under the hood, the `chat` CLI launches and uses `transformers serve`. This server is also an MCP client, which can receive information available MCP servers (i.e. tools), massage their information into the model prompt, and prepare calls to these tools when the model commands to do so. Naturally, this requires a model that is trained to use tools.
-
-At the moment, MCP tool usage in `transformers` has the following constraints:
-- `chat` can't handle tools, but the [`tiny-agents`](https://huggingface.co/blog/python-tiny-agents) CLI can;
-- Only the `qwen` family of models is supported.
-
-The first step to use MCP tools is to let the model know which tools are available. As an example, let's consider a `tiny-agents` configuration file with a reference to an [image generation MCP server](https://evalstate-flux1-schnell.hf.space/).
-
-> [!TIP]
-> Many Hugging Face Spaces can be used as MCP servers. You can find all compatible Spaces [here](https://huggingface.co/spaces?filter=mcp-server).
-
-```json
-{
-    "model": "http://localhost:8000",
-    "provider": "local",
-    "servers": [
-        {
-            "type": "sse",
-            "config": {
-                "url": "https://evalstate-flux1-schnell.hf.space/gradio_api/mcp/sse"
-            }
-        }
-    ]
-}
-```
-
-You can then launch your `tiny-agents` chat interface with the following command.
-
-```bash
-tiny-agents run path/to/your/config.json
-```
-
-If you have a server (from `transformers serve`) running in the background, you're ready to use MCP tools from a local model! For instance, here's the example of a chat session:
-
-```bash
-Agent loaded with 1 tools:
- • flux1_schnell_infer
-»  Generate an image of a cat on the moon
-<Tool req_0_tool_call>flux1_schnell_infer {"prompt": "a cat on the moon", "seed": 42, "randomize_seed": true, "width": 1024, "height": 1024, "num_inference_steps": 4}
-
-Tool req_0_tool_call
-[Binary Content: Image image/webp, 57732 bytes]
-The task is complete and the content accessible to the User
-Image URL: https://evalstate-flux1-schnell.hf.space/gradio_api/file=/tmp/gradio/3dbddc0e53b5a865ed56a4e3dbdd30f3f61cf3b8aabf1b456f43e5241bd968b8/image.webp
-380576952
-
-I have generated an image of a cat on the moon using the Flux 1 Schnell Image Generator. The image is 1024x1024 pixels and was created with 4 inference steps. Let me know if you would like to make any changes or need further assistance!
-```
+The chat is implemented on top of the [AutoClass](./model_doc/auto), using tooling from [text generation](./llm_tutorial) and [chat](./chat_templating). It uses the `transformers serve` CLI under the hood ([docs](./serving.md#serve-cli)).
 
 
 ## TextGenerationPipeline

docs/source/en/internal/generation_utils.md

@@ -356,66 +356,93 @@ A [`Constraint`] can be used to force the generation to include specific tokens
 
 ## Caches
 
-[[autodoc]] Cache
-    - update
-
-[[autodoc]] CacheConfig
-	- update
-
-[[autodoc]] QuantizedCacheConfig
-	- validate
-
-[[autodoc]] DynamicCache
+[[autodoc]] CacheLayerMixin
     - update
     - get_seq_length
+    - get_mask_sizes
+    - get_max_cache_shape
+    - reset
     - reorder_cache
+
+[[autodoc]] DynamicLayer
+    - update
+    - crop
+    - batch_repeat_interleave
+    - batch_select_indices
+
+[[autodoc]] StaticLayer
+    - update
+
+[[autodoc]] SlidingWindowLayer
+    - update
+
+[[autodoc]] CacheProcessor
+    - pre_update
+    - post_update
+
+[[autodoc]] OffloadedCacheProcessor
+    - pre_update
+
+[[autodoc]] QuantizedCacheProcessor
+    - post_update
+
+[[autodoc]] QuantoQuantizedCacheProcessor
+    - post_update
+
+[[autodoc]] HQQQuantizedCacheProcessor
+    - post_update
+
+[[autodoc]] Cache
+    - update
+    - get_seq_length
+    - get_mask_sizes
+    - get_max_cache_shape
+    - reset
+    - reorder_cache
+    - crop
+    - batch_repeat_interleave
+    - batch_select_indices
+
+[[autodoc]] DynamicCache
     - to_legacy_cache
     - from_legacy_cache
 
 [[autodoc]] QuantizedCache
-    - update
-    - get_seq_length
 
 [[autodoc]] QuantoQuantizedCache
 
+[[autodoc]] QuantoQuantizedCacheProcessor
+
 [[autodoc]] HQQQuantizedCache
 
+[[autodoc]] HQQQuantizedCacheProcessor
+
 [[autodoc]] OffloadedCache
-    - update
-    - prefetch_layer
-    - evict_previous_layer
 
 [[autodoc]] StaticCache
-    - update
-    - get_seq_length
-    - reset
 
 [[autodoc]] OffloadedStaticCache
-    - update
-    - get_seq_length
-    - reset
 
 [[autodoc]] HybridCache
-    - update
-    - get_seq_length
-    - reset
+
+[[autodoc]] HybridChunkedCache
 
 [[autodoc]] SlidingWindowCache
-    - update
-    - reset
 
 [[autodoc]] EncoderDecoderCache
-    - get_seq_length
     - to_legacy_cache
     - from_legacy_cache
-    - reset
-    - reorder_cache
 
 [[autodoc]] MambaCache
     - update_conv_state
     - update_ssm_state
     - reset
 
+[[autodoc]] CacheConfig
+
+[[autodoc]] QuantizedCacheConfig
+
+
 ## Watermark Utils
 
 [[autodoc]] WatermarkingConfig

docs/source/en/internal/model_debugging_utils.md

@@ -247,3 +247,114 @@ first and last layer will be shown. This is useful when some layers (typically c
 layers.
 
 [[autodoc]] model_addition_debugger_context
+
+## Analyzer of skipped tests
+
+### Scan skipped tests - for model adders and maintainers
+
+This small util is a power user tool intended for model adders and maintainers. It lists all test methods
+existing in `test_modeling_common.py`, inherited by all model tester classes, and scans the repository to measure
+how many tests are being skipped and for which models. 
+
+### Rationale
+
+When porting models to transformers, tests fail as they should, and sometimes `test_modeling_common` feels irreconcilable with the peculiarities of our brand new model. But how can we be sure we're not breaking everything by adding a seemingly innocent skip?
+
+This utility:
+- scans all test_modeling_common methods
+- looks for times where a method is skipped
+- returns a summary json you can load as a DataFrame/inspect
+
+**For instance test_inputs_embeds is skipped in a whooping 39% proportion at the time of writing this util.**
+
+![download-icon](https://huggingface.co/datasets/huggingface/documentation-images/resolve/f7f671f69b88ce4967e19179172c248958d35742/transformers/tests_skipped_visualisation.png)
+
+
+### Usage 
+
+You can run the skipped test analyzer in two ways:
+
+#### Full scan (default)
+
+From the root of `transformers` repo, scans all common test methods and outputs the results to a JSON file (default: `all_tests_scan_result.json`).
+
+```bash
+python utils/scan_skipped_tests.py --output_dir path/to/output
+```
+
+- `--output_dir` (optional): Directory where the JSON results will be saved. Defaults to the current directory.
+
+**Example output:**
+
+```
+🔬 Parsing 331 model test files once each...
+📝 Aggregating 224 tests...
+  (224/224) test_update_candidate_strategy_with_matches_1es_3d_is_nonecodet_schedule_fa_kwargs
+✅ Scan complete.
+
+📄 JSON saved to /home/pablo/git/transformers/all_tests_scan_result.json
+
+```
+
+And it will generate `all_tests_scan_result.json` file that you can inspect. The JSON is indexed by method name, and each entry follows this schema, indicating the origin as well (from `common`or `GenerationMixin`.)
+
+```json
+{
+  "<method_name>": {
+    "origin": "<test suite>"
+    "models_ran": ["<model_name>", ...],
+    "models_skipped": ["<model_name>", ...],
+    "skipped_proportion": <float>,
+    "reasons_skipped": ["<model_name>: <reason>",
+      ...
+    ]
+  },
+  ...
+}
+```
+
+Which you can visualise as above with e.g. `pandas`
+
+```python
+df = pd.read_json('all_tests_scan_result.json').T
+df.sort_values(by=['skipped_proportion'], ascending=False)
+
+```
+
+### Scan a single test method
+
+You can focus on a specific test method using `--test_method_name`:
+
+```bash
+$ python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds --output_dir path/to/output
+```
+
+- `--test_method_name`: Name of the test method to scan (e.g., `test_inputs_embeds`).
+- `--output_dir` (optional): Directory where the JSON result will be saved.
+
+**Example output:**
+
+```bash
+$ python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds
+
+🔬 Parsing 331 model test files once each...
+
+== test_inputs_embeds ==
+
+Ran    : 199/323
+Skipped : 124/323 (38.4%)
+ - aimv2: Aimv2 does not use inputs_embeds
+ - align: Inputs_embeds is tested in individual model tests
+ - altclip: Inputs_embeds is tested in individual model tests
+ - audio_spectrogram_transformer: AST does not use inputs_embeds
+ - beit: BEiT does not use inputs_embeds
+ - bit: Bit does not use inputs_embeds
+ - blip: Blip does not use inputs_embeds
+ - blip_2: Inputs_embeds is tested in individual model tests
+ - bridgetower: 
+ - canine: CANINE does not have a get_input_embeddings() method.
+ - ...
+
+📄 JSON saved to /home/pablo/git/transformers/scan_test_inputs_embeds.json
+
+```

docs/source/en/kv_cache.md

@@ -44,7 +44,7 @@ import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map="auto")
 inputs = tokenizer("I like rock music because", return_tensors="pt").to(model.device)
 
 model.generate(**inputs, do_sample=False, max_new_tokens=20, use_cache=False)
@@ -59,7 +59,7 @@ import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map="auto")
 inputs = tokenizer("I like rock music because", return_tensors="pt").to(model.device)
 
 past_key_values = DynamicCache()
@@ -134,7 +134,7 @@ The [`QuantizedCache`] reduces memory requirements by quantizing the KV values t
 > [!WARNING]
 > Quantizing the cache can harm latency if the context length is short and there is enough GPU memory available for generation without enabling cache quantization. Try to find a balance between memory efficiency and latency.
 
-Enable [`QuantizedCache`] by configuring `cache_implementation="quantized"` in [`GenerationConfig`], and indicate the quantization backend in [`QuantizedCacheConfig`]. Any additional quantization related parameters should also be passed either as a dict or an instance of [`QuantizedCacheConfig`]. You should use the default values for these additional parameters unless you're running out-of-memory. In that case, consider decreasing the residual length.
+Enable [`QuantizedCache`] by configuring `cache_implementation="quantized"` in [`GenerationConfig`], and the quantization backend, as well as any additional quantization related parameters should also be passed either as a dict. You should use the default values for these additional parameters unless you're running out-of-memory. In that case, consider decreasing the residual length.
 
 <hfoptions id="quantized-cache">
 <hfoption id="HQQQuantizedCache">
@@ -142,13 +142,14 @@ Enable [`QuantizedCache`] by configuring `cache_implementation="quantized"` in [
 For [`HQQQuantizedCache`], we recommend setting the `axis-key` and `axis-value` parameters to `1`.
 
 ```py
-from transformers import AutoTokenizer, AutoModelForCausalLM, HQQQuantizedCache, QuantizedCacheConfig
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, HQQQuantizedCache
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map="auto")
 inputs = tokenizer("I like rock music because", return_tensors="pt").to(model.device)
 
-out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="quantized", cache_config={"axis-key": 1, "axis-value": 1, "backend": "hqq"})
+out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="quantized", cache_config={"backend": "HQQ"})
 print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
 I like rock music because it's loud and energetic. It's a great way to express myself and rel
 ```
@@ -159,13 +160,14 @@ I like rock music because it's loud and energetic. It's a great way to express m
 For [`QuantoQuantizedCache`], we recommend setting the `axis-key` and `axis-value` parameters to `0`.
 
 ```py
-from transformers import AutoTokenizer, AutoModelForCausalLM, QuantoQuantizedCache, QuantizedCacheConfig
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, QuantoQuantizedCache
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map="auto")
 inputs = tokenizer("I like rock music because", return_tensors="pt").to(model.device)
 
-out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="quantized", cache_config={"nbits": 4, "axis-key": 0, "axis-value": 0, "backend": "quanto"})
+out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="quantized", cache_config={"nbits": 4, "backend": "quanto"})
 print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
 I like rock music because it's loud and energetic. It's a great way to express myself and rel
 ```
@@ -207,14 +209,14 @@ import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map="auto")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map={"": 0})
 inputs = tokenizer("Hello, my name is", return_tensors="pt").to(model.device)
 
 out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="offloaded_static")
 tokenizer.batch_decode(out, skip_special_tokens=True)[0]
 "Hello, my name is [Your Name], and I am a [Your Profession] with [Number of Years] of"
 ```
-Cache offloading requires a CUDA GPU.
+Cache offloading requires a CUDA GPU or Intel XPU.
 
 ### Sliding window cache
 
@@ -227,7 +229,7 @@ import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
 tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
-model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", torch_dtype=torch.float16).to("cuda:0")
+model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", torch_dtype=torch.float16, device_map="auto")
 inputs = tokenizer("Yesterday I was on a rock concert and.", return_tensors="pt").to(model.device)
 
 out = model.generate(**inputs, do_sample=False, max_new_tokens=30, cache_implementation="sliding_window")
@@ -273,7 +275,6 @@ from transformers.cache_utils import (
     StaticCache,
     SlidingWindowCache,
     QuantoQuantizedCache,
-    QuantizedCacheConfig,
 )
 
 model_id = "meta-llama/Llama-2-7b-chat-hf"
@@ -306,15 +307,15 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, DynamicCache, StaticCache
 
 model_id = "meta-llama/Llama-2-7b-chat-hf"
-model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="cuda")
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map={"": 0})
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 
 # Init StaticCache with big enough max-length (1024 tokens for the below example)
 # You can also init a DynamicCache, if that suits you better
-prompt_cache = StaticCache(config=model.config, max_batch_size=1, max_cache_len=1024, device="cuda", dtype=torch.bfloat16)
+prompt_cache = StaticCache(config=model.config, max_batch_size=1, max_cache_len=1024, device=model.device.type, dtype=torch.bfloat16)
 
 INITIAL_PROMPT = "You are a helpful assistant. "
-inputs_initial_prompt = tokenizer(INITIAL_PROMPT, return_tensors="pt").to("cuda")
+inputs_initial_prompt = tokenizer(INITIAL_PROMPT, return_tensors="pt").to(model.device.type)
 # This is the common prompt cached, we need to run forward without grad to be able to copy
 with torch.no_grad():
      prompt_cache = model(**inputs_initial_prompt, past_key_values = prompt_cache).past_key_values
@@ -322,7 +323,7 @@ with torch.no_grad():
 prompts = ["Help me to write a blogpost about travelling.", "What is the capital of France?"]
 responses = []
 for prompt in prompts:
-    new_inputs = tokenizer(INITIAL_PROMPT + prompt, return_tensors="pt").to("cuda")
+    new_inputs = tokenizer(INITIAL_PROMPT + prompt, return_tensors="pt").to(model.device.type)
     past_key_values = copy.deepcopy(prompt_cache)
     outputs = model.generate(**new_inputs, past_key_values=past_key_values,max_new_tokens=20)
     response = tokenizer.batch_decode(outputs)[0]

docs/source/en/llm_optims.md

@@ -341,7 +341,7 @@ A known issue with transformer models is that the self-attention mechanism grows
 
 FlashAttention and [FlashAttention-2](./perf_infer_gpu_one#flashattention-2) break up the attention computation into smaller chunks and reduces the number of intermediate read/write operations to the GPU memory to speed up inference. FlashAttention-2 improves on the original FlashAttention algorithm by also parallelizing over sequence length dimension and better partitioning work on the hardware to reduce synchronization and communication overhead.
 
-To use FlashAttention-2, set [attn_implementation](https://hf.co/docs/transformers/main/en/main_classes/text_generation#transformers.PreTrainedModel.from_pretrained.attn_implementation) to `"flash_attention_2"` in [`~PreTrainedModel.from_pretrained`].
+To use FlashAttention-2, set [attn_implementation](https://hf.co/docs/transformers/main/en/main_classes/text_generation#transformers.PreTrainedModel.from_pretrained.attn_implementation) to `"flash_attention_2"` in [`~PreTrainedModel.from_pretrained`] or set with `model.set_attention_implementation("flash_attention_2")` to dynamically update the [attention interface](./attention_interface) after the model is loaded.
 
 ```py
 from transformers import AutoModelForCausalLM, BitsAndBytesConfig
@@ -353,6 +353,14 @@ model = AutoModelForCausalLM.from_pretrained(
     torch_dtype=torch.bfloat16,
     attn_implementation="flash_attention_2",
 )
+
+# Change the model's attention dynamically after loading
+model = AutoModelForCausalLM.from_pretrained(
+    "google/gemma-2b",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16
+)
+model.set_attention_implementation("flash_attention_2")
 ```
 
 ### PyTorch scaled dot product attention
@@ -360,7 +368,7 @@ model = AutoModelForCausalLM.from_pretrained(
 Scaled dot product attention (SDPA) is automatically enabled in PyTorch 2.0 and it supports FlashAttention, xFormers, and PyTorch's C++ implementation. SDPA chooses the most performant attention algorithm if you're using a CUDA backend. For other backends, SDPA defaults to the PyTorch C++ implementation.
 
 > [!TIP]
-> SDPA automaticallysupports FlashAttention-2 as long as you have the latest PyTorch version installed.
+> SDPA automatically supports FlashAttention-2 as long as you have the latest PyTorch version installed.
 
 Use the [torch.nn.attention.sdpa_kernel](https://pytorch.org/docs/stable/generated/torch.nn.attention.sdpa_kernel.html) context manager to explicitly enable or disable any of the four attention algorithms. For example, use `SDPBackend.FLASH_ATTENTION` to enable FlashAttention.
 

docs/source/en/model_doc/aimv2.md

@@ -0,0 +1,104 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# AIMv2
+
+## Overview
+
+The AIMv2 model was proposed in [Multimodal Autoregressive Pre-training of Large Vision Encoders](https://arxiv.org/abs/2411.14402) by Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, Alexander T Toshev, Marcin Eichner, Moin Nabi, Yinfei Yang, Joshua M. Susskind, Alaaeldin El-Nouby.
+
+The abstract from the paper is the following:
+
+*We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Furthermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal image understanding across diverse settings.*
+
+
+This model was contributed by [Yaswanth Gali](https://huggingface.co/yaswanthgali).
+The original code can be found [here](https://github.com/apple/ml-aim).
+
+## Usage Example
+
+Here is an example of Image Feature Extraction using specific checkpoints on resized images and native resolution images:
+
+```python
+import requests
+from PIL import Image
+from transformers import AutoImageProcessor, AutoModel
+
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image = Image.open(requests.get(url, stream=True).raw)
+
+processor = AutoImageProcessor.from_pretrained("apple/aimv2-large-patch14-native")
+model = AutoModel.from_pretrained("apple/aimv2-large-patch14-native")
+
+inputs = processor(images=image, return_tensors="pt")
+outputs = model(**inputs)
+```
+
+Here is an example of a checkpoint performing zero-shot classification:
+
+```python
+import requests
+from PIL import Image
+from transformers import AutoProcessor, AutoModel
+
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image = Image.open(requests.get(url, stream=True).raw)
+text = ["Picture of a dog.", "Picture of a cat.", "Picture of a horse."]
+
+processor = AutoProcessor.from_pretrained("apple/aimv2-large-patch14-224-lit")
+model = AutoModel.from_pretrained("apple/aimv2-large-patch14-224-lit")
+
+inputs = processor(
+    images=image,
+    text=text,
+    add_special_tokens=True,
+    truncation=True,
+    padding=True,
+    return_tensors="pt",
+)
+outputs = model(**inputs)
+probs = outputs.logits_per_image.softmax(dim=-1)
+```
+
+## Aimv2Config
+
+[[autodoc]] Aimv2Config
+
+## Aimv2TextConfig
+
+[[autodoc]] Aimv2TextConfig
+
+## Aimv2VisionConfig
+
+[[autodoc]] Aimv2VisionConfig
+
+## Aimv2Model
+
+[[autodoc]] Aimv2Model
+    - forward
+
+## Aimv2VisionModel
+
+[[autodoc]] Aimv2VisionModel
+    - forward
+
+## Aimv2TextModel
+
+[[autodoc]] Aimv2TextModel
+    - forward
+
+</pt>
+<tf>

docs/source/en/model_doc/auto.md

@@ -258,6 +258,10 @@ The following auto classes are available for the following computer vision tasks
 
 [[autodoc]] AutoModelForKeypointDetection
 
+### AutoModelForKeypointMatching
+
+[[autodoc]] AutoModelForKeypointMatching
+
 ### AutoModelForMaskedImageModeling
 
 [[autodoc]] AutoModelForMaskedImageModeling

docs/source/en/model_doc/camembert.md

@@ -14,49 +14,105 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# CamemBERT
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+	<div class="flex flex-wrap space-x-1">
+		<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+		<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+    <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+	</div>
 </div>
 
-## Overview
+# CamemBERT
 
-The CamemBERT model was proposed in [CamemBERT: a Tasty French Language Model](https://huggingface.co/papers/1911.03894) by
-[Louis Martin](https://huggingface.co/louismartin), [Benjamin Muller](https://huggingface.co/benjamin-mlr), [Pedro Javier Ortiz Suárez](https://huggingface.co/pjox), Yoann Dupont, Laurent Romary, Éric Villemonte de la
-Clergerie, [Djamé Seddah](https://huggingface.co/Djame), and [Benoît Sagot](https://huggingface.co/sagot). It is based on Facebook's RoBERTa model released in 2019. It is a model
-trained on 138GB of French text.
+[CamemBERT](https://huggingface.co/papers/1911.03894) is a language model based on [RoBERTa](./roberta), but trained specifically on French text from the OSCAR dataset, making it more effective for French language tasks.
 
-The abstract from the paper is the following:
+What sets CamemBERT apart is that it learned from a huge, high quality collection of French data, as opposed to mixing lots of languages. This helps it really understand French better than many multilingual models.
 
-*Pretrained language models are now ubiquitous in Natural Language Processing. Despite their success, most available
-models have either been trained on English data or on the concatenation of data in multiple languages. This makes
-practical use of such models --in all languages except English-- very limited. Aiming to address this issue for French,
-we release CamemBERT, a French version of the Bi-directional Encoders for Transformers (BERT). We measure the
-performance of CamemBERT compared to multilingual models in multiple downstream tasks, namely part-of-speech tagging,
-dependency parsing, named-entity recognition, and natural language inference. CamemBERT improves the state of the art
-for most of the tasks considered. We release the pretrained model for CamemBERT hoping to foster research and
-downstream applications for French NLP.*
+Common applications of CamemBERT include masked language modeling (Fill-mask prediction), text classification (sentiment analysis), token classification (entity recognition) and sentence pair classification (entailment tasks).
 
-This model was contributed by [the ALMAnaCH team (Inria)](https://huggingface.co/almanach). The original code can be found [here](https://camembert-model.fr/).
+You can find all the original CamemBERT checkpoints under the [ALMAnaCH](https://huggingface.co/almanach/models?search=camembert) organization.
 
-<Tip>
+> [!TIP]
+> This model was contributed by the [ALMAnaCH (Inria)](https://huggingface.co/almanach) team.
+>
+> Click on the CamemBERT models in the right sidebar for more examples of how to apply CamemBERT to different NLP tasks.
 
-This implementation is the same as RoBERTa. Refer to the [documentation of RoBERTa](roberta) for usage examples as well 
-as the information relative to the inputs and outputs.
+The examples below demonstrate how to predict the `<mask>` token with [`Pipeline`], [`AutoModel`], and from the command line.
 
-</Tip>
+<hfoptions id="usage">
 
-## Resources
+<hfoption id="Pipeline">
 
-- [Text classification task guide](../tasks/sequence_classification)
-- [Token classification task guide](../tasks/token_classification)
-- [Question answering task guide](../tasks/question_answering)
-- [Causal language modeling task guide](../tasks/language_modeling)
-- [Masked language modeling task guide](../tasks/masked_language_modeling)
-- [Multiple choice task guide](../tasks/multiple_choice)
+```python
+import torch
+from transformers import pipeline
+
+pipeline = pipeline("fill-mask", model="camembert-base", torch_dtype=torch.float16, device=0)
+pipeline("Le camembert est un délicieux fromage <mask>.")
+```
+</hfoption> 
+
+<hfoption id="AutoModel">
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+
+tokenizer = AutoTokenizer.from_pretrained("camembert-base")
+model = AutoModelForMaskedLM.from_pretrained("camembert-base", torch_dtype="auto", device_map="auto", attn_implementation="sdpa")
+inputs = tokenizer("Le camembert est un délicieux fromage <mask>.", return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    predictions = outputs.logits
+
+masked_index = torch.where(inputs['input_ids'] == tokenizer.mask_token_id)[1]
+predicted_token_id = predictions[0, masked_index].argmax(dim=-1)
+predicted_token = tokenizer.decode(predicted_token_id)
+
+print(f"The predicted token is: {predicted_token}")
+```
+</hfoption> 
+
+<hfoption id="transformers CLI">
+
+```bash
+echo -e "Le camembert est un délicieux fromage <mask>." | transformers run --task fill-mask --model camembert-base --device 0
+```
+
+</hfoption> 
+
+</hfoptions> 
+
+
+Quantization reduces the memory burden of large models by representing weights in lower precision. Refer to the [Quantization](../quantization/overview) overview for available options.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes) quantization to quantize the weights to 8-bits.
+  
+```python
+from transformers import AutoTokenizer, AutoModelForMaskedLM, BitsAndBytesConfig
+import torch
+
+quant_config = BitsAndBytesConfig(load_in_8bit=True)
+model = AutoModelForMaskedLM.from_pretrained(
+    "almanach/camembert-large",
+    quantization_config=quant_config,
+    device_map="auto"
+)
+tokenizer = AutoTokenizer.from_pretrained("almanach/camembert-large")
+
+inputs = tokenizer("Le camembert est un délicieux fromage <mask>.", return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    predictions = outputs.logits
+
+masked_index = torch.where(inputs["input_ids"] == tokenizer.mask_token_id)[1]
+predicted_token_id = predictions[0, masked_index].argmax(dim=-1)
+predicted_token = tokenizer.decode(predicted_token_id)
+
+print(f"The predicted token is: {predicted_token}")
+```
 
 ## CamembertConfig
 
@@ -137,5 +193,4 @@ as the information relative to the inputs and outputs.
 [[autodoc]] TFCamembertForQuestionAnswering
 
 </tf>
-</frameworkcontent>
-
+</frameworkcontent>

docs/source/en/model_doc/deepseek_v2.md

@@ -0,0 +1,49 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# DeepSeek-V2
+
+## Overview
+
+The DeepSeek-V2 model was proposed in [DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model](https://arxiv.org/abs/2405.04434) by DeepSeek-AI Team.
+
+The abstract from the paper is the following:
+We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.
+
+This model was contributed by [VladOS95-cyber](https://github.com/VladOS95-cyber).
+The original code can be found [here](https://huggingface.co/deepseek-ai/DeepSeek-V2).
+
+### Usage tips
+The model uses Multi-head Latent Attention (MLA) and DeepSeekMoE architectures for efficient inference and cost-effective training. It employs an auxiliary-loss-free strategy for load balancing and multi-token prediction training objective. The model can be used for various language tasks after being pre-trained on 14.8 trillion tokens and going through Supervised Fine-Tuning and Reinforcement Learning stages.
+
+## DeepseekV2Config
+
+[[autodoc]] DeepseekV2Config
+
+## DeepseekV2Model
+
+[[autodoc]] DeepseekV2Model
+    - forward
+
+## DeepseekV2ForCausalLM
+
+[[autodoc]] DeepseekV2ForCausalLM
+    - forward
+
+## DeepseekV2ForSequenceClassification
+
+[[autodoc]] DeepseekV2ForSequenceClassification
+    - forward

docs/source/en/model_doc/dia.md

@@ -44,7 +44,7 @@ tokens and decodes them back into audio.
 from transformers import AutoProcessor, DiaForConditionalGeneration
 
 torch_device = "cuda"
-model_checkpoint = "buttercrab/dia-v1-1.6b"
+model_checkpoint = "nari-labs/Dia-1.6B-0626"
 
 text = ["[S1] Dia is an open weights text to dialogue model."]
 processor = AutoProcessor.from_pretrained(model_checkpoint)
@@ -66,7 +66,7 @@ from datasets import load_dataset, Audio
 from transformers import AutoProcessor, DiaForConditionalGeneration
 
 torch_device = "cuda"
-model_checkpoint = "buttercrab/dia-v1-1.6b"
+model_checkpoint = "nari-labs/Dia-1.6B-0626"
 
 ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
 ds = ds.cast_column("audio", Audio(sampling_rate=44100))
@@ -93,7 +93,7 @@ from datasets import load_dataset, Audio
 from transformers import AutoProcessor, DiaForConditionalGeneration
 
 torch_device = "cuda"
-model_checkpoint = "buttercrab/dia-v1-1.6b"
+model_checkpoint = "nari-labs/Dia-1.6B-0626"
 
 ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
 ds = ds.cast_column("audio", Audio(sampling_rate=44100))

docs/source/en/model_doc/doge.md

@@ -0,0 +1,103 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Doge
+
+
+## Overview
+
+Doge is a series of small language models based on the [Doge](https://github.com/SmallDoges/small-doge) architecture, aiming to combine the advantages of state-space and self-attention algorithms, calculate dynamic masks from cached value states using the zero-order hold method, and solve the problem of existing mainstream language models getting lost in context. It uses the `wsd_scheduler` scheduler to pre-train on the `smollm-corpus`, and can continue training on new datasets or add sparse activation feedforward networks from stable stage checkpoints.
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/refs%2Fpr%2F426/transformers/model_doc/doge_architecture.png" alt="drawing" width="600"/>
+
+As shown in the figure below, the sequence transformation part of the Doge architecture uses `Dynamic Mask Attention`, which can be understood as using self-attention related to value states during training, and using state-space without past state decay during inference, to solve the problem of existing Transformers or SSMs getting lost in long text. The state transformation part of Doge uses `Cross Domain Mixture of Experts`, which consists of dense linear layers and sparse embedding layers, and can additionally increase sparse parameters to continue training from dense weight checkpoints without retraining the entire model, thereby reducing the cost of continuous iteration of the model. In addition, Doge also uses `RMSNorm` and `Residual` with learnable parameters to adapt the gradient range of deep models.
+
+Checkout all Doge model checkpoints [here](https://huggingface.co/collections/SmallDoge/doge-slm-679cc991f027c4a3abbded4a).
+
+
+## Usage
+
+<details>
+<summary>Using Doge-Base for text generation</summary>
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("SmallDoge/Doge-20M")
+model = AutoModelForCausalLM.from_pretrained("SmallDoge/Doge-20M")
+inputs = tokenizer("Hey how are you doing?", return_tensors="pt")
+
+outputs = model.generate(**inputs, max_new_tokens=100)
+print(tokenizer.batch_decode(outputs))
+```
+</details>
+
+<details>
+<summary>Using Doge-Instruct for question answering</summary>
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, TextStreamer
+
+tokenizer = AutoTokenizer.from_pretrained("SmallDoge/Doge-20M-Instruct")
+model = AutoModelForCausalLM.from_pretrained("SmallDoge/Doge-20M-Instruct")
+
+generation_config = GenerationConfig(
+      max_new_tokens=100, 
+      use_cache=True, 
+      do_sample=True, 
+      temperature=0.8, 
+      top_p=0.9,
+      repetition_penalty=1.0
+)
+steamer = TextStreamer(tokenizer=tokenizer, skip_prompt=True)
+
+prompt = "Hi, how are you doing today?"
+conversation = [
+      {"role": "user", "content": prompt}
+]
+inputs = tokenizer.apply_chat_template(
+    conversation=conversation,
+    tokenize=True,
+    return_tensors="pt",
+)
+
+outputs = model.generate(
+    inputs, 
+    tokenizer=tokenizer,
+    generation_config=generation_config, 
+    streamer=steamer
+)
+```
+</details>
+
+## DogeConfig
+
+[[autodoc]] DogeConfig
+
+## DogeModel
+
+[[autodoc]] DogeModel
+    - forward
+
+## DogeForCausalLM
+
+[[autodoc]] DogeForCausalLM
+    - forward
+
+## DogeForSequenceClassification
+
+[[autodoc]] DogeForSequenceClassification
+    - forward

docs/source/en/model_doc/efficientloftr.md

@@ -0,0 +1,114 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the MIT License; you may not use this file except in compliance with
+the License.
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+
+-->
+
+# EfficientLoFTR
+
+<div class="flex flex-wrap space-x-1">
+<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+</div>
+
+## Overview
+
+The EfficientLoFTR model was proposed in [Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed](https://arxiv.org/abs/2403.04765) by Yifan Wang, Xingyi He, Sida Peng, Dongli Tan and Xiaowei Zhou.
+
+This model consists of matching two images together by finding pixel correspondences. It can be used to estimate the pose between them. 
+This model is useful for tasks such as image matching, homography estimation, etc.
+
+The abstract from the paper is the following:
+
+*We present a novel method for efficiently producing semidense matches across images. Previous detector-free matcher 
+LoFTR has shown remarkable matching capability in handling large-viewpoint change and texture-poor scenarios but suffers
+from low efficiency. We revisit its design choices and derive multiple improvements for both efficiency and accuracy. 
+One key observation is that performing the transformer over the entire feature map is redundant due to shared local 
+information, therefore we propose an aggregated attention mechanism with adaptive token selection for efficiency. 
+Furthermore, we find spatial variance exists in LoFTR’s fine correlation module, which is adverse to matching accuracy. 
+A novel two-stage correlation layer is proposed to achieve accurate subpixel correspondences for accuracy improvement. 
+Our efficiency optimized model is ∼ 2.5× faster than LoFTR which can even surpass state-of-the-art efficient sparse 
+matching pipeline SuperPoint + LightGlue. Moreover, extensive experiments show that our method can achieve higher 
+accuracy compared with competitive semi-dense matchers, with considerable efficiency benefits. This opens up exciting 
+prospects for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction. 
+Project page: [https://zju3dv.github.io/efficientloftr/](https://zju3dv.github.io/efficientloftr/).*
+
+## How to use
+
+Here is a quick example of using the model. 
+```python
+import torch
+
+from transformers import AutoImageProcessor, AutoModelForKeypointMatching
+from transformers.image_utils import load_image
+
+
+image1 = load_image("https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/united_states_capitol_98169888_3347710852.jpg")
+image2 = load_image("https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/united_states_capitol_26757027_6717084061.jpg")
+
+images = [image1, image2]
+
+processor = AutoImageProcessor.from_pretrained("stevenbucaille/efficientloftr")
+model = AutoModelForKeypointMatching.from_pretrained("stevenbucaille/efficientloftr")
+
+inputs = processor(images, return_tensors="pt")
+with torch.no_grad():
+    outputs = model(**inputs)
+```
+
+You can use the `post_process_keypoint_matching` method from the `ImageProcessor` to get the keypoints and matches in a more readable format:
+
+```python
+image_sizes = [[(image.height, image.width) for image in images]]
+outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
+for i, output in enumerate(outputs):
+    print("For the image pair", i)
+    for keypoint0, keypoint1, matching_score in zip(
+            output["keypoints0"], output["keypoints1"], output["matching_scores"]
+    ):
+        print(
+            f"Keypoint at coordinate {keypoint0.numpy()} in the first image matches with keypoint at coordinate {keypoint1.numpy()} in the second image with a score of {matching_score}."
+        )
+```
+
+From the post processed outputs, you can visualize the matches between the two images using the following code:
+```python
+images_with_matching = processor.visualize_keypoint_matching(images, outputs)
+```
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/632885ba1558dac67c440aa8/2nJZQlFToCYp_iLurvcZ4.png)
+
+This model was contributed by [stevenbucaille](https://huggingface.co/stevenbucaille).
+The original code can be found [here](https://github.com/zju3dv/EfficientLoFTR).
+
+## EfficientLoFTRConfig
+
+[[autodoc]] EfficientLoFTRConfig
+
+## EfficientLoFTRImageProcessor
+
+[[autodoc]] EfficientLoFTRImageProcessor
+
+- preprocess
+- post_process_keypoint_matching
+- visualize_keypoint_matching
+
+## EfficientLoFTRModel
+
+[[autodoc]] EfficientLoFTRModel
+
+- forward
+
+## EfficientLoFTRForKeypointMatching
+
+[[autodoc]] EfficientLoFTRForKeypointMatching
+
+- forward

docs/source/en/model_doc/encoder-decoder.md

@@ -14,115 +14,88 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# Encoder Decoder Models
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
-<img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAC0AAAAtCAMAAAANxBKoAAAC7lBMVEUAAADg5vYHPVgAoJH+/v76+v39/f9JbLP///9+AIgAnY3///+mcqzt8fXy9fgkXa3Ax9709fr+///9/f8qXq49qp5AaLGMwrv8/P0eW60VWawxYq8yqJzG2dytt9Wyu9elzci519Lf3O3S2efY3OrY0+Xp7PT///////+dqNCexMc6Z7AGpJeGvbenstPZ5ejQ1OfJzOLa7ejh4+/r8fT29vpccbklWK8PVa0AS6ghW63O498vYa+lsdKz1NDRt9Kw1c672tbD3tnAxt7R6OHp5vDe7OrDyuDn6vLl6/EAQKak0MgATakkppo3ZK/Bz9y8w9yzu9jey97axdvHzeG21NHH4trTwthKZrVGZLSUSpuPQJiGAI+GAI8SWKydycLL4d7f2OTi1+S9xNzL0ePT6OLGzeEAo5U0qJw/aLEAo5JFa7JBabEAp5Y4qZ2QxLyKmsm3kL2xoMOehrRNb7RIbbOZgrGre68AUqwAqZqNN5aKJ5N/lMq+qsd8kMa4pcWzh7muhLMEV69juq2kbKqgUaOTR5uMMZWLLZSGAI5VAIdEAH+ovNDHuNCnxcy3qcaYx8K8msGplrx+wLahjbYdXrV6vbMvYK9DrZ8QrZ8tqJuFms+Sos6sw8ecy8RffsNVeMCvmb43aLltv7Q4Y7EZWK4QWa1gt6meZKUdr6GOAZVeA4xPAISyveLUwtivxtKTpNJ2jcqfvcltiMiwwcfAoMVxhL+Kx7xjdrqTe60tsaNQs6KaRKACrJ6UTZwkqpqTL5pkHY4AloSgsd2ptNXPvNOOncuxxsqFl8lmg8apt8FJcr9EbryGxLqlkrkrY7dRa7ZGZLQ5t6iXUZ6PPpgVpZeJCJFKAIGareTa0+KJod3H0deY2M+esM25usmYu8d2zsJOdcBVvrCLbqcAOaaHaKQAMaScWqKBXqCXMJ2RHpiLF5NmJZAdAHN2kta11dKu1M+DkcZLdb+Mcql3TppyRJdzQ5ZtNZNlIY+DF4+voCOQAAAAZ3RSTlMABAT+MEEJ/RH+/TP+Zlv+pUo6Ifz8+fco/fz6+evr39S9nJmOilQaF/7+/f38+smmoYp6b1T+/v7++vj189zU0tDJxsGzsrKSfv34+Pf27dDOysG9t6+n/vv6+vr59uzr1tG+tZ6Qg9Ym3QAABR5JREFUSMeNlVVUG1EQhpcuxEspXqS0SKEtxQp1d3d332STTRpIQhIISQgJhODu7lAoDoUCpe7u7u7+1puGpqnCPOyZvffbOXPm/PsP9JfQgyCC+tmTABTOcbxDz/heENS7/1F+9nhvkHePG0wNDLbGWwdXL+rbLWvpmZHXD8+gMfBjTh+aSe6Gnn7lwQIOTR0c8wfX3PWgv7avbdKwf/ZoBp1Gp/PvuvXW3vw5ib7emnTW4OR+3D4jB9vjNJ/7gNvfWWeH/TO/JyYrsiKCRjVEZA3UB+96kON+DxOQ/NLE8PE5iUYgIXjFnCOlxEQMaSGVxjg4gxOnEycGz8bptuNjVx08LscIgrzH3umcn+KKtiBIyvzOO2O99aAdR8cF19oZalnCtvREUw79tCd5sow1g1UKM6kXqUx4T8wsi3sTjJ3yzDmmhenLXLpo8u45eG5y4Vvbk6kkC4LLtJMowkSQxmk4ggVJEG+7c6QpHT8vvW9X7/o7+3ELmiJi2mEzZJiz8cT6TBlanBk70cB5GGIGC1gRDdZ00yADLW1FL6gqhtvNXNG5S9gdSrk4M1qu7JAsmYshzDS4peoMrU/gT7qQdqYGZaYhxZmVbGJAm/CS/HloWyhRUlknQ9KYcExTwS80d3VNOxUZJpITYyspl0LbhArhpZCD9cRWEQuhYkNGMHToQ/2Cs6swJlb39CsllxdXX6IUKh/H5jbnSsPKjgmoaFQ1f8wRLR0UnGE/RcDEjj2jXG1WVTwUs8+zxfcrVO+vSsuOpVKxCfYZiQ0/aPKuxQbQ8lIz+DClxC8u+snlcJ7Yr1z1JPqUH0V+GDXbOwAib931Y4Imaq0NTIXPXY+N5L18GJ37SVWu+hwXff8l72Ds9XuwYIBaXPq6Shm4l+Vl/5QiOlV+uTk6YR9PxKsI9xNJny31ygK1e+nIRC1N97EGkFPI+jCpiHe5PCEy7oWqWSwRrpOvhFzcbTWMbm3ZJAOn1rUKpYIt/lDhW/5RHHteeWFN60qo98YJuoq1nK3uW5AabyspC1BcIEpOhft+SZAShYoLSvnmSfnYADUERP5jJn2h5XtsgCRuhYQqAvwTwn33+YWEKUI72HX5AtfSAZDe8F2DtPPm77afhl0EkthzuCQU0BWApgQIH9+KB0JhopMM7bJrdTRoleM2JAVNMyPF+wdoaz+XJpGoVAQ7WXUkcV7gT3oUZyi/ISIJAVKhgNp+4b4veCFhYVJw4locdSjZCp9cPUhLF9EZ3KKzURepMEtCDPP3VcWFx4UIiZIklIpFNfHpdEafIF2aRmOcrUmjohbT2WUllbmRvgfbythbQO3222fpDJoufaQPncYYuqoGtUEsCJZL6/3PR5b4syeSjZMQG/T2maGANlXT2v8S4AULWaUkCxfLyW8iW4kdka+nEMjxpL2NCwsYNBp+Q61PF43zyDg9Bm9+3NNySn78jMZUUkumqE4Gp7JmFOdP1vc8PpRrzj9+wPinCy8K1PiJ4aYbnTYpCCbDkBSbzhu2QJ1Gd82t8jI8TH51+OzvXoWbnXUOBkNW+0mWFwGcGOUVpU81/n3TOHb5oMt2FgYGjzau0Nif0Ss7Q3XB33hjjQHjHA5E5aOyIQc8CBrLdQSs3j92VG+3nNEjbkbdbBr9zm04ruvw37vh0QKOdeGIkckc80fX3KH/h7PT4BOjgCty8VZ5ux1MoO5Cf5naca2LAsEgehI+drX8o/0Nu+W0m6K/I9gGPd/dfx/EN/wN62AhsBWuAAAAAElFTkSuQmCC
-">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+        <img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
+        ">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
 </div>
 
-## Overview
+# Encoder Decoder Models
 
-The [`EncoderDecoderModel`] can be used to initialize a sequence-to-sequence model with any
-pretrained autoencoding model as the encoder and any pretrained autoregressive model as the decoder.
+[`EncoderDecoderModel`](https://huggingface.co/papers/1706.03762) initializes a sequence-to-sequence model with any pretrained autoencoder and pretrained autoregressive model. It is effective for sequence generation tasks as demonstrated in [Text Summarization with Pretrained Encoders](https://huggingface.co/papers/1908.08345) which uses [`BertModel`] as the encoder and decoder.
 
-The effectiveness of initializing sequence-to-sequence models with pretrained checkpoints for sequence generation tasks
-was shown in [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://huggingface.co/papers/1907.12461) by
-Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
+> [!TIP]
+> This model was contributed by [thomwolf](https://huggingface.co/thomwolf) and the TensorFlow/Flax version by [ydshieh](https://huggingface.co/ydshieh).
+>
+> Click on the Encoder Decoder models in the right sidebar for more examples of how to apply Encoder Decoder to different language tasks.
 
-After such an [`EncoderDecoderModel`] has been trained/fine-tuned, it can be saved/loaded just like
-any other models (see the examples for more information).
+The example below demonstrates how to generate text with [`Pipeline`], [`AutoModel`], and from the command line.
 
-An application of this architecture could be to leverage two pretrained [`BertModel`] as the encoder
-and decoder for a summarization model as was shown in: [Text Summarization with Pretrained Encoders](https://huggingface.co/papers/1908.08345) by Yang Liu and Mirella Lapata.
-
-## Randomly initializing `EncoderDecoderModel` from model configurations.
-
-[`EncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default [`BertModel`] configuration for the encoder and the default [`BertForCausalLM`] configuration for the decoder.
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
 ```python
->>> from transformers import BertConfig, EncoderDecoderConfig, EncoderDecoderModel
+from transformers import pipeline
 
->>> config_encoder = BertConfig()
->>> config_decoder = BertConfig()
+summarizer = pipeline(
+    "summarization",
+    model="patrickvonplaten/bert2bert-cnn_dailymail-fp16",
+    device=0
+)
 
->>> config = EncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
->>> model = EncoderDecoderModel(config=config)
+text = "Plants create energy through a process known as photosynthesis. This involves capturing sunlight and converting carbon dioxide and water into glucose and oxygen."
+print(summarizer(text))
 ```
 
-## Initialising `EncoderDecoderModel` from a pretrained encoder and a pretrained decoder.
-
-[`EncoderDecoderModel`] can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained auto-encoding model, *e.g.* BERT, can serve as the encoder and both pretrained auto-encoding models, *e.g.* BERT, pretrained causal language models, *e.g.* GPT2, as well as the pretrained decoder part of sequence-to-sequence models, *e.g.* decoder of BART, can be used as the decoder.
-Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
-Initializing [`EncoderDecoderModel`] from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in [the *Warm-starting-encoder-decoder blog post*](https://huggingface.co/blog/warm-starting-encoder-decoder).
-To do so, the `EncoderDecoderModel` class provides a [`EncoderDecoderModel.from_encoder_decoder_pretrained`] method.
+</hfoption>
+<hfoption id="AutoModel">
 
 ```python
->>> from transformers import EncoderDecoderModel, BertTokenizer
+import torch  
+from transformers import AutoModelForCausalLM, AutoTokenizer  
 
->>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
->>> model = EncoderDecoderModel.from_encoder_decoder_pretrained("google-bert/bert-base-uncased", "google-bert/bert-base-uncased")
+tokenizer = AutoTokenizer.from_pretrained("patrickvonplaten/bert2bert-cnn_dailymail-fp16")
+model = AutoModelForCausalLM.from_pretrained("patrickvonplaten/bert2bert-cnn_dailymail-fp16", torch_dtype=torch.bfloat16, device_map="auto",attn_implementation="sdpa")  
+
+text = "Plants create energy through a process known as photosynthesis. This involves capturing sunlight and converting carbon dioxide and water into glucose and oxygen."
+
+inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(model.device)
+
+summary = model.generate(**inputs, max_length=60, num_beams=4, early_stopping=True)
+print(tokenizer.decode(summary[0], skip_special_tokens=True))
 ```
 
-## Loading an existing `EncoderDecoderModel` checkpoint and perform inference.
+</hfoption>
+<hfoption id="transformers CLI">
 
-To load fine-tuned checkpoints of the `EncoderDecoderModel` class, [`EncoderDecoderModel`] provides the `from_pretrained(...)` method just like any other model architecture in Transformers.
+```bash
+echo -e "Plants create energy through a process known as photosynthesis. This involves capturing sunlight and converting carbon dioxide and water into glucose and oxygen." | transformers-cli run --task summarization --model "patrickvonplaten/bert2bert-cnn_dailymail-fp16" --device 0
+```
 
-To perform inference, one uses the [`generate`] method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
+</hfoption>
+</hfoptions>
+
+## Notes
+
+- [`EncoderDecoderModel`] can be initialized using any pretrained encoder and decoder. But depending on the decoder architecture, the cross-attention layers may be randomly initialized.
+
+These models require downstream fine-tuning, as discussed in this [blog post](https://huggingface.co/blog/warm-starting-encoder-decoder). Use [`~EncoderDecoderModel.from_encoder_decoder_pretrained`] to combine encoder and decoder checkpoints.
 
 ```python
->>> from transformers import AutoTokenizer, EncoderDecoderModel
+from transformers import EncoderDecoderModel, BertTokenizer
 
->>> # load a fine-tuned seq2seq model and corresponding tokenizer
->>> model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2bert_cnn_daily_mail")
->>> tokenizer = AutoTokenizer.from_pretrained("patrickvonplaten/bert2bert_cnn_daily_mail")
-
->>> # let's perform inference on a long piece of text
->>> ARTICLE_TO_SUMMARIZE = (
-...     "PG&E stated it scheduled the blackouts in response to forecasts for high winds "
-...     "amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were "
-...     "scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."
-... )
->>> input_ids = tokenizer(ARTICLE_TO_SUMMARIZE, return_tensors="pt").input_ids
-
->>> # autoregressively generate summary (uses greedy decoding by default)
->>> generated_ids = model.generate(input_ids)
->>> generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
->>> print(generated_text)
-nearly 800 thousand customers were affected by the shutoffs. the aim is to reduce the risk of wildfires. nearly 800, 000 customers were expected to be affected by high winds amid dry conditions. pg & e said it scheduled the blackouts to last through at least midday tomorrow.
+tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
+model = EncoderDecoderModel.from_encoder_decoder_pretrained(
+    "google-bert/bert-base-uncased", 
+    "google-bert/bert-base-uncased"
+)
 ```
 
-## Loading a PyTorch checkpoint into `TFEncoderDecoderModel`.
-
-[`TFEncoderDecoderModel.from_pretrained`] currently doesn't support initializing the model from a
-pytorch checkpoint. Passing `from_pt=True` to this method will throw an exception. If there are only pytorch
-checkpoints for a particular encoder-decoder model, a workaround is:
-
-```python
->>> # a workaround to load from pytorch checkpoint
->>> from transformers import EncoderDecoderModel, TFEncoderDecoderModel
-
->>> _model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2bert-cnn_dailymail-fp16")
-
->>> _model.encoder.save_pretrained("./encoder")
->>> _model.decoder.save_pretrained("./decoder")
-
->>> model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
-...     "./encoder", "./decoder", encoder_from_pt=True, decoder_from_pt=True
-... )
->>> # This is only for copying some specific attributes of this particular model.
->>> model.config = _model.config
-```
-
-## Training
-
-Once the model is created, it can be fine-tuned similar to BART, T5 or any other encoder-decoder model.
-As you can see, only 2 inputs are required for the model in order to compute a loss: `input_ids` (which are the
-`input_ids` of the encoded input sequence) and `labels` (which are the `input_ids` of the encoded
-target sequence).
+- Encoder Decoder models can be fine-tuned like BART, T5 or any other encoder-decoder model. Only 2 inputs are required to compute a loss, `input_ids` and `labels`. Refer to this [notebook](https://colab.research.google.com/drive/1WIk2bxglElfZewOHboPFNj8H44_VAyKE?usp=sharing#scrollTo=ZwQIEhKOrJpl) for a more detailed training example.
 
 ```python
 >>> from transformers import BertTokenizer, EncoderDecoderModel
@@ -147,11 +120,42 @@ target sequence).
 >>> loss = model(input_ids=input_ids, labels=labels).loss
 ```
 
-Detailed [colab](https://colab.research.google.com/drive/1WIk2bxglElfZewOHboPFNj8H44_VAyKE?usp=sharing#scrollTo=ZwQIEhKOrJpl) for training.
+- [`EncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config as shown below.
 
-This model was contributed by [thomwolf](https://github.com/thomwolf). This model's TensorFlow and Flax versions
-were contributed by [ydshieh](https://github.com/ydshieh).
+```python
+>>> from transformers import BertConfig, EncoderDecoderConfig, EncoderDecoderModel
 
+>>> config_encoder = BertConfig()
+>>> config_decoder = BertConfig()
+
+>>> config = EncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
+>>> model = EncoderDecoderModel(config=config)
+```
+
+- The Encoder Decoder Model can also be used for translation as shown below.
+
+```python
+from transformers import AutoTokenizer, EncoderDecoderModel  
+
+# Load a pre-trained translation model  
+model_name = "google/bert2bert_L-24_wmt_en_de" 
+tokenizer = AutoTokenizer.from_pretrained(model_name, pad_token="<pad>", eos_token="</s>", bos_token="<s>")  
+model = EncoderDecoderModel.from_pretrained(model_name)  
+
+# Input sentence to translate  
+input_text = "Plants create energy through a process known as"  
+
+# Encode the input text  
+inputs = tokenizer(input_text, return_tensors="pt", add_special_tokens=False).input_ids  
+
+# Generate the translated output  
+outputs = model.generate(inputs)[0]  
+
+# Decode the output tokens to get the translated sentence  
+translated_text = tokenizer.decode(outputs, skip_special_tokens=True)  
+
+print("Translated text:", translated_text)  
+```
 
 ## EncoderDecoderConfig
 

docs/source/en/model_doc/eomt.md

@@ -74,20 +74,16 @@ inputs = processor(
     return_tensors="pt",
 )
 
-# Remove Patch Offsets from inputs — only used later for post-processing.
-patch_offsets = inputs.pop("patch_offsets")
-
 with torch.inference_mode():
     outputs = model(**inputs)
 
 # Prepare the original image size in the format (height, width)
-original_image_sizes = [(image.height, image.width)]
+target_sizes = [(image.height, image.width)]
 
 # Post-process the model outputs to get final segmentation prediction
 preds = processor.post_process_semantic_segmentation(
     outputs,
-    patch_offsets=patch_offsets,
-    original_image_sizes=original_image_sizes,
+    target_sizes=target_sizes,
 )
 
 # Visualize the segmentation mask
@@ -130,12 +126,12 @@ with torch.inference_mode():
     outputs = model(**inputs)
 
 # Prepare the original image size in the format (height, width)
-original_image_sizes = [(image.height, image.width)]
+target_sizes = [(image.height, image.width)]
 
 # Post-process the model outputs to get final segmentation prediction
 preds = processor.post_process_instance_segmentation(
     outputs,
-    original_image_sizes=original_image_sizes,
+    target_sizes=target_sizes,
 )
 
 # Visualize the segmentation mask
@@ -173,12 +169,12 @@ with torch.inference_mode():
     outputs = model(**inputs)
 
 # Prepare the original image size in the format (height, width)
-original_image_sizes = [(image.height, image.width)]
+target_sizes = [(image.height, image.width)]
 
 # Post-process the model outputs to get final segmentation prediction
 preds = processor.post_process_panoptic_segmentation(
     outputs,
-    original_image_sizes=original_image_sizes,
+    target_sizes=target_sizes,
 )
 
 # Visualize the panoptic segmentation mask

docs/source/en/model_doc/ernie4_5.md

@@ -0,0 +1,99 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="Tensor parallelism" src="https://img.shields.io/badge/Tensor%20parallelism-06b6d4?style=flat&logoColor=white">
+    </div>
+</div>
+
+# Ernie 4.5
+
+## Overview
+
+The Ernie 4.5 model was released in the [Ernie 4.5 Model Family](https://ernie.baidu.com/blog/posts/ernie4.5/) release by baidu.
+This family of models contains multiple different architectures and model sizes. This model in specific targets the base text
+model without mixture of experts (moe) with 0.3B parameters in total. It uses the standard [Llama](./llama.md) at its core.
+
+Other models from the family can be found at [Ernie 4.5 MoE](./ernie4_5_moe.md).
+
+<div class="flex justify-center">
+    <img src="https://ernie.baidu.com/blog/posts/ernie4.5/overview.png"/>
+</div>
+
+
+## Usage Tips
+
+### Generate text
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "baidu/ERNIE-4.5-0.3B-PT"
+
+# load the tokenizer and the model
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+)
+
+# prepare the model input
+inputs = tokenizer("Hey, are you conscious? Can you talk to me?", return_tensors="pt")
+prompt = "Hey, are you conscious? Can you talk to me?"
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+model_inputs = tokenizer([text], add_special_tokens=False, return_tensors="pt").to(model.device)
+
+# conduct text completion
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=32,
+)
+output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
+
+# decode the generated ids
+generate_text = tokenizer.decode(output_ids, skip_special_tokens=True)
+```
+
+This model was contributed by [Anton Vlasjuk](https://huggingface.co/AntonV).
+The original code can be found [here](https://github.com/PaddlePaddle/ERNIE).
+
+
+## Ernie4_5Config
+
+[[autodoc]] Ernie4_5Config
+
+## Ernie4_5Model
+
+[[autodoc]] Ernie4_5Model
+    - forward
+
+## Ernie4_5ForCausalLM
+
+[[autodoc]] Ernie4_5ForCausalLM
+    - forward

docs/source/en/model_doc/ernie4_5_moe.md

@@ -0,0 +1,183 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="Tensor parallelism" src="https://img.shields.io/badge/Tensor%20parallelism-06b6d4?style=flat&logoColor=white">
+    </div>
+</div>
+
+# Ernie 4.5 MoE
+
+## Overview
+
+The Ernie 4.5 MoE model was released in the [Ernie 4.5 Model Family](https://ernie.baidu.com/blog/posts/ernie4.5/) release by baidu.
+This family of models contains multiple different architectures and model sizes. This model in specific targets the base text
+model with mixture of experts (moe) - one with 21B total, 3B active parameters and another one with 300B total, 47B active parameters.
+It uses the standard [Llama](./llama.md) at its core combined with a specialized MoE based on [Mixtral](./mixtral.md) with additional shared
+experts.
+
+Other models from the family can be found at [Ernie 4.5](./ernie4_5.md).
+
+<div class="flex justify-center">
+    <img src="https://ernie.baidu.com/blog/posts/ernie4.5/overview.png"/>
+</div>
+
+
+## Usage Tips
+
+### Generate text
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "baidu/ERNIE-4.5-21B-A3B-PT"
+
+# load the tokenizer and the model
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+)
+
+# prepare the model input
+inputs = tokenizer("Hey, are you conscious? Can you talk to me?", return_tensors="pt")
+prompt = "Hey, are you conscious? Can you talk to me?"
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+model_inputs = tokenizer([text], add_special_tokens=False, return_tensors="pt").to(model.device)
+
+# conduct text completion
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=32,
+)
+output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
+
+# decode the generated ids
+generate_text = tokenizer.decode(output_ids, skip_special_tokens=True)
+```
+
+### Distributed Generation with Tensor Parallelism
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "baidu/ERNIE-4.5-21B-A3B-PT"
+
+# load the tokenizer and the model
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+    tp_plan="auto",
+)
+
+# prepare the model input
+inputs = tokenizer("Hey, are you conscious? Can you talk to me?", return_tensors="pt")
+prompt = "Hey, are you conscious? Can you talk to me?"
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+model_inputs = tokenizer([text], add_special_tokens=False, return_tensors="pt").to(model.device)
+
+# conduct text completion
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=32,
+)
+output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
+
+# decode the generated ids
+generate_text = tokenizer.decode(output_ids, skip_special_tokens=True)
+```
+
+### Quantization with Bitsandbytes
+
+```python
+import torch
+from transformers import BitsAndBytesConfig, AutoModelForCausalLM, AutoTokenizer
+
+model_name = "baidu/ERNIE-4.5-21B-A3B-PT"
+
+# load the tokenizer and the model
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    device_map="auto",
+    quantization_config=BitsAndBytesConfig(load_in_4bit=True),
+)
+
+# prepare the model input
+inputs = tokenizer("Hey, are you conscious? Can you talk to me?", return_tensors="pt")
+prompt = "Hey, are you conscious? Can you talk to me?"
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+model_inputs = tokenizer([text], add_special_tokens=False, return_tensors="pt").to(model.device)
+
+# conduct text completion
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=32,
+)
+output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
+
+# decode the generated ids
+generate_text = tokenizer.decode(output_ids, skip_special_tokens=True)
+```
+
+This model was contributed by [Anton Vlasjuk](https://huggingface.co/AntonV).
+The original code can be found [here](https://github.com/PaddlePaddle/ERNIE).
+
+
+## Ernie4_5_MoEConfig
+
+[[autodoc]] Ernie4_5_MoEConfig
+
+## Ernie4_5_MoEModel
+
+[[autodoc]] Ernie4_5_MoEModel
+    - forward
+
+## Ernie4_5_MoEForCausalLM
+
+[[autodoc]] Ernie4_5_MoEForCausalLM
+    - forward
+    - generate

docs/source/en/model_doc/falcon_mamba.md

@@ -110,6 +110,13 @@ outputs = model.generate(**inputs, max_new_tokens=100)
 print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```
 
+## FalconMambaCache
+
+[[autodoc]] FalconMambaCache
+    - update_conv_state
+    - update_ssm_state
+    - reset
+
 ## FalconMambaConfig
 
 [[autodoc]] FalconMambaConfig

docs/source/en/model_doc/gemma3.md

@@ -267,3 +267,8 @@ visualizer("<img>What is shown in this image?")
 
 [[autodoc]] Gemma3ForConditionalGeneration
     - forward
+
+## Gemma3ForSequenceClassification
+
+[[autodoc]] Gemma3ForSequenceClassification
+    - forward

docs/source/en/model_doc/gemma3n.md

@@ -29,7 +29,7 @@ rendered properly in your Markdown viewer.
 Gemma3n is a multimodal model with pretrained and instruction-tuned variants, available in E4B and E2B sizes. While
 large portions of the language model architecture are shared with prior Gemma releases, there are many new additions in
 this model, including [Alternating Updates][altup] (AltUp), [Learned Augmented Residual Layer][laurel] (LAuReL),
-[MatFormer][matformer], Per-Layer Embeddings (PLE), activation sparsity, and KV cache sharing. The language model uses
+[MatFormer][matformer], Per-Layer Embeddings (PLE), [Activation Sparsity with Statistical Top-k][spark-transformer], and KV cache sharing. The language model uses
 a similar attention pattern to [Gemma 3](./gemma3.md) with alternating 4 local sliding window self-attention layers for
 every global self-attention layer with a maximum context length of 32k tokens. Gemma 3n introduces
 [MobileNet v5][mobilenetv5] as the vision encoder, using a default resolution of 768x768 pixels, and adds a newly
@@ -121,7 +121,7 @@ echo -e "Plants create energy through a process known as" | transformers run --t
 ## Notes
 
 -   Use [`Gemma3nForConditionalGeneration`] for image-audio-and-text, image-and-text, image-and-audio, audio-and-text,
-    image-only and aduio-only inputs.
+    image-only and audio-only inputs.
 -   Gemma 3n supports multiple images per input, but make sure the images are correctly batched before passing them to
     the processor. Each batch should be a list of one or more images.
 
@@ -201,4 +201,5 @@ echo -e "Plants create energy through a process known as" | transformers run --t
 [gemma3n-collection]: https://huggingface.co/collections/google/gemma-3n
 [laurel]: https://arxiv.org/abs/2411.07501
 [matformer]: https://arxiv.org/abs/2310.07707
+[spark-transformer]: https://arxiv.org/abs/2506.06644
 [usm]: https://arxiv.org/abs/2303.01037

docs/source/en/model_doc/glm4.md

@@ -18,7 +18,37 @@ rendered properly in your Markdown viewer.
 
 ## Overview
 
-To be released with the official model launch.
+The GLM family welcomes new members [GLM-4-0414](https://arxiv.org/pdf/2406.12793) series models.
+
+The **GLM-4-32B-0414** series models, featuring 32 billion parameters. Its performance is comparable to OpenAI’s GPT
+series and DeepSeek’s V3/R1 series. It also supports very user-friendly local deployment features. GLM-4-32B-Base-0414
+was pre-trained on 15T of high-quality data, including substantial reasoning-type synthetic data. This lays the
+foundation for subsequent reinforcement learning extensions. In the post-training stage, we employed human preference
+alignment for dialogue scenarios. Additionally, using techniques like rejection sampling and reinforcement learning, we
+enhanced the model’s performance in instruction following, engineering code, and function calling, thus strengthening
+the atomic capabilities required for agent tasks. GLM-4-32B-0414 achieves good results in engineering code, Artifact
+generation, function calling, search-based Q&A, and report generation. In particular, on several benchmarks, such as
+code generation or specific Q&A tasks, GLM-4-32B-Base-0414 achieves comparable performance with those larger models like
+GPT-4o and DeepSeek-V3-0324 (671B).
+
+**GLM-Z1-32B-0414** is a reasoning model with deep thinking capabilities. This was developed based on GLM-4-32B-0414
+through cold start, extended reinforcement learning, and further training on tasks including mathematics, code, and
+logic. Compared to the base model, GLM-Z1-32B-0414 significantly improves mathematical abilities and the capability to
+solve complex tasks. During training, we also introduced general reinforcement learning based on pairwise ranking
+feedback, which enhances the model's general capabilities.
+
+**GLM-Z1-Rumination-32B-0414** is a deep reasoning model with rumination capabilities (against OpenAI's Deep Research).
+Unlike typical deep thinking models, the rumination model is capable of deeper and longer thinking to solve more
+open-ended and complex problems (e.g., writing a comparative analysis of AI development in two cities and their future
+development plans). Z1-Rumination is trained through scaling end-to-end reinforcement learning with responses graded by
+the ground truth answers or rubrics and can make use of search tools during its deep thinking process to handle complex
+tasks. The model shows significant improvements in research-style writing and complex tasks.
+
+Finally, **GLM-Z1-9B-0414** is a surprise. We employed all the aforementioned techniques to train a small model (9B).
+GLM-Z1-9B-0414 exhibits excellent capabilities in mathematical reasoning and general tasks. Its overall performance is
+top-ranked among all open-source models of the same size. Especially in resource-constrained scenarios, this model
+achieves an excellent balance between efficiency and effectiveness, providing a powerful option for users seeking
+lightweight deployment.
 
 ## Glm4Config
 

docs/source/en/model_doc/glm4_moe.md

@@ -0,0 +1,35 @@
+<!--Copyright 2025 The ZhipuAI Inc. and The HuggingFace Inc. team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Glm4Moe
+
+## Overview
+
+This will update After model release.
+
+## Glm4MoeConfig
+
+[[autodoc]] Glm4MoeConfig
+
+## Glm4MoeModel
+
+[[autodoc]] Glm4MoeModel
+    - forward
+
+## Glm4MoeForCausalLM
+
+[[autodoc]] Glm4MoeForCausalLM
+    - forward

docs/source/en/model_doc/glm4v.md

@@ -23,6 +23,29 @@ rendered properly in your Markdown viewer.
 
 # GLM-4.1V
 
+## Overview
+
+**GLM-4.1V-9B-Thinking** is a bilingual vision-language model optimized for reasoning, built on GLM-4-9B. It introduces
+a "thinking paradigm" with reinforcement learning, achieving state-of-the-art results among 10B-class models and
+rivaling 72B-scale models. It supports 64k context, 4K resolution, and arbitrary aspect ratios, with an open-source base
+model for further research. You can check our paper [here](https://huggingface.co/papers/2507.01006). and below is a abstract.
+
+*We present GLM-4.1V-Thinking, a vision-language model (VLM) designed to advance general-purpose multimodal understanding
+and reasoning. In this report, we share our key findings in the development of the reasoning-centric training framework.
+We first develop a capable vision foundation model with significant potential through large-scale pre-training, which
+arguably sets the upper bound for the final performance. We then propose Reinforcement Learning with Curriculum
+Sampling (RLCS) to unlock the full potential of the model, leading to comprehensive capability enhancement across a
+diverse range of tasks, including STEM problem solving, video understanding, content recognition, coding, grounding,
+GUI-based agents, and long document understanding. We open-source GLM-4.1V-9B-Thinking, which achieves state-of-the-art
+performance among models of comparable size. In a comprehensive evaluation across 28 public benchmarks, our model
+outperforms Qwen2.5-VL-7B on nearly all tasks and achieves comparable or even superior performance on 18 benchmarks
+relative to the significantly larger Qwen2.5-VL-72B. Notably, GLM-4.1V-9B-Thinking also demonstrates competitive or
+superior performance compared to closed-source models such as GPT-4o on challenging tasks including long document
+understanding and STEM reasoning, further underscoring its strong capabilities. Code, models and more information
+are released at https://github.com/THUDM/GLM-4.1V-Thinking.*
+
+## Usage
+
 The example below demonstrates how to generate text based on an image with [`Pipeline`] or the [`AutoModel`] class.
 
 <hfoptions id="usage">

docs/source/en/model_doc/gpt2.md

@@ -57,7 +57,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2", torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa")
 tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
 
-input_ids = tokenzier("Hello, I'm a language model". return_tensors="pt").to("cuda")
+input_ids = tokenizer("Hello, I'm a language model", return_tensors="pt").to("cuda")
 
 output = model.generate(**input_ids, cache_implementation="static")
 print(tokenizer.decode(output[0], skip_special_tokens=True))

docs/source/en/model_doc/ijepa.md

@@ -1,4 +1,4 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
 the License. You may obtain a copy of the License at
@@ -14,53 +14,107 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# I-JEPA
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
 </div>
 
-## Overview
+# I-JEPA
 
-The I-JEPA model was proposed in [Image-based Joint-Embedding Predictive Architecture](https://huggingface.co/papers/2301.08243) by Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, Nicolas Ballas.
-I-JEPA is a self-supervised learning method that predicts the representations of one part of an image based on other parts of the same image. This approach focuses on learning semantic features without relying on pre-defined invariances from hand-crafted data transformations, which can bias specific tasks, or on filling in pixel-level details, which often leads to less meaningful representations.
+[I-JEPA](https://huggingface.co/papers/2301.08243) is a self-supervised learning method that learns semantic image representations by predicting parts of an image from other parts of the image. It compares the abstract representations of the image (rather than pixel level comparisons), which avoids the typical pitfalls of data augmentation bias and pixel-level details that don't capture semantic meaning.
 
-The abstract from the paper is the following:
+You can find the original I-JEPA checkpoints under the [AI at Meta](https://huggingface.co/facebook/models?search=ijepa) organization.
+> [!TIP]
+> This model was contributed by [jmtzt](https://huggingface.co/jmtzt).
 
-This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image- based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample tar- get blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transform- ers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.
 
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/ijepa_architecture.jpg"
-alt="drawing" width="600"/>
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/ijepa_architecture.jpg">
 
-<small> I-JEPA architecture. Taken from the <a href="https://huggingface.co/papers/2301.08243">original paper.</a> </small>
 
-This model was contributed by [jmtzt](https://huggingface.co/jmtzt).
-The original code can be found [here](https://github.com/facebookresearch/ijepa).
+> Click on the I-JEPA models in the right sidebar for more examples of how to apply I-JEPA to different image representation and classification tasks.
 
-## How to use
+The example below demonstrates how to extract image features with [`Pipeline`] or the [`AutoModel`] class.
 
-Here is how to use this model for image feature extraction:
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-```python
+```py
+import torch
+from transformers import pipeline
+feature_extractor = pipeline(
+    task="image-feature-extraction",
+    model="facebook/ijepa_vith14_1k",
+    device=0,
+    torch_dtype=torch.bfloat16
+)
+features = feature_extractor("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg", return_tensors=True)  
+
+print(f"Feature shape: {features.shape}")
+
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```py
 import requests
 import torch
 from PIL import Image
 from torch.nn.functional import cosine_similarity
+from transformers import AutoModel, AutoProcessor  
 
-from transformers import AutoModel, AutoProcessor
+url_1 = "http://images.cocodataset.org/val2017/000000039769.jpg"  
+url_2 = "http://images.cocodataset.org/val2017/000000219578.jpg"
+image_1 = Image.open(requests.get(url_1, stream=True).raw)
+image_2 = Image.open(requests.get(url_2, stream=True).raw)
+
+processor = AutoProcessor.from_pretrained("facebook/ijepa_vith14_1k")  
+model = AutoModel.from_pretrained("facebook/ijepa_vith14_1k", torch_dtype="auto", attn_implementation="sdpa")  
+
+
+def infer(image):  
+    inputs = processor(image, return_tensors="pt")  
+    outputs = model(**inputs)  
+    return outputs.last_hidden_state.mean(dim=1)  
+
+
+embed_1 = infer(image_1)  
+embed_2 = infer(image_2)  
+
+similarity = cosine_similarity(embed_1, embed_2)  
+print(similarity)
+```
+</hfoption>
+</hfoptions>
+
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to 4-bits.
+
+```py
+import torch
+from transformers import BitsAndBytesConfig, AutoModel, AutoProcessor
+from datasets import load_dataset
+
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_use_double_quant=True,
+)
 
 url_1 = "http://images.cocodataset.org/val2017/000000039769.jpg"
 url_2 = "http://images.cocodataset.org/val2017/000000219578.jpg"
 image_1 = Image.open(requests.get(url_1, stream=True).raw)
 image_2 = Image.open(requests.get(url_2, stream=True).raw)
 
-model_id = "facebook/ijepa_vith14_1k"
-processor = AutoProcessor.from_pretrained(model_id)
-model = AutoModel.from_pretrained(model_id)
+processor = AutoProcessor.from_pretrained("facebook/ijepa_vitg16_22k")
+model = AutoModel.from_pretrained("facebook/ijepa_vitg16_22k", quantization_config=quantization_config, torch_dtype="auto", attn_implementation="sdpa")
+
 
-@torch.no_grad()
 def infer(image):
     inputs = processor(image, return_tensors="pt")
     outputs = model(**inputs)
@@ -74,15 +128,6 @@ similarity = cosine_similarity(embed_1, embed_2)
 print(similarity)
 ```
 
-## Resources
-
-A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with I-JEPA.
-
-<PipelineTag pipeline="image-classification"/>
-
-- [`IJepaForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
-- See also: [Image classification task guide](../tasks/image_classification)
-
 ## IJepaConfig
 
 [[autodoc]] IJepaConfig
@@ -95,4 +140,5 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 ## IJepaForImageClassification
 
 [[autodoc]] IJepaForImageClassification
-    - forward
+    - forward
+

docs/source/en/model_doc/led.md

@@ -14,62 +14,135 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# LED
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+           <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+            <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+    </div>
 </div>
 
-## Overview
+# LED
 
-The LED model was proposed in [Longformer: The Long-Document Transformer](https://huggingface.co/papers/2004.05150) by Iz
-Beltagy, Matthew E. Peters, Arman Cohan.
+[Longformer-Encoder-Decoder (LED)](https://huggingface.co/papers/2004.05150) is an encoder-decoder  transformer model for sequence-to-sequence tasks like summarization. It extends [Longformer](.longformer), an encoder-only model designed to handle long inputs, by adding a decoder layer. The decoder uses full self-attention on the encoded tokens and previously decoded locations. Because of Longformer's linear self-attention mechanism, LED is more efficient than standard encoder-decoder models when processing long sequences.
 
-The abstract from the paper is the following:
+You can find all the original [LED] checkpoints under the [Ai2](https://huggingface.co/allenai/models?search=led) organization.
 
-*Transformer-based models are unable to process long sequences due to their self-attention operation, which scales
-quadratically with the sequence length. To address this limitation, we introduce the Longformer with an attention
-mechanism that scales linearly with sequence length, making it easy to process documents of thousands of tokens or
-longer. Longformer's attention mechanism is a drop-in replacement for the standard self-attention and combines a local
-windowed attention with a task motivated global attention. Following prior work on long-sequence transformers, we
-evaluate Longformer on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In
-contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream tasks. Our
-pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on
-WikiHop and TriviaQA. We finally introduce the Longformer-Encoder-Decoder (LED), a Longformer variant for supporting
-long document generative sequence-to-sequence tasks, and demonstrate its effectiveness on the arXiv summarization
-dataset.*
+> [!TIP]
+> This model was contributed by [patrickvonplaten](https://huggingface.co/patrickvonplaten).
+>
+> Click on the LED models in the right sidebar for more examples of how to apply LED to different language tasks.
 
-## Usage tips
+The example below demonstrates how to summarize text with [`Pipeline`], [`AutoModel`], and from the command line.
 
-- [`LEDForConditionalGeneration`] is an extension of
-  [`BartForConditionalGeneration`] exchanging the traditional *self-attention* layer with
-  *Longformer*'s *chunked self-attention* layer. [`LEDTokenizer`] is an alias of
-  [`BartTokenizer`].
-- LED works very well on long-range *sequence-to-sequence* tasks where the `input_ids` largely exceed a length of
-  1024 tokens.
-- LED pads the `input_ids` to be a multiple of `config.attention_window` if required. Therefore a small speed-up is
-  gained, when [`LEDTokenizer`] is used with the `pad_to_multiple_of` argument.
-- LED makes use of *global attention* by means of the `global_attention_mask` (see
-  [`LongformerModel`]). For summarization, it is advised to put *global attention* only on the first
-  `<s>` token. For question answering, it is advised to put *global attention* on all tokens of the question.
-- To fine-tune LED on all 16384, *gradient checkpointing* can be enabled in case training leads to out-of-memory (OOM)
-  errors. This can be done by executing `model.gradient_checkpointing_enable()`. 
- Moreover, the `use_cache=False`
-  flag can be used to disable the caching mechanism to save memory.
-- LED is a model with absolute position embeddings so it's usually advised to pad the inputs on the right rather than
-  the left.
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-This model was contributed by [patrickvonplaten](https://huggingface.co/patrickvonplaten).
+```python
+import torch
+from transformers import pipeline
+
+pipeline = pipeline(
+    task="summarization",
+    model="allenai/led-base-16384",
+    torch_dtype=torch.float16,
+    device=0
+)
+pipeline("""Plants are among the most remarkable and essential life forms on Earth, possessing a unique ability to produce their own food through a process known as photosynthesis. This complex biochemical process is fundamental not only to plant life but to virtually all life on the planet.
+Through photosynthesis, plants capture energy from sunlight using a green pigment called chlorophyll, which is located in specialized cell structures called chloroplasts. In the presence of light, plants absorb carbon dioxide from the atmosphere through small pores in their leaves called stomata, and take in water from the soil through their root systems.
+These ingredients are then transformed into glucose, a type of sugar that serves as a source of chemical energy, and oxygen, which is released as a byproduct into the atmosphere. The glucose produced during photosynthesis is not just used immediately; plants also store it as starch or convert it into other organic compounds like cellulose, which is essential for building their cellular structure.
+This energy reserve allows them to grow, develop leaves, produce flowers, bear fruit, and carry out various physiological processes throughout their lifecycle.""")
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+
+tokenizer = AutoTokenizer.from_pretrained(
+    "allenai/led-base-16384"
+)
+model = AutoModelForSeq2SeqLM.from_pretrained(
+    "allenai/led-base-16384",
+    torch_dtype=torch.float16,
+    device_map="auto"
+)
+
+input_text = """Plants are among the most remarkable and essential life forms on Earth, possessing a unique ability to produce their own food through a process known as photosynthesis. This complex biochemical process is fundamental not only to plant life but to virtually all life on the planet.
+Through photosynthesis, plants capture energy from sunlight using a green pigment called chlorophyll, which is located in specialized cell structures called chloroplasts. In the presence of light, plants absorb carbon dioxide from the atmosphere through small pores in their leaves called stomata, and take in water from the soil through their root systems.
+These ingredients are then transformed into glucose, a type of sugar that serves as a source of chemical energy, and oxygen, which is released as a byproduct into the atmosphere. The glucose produced during photosynthesis is not just used immediately; plants also store it as starch or convert it into other organic compounds like cellulose, which is essential for building their cellular structure.
+This energy reserve allows them to grow, develop leaves, produce flowers, bear fruit, and carry out various physiological processes throughout their lifecycle."""
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+# Place global attention on the first token
+global_attention_mask = torch.zeros_like(input_ids.input_ids).to("cuda")
+global_attention_mask[:, 0] = 1
+
+output = model.generate(**input_ids, global_attention_mask=global_attention_mask, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+</hfoption>
+<hfoption id="transformers-cli">
+
+```bash
+!echo -e "Plants are among the most remarkable and essential life forms on Earth, possessing a unique ability to produce their own food through a process known as photosynthesis. This complex biochemical process is fundamental not only to plant life but to virtually all life on the planet. Through photosynthesis, plants capture energy from sunlight using a green pigment called chlorophyll, which is located in specialized cell structures called chloroplasts." | transformers-cli run --task summarization --model allenai/led-base-16384 --device 0
+```
+</hfoption>
+</hfoptions>
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to int4.
+
+```python
+import torch
+from transformers import BitsAndBytesConfig, AutoModelForSeq2SeqLM, AutoTokenizer
+
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_quant_type="nf4"
+)
+model = AutoModelForSeq2SeqLM.from_pretrained(
+    "allenai/led-large-16384",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    quantization_config=quantization_config
+)
+
+tokenizer = AutoTokenizer.from_pretrained(
+    "allenai/led-large-16384"
+)
+
+input_text = """Plants are among the most remarkable and essential life forms on Earth, possessing a unique ability to produce their own food through a process known as photosynthesis. This complex biochemical process is fundamental not only to plant life but to virtually all life on the planet.
+Through photosynthesis, plants capture energy from sunlight using a green pigment called chlorophyll, which is located in specialized cell structures called chloroplasts. In the presence of light, plants absorb carbon dioxide from the atmosphere through small pores in their leaves called stomata, and take in water from the soil through their root systems.
+These ingredients are then transformed into glucose, a type of sugar that serves as a source of chemical energy, and oxygen, which is released as a byproduct into the atmosphere. The glucose produced during photosynthesis is not just used immediately; plants also store it as starch or convert it into other organic compounds like cellulose, which is essential for building their cellular structure.
+This energy reserve allows them to grow, develop leaves, produce flowers, bear fruit, and carry out various physiological processes throughout their lifecycle."""
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+# Place global attention on the first token
+global_attention_mask = torch.zeros_like(input_ids.input_ids).to("cuda")
+global_attention_mask[:, 0] = 1
+
+output = model.generate(**input_ids, global_attention_mask=global_attention_mask, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+## Notes
+
+- [`LEDForConditionalGeneration`] is an extension of [`BartForConditionalGeneration`] exchanging the traditional self-attention layer with Longformer's chunked self-attention layer. [`LEDTokenizer`] is an alias of [`BartTokenizer`].
+- LED pads the `input_ids` to be a multiple of `config.attention_window` if required. A small speedup is gained when [`LEDTokenizer`] is used with the `pad_to_multiple_of` argument.
+- LED works best on long-range sequence-to-sequence tasks where the `input_ids` are significantly longer than 1024 tokens.
+- LED uses global attention by means of the `global_attention_mask` (see [`LongformerModel`]). For summarization, it is advised to put global attention only on the first `<s>` token. For question answering, it is advised to put global attention on all tokens of the question.
+- To fine-tune LED on all 16384 parameters, gradient checkpointing can be enabled in case training leads to out-of-memory (OOM) errors. Enable gradient checkpointing by adding `model.gradient_checkpointing_enable()` and setting `use_cache=False` to disable the caching mechanism to save memory.
+- Inputs should be padded on the right because LED uses absolute position embeddings.
 
 ## Resources
 
-- [A notebook showing how to evaluate LED](https://colab.research.google.com/drive/12INTTR6n64TzS4RrXZxMSXfrOd9Xzamo?usp=sharing).
-- [A notebook showing how to fine-tune LED](https://colab.research.google.com/drive/12LjJazBl7Gam0XBPy_y0CTOJZeZ34c2v?usp=sharing).
-- [Text classification task guide](../tasks/sequence_classification)
-- [Question answering task guide](../tasks/question_answering)
-- [Translation task guide](../tasks/translation)
-- [Summarization task guide](../tasks/summarization)
+- Read the [LED on Arxiv notebook](https://colab.research.google.com/drive/12INTTR6n64TzS4RrXZxMSXfrOd9Xzamo?usp=sharing) to see how LED can achieve state-of-the-art performance on Arxiv article summarization.
+- Read the [Fine-tune LED notebook](https://colab.research.google.com/drive/12LjJazBl7Gam0XBPy_y0CTOJZeZ34c2v?usp=sharing) to learn how to fine-tune LED on PubMed articles.
 
 ## LEDConfig
 

docs/source/en/model_doc/lfm2.md

@@ -0,0 +1,84 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+<div class="flex flex-wrap space-x-1">
+<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+</div>
+
+# LFM2
+
+## Overview
+
+[LFM2](https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models) represents a new generation of Liquid Foundation Models developed by [Liquid AI](https://liquid.ai/), specifically designed for edge AI and on-device deployment. 
+
+The models are available in three sizes (350M, 700M, and 1.2B parameters) and are engineered to run efficiently on CPU, GPU, and NPU hardware, making them particularly well-suited for applications requiring low latency, offline operation, and privacy.
+
+## Architecture
+
+The architecture consists of 16 blocks total: 10 double-gated short-range convolution blocks and 6 blocks of grouped query attention. This design stems from the concept of dynamical systems, where linear operations are modulated by input-dependent gates, allowing for "liquid" dynamics that can adapt in real-time. The short convolutions are particularly optimized for embedded SoC CPUs, making them ideal for devices that require fast, local inference without relying on cloud connectivity.
+
+The key architectural innovation of LFM2 lies in its systematic approach to balancing quality, latency, and memory efficiency through our STAR neural architecture search engine. Using STAR, Liquid AI optimized the models for real-world performance on embedded hardware, measuring actual peak memory usage and inference speed on Qualcomm Snapdragon processors. This results in models that achieve 2x faster decode and prefill performance compared to similar-sized models, while maintaining superior benchmark performance across knowledge, mathematics, instruction following, and multilingual tasks.
+
+## Example
+
+The following example shows how to generate an answer using the `AutoModelForCausalLM` class.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# Load model and tokenizer
+model_id = "LiquidAI/LFM2-1.2B"
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    device_map="auto",
+    torch_dtype="bfloat16",
+)
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+# Generate answer
+prompt = "What is C. elegans?"
+input_ids = tokenizer.apply_chat_template(
+    [{"role": "user", "content": prompt}],
+    add_generation_prompt=True,
+    return_tensors="pt",
+    tokenize=True,
+)
+
+output = model.generate(
+    input_ids,
+    do_sample=True,
+    temperature=0.3,
+    min_p=0.15,
+    repetition_penalty=1.05,
+    max_new_tokens=512,
+)
+
+print(tokenizer.decode(output[0], skip_special_tokens=False))
+```
+
+## Lfm2Config
+
+[[autodoc]] Lfm2Config
+
+## Lfm2Model
+
+[[autodoc]] Lfm2Model
+    - forward
+
+## Lfm2ForCausalLM
+
+[[autodoc]] Lfm2ForCausalLM
+    - forward

docs/source/en/model_doc/lightglue.md

@@ -10,37 +10,31 @@ specific language governing permissions and limitations under the License.
 ⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
 rendered properly in your Markdown viewer.
 
-
 -->
 
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white" >
+    </div>
+</div>
+
 # LightGlue
 
-## Overview
+[LightGlue](https://arxiv.org/abs/2306.13643) is a deep neural network that learns to match local features across images. It revisits multiple design decisions of SuperGlue and derives simple but effective improvements. Cumulatively, these improvements make LightGlue more efficient - in terms of both memory and computation, more accurate, and much easier to train. Similar to [SuperGlue](https://huggingface.co/magic-leap-community/superglue_outdoor), this model consists of matching two sets of local features extracted from two images, with the goal of being faster than SuperGlue. Paired with the [SuperPoint model](https://huggingface.co/magic-leap-community/superpoint), it can be used to match two images and estimate the pose between them.
 
-The LightGlue model was proposed in [LightGlue: Local Feature Matching at Light Speed](https://arxiv.org/abs/2306.13643)
-by Philipp Lindenberger, Paul-Edouard Sarlin and Marc Pollefeys.
+You can find all the original LightGlue checkpoints under the [ETH-CVG](https://huggingface.co/ETH-CVG) organization.
 
-Similar to [SuperGlue](https://huggingface.co/magic-leap-community/superglue_outdoor), this model consists of matching
-two sets of local features extracted from two images, its goal is to be faster than SuperGlue. Paired with the 
-[SuperPoint model](https://huggingface.co/magic-leap-community/superpoint), it can be used to match two images and 
-estimate the pose between them. This model is useful for tasks such as image matching, homography estimation, etc.
+> [!TIP]
+> This model was contributed by [stevenbucaille](https://huggingface.co/stevenbucaille).
+>
+> Click on the LightGlue models in the right sidebar for more examples of how to apply LightGlue to different computer vision tasks.
 
-The abstract from the paper is the following:
+The example below demonstrates how to match keypoints between two images with the [`AutoModel`] class.
 
-*We introduce LightGlue, a deep neural network that learns to match local features across images. We revisit multiple
-design decisions of SuperGlue, the state of the art in sparse matching, and derive simple but effective improvements. 
-Cumulatively, they make LightGlue more efficient - in terms of both memory and computation, more accurate, and much
-easier to train. One key property is that LightGlue is adaptive to the difficulty of the problem: the inference is much
-faster on image pairs that are intuitively easy to match, for example because of a larger visual overlap or limited
-appearance change. This opens up exciting prospects for deploying deep matchers in latency-sensitive applications like
-3D reconstruction. The code and trained models are publicly available at this [https URL](https://github.com/cvg/LightGlue)*
+<hfoptions id="usage">
+<hfoption id="AutoModel">
 
-## How to use
-
-Here is a quick example of using the model. Since this model is an image matching model, it requires pairs of images to be matched. 
-The raw outputs contain the list of keypoints detected by the keypoint detector as well as the list of matches with their corresponding 
-matching scores.
-```python
+```py
 from transformers import AutoImageProcessor, AutoModel
 import torch
 from PIL import Image
@@ -59,31 +53,70 @@ model = AutoModel.from_pretrained("ETH-CVG/lightglue_superpoint")
 inputs = processor(images, return_tensors="pt")
 with torch.no_grad():
     outputs = model(**inputs)
-```
 
-You can use the `post_process_keypoint_matching` method from the `LightGlueImageProcessor` to get the keypoints and matches in a readable format:
-```python
+# Post-process to get keypoints and matches
 image_sizes = [[(image.height, image.width) for image in images]]
-outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
-for i, output in enumerate(outputs):
-    print("For the image pair", i)
-    for keypoint0, keypoint1, matching_score in zip(
-            output["keypoints0"], output["keypoints1"], output["matching_scores"]
-    ):
-        print(
-            f"Keypoint at coordinate {keypoint0.numpy()} in the first image matches with keypoint at coordinate {keypoint1.numpy()} in the second image with a score of {matching_score}."
-        )
+processed_outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
 ```
 
-You can visualize the matches between the images by providing the original images as well as the outputs to this method:
-```python
-processor.plot_keypoint_matching(images, outputs)
-```
+</hfoption>
+</hfoptions>
 
-![image/png](https://cdn-uploads.huggingface.co/production/uploads/632885ba1558dac67c440aa8/duPp09ty8NRZlMZS18ccP.png)
+## Notes
 
-This model was contributed by [stevenbucaille](https://huggingface.co/stevenbucaille).
-The original code can be found [here](https://github.com/cvg/LightGlue).
+- LightGlue is adaptive to the task difficulty. Inference is much faster on image pairs that are intuitively easy to match, for example, because of a larger visual overlap or limited appearance change.
+
+    ```py
+    from transformers import AutoImageProcessor, AutoModel
+    import torch
+    from PIL import Image
+    import requests
+    
+    processor = AutoImageProcessor.from_pretrained("ETH-CVG/lightglue_superpoint")
+    model = AutoModel.from_pretrained("ETH-CVG/lightglue_superpoint")
+    
+    # LightGlue requires pairs of images
+    images = [image1, image2]
+    inputs = processor(images, return_tensors="pt")
+    outputs = model(**inputs)
+    
+    # Extract matching information
+    keypoints0 = outputs.keypoints0  # Keypoints in first image
+    keypoints1 = outputs.keypoints1  # Keypoints in second image
+    matches = outputs.matches        # Matching indices
+    matching_scores = outputs.matching_scores  # Confidence scores
+    ```
+
+- The model outputs matching indices, keypoints, and confidence scores for each match, similar to SuperGlue but with improved efficiency.
+- For better visualization and analysis, use the [`LightGlueImageProcessor.post_process_keypoint_matching`] method to get matches in a more readable format.
+
+    ```py
+    # Process outputs for visualization
+    image_sizes = [[(image.height, image.width) for image in images]]
+    processed_outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
+    
+    for i, output in enumerate(processed_outputs):
+        print(f"For the image pair {i}")
+        for keypoint0, keypoint1, matching_score in zip(
+                output["keypoints0"], output["keypoints1"], output["matching_scores"]
+        ):
+            print(f"Keypoint at {keypoint0.numpy()} matches with keypoint at {keypoint1.numpy()} with score {matching_score}")
+    ```
+
+- Visualize the matches between the images using the built-in plotting functionality.
+
+    ```py
+    # Easy visualization using the built-in plotting method
+    processor.plot_keypoint_matching(images, processed_outputs)
+    ```
+
+<div class="flex justify-center">
+    <img src="https://cdn-uploads.huggingface.co/production/uploads/632885ba1558dac67c440aa8/duPp09ty8NRZlMZS18ccP.png">
+</div>
+
+## Resources
+
+- Refer to the [original LightGlue repository](https://github.com/cvg/LightGlue) for more examples and implementation details.
 
 ## LightGlueConfig
 
@@ -97,8 +130,13 @@ The original code can be found [here](https://github.com/cvg/LightGlue).
 - post_process_keypoint_matching
 - plot_keypoint_matching
 
+<frameworkcontent>
+<pt>
 ## LightGlueForKeypointMatching
 
 [[autodoc]] LightGlueForKeypointMatching
 
 - forward
+
+</pt>
+</frameworkcontent>

docs/source/en/model_doc/llava_next.md

@@ -14,287 +14,178 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# LLaVA-NeXT
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+  </div>
 </div>
 
-## Overview
+# LLaVA-NeXT
 
-The LLaVA-NeXT model was proposed in [LLaVA-NeXT: Improved reasoning, OCR, and world knowledge](https://llava-vl.github.io/blog/2024-01-30-llava-next/) by Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, Yong Jae Lee. LLaVa-NeXT (also called LLaVa-1.6) improves upon [LLaVa](llava) by increasing the input image resolution and training on an improved visual instruction tuning dataset to improve OCR and common sense reasoning.
+[LLaVA‑NeXT](https://llava-vl.github.io/blog/2024-05-10-llava-next-stronger-llms/) improves on [Llava](./llava) by increasing the input image resolution by 4x more pixels and supporting 3 aspect ratios (up to 672x672, 336x1344, 1344x336) to better grasp visual details. It is also trained on an improved visual instruction tuning dataset covering more scenarios and applications to improve OCR and common sense reasoning.
 
-The introduction from the blog is the following:
+You can find all the original LLaVA‑NeXT checkpoints under the [LLaVA-NeXT](https://huggingface.co/collections/llava-hf/llava-next-65f75c4afac77fd37dbbe6cf) collection.
 
-*In October 2023, we released LLaVA-1.5 with a simple and efficient design along with great performance on a benchmark suite of 12 datasets. It has since served as the foundation of many comprehensive studies of data, model, and capabilities of large multimodal models (LMM), and has enabled various new applications.
+> [!TIP]
+> This model was contributed by [nielsr](https://huggingface.co/nielsr).
+>
+> Click on the LLaVA‑NeXT models in the right sidebar for more examples of how to apply Llava-NeXT to different multimodal tasks.
 
-Today, we are thrilled to present LLaVA-NeXT, with improved reasoning, OCR, and world knowledge. LLaVA-NeXT even exceeds Gemini Pro on several benchmarks.
+The example below demonstrates how to generate text based on an image with [`Pipeline`] or the [`AutoModel`] class.
 
-Compared with LLaVA-1.5, LLaVA-NeXT has several improvements:
+<hfoptions id="usage">
 
-Increasing the input image resolution to 4x more pixels. This allows it to grasp more visual details. It supports three aspect ratios, up to 672x672, 336x1344, 1344x336 resolution.
-Better visual reasoning and OCR capability with an improved visual instruction tuning data mixture.
-Better visual conversation for more scenarios, covering different applications. Better world knowledge and logical reasoning.
-Efficient deployment and inference with SGLang.
-Along with performance improvements, LLaVA-NeXT maintains the minimalist design and data efficiency of LLaVA-1.5. It re-uses the pretrained connector of LLaVA-1.5, and still uses less than 1M visual instruction tuning samples. The largest 34B variant finishes training in ~1 day with 32 A100s.*
-
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/llava_next_overview.png"
-alt="drawing" width="600"/>
-
-<small> LLaVa-NeXT incorporates a higher input resolution by encoding various patches of the input image. Taken from the <a href="https://huggingface.co/papers/2310.03744">original paper.</a> </small>
-
-This model was contributed by [nielsr](https://huggingface.co/nielsr).
-The original code can be found [here](https://github.com/haotian-liu/LLaVA/tree/main).
-
-## Usage tips
-
-- We advise users to use `padding_side="left"` when computing batched generation as it leads to more accurate results. Simply make sure to call `processor.tokenizer.padding_side = "left"` before generating.
-
-<Tip warning={true}>
-
-- Llava-Next uses different number of patches for images and thus has to pad the inputs inside modeling code, aside from the padding done when processing the inputs. The default setting is "left-padding" if model is in `eval()` mode, otherwise "right-padding".
-
-</Tip>
-
-
-> [!NOTE]
-> LLaVA models after release v4.46 will raise warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you.
-Adding these attributes means that LLaVA will try to infer the number of image tokens required per image and expand the text with as many `<image>` placeholders as there will be tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there will be failure when merging the embeddings.
-The attributes can be obtained from model config, as `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`. The `num_additional_image_tokens` should be `1` if the vision backbone adds a CLS token or `0` if nothing extra is added to the vision patches.
-
-
-### Formatting Prompts with Chat Templates  
-
-Each **checkpoint** is trained with a specific prompt format, depending on the underlying large language model backbone. To ensure correct formatting, use the processor’s `apply_chat_template` method.  
-
-**Important:**  
-- You must construct a conversation history — passing a plain string won't work.  
-- Each message should be a dictionary with `"role"` and `"content"` keys.  
-- The `"content"` should be a list of dictionaries for different modalities like `"text"` and `"image"`.  
-
-
-Here’s an example of how to structure your input. We will use [llava-v1.6-mistral-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-mistral-7b-hf) and a conversation history of text and image.
+<hfoption id="Pipeline">
 
 ```python
-from transformers import LlavaNextProcessor
+import torch  
+from transformers import pipeline  
 
-processor = LlavaNextProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
-
-conversation = [
-    {
-        "role": "user",
-        "content": [
-            {"type": "image"},
-            {"type": "text", "text": "What’s shown in this image?"},
-        ],
-    },
-    {
-        "role": "assistant",
-        "content": [{"type": "text", "text": "This image shows a red stop sign."},]
-    },
-    {
-
-        "role": "user",
-        "content": [
-            {"type": "text", "text": "Describe the image in more details."},
-        ],
-    },
-]
-
-text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
-
-# Note that the template simply formats your prompt, you still have to tokenize it and obtain pixel values for your images
-print(text_prompt)
->>> "[INST] <image>\nWhat's shown in this image? [/INST] This image shows a red stop sign. [INST] Describe the image in more details. [/INST]"
+pipeline = pipeline(  
+    task="image-text-to-text",  
+    model="llava-hf/llava-v1.6-mistral-7b-hf",  
+    device=0,  
+    torch_dtype=torch.bfloat16  
+)  
+messages = [  
+    {  
+        "role": "user",  
+        "content": [  
+            {  
+                "type": "image",  
+                "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg",  
+            },  
+            { "type": "text", "text": "Describe this image."},  
+        ]  
+    }  
+]  
+pipeline(text=messages, max_new_tokens=20, return_full_text=False)
 ```
 
-- If you want to construct a chat prompt yourself, below is a list of possible formats
-.
-[llava-v1.6-mistral-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-mistral-7b-hf) requires the following format:
-```bash
-"[INST] <image>\nWhat is shown in this image? [/INST]"
+</hfoption>
+
+<hfoption id="AutoModel">
+
+```python
+import torch  
+import requests  
+from PIL import Image  
+from transformers import AutoProcessor, LlavaNextForConditionalGeneration  
+
+processor = AutoProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")  
+model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16).to("cuda")  
+
+url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"  
+image = Image.open(requests.get(url, stream=True).raw)  
+
+conversation = [  
+    {  
+        "role": "user",  
+        "content": [  
+            {"type": "image"},  
+            {"type": "text", "text": "What is shown in this image?"},  
+        ],  
+    },  
+]  
+prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)  
+inputs = processor(image, prompt, return_tensors="pt").to("cuda")  
+output = model.generate(**inputs, max_new_tokens=100)  
+print(processor.decode(output[0], skip_special_tokens=True))  
 ```
 
-[llava-v1.6-vicuna-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-vicuna-7b-hf) and [llava-v1.6-vicuna-13b-hf](https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf) require the following format:
-```bash
-"A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: <image>\nWhat is shown in this image? ASSISTANT:"
+</hfoption>
+
+</hfoptions>
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to int4.
+
+```python
+import torch  
+import requests  
+from PIL import Image  
+from transformers import AutoModelForImageTextToText, AutoProcessor, BitsAndBytesConfig  
+
+quant_config = BitsAndBytesConfig(  
+    load_in_4bit=True,  
+    bnb_4bit_compute_dtype=torch.float16,  
+    bnb_4bit_quant_type="nf4"  
+)  
+
+processor = AutoProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")  
+model = AutoModelForImageTextToText.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", quantization_config=quant_config, device_map="auto")  
+
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/llava_next_ocr.png"  
+image = Image.open(requests.get(url, stream=True).raw)  
+
+conversation = [  
+    {  
+        "role": "user",  
+        "content": [  
+            {"type": "image"},  
+            {"type": "text", "text": "What does this chart show?"},  
+        ],  
+    },  
+]  
+prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)  
+inputs = processor(image, prompt, return_tensors="pt").to("cuda")  
+
+with torch.inference_mode():  
+    output = model.generate(**inputs, max_new_tokens=100)  
+print(processor.decode(output[0], skip_special_tokens=True))  
 ```
 
-[llava-v1.6-34b-hf](https://huggingface.co/llava-hf/llava-v1.6-34b-hf) requires the following format:
-```bash
-"<|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nWhat is shown in this image?<|im_end|><|im_start|>assistant\n"
+
+## Notes
+
+* Different checkpoints (Mistral, Vicuna, etc.) require a specific prompt format depending on the underlying LLM. Always use [`~ProcessorMixin.apply_chat_template`] to ensure correct formatting. Refer to the [Templates](../chat_templating) guide for more details.
+
+* Set `padding_side="left"` during batched generation for more accurate results.
+
+```py
+processor.tokenizer.padding_side = "left"
 ```
 
-[llama3-llava-next-8b-hf](https://huggingface.co/llava-hf/llava-next-8b-hf) requires the following format:
+* LLaVA-NeXT uses different numbers of patches for images and pads the inputs inside the modeling code except when padding is done during processing. The default setting is *left-padding* if the model is in `eval()` mode, otherwise it is *right-padding*.
 
-```bash
-"<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful language and vision assistant. You are able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language.<|eot_id|><|start_header_id|><|start_header_id|>user<|end_header_id|>\n\n<image>\nWhat is shown in this image?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
-```
+* LLaVA models after v4.46 raises warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}`, and `processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add these attributes to the processor if you own the model checkpoint or open a PR if it isn't.
 
-[llava-next-72b-hf](https://huggingface.co/llava-hf/llava-next-72b-hf) and [llava-next-110b-hf](https://huggingface.co/llava-hf/llava-next-110b-hf) require the following format:
+  Adding these attributes means LLaVA will try to infer the number of image tokens required per image and expand the text with the same number of `<image>` token placeholders. There are usually ~500 tokens per image, so make sure the text is not truncated because it will cause a failure when merging the embeddings. The attributes can be found in `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`.
 
-```bash
-"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<image>\nWhat is shown in this image?<|im_end|>\n<|im_start|>assistant\n"
-```
+  The `num_additional_image_tokens` should be `1` if the vision backbone adds a `CLS` token or `0` if nothing extra is added.
 
-🚀 **Bonus:** If you're using `transformers>=4.49.0`, you can also get a vectorized output from `apply_chat_template`. See the **Usage Examples** below for more details on how to use it.
-
-
-
-## Usage example
-
-### Single image inference
-
-Here's how to load the model and perform inference in half-precision (`torch.float16`):
+* The example below demonstrates inference with multiple input images.
 
 ```python
 from transformers import LlavaNextProcessor, LlavaNextForConditionalGeneration
-import torch
 from PIL import Image
-import requests
+import requests, torch
 
 processor = LlavaNextProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
+model = LlavaNextForConditionalGeneration.from_pretrained(
+    "llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16
+).to("cuda")
 
-model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16)
-model.to("cuda:0")
+# Load multiple images
+url1 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/llava_next_ocr.png"
+url2 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/llava_next_comparison.png"
 
-# prepare image and text prompt, using the appropriate prompt template
-url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
-image = Image.open(requests.get(url, stream=True).raw)
+image1 = Image.open(requests.get(url1, stream=True).raw)
+image2 = Image.open(requests.get(url2, stream=True).raw)
 
 conversation = [
-    {
-        "role": "user",
-        "content": [
-            {"type": "image"},
-            {"type": "text", "text": "What is shown in this image?"},
-        ],
-    },
+    {"role": "user", "content": [{"type": "image"}, {"type": "image"}, {"type": "text", "text": "Compare these two images and describe the differences."}]}
 ]
 prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
-inputs = processor(image, prompt, return_tensors="pt").to("cuda:0")
+inputs = processor([image1, image2], prompt, return_tensors="pt").to("cuda")
 
-# autoregressively complete prompt
 output = model.generate(**inputs, max_new_tokens=100)
-
 print(processor.decode(output[0], skip_special_tokens=True))
 ```
 
-### Multi image inference
-
-LLaVa-Next can perform inference with multiple images as input, where images either belong to the same prompt or different prompts (in batched inference). Here is how you can do it:
-
-```python
-import requests
-from PIL import Image
-import torch
-from transformers import AutoProcessor, AutoModelForImageTextToText
-
-# Load the model in half-precision
-model = AutoModelForImageTextToText.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16, device_map="auto")
-processor = AutoProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
-
-# Get three different images
-url = "https://www.ilankelman.org/stopsigns/australia.jpg"
-image_stop = Image.open(requests.get(url, stream=True).raw)
-
-url = "http://images.cocodataset.org/val2017/000000039769.jpg"
-image_cats = Image.open(requests.get(url, stream=True).raw)
-
-url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.jpg"
-image_snowman = Image.open(requests.get(url, stream=True).raw)
-
-# Prepare a batch of two prompts, where the first one is a multi-turn conversation and the second is not
-conversation_1 = [
-    {
-        "role": "user",
-        "content": [
-            {"type": "image"},
-            {"type": "text", "text": "What is shown in this image?"},
-            ],
-    },
-    {
-        "role": "assistant",
-        "content": [
-            {"type": "text", "text": "There is a red stop sign in the image."},
-            ],
-    },
-    {
-        "role": "user",
-        "content": [
-            {"type": "image"},
-            {"type": "text", "text": "What about this image? How many cats do you see?"},
-            ],
-    },
-]
-
-conversation_2 = [
-    {
-        "role": "user",
-        "content": [
-            {"type": "image"},
-            {"type": "text", "text": "What is shown in this image?"},
-            ],
-    },
-]
-
-prompt_1 = processor.apply_chat_template(conversation_1, add_generation_prompt=True)
-prompt_2 = processor.apply_chat_template(conversation_2, add_generation_prompt=True)
-prompts = [prompt_1, prompt_2]
-
-# We can simply feed images in the order they have to be used in the text prompt
-# Each "<image>" token uses one image leaving the next for the subsequent "<image>" tokens
-inputs = processor(images=[image_stop, image_cats, image_snowman], text=prompts, padding=True, return_tensors="pt").to(model.device)
-
-# Generate
-generate_ids = model.generate(**inputs, max_new_tokens=30)
-processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
-```
-
-## Model optimization
-
-### Quantization using Bitsandbytes
-
-The model can be loaded in 8 or 4 bits, greatly reducing the memory requirements while maintaining the performance of the original model. First make sure to install bitsandbytes, `pip install bitsandbytes`, and to have access to a GPU/accelerator that is supported by the library.
-
-<Tip>
-
-bitsandbytes is being refactored to support multiple backends beyond CUDA. Currently, ROCm (AMD GPU) and Intel CPU implementations are mature, with Intel XPU in progress and Apple Silicon support expected by Q4/Q1. For installation instructions and the latest backend updates, visit [this link](https://huggingface.co/docs/bitsandbytes/main/en/installation#multi-backend).
-
-We value your feedback to help identify bugs before the full release! Check out [these docs](https://huggingface.co/docs/bitsandbytes/main/en/non_cuda_backends) for more details and feedback links.
-
-</Tip>
-
-Simply change the snippet above with:
-
-```python
-from transformers import AutoModelForImageTextToText, BitsAndBytesConfig
-
-# specify how to quantize the model
-quantization_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_quant_type="nf4",
-    bnb_4bit_compute_dtype=torch.float16,
-)
-
-model = AutoModelForImageTextToText.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", quantization_config=quantization_config, device_map="auto")
-```
-
-### Use Flash-Attention 2 to further speed-up generation
-
-First make sure to install flash-attn. Refer to the [original repository of Flash Attention](https://github.com/Dao-AILab/flash-attention) regarding that package installation. Simply change the snippet above with:
-
-```python
-from transformers import AutoModelForImageTextToText
-
-model = AutoModelForImageTextToText.from_pretrained(
-    model_id,
-    torch_dtype=torch.float16,
-    use_flash_attention_2=True
-).to(0)
-```
 
 ## LlavaNextConfig
 

docs/source/en/model_doc/mamba.md

@@ -28,6 +28,7 @@ You can find all the original Mamba checkpoints under the [State Space Models](h
 
 
 > [!TIP]
+> This model was contributed by [Molbap](https://huggingface.co/Molbap) and [AntonV](https://huggingface.co/AntonV).
 > Click on the Mamba models in the right sidebar for more examples of how to apply Mamba to different language tasks.
 
 The example below demonstrates how to generate text with [`Pipeline`], [`AutoModel`], and from the command line.
@@ -115,6 +116,13 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
   trainer.train()
    ```
 
+## MambaCache
+
+[[autodoc]] MambaCache
+    - update_conv_state
+    - update_ssm_state
+    - reset
+
 ## MambaConfig
 
 [[autodoc]] MambaConfig

docs/source/en/model_doc/mamba2.md

@@ -26,6 +26,7 @@ rendered properly in your Markdown viewer.
 You can find all the original Mamba 2 checkpoints under the [State Space Models](https://huggingface.co/state-spaces) organization, but the examples shown below use [mistralai/Mamba-Codestral-7B-v0.1](https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1) because a Hugging Face implementation isn't supported yet for the original checkpoints.
 
 > [!TIP]
+> This model was contributed by [ArthurZ](https://huggingface.co/ArthurZ).
 > Click on the Mamba models in the right sidebar for more examples of how to apply Mamba to different language tasks.
 
 The example below demonstrates how to generate text with [`Pipeline`], [`AutoModel`], and from the command line.

docs/source/en/model_doc/marian.md

@@ -14,159 +14,138 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# MarianMT
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
-<img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+        <img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
 ">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
 </div>
 
-## Overview
-
-A framework for translation models, using the same models as BART. Translations should be similar, but not identical to output in the test set linked to in each model card.
-This model was contributed by [sshleifer](https://huggingface.co/sshleifer).
+# MarianMT
 
 
-## Implementation Notes
 
-- Each model is about 298 MB on disk, there are more than 1,000 models.
-- The list of supported language pairs can be found [here](https://huggingface.co/Helsinki-NLP).
-- Models were originally trained by [Jörg Tiedemann](https://researchportal.helsinki.fi/en/persons/j%C3%B6rg-tiedemann) using the [Marian](https://marian-nmt.github.io/) C++ library, which supports fast training and translation.
-- All models are transformer encoder-decoders with 6 layers in each component. Each model's performance is documented
-  in a model card.
-- The 80 opus models that require BPE preprocessing are not supported.
-- The modeling code is the same as [`BartForConditionalGeneration`] with a few minor modifications:
+[MarianMT](https://huggingface.co/papers/1804.00344) is a machine translation model trained with the Marian framework which is written in pure C++. The framework includes its own custom auto-differentiation engine and efficient meta-algorithms to train encoder-decoder models like BART.
 
-  - static (sinusoid) positional embeddings (`MarianConfig.static_position_embeddings=True`)
-  - no layernorm_embedding (`MarianConfig.normalize_embedding=False`)
-  - the model starts generating with `pad_token_id` (which has 0 as a token_embedding) as the prefix (Bart uses
-    `<s/>`),
-- Code to bulk convert models can be found in `convert_marian_to_pytorch.py`.
+All MarianMT models are transformer encoder-decoders with 6 layers in each component, use static sinusoidal positional embeddings, don't have a layernorm embedding, and the model starts generating with the prefix `pad_token_id` instead of `<s/>`.
 
 
-## Naming
 
-- All model names use the following format: `Helsinki-NLP/opus-mt-{src}-{tgt}`
-- The language codes used to name models are inconsistent. Two digit codes can usually be found [here](https://developers.google.com/admin-sdk/directory/v1/languages), three digit codes require googling "language
-  code {code}".
-- Codes formatted like `es_AR` are usually `code_{region}`. That one is Spanish from Argentina.
-- The models were converted in two stages. The first 1000 models use ISO-639-2 codes to identify languages, the second
-  group use a combination of ISO-639-5 codes and ISO-639-2 codes.
+You can find all the original MarianMT checkpoints under the [Language Technology Research Group at the University of Helsinki](https://huggingface.co/Helsinki-NLP/models?search=opus-mt) organization.
 
 
-## Examples
+> [!TIP]
+> This model was contributed by [sshleifer](https://huggingface.co/sshleifer).
+>
+> Click on the MarianMT models in the right sidebar for more examples of how to apply MarianMT to translation tasks.
 
-- Since Marian models are smaller than many other translation models available in the library, they can be useful for
-  fine-tuning experiments and integration tests.
-- [Fine-tune on GPU](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/train_distil_marian_enro.sh)
 
-## Multilingual Models
+The example below demonstrates how to translate text using [`Pipeline`] or the [`AutoModel`] class.
 
-- All model names use the following format: `Helsinki-NLP/opus-mt-{src}-{tgt}`:
-- If a model can output multiple languages, and you should specify a language code by prepending the desired output
-  language to the `src_text`.
-- You can see a models's supported language codes in its model card, under target constituents, like in [opus-mt-en-roa](https://huggingface.co/Helsinki-NLP/opus-mt-en-roa).
-- Note that if a model is only multilingual on the source side, like `Helsinki-NLP/opus-mt-roa-en`, no language
-  codes are required.
-
-New multi-lingual models from the [Tatoeba-Challenge repo](https://github.com/Helsinki-NLP/Tatoeba-Challenge)
-require 3 character language codes:
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
 ```python
->>> from transformers import MarianMTModel, MarianTokenizer
 
->>> src_text = [
-...     ">>fra<< this is a sentence in english that we want to translate to french",
-...     ">>por<< This should go to portuguese",
-...     ">>esp<< And this to Spanish",
-... ]
+import torch
+from transformers import pipeline
 
->>> model_name = "Helsinki-NLP/opus-mt-en-roa"
->>> tokenizer = MarianTokenizer.from_pretrained(model_name)
->>> print(tokenizer.supported_language_codes)
-['>>zlm_Latn<<', '>>mfe<<', '>>hat<<', '>>pap<<', '>>ast<<', '>>cat<<', '>>ind<<', '>>glg<<', '>>wln<<', '>>spa<<', '>>fra<<', '>>ron<<', '>>por<<', '>>ita<<', '>>oci<<', '>>arg<<', '>>min<<']
+pipeline = pipeline("translation_en_to_de", model="Helsinki-NLP/opus-mt-en-de", torch_dtype=torch.float16, device=0)
+pipeline("Hello, how are you?")
 
->>> model = MarianMTModel.from_pretrained(model_name)
->>> translated = model.generate(**tokenizer(src_text, return_tensors="pt", padding=True))
->>> [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
-["c'est une phrase en anglais que nous voulons traduire en français",
- 'Isto deve ir para o português.',
- 'Y esto al español']
 ```
 
-Here is the code to see all available pretrained models on the hub:
+</hfoption>
+
+<hfoption id="AutoModel">
 
 ```python
-from huggingface_hub import list_models
 
-model_list = list_models()
-org = "Helsinki-NLP"
-model_ids = [x.id for x in model_list if x.id.startswith(org)]
-suffix = [x.split("/")[1] for x in model_ids]
-old_style_multi_models = [f"{org}/{s}" for s in suffix if s != s.lower()]
+import torch
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-de")
+model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-de", torch_dtype=torch.float16, attn_implementation="sdpa", device_map="auto")
+
+inputs = tokenizer("Hello, how are you?", return_tensors="pt").to("cuda")
+outputs = model.generate(**inputs, cache_implementation="static")
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+
 ```
 
-## Old Style Multi-Lingual Models
+</hfoption>
+</hfoptions>
 
-These are the old style multi-lingual models ported from the OPUS-MT-Train repo: and the members of each language
-group:
-
-```python no-style
-['Helsinki-NLP/opus-mt-NORTH_EU-NORTH_EU',
- 'Helsinki-NLP/opus-mt-ROMANCE-en',
- 'Helsinki-NLP/opus-mt-SCANDINAVIA-SCANDINAVIA',
- 'Helsinki-NLP/opus-mt-de-ZH',
- 'Helsinki-NLP/opus-mt-en-CELTIC',
- 'Helsinki-NLP/opus-mt-en-ROMANCE',
- 'Helsinki-NLP/opus-mt-es-NORWAY',
- 'Helsinki-NLP/opus-mt-fi-NORWAY',
- 'Helsinki-NLP/opus-mt-fi-ZH',
- 'Helsinki-NLP/opus-mt-fi_nb_no_nn_ru_sv_en-SAMI',
- 'Helsinki-NLP/opus-mt-sv-NORWAY',
- 'Helsinki-NLP/opus-mt-sv-ZH']
-GROUP_MEMBERS = {
- 'ZH': ['cmn', 'cn', 'yue', 'ze_zh', 'zh_cn', 'zh_CN', 'zh_HK', 'zh_tw', 'zh_TW', 'zh_yue', 'zhs', 'zht', 'zh'],
- 'ROMANCE': ['fr', 'fr_BE', 'fr_CA', 'fr_FR', 'wa', 'frp', 'oc', 'ca', 'rm', 'lld', 'fur', 'lij', 'lmo', 'es', 'es_AR', 'es_CL', 'es_CO', 'es_CR', 'es_DO', 'es_EC', 'es_ES', 'es_GT', 'es_HN', 'es_MX', 'es_NI', 'es_PA', 'es_PE', 'es_PR', 'es_SV', 'es_UY', 'es_VE', 'pt', 'pt_br', 'pt_BR', 'pt_PT', 'gl', 'lad', 'an', 'mwl', 'it', 'it_IT', 'co', 'nap', 'scn', 'vec', 'sc', 'ro', 'la'],
- 'NORTH_EU': ['de', 'nl', 'fy', 'af', 'da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],
- 'SCANDINAVIA': ['da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],
- 'SAMI': ['se', 'sma', 'smj', 'smn', 'sms'],
- 'NORWAY': ['nb_NO', 'nb', 'nn_NO', 'nn', 'nog', 'no_nb', 'no'],
- 'CELTIC': ['ga', 'cy', 'br', 'gd', 'kw', 'gv']
-}
-```
-
-Example of translating english to many romance languages, using old-style 2 character language codes
 
+Use the [AttentionMaskVisualizer](https://github.com/huggingface/transformers/blob/beb9b5b02246b9b7ee81ddf938f93f44cfeaad19/src/transformers/utils/attention_visualizer.py#L139) to better understand what tokens the model can and cannot attend to.
 
 ```python
->>> from transformers import MarianMTModel, MarianTokenizer
+from transformers.utils.attention_visualizer import AttentionMaskVisualizer
 
->>> src_text = [
-...     ">>fr<< this is a sentence in english that we want to translate to french",
-...     ">>pt<< This should go to portuguese",
-...     ">>es<< And this to Spanish",
-... ]
-
->>> model_name = "Helsinki-NLP/opus-mt-en-ROMANCE"
->>> tokenizer = MarianTokenizer.from_pretrained(model_name)
-
->>> model = MarianMTModel.from_pretrained(model_name)
->>> translated = model.generate(**tokenizer(src_text, return_tensors="pt", padding=True))
->>> tgt_text = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
-["c'est une phrase en anglais que nous voulons traduire en français",
- 'Isto deve ir para o português.',
- 'Y esto al español']
+visualizer = AttentionMaskVisualizer("Helsinki-NLP/opus-mt-en-de")
+visualizer("Hello, how are you?")
 ```
+<div class="flex justify-center">
+   <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/marianmt-attn-mask.png"/>
+</div>
 
-## Resources
+## Notes
 
-- [Translation task guide](../tasks/translation)
-- [Summarization task guide](../tasks/summarization)
-- [Causal language modeling task guide](../tasks/language_modeling)
+- MarianMT models are ~298MB on disk and there are more than 1000 models. Check this [list](https://huggingface.co/Helsinki-NLP) for supported language pairs. The language codes may be inconsistent. Two digit codes can be found [here](https://developers.google.com/admin-sdk/directory/v1/languages) while three digit codes may require further searching.
+- Models that require BPE preprocessing are not supported.
+- All model names use the following format: `Helsinki-NLP/opus-mt-{src}-{tgt}`. Language codes formatted like `es_AR` usually refer to the `code_{region}`. For example, `es_AR` refers to Spanish from Argentina.
+- If a model can output multiple languages, prepend the desired output language to `src_txt` as shown below. New multilingual models from the [Tatoeba-Challenge](https://github.com/Helsinki-NLP/Tatoeba-Challenge) require 3 character language codes.
+
+```python
+
+from transformers import MarianMTModel, MarianTokenizer
+
+# Model trained on multiple source languages → multiple target languages
+# Example: multilingual to Arabic (arb)
+model_name = "Helsinki-NLP/opus-mt-mul-mul"  # Tatoeba Challenge model
+tokenizer = MarianTokenizer.from_pretrained(model_name)
+model = MarianMTModel.from_pretrained(model_name)
+
+# Prepend the desired output language code (3-letter ISO 639-3)
+src_texts = ["arb>> Hello, how are you today?"]
+
+# Tokenize and translate
+inputs = tokenizer(src_texts, return_tensors="pt", padding=True, truncation=True)
+translated = model.generate(**inputs)
+
+# Decode and print result
+translated_texts = tokenizer.batch_decode(translated, skip_special_tokens=True)
+print(translated_texts[0])
+
+```
+   
+- Older multilingual models use 2 character language codes.
+
+```python
+
+from transformers import MarianMTModel, MarianTokenizer
+
+# Example: older multilingual model (like en → many)
+model_name = "Helsinki-NLP/opus-mt-en-ROMANCE"  # English → French, Spanish, Italian, etc.
+tokenizer = MarianTokenizer.from_pretrained(model_name)
+model = MarianMTModel.from_pretrained(model_name)
+
+# Prepend the 2-letter ISO 639-1 target language code (older format)
+src_texts = [">>fr<< Hello, how are you today?"]
+
+# Tokenize and translate
+inputs = tokenizer(src_texts, return_tensors="pt", padding=True, truncation=True)
+translated = model.generate(**inputs)
+
+# Decode and print result
+translated_texts = tokenizer.batch_decode(translated, skip_special_tokens=True)
+print(translated_texts[0])
+
+```
 
 ## MarianConfig
 

docs/source/en/model_doc/mistral.md

@@ -139,6 +139,10 @@ Use the [AttentionMaskVisualizer](https://github.com/huggingface/transformers/bl
 
 [[autodoc]] MistralConfig
 
+## MistralCommonTokenizer
+
+[[autodoc]] MistralCommonTokenizer
+
 ## MistralModel
 
 [[autodoc]] MistralModel

docs/source/en/model_doc/mistral3.md

@@ -227,6 +227,10 @@ This example also how to use `BitsAndBytes` to load the model in 4bit quantizati
 
 [[autodoc]] Mistral3Config
 
+## MistralCommonTokenizer
+
+[[autodoc]] MistralCommonTokenizer
+
 ## Mistral3Model
 
 [[autodoc]] Mistral3Model

docs/source/en/model_doc/mixtral.md

@@ -197,6 +197,10 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 
 [[autodoc]] MixtralConfig
 
+## MistralCommonTokenizer
+
+[[autodoc]] MistralCommonTokenizer
+
 ## MixtralModel
 
 [[autodoc]] MixtralModel

docs/source/en/model_doc/mobilenet_v2.md

@@ -114,6 +114,7 @@ print(f"The predicted class label is: {predicted_class_label}")
 
 [[autodoc]] MobileNetV2ImageProcessor
     - preprocess
+    - post_process_semantic_segmentation
 
 ## MobileNetV2ImageProcessorFast
 

docs/source/en/model_doc/modernbert-decoder.md

@@ -0,0 +1,188 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+    <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+    <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+  </div>
+</div>
+
+# ModernBERT Decoder
+
+ModernBERT Decoder has the same architecture as [ModernBERT](https://huggingface.co/papers/2412.13663) but it is trained from scratch with a causal language modeling objective from the [Ettin paper](https://huggingface.co/papers/2507.11412). This allows for using the same architecture to compare encoders and decoders. This model is the decoder architecture implementation of ModernBERT, designed for autoregressive text generation tasks.
+
+ModernBERT Decoder uses sliding window attention and rotary positional embeddings for efficiency and to handle longer sequences.
+
+You can find all the original ModernBERT Decoder checkpoints under the [jhu-clsp](https://huggingface.co/collections/jhu-clsp/encoders-vs-decoders-the-ettin-suite-686303e16142257eed8e6aeb) collection.
+
+> [!TIP]
+> This model was contributed by [orionw](https://huggingface.co/orionweller).
+>
+> Click on the ModernBERT Decoder models in the right sidebar for more examples of how to apply ModernBERT Decoder to different text generation tasks.
+
+The example below demonstrates how to use ModernBERT Decoder for text generation with [`Pipeline`], [`AutoModel`] (with and without quantization), and from the command line. 
+
+<hfoptions id="usage">
+<hfoption id="Pipeline">
+
+```py
+import torch
+from transformers import pipeline
+
+generator = pipeline(
+    task="text-generation",
+    model="jhu-clsp/ettin-decoder-17m",
+    torch_dtype=torch.float16,
+    device=0
+)
+generator("The future of artificial intelligence is", max_length=50, num_return_sequences=1)
+
+# For sequence classification
+classifier = pipeline(
+    task="text-classification",
+    model="jhu-clsp/ettin-decoder-17m",
+    torch_dtype=torch.float16,
+    device=0
+)
+classifier("This movie is really great!")
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/ettin-decoder-17m")
+model = AutoModelForCausalLM.from_pretrained(
+    "jhu-clsp/ettin-decoder-17m",
+    torch_dtype=torch.float16,
+    device_map="auto",
+)
+
+prompt = "The future of artificial intelligence is"
+inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    outputs = model.generate(
+        **inputs,
+        max_length=50,
+        num_return_sequences=1,
+        temperature=0.7,
+        do_sample=True,
+        pad_token_id=tokenizer.eos_token_id
+    )
+
+generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+print(f"Generated text: {generated_text}")
+
+# For sequence classification
+from transformers import AutoModelForSequenceClassification
+
+classifier_model = AutoModelForSequenceClassification.from_pretrained(
+    "jhu-clsp/ettin-decoder-17m",
+    torch_dtype=torch.float16,
+    device_map="auto",
+    num_labels=2
+)
+
+text = "This movie is really great!"
+inputs = tokenizer(text, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    outputs = classifier_model(**inputs)
+    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
+    predicted_class = torch.argmax(predictions, dim=-1)
+
+print(f"Predicted class: {predicted_class.item()}")
+print(f"Prediction probabilities: {predictions}")
+```
+
+</hfoption>
+
+<hfoption id="AutoModel (w/quantization)">
+
+```
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+quantization_config = BitsAndBytesConfig(
+    load_in_8bit=True,
+)
+
+tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/ettin-decoder-1b")
+model = AutoModelForCausalLM.from_pretrained(
+    "jhu-clsp/ettin-decoder-1b",
+    torch_dtype=torch.float16,
+    device_map="auto",
+    quantization_config=quantization_config
+)
+
+prompt = "The future of artificial intelligence is"
+inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    outputs = model.generate(
+        **inputs,
+        max_length=50,
+        num_return_sequences=1,
+        temperature=0.7,
+        do_sample=True,
+        pad_token_id=tokenizer.eos_token_id
+    )
+
+generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+print(f"Generated text: {generated_text}")
+```
+</hfoption>
+
+<hfoption id="transformers CLI">
+
+```bash
+echo "The future of artificial intelligence is" | transformers run --task text-generation --model jhu-clsp/ettin-decoder-17m --device 0
+```
+
+</hfoption>
+</hfoptions>
+
+
+## ModernBertDecoderConfig
+
+[[autodoc]] ModernBertDecoderConfig
+
+<frameworkcontent>
+<pt>
+
+## ModernBertDecoderModel
+
+[[autodoc]] ModernBertDecoderModel
+    - forward
+
+## ModernBertDecoderForCausalLM
+
+[[autodoc]] ModernBertDecoderForCausalLM
+    - forward
+
+## ModernBertDecoderForSequenceClassification
+
+[[autodoc]] ModernBertDecoderForSequenceClassification
+    - forward
+
+</pt>
+</frameworkcontent>

docs/source/en/model_doc/olmoe.md

@@ -14,27 +14,89 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# OLMoE
-
+<div style="float: right;">
 <div class="flex flex-wrap space-x-1">
 <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
 <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
 <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
 </div>
+</div>
 
-## Overview
+# OLMoE
 
-The OLMoE model was proposed in [OLMoE: Open Mixture-of-Experts Language Models](https://huggingface.co/papers/2409.02060) by Niklas Muennighoff, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Jacob Morrison, Sewon Min, Weijia Shi, Pete Walsh, Oyvind Tafjord, Nathan Lambert, Yuling Gu, Shane Arora, Akshita Bhagia, Dustin Schwenk, David Wadden, Alexander Wettig, Binyuan Hui, Tim Dettmers, Douwe Kiela, Ali Farhadi, Noah A. Smith, Pang Wei Koh, Amanpreet Singh, Hannaneh Hajishirzi.
+[OLMoE](https://huggingface.co/papers/2409.02060) is a sparse Mixture-of-Experts (MoE) language model with 7B parameters but only 1B parameters are used per input token. It has similar inference costs as dense models but trains ~3x faster. OLMoE uses fine-grained routing with 64 small experts in each layer and uses a dropless token-based routing algorithm.
 
-OLMoE is a series of **O**pen **L**anguage **Mo**dels using sparse **M**ixture-**o**f-**E**xperts designed to enable the science of language models. We release all code, checkpoints, logs, and details involved in training these models.
+You can find all the original OLMoE checkpoints under the [OLMoE](https://huggingface.co/collections/allenai/olmoe-november-2024-66cf678c047657a30c8cd3da) collection.
 
-The abstract from the paper is the following:
+> [!TIP]
+> This model was contributed by [Muennighoff](https://hf.co/Muennighoff).
+>
+> Click on the OLMoE models in the right sidebar for more examples of how to apply OLMoE to different language tasks.
 
-*We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2-13B-Chat and DeepSeekMoE-16B. We present various experiments on MoE training, analyze routing in our model showing high specialization, and open-source all aspects of our work: model weights, training data, code, and logs.*
+The example below demonstrates how to generate text with [`Pipeline`] or the [`AutoModel`] class.
 
-This model was contributed by [Muennighoff](https://hf.co/Muennighoff).
-The original code can be found [here](https://github.com/allenai/OLMoE).
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
+```py
+import torch
+from transformers import pipeline
+
+pipe = pipeline(
+    task="text-generation",
+    model="allenai/OLMoE-1B-7B-0125",
+    torch_dtype=torch.float16,
+    device=0,
+)
+
+result = pipe("Dionysus is the god of")
+print(result)
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+model = AutoModelForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0924", attn_implementation="sdpa", torch_dtype="auto", device_map="auto").to(device)
+tokenizer = AutoTokenizer.from_pretrained("allenai/OLMoE-1B-7B-0924")
+
+inputs = tokenizer("Bitcoin is", return_tensors="pt")
+inputs = {k: v.to(device) for k, v in inputs.items()}
+output = model.generate(**inputs, max_length=64)
+print(tokenizer.decode(output[0]))
+```
+
+## Quantization
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to 4-bits.
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+quantization_config = BitsAndBytesConfig(
+   load_in_4bit=True,
+   bnb_4bit_compute_dtype=torch.float16,
+   bnb_4bit_use_double_quant=True,
+   bnb_4bit_quant_type="nf4"
+)
+
+model = AutoModelForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0924", attn_implementation="sdpa", torch_dtype="auto", device_map="auto", quantization_config=quantization_config).to(device)
+tokenizer = AutoTokenizer.from_pretrained("allenai/OLMoE-1B-7B-0924")
+
+inputs = tokenizer("Bitcoin is", return_tensors="pt")
+inputs = {k: v.to(device) for k, v in inputs.items()}
+output = model.generate(**inputs, max_length=64)
+print(tokenizer.decode(output[0]))
+```
 
 ## OlmoeConfig
 

docs/source/en/model_doc/perception_lm.md

@@ -0,0 +1,68 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PerceptionLM
+
+## Overview
+
+The PerceptionLM model was proposed in [PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding](https://ai.meta.com/research/publications/perceptionlm-open-access-data-and-models-for-detailed-visual-understanding/) by Jang Hyun Cho et al. It's a fully open, reproducible model for transparent research in image and video understanding. PLM consists of
+a vision encoder with a small scale (<8B parameters) LLM decoder.
+
+The abstract from the paper is the following:
+
+*Vision-language models are integral to computer vision research, yet many high-performing models
+remain closed-source, obscuring their data, design and training recipe. The research community
+has responded by using distillation from black-box models to label training data, achieving strong
+benchmark results, at the cost of measurable scientific progress. However, without knowing the details
+of the teacher model and its data sources, scientific progress remains difficult to measure. In this
+paper, we study building a Perception Language Model (PLM) in a fully open and reproducible
+framework for transparent research in image and video understanding. We analyze standard training
+pipelines without distillation from proprietary models and explore large-scale synthetic data to identify
+critical data gaps, particularly in detailed video understanding. To bridge these gaps, we release 2.8M
+human-labeled instances of fine-grained video question-answer pairs and spatio-temporally grounded
+video captions. Additionally, we introduce PLM–VideoBench, a suite for evaluating challenging video
+understanding tasks focusing on the ability to reason about “what”, “where”, “when”, and “how” of a
+video. We make our work fully reproducible by providing data, training recipes, code & models.*
+
+
+This model was contributed by [shumingh](https://huggingface.co/shumingh).
+The original code can be found [here](https://github.com/facebookresearch/perception_models).
+
+
+## PerceptionLMConfig
+
+[[autodoc]] PerceptionLMConfig
+
+## PerceptionLMProcessor
+
+[[autodoc]] PerceptionLMProcessor
+
+## PerceptionLMImageProcessorFast
+
+[[autodoc]] PerceptionLMImageProcessorFast
+
+## PerceptionLMVideoProcessor
+
+[[autodoc]] PerceptionLMVideoProcessor
+
+## PerceptionLMModel
+
+[[autodoc]] PerceptionLMModel
+
+## PerceptionLMForConditionalGeneration
+
+[[autodoc]] PerceptionLMForConditionalGeneration
+    - forward

docs/source/en/model_doc/phi4_multimodal.md

@@ -9,44 +9,53 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Phi4 Multimodal
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-EE4C2C?logo=pytorch&logoColor=white&style=flat">
+  </div>
+</div>
 
-## Overview
+## Phi4 Multimodal
 
-Phi4 Multimodal is a lightweight open multimodal foundation model that leverages the language, vision, and speech research and datasets used for Phi-3.5 and 4.0 models. The model processes text, image, and audio inputs, generating text outputs, and comes with 128K token context length. The model underwent an enhancement process, incorporating both supervised fine-tuning, direct preference optimization and RLHF (Reinforcement Learning from Human Feedback) to support precise instruction adherence and safety measures. The languages that each modal supports are the following:
+[Phi4 Multimodal](https://huggingface.co/papers/2503.01743) is a multimodal model capable of text, image, and speech and audio inputs or any combination of these. It features a mixture of LoRA adapters for handling different inputs, and each input is routed to the appropriate encoder.
 
-- Text: Arabic, Chinese, Czech, Danish, Dutch, English, Finnish, French, German, Hebrew, Hungarian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish, Ukrainian
-- Vision: English
-- Audio: English, Chinese, German, French, Italian, Japanese, Spanish, Portuguese
+You can find all the original Phi4 Multimodal checkpoints under the [Phi4](https://huggingface.co/collections/microsoft/phi-4-677e9380e514feb5577a40e4) collection.
 
-This model was contributed by [Cyril Vallez](https://huggingface.co/cyrilvallez). The most recent code can be
-found [here](https://github.com/huggingface/transformers/blob/main/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py).
+> [!TIP]
+> This model was contributed by [cyrilvallez](https://huggingface.co/cyrilvallez).
+>
+> Click on the Phi-4 Multimodal in the right sidebar for more examples of how to apply Phi-4 Multimodal to different tasks.
 
+The example below demonstrates how to generate text based on an image with [`Pipeline`] or the [`AutoModel`] class.
 
-## Usage tips
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-`Phi4-multimodal-instruct` can be found on the [Huggingface Hub](https://huggingface.co/microsoft/Phi-4-multimodal-instruct)
+```python
+from transformers import pipeline
+generator = pipeline("text-generation", model="microsoft/Phi-4-multimodal-instruct", torch_dtype="auto", device=0)
 
-In the following, we demonstrate how to use it for inference depending on the input modalities (text, image, audio).
+prompt = "Explain the concept of multimodal AI in simple terms."
+
+result = generator(prompt, max_length=50)
+print(result[0]['generated_text'])
+```
+
+</hfoption>
+<hfoption id="AutoModel">
 
 ```python
 import torch
 from transformers import AutoModelForCausalLM, AutoProcessor, GenerationConfig
 
-
-# Define model path
 model_path = "microsoft/Phi-4-multimodal-instruct"
 device = "cuda:0"
 
-# Load model and processor
 processor = AutoProcessor.from_pretrained(model_path)
-model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device,  torch_dtype=torch.float16)
+model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device, torch_dtype=torch.float16)
 
-# Optional: load the adapters (note that without them, the base model will very likely not work well)
-model.load_adapter(model_path, adapter_name="speech", device_map=device, adapter_kwargs={"subfolder": 'speech-lora'})
 model.load_adapter(model_path, adapter_name="vision", device_map=device, adapter_kwargs={"subfolder": 'vision-lora'})
 
-# Part : Image Processing
 messages = [
     {
         "role": "user",
@@ -57,7 +66,7 @@ messages = [
     },
 ]
 
-model.set_adapter("vision") # if loaded, activate the vision adapter
+model.set_adapter("vision")
 inputs = processor.apply_chat_template(
     messages,
     add_generation_prompt=True,
@@ -66,7 +75,6 @@ inputs = processor.apply_chat_template(
     return_tensors="pt",
 ).to(device)
 
-# Generate response
 generate_ids = model.generate(
     **inputs,
     max_new_tokens=1000,
@@ -77,10 +85,27 @@ response = processor.batch_decode(
     generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
 )[0]
 print(f'>>> Response\n{response}')
+```
 
+</hfoption>
+</hfoptions>
 
-# Part 2: Audio Processing
-model.set_adapter("speech") # if loaded, activate the speech adapter
+## Notes
+
+The example below demonstrates inference with an audio and text input.
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoProcessor, GenerationConfig
+
+model_path = "microsoft/Phi-4-multimodal-instruct"
+device = "cuda:0"
+
+processor = AutoProcessor.from_pretrained(model_path)
+model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device,  torch_dtype=torch.float16)
+
+model.load_adapter(model_path, adapter_name="speech", device_map=device, adapter_kwargs={"subfolder": 'speech-lora'})
+model.set_adapter("speech")
 audio_url = "https://upload.wikimedia.org/wikipedia/commons/b/b0/Barbara_Sahakian_BBC_Radio4_The_Life_Scientific_29_May_2012_b01j5j24.flac"
 messages = [
     {
@@ -110,6 +135,7 @@ response = processor.batch_decode(
     generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
 )[0]
 print(f'>>> Response\n{response}')
+
 ```
 
 ## Phi4MultimodalFeatureExtractor

docs/source/en/model_doc/pixtral.md

@@ -86,6 +86,10 @@ output = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up
 
 [[autodoc]] PixtralVisionConfig
 
+## MistralCommonTokenizer
+
+[[autodoc]] MistralCommonTokenizer
+
 ## PixtralVisionModel
 
 [[autodoc]] PixtralVisionModel

docs/source/en/model_doc/sam.md

@@ -25,7 +25,7 @@ rendered properly in your Markdown viewer.
 
 SAM (Segment Anything Model) was proposed in [Segment Anything](https://huggingface.co/papers/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
 
-The model can be used to predict segmentation masks of any object of interest given an input image. 
+The model can be used to predict segmentation masks of any object of interest given an input image.
 
 ![example image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/sam-output.png)
 
@@ -37,9 +37,9 @@ Tips:
 
 - The model predicts binary masks that states the presence or not of the object of interest given an image.
 - The model predicts much better results if input 2D points and/or input bounding boxes are provided
-- You can prompt multiple points for the same image, and predict a single mask. 
+- You can prompt multiple points for the same image, and predict a single mask.
 - Fine-tuning the model is not supported yet
-- According to the paper, textual input should be also supported. However, at this time of writing this seems not to be supported according to [the official repository](https://github.com/facebookresearch/segment-anything/issues/4#issuecomment-1497626844). 
+- According to the paper, textual input should be also supported. However, at this time of writing this seems not to be supported according to [the official repository](https://github.com/facebookresearch/segment-anything/issues/4#issuecomment-1497626844).
 
 
 This model was contributed by [ybelkada](https://huggingface.co/ybelkada) and [ArthurZ](https://huggingface.co/ArthurZ).
@@ -149,6 +149,11 @@ alt="drawing" width="900"/>
 [[autodoc]] SamImageProcessor
 
 
+## SamImageProcessorFast
+
+[[autodoc]] SamImageProcessorFast
+
+
 ## SamVisionModel
 
 [[autodoc]] SamVisionModel

docs/source/en/model_doc/speech_to_text_2.md

@@ -61,19 +61,16 @@ predicted token ids.
 - Step-by-step Speech Translation
 
 ```python
->>> import torch
 >>> from transformers import Speech2Text2Processor, SpeechEncoderDecoderModel
 >>> from datasets import load_dataset
->>> from torchcodec.decoders import AudioDecoder
 
 >>> model = SpeechEncoderDecoderModel.from_pretrained("facebook/s2t-wav2vec2-large-en-de")
 >>> processor = Speech2Text2Processor.from_pretrained("facebook/s2t-wav2vec2-large-en-de")
 
 
->>> def map_to_array(batch):
-...     decoder = AudioDecoder(batch["file"])
-...     batch["speech"] = decoder.get_all_samples().data
-...     return batch
+>>> def map_to_array(example):
+...     example["speech"] = example["audio"]["array"]
+...     return example
 
 
 >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")

docs/source/en/model_doc/superglue.md

@@ -10,40 +10,31 @@ specific language governing permissions and limitations under the License.
 ⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
 rendered properly in your Markdown viewer.
 
-
 -->
 
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white" >
+    </div>
+</div>
+
 # SuperGlue
 
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-</div>
+[SuperGlue](https://huggingface.co/papers/1911.11763) is a neural network that matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs are predicted by a graph neural network. SuperGlue introduces a flexible context aggregation mechanism based on attention, enabling it to reason about the underlying 3D scene and feature assignments jointly. Paired with the [SuperPoint model](https://huggingface.co/magic-leap-community/superpoint), it can be used to match two images and estimate the pose between them. This model is useful for tasks such as image matching, homography estimation, etc.
 
-## Overview
+You can find all the original SuperGlue checkpoints under the [Magic Leap Community](https://huggingface.co/magic-leap-community) organization.
 
-The SuperGlue model was proposed in [SuperGlue: Learning Feature Matching with Graph Neural Networks](https://huggingface.co/papers/1911.11763) by Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz and Andrew Rabinovich.
+> [!TIP]
+> This model was contributed by [stevenbucaille](https://huggingface.co/stevenbucaille).
+>
+> Click on the SuperGlue models in the right sidebar for more examples of how to apply SuperGlue to different computer vision tasks.
 
-This model consists of matching two sets of interest points detected in an image. Paired with the 
-[SuperPoint model](https://huggingface.co/magic-leap-community/superpoint), it can be used to match two images and 
-estimate the pose between them. This model is useful for tasks such as image matching, homography estimation, etc.
+The example below demonstrates how to match keypoints between two images with the [`AutoModel`] class.
 
-The abstract from the paper is the following:
+<hfoptions id="usage">
+<hfoption id="AutoModel">
 
-*This paper introduces SuperGlue, a neural network that matches two sets of local features by jointly finding correspondences 
-and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs 
-are predicted by a graph neural network. We introduce a flexible context aggregation mechanism based on attention, enabling 
-SuperGlue to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, 
-our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from image 
-pairs. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on the task of pose estimation in 
-challenging real-world indoor and outdoor environments. The proposed method performs matching in real-time on a modern GPU and 
-can be readily integrated into modern SfM or SLAM systems. The code and trained weights are publicly available at this [URL](https://github.com/magicleap/SuperGluePretrainedNetwork).*
-
-## How to use
-
-Here is a quick example of using the model. Since this model is an image matching model, it requires pairs of images to be matched. 
-The raw outputs contain the list of keypoints detected by the keypoint detector as well as the list of matches with their corresponding 
-matching scores.
-```python
+```py
 from transformers import AutoImageProcessor, AutoModel
 import torch
 from PIL import Image
@@ -52,7 +43,7 @@ import requests
 url_image1 = "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/united_states_capitol_98169888_3347710852.jpg"
 image1 = Image.open(requests.get(url_image1, stream=True).raw)
 url_image2 = "https://raw.githubusercontent.com/magicleap/SuperGluePretrainedNetwork/refs/heads/master/assets/phototourism_sample_images/united_states_capitol_26757027_6717084061.jpg"
-image_2 = Image.open(requests.get(url_image2, stream=True).raw)
+image2 = Image.open(requests.get(url_image2, stream=True).raw)
 
 images = [image1, image2]
 
@@ -62,67 +53,97 @@ model = AutoModel.from_pretrained("magic-leap-community/superglue_outdoor")
 inputs = processor(images, return_tensors="pt")
 with torch.no_grad():
     outputs = model(**inputs)
-```
 
-You can use the `post_process_keypoint_matching` method from the `SuperGlueImageProcessor` to get the keypoints and matches in a more readable format:
-
-```python
+# Post-process to get keypoints and matches
 image_sizes = [[(image.height, image.width) for image in images]]
-outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
-for i, output in enumerate(outputs):
-    print("For the image pair", i)
-    for keypoint0, keypoint1, matching_score in zip(
-            output["keypoints0"], output["keypoints1"], output["matching_scores"]
-    ):
-        print(
-            f"Keypoint at coordinate {keypoint0.numpy()} in the first image matches with keypoint at coordinate {keypoint1.numpy()} in the second image with a score of {matching_score}."
+processed_outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
+```
+
+</hfoption>
+</hfoptions>
+
+## Notes
+
+- SuperGlue performs feature matching between two images simultaneously, requiring pairs of images as input.
+
+    ```python
+    from transformers import AutoImageProcessor, AutoModel
+    import torch
+    from PIL import Image
+    import requests
+    
+    processor = AutoImageProcessor.from_pretrained("magic-leap-community/superglue_outdoor")
+    model = AutoModel.from_pretrained("magic-leap-community/superglue_outdoor")
+    
+    # SuperGlue requires pairs of images
+    images = [image1, image2]
+    inputs = processor(images, return_tensors="pt")
+    outputs = model(**inputs)
+    
+    # Extract matching information
+    keypoints0 = outputs.keypoints0  # Keypoints in first image
+    keypoints1 = outputs.keypoints1  # Keypoints in second image
+    matches = outputs.matches        # Matching indices
+    matching_scores = outputs.matching_scores  # Confidence scores
+    ```
+
+- The model outputs matching indices, keypoints, and confidence scores for each match.
+- For better visualization and analysis, use the [`SuperGlueImageProcessor.post_process_keypoint_matching`] method to get matches in a more readable format.
+
+    ```py
+    # Process outputs for visualization
+    image_sizes = [[(image.height, image.width) for image in images]]
+    processed_outputs = processor.post_process_keypoint_matching(outputs, image_sizes, threshold=0.2)
+    
+    for i, output in enumerate(processed_outputs):
+        print(f"For the image pair {i}")
+        for keypoint0, keypoint1, matching_score in zip(
+                output["keypoints0"], output["keypoints1"], output["matching_scores"]
+        ):
+            print(f"Keypoint at {keypoint0.numpy()} matches with keypoint at {keypoint1.numpy()} with score {matching_score}")
+    ```
+
+- The example below demonstrates how to visualize matches between two images.
+
+    ```py
+    import matplotlib.pyplot as plt
+    import numpy as np
+
+    # Create side by side image
+    merged_image = np.zeros((max(image1.height, image2.height), image1.width + image2.width, 3))
+    merged_image[: image1.height, : image1.width] = np.array(image1) / 255.0
+    merged_image[: image2.height, image1.width :] = np.array(image2) / 255.0
+    plt.imshow(merged_image)
+    plt.axis("off")
+
+    # Retrieve the keypoints and matches
+    output = processed_outputs[0]
+    keypoints0 = output["keypoints0"]
+    keypoints1 = output["keypoints1"]
+    matching_scores = output["matching_scores"]
+
+    # Plot the matches
+    for keypoint0, keypoint1, matching_score in zip(keypoints0, keypoints1, matching_scores):
+        plt.plot(
+            [keypoint0[0], keypoint1[0] + image1.width],
+            [keypoint0[1], keypoint1[1]],
+            color=plt.get_cmap("RdYlGn")(matching_score.item()),
+            alpha=0.9,
+            linewidth=0.5,
         )
+        plt.scatter(keypoint0[0], keypoint0[1], c="black", s=2)
+        plt.scatter(keypoint1[0] + image1.width, keypoint1[1], c="black", s=2)
 
-```
+    plt.savefig("matched_image.png", dpi=300, bbox_inches='tight')
+    ```
 
-From the outputs, you can visualize the matches between the two images using the following code:
-```python
-import matplotlib.pyplot as plt
-import numpy as np
+<div class="flex justify-center">
+    <img src="https://cdn-uploads.huggingface.co/production/uploads/632885ba1558dac67c440aa8/01ZYaLB1NL5XdA8u7yCo4.png">
+</div>
 
-# Create side by side image
-merged_image = np.zeros((max(image1.height, image2.height), image1.width + image2.width, 3))
-merged_image[: image1.height, : image1.width] = np.array(image1) / 255.0
-merged_image[: image2.height, image1.width :] = np.array(image2) / 255.0
-plt.imshow(merged_image)
-plt.axis("off")
+## Resources
 
-# Retrieve the keypoints and matches
-output = outputs[0]
-keypoints0 = output["keypoints0"]
-keypoints1 = output["keypoints1"]
-matching_scores = output["matching_scores"]
-keypoints0_x, keypoints0_y = keypoints0[:, 0].numpy(), keypoints0[:, 1].numpy()
-keypoints1_x, keypoints1_y = keypoints1[:, 0].numpy(), keypoints1[:, 1].numpy()
-
-# Plot the matches
-for keypoint0_x, keypoint0_y, keypoint1_x, keypoint1_y, matching_score in zip(
-        keypoints0_x, keypoints0_y, keypoints1_x, keypoints1_y, matching_scores
-):
-    plt.plot(
-        [keypoint0_x, keypoint1_x + image1.width],
-        [keypoint0_y, keypoint1_y],
-        color=plt.get_cmap("RdYlGn")(matching_score.item()),
-        alpha=0.9,
-        linewidth=0.5,
-    )
-    plt.scatter(keypoint0_x, keypoint0_y, c="black", s=2)
-    plt.scatter(keypoint1_x + image1.width, keypoint1_y, c="black", s=2)
-
-# Save the plot
-plt.savefig("matched_image.png", dpi=300, bbox_inches='tight')
-plt.close()
-```
-
-![image/png](https://cdn-uploads.huggingface.co/production/uploads/632885ba1558dac67c440aa8/01ZYaLB1NL5XdA8u7yCo4.png)
-
-This model was contributed by [stevenbucaille](https://huggingface.co/stevenbucaille).
-The original code can be found [here](https://github.com/magicleap/SuperGluePretrainedNetwork).
+- Refer to the [original SuperGlue repository](https://github.com/magicleap/SuperGluePretrainedNetwork) for more examples and implementation details.
 
 ## SuperGlueConfig
 
@@ -133,10 +154,15 @@ The original code can be found [here](https://github.com/magicleap/SuperGluePret
 [[autodoc]] SuperGlueImageProcessor
 
 - preprocess
+- post_process_keypoint_matching
 
+<frameworkcontent>
+<pt>
 ## SuperGlueForKeypointMatching
 
 [[autodoc]] SuperGlueForKeypointMatching
 
 - forward
-- post_process_keypoint_matching
+
+</pt>
+</frameworkcontent>

docs/source/en/model_doc/switch_transformers.md

@@ -14,35 +14,90 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# SwitchTransformers
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
 </div>
 
-## Overview
+# Switch Transformers
 
-The SwitchTransformers model was proposed in [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://huggingface.co/papers/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer.
+[Switch Transformers](https://huggingface.co/papers/2101.03961) is a sparse T5 model where the MLP layer is replaced by a Mixture-of-Experts (MoE). A routing mechanism associates each token with an expert and each expert is a dense MLP. Sparsity enables better scaling and the routing mechanism allows the model to select relevant weights on the fly which increases model capacity.
 
-The Switch Transformer model uses a sparse T5 encoder-decoder architecture, where the MLP are replaced by a Mixture of Experts (MoE). A routing mechanism (top 1 in this case) associates each token to one of the expert, where each expert is a dense MLP. While switch transformers have a lot more weights than their equivalent dense models, the sparsity allows better scaling and better finetuning performance at scale.
-During a forward pass, only a fraction of the weights are used. The routing mechanism allows the model to select relevant weights on the fly which increases the model capacity without increasing the number of operations.
+You can find all the original Switch Transformers checkpoints under the [Switch Transformer](https://huggingface.co/collections/google/switch-transformers-release-6548c35c6507968374b56d1f) collection.
 
-The abstract from the paper is the following:
 
-*In deep learning, models typically reuse the same parameters for all inputs. Mixture of Experts (MoE) defies this and instead selects different parameters for each incoming example. The result is a sparsely-activated model -- with outrageous numbers of parameters -- but a constant computational cost. However, despite several notable successes of MoE, widespread adoption has been hindered by complexity, communication costs and training instability -- we address these with the Switch Transformer. We simplify the MoE routing algorithm and design intuitive improved models with reduced communication and computational costs. Our proposed training techniques help wrangle the instabilities and we show large sparse models may be trained, for the first time, with lower precision (bfloat16) formats. We design models based off T5-Base and T5-Large to obtain up to 7x increases in pre-training speed with the same computational resources. These improvements extend into multilingual settings where we measure gains over the mT5-Base version across all 101 languages. Finally, we advance the current scale of language models by pre-training up to trillion parameter models on the "Colossal Clean Crawled Corpus" and achieve a 4x speedup over the T5-XXL model.*
+> [!TIP]
+> This model was contributed by [ybelkada](https://huggingface.co/ybelkada) and [ArthurZ](https://huggingface.co/ArthurZ).
+>
+> Click on the Switch Transformers models in the right sidebar for more examples of how to apply Switch Transformers to different natural language tasks.
 
-This model was contributed by [Younes Belkada](https://huggingface.co/ybelkada) and [Arthur Zucker](https://huggingface.co/ArthurZ).
-The original code can be found [here](https://github.com/google/flaxformer/tree/main/flaxformer/architectures/moe).
+The example below demonstrates how to predict the masked token with [`Pipeline`], [`AutoModel`], and from the command line.
 
-## Usage tips
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-- SwitchTransformers uses the [`T5Tokenizer`], which can be loaded directly from each model's repository.
-- The released weights are pretrained on English [Masked Language Modeling](https://moon-ci-docs.huggingface.co/docs/transformers/pr_19323/en/glossary#general-terms) task, and should be finetuned.
+```python
+import torch
+from transformers import pipeline
 
-## Resources
+pipeline = pipeline(
+    task="text2text-generation", 
+    model="google/switch-base-8",
+    torch_dtype=torch.float16,
+    device=0
+)
+print(pipeline("The capital of France is <extra_id_0>."))
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```python
+import torch
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8")
+model = AutoModelForSeq2SeqLM.from_pretrained("google/switch-base-8", device_map="auto", torch_dtype=torch.float16)
+
+input_text = "The capital of France is <extra_id_0>."
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(0)
+
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
+
+</hfoption>
+<hfoption id="transformers CLI">
+
+```bash
+echo -e "The capital of France is <extra_id_0>." | transformers run --task text2text-generation --model google/switch-base-8 --device 0
+# [{'generated_text': 'Paris.'}]
+```
+
+</hfoption>
+</hfoptions>
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes/) to only quantize the weights to 8-bits.
+
+```py
+# pip install bitsandbytes
+import torch
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, BitsAndBytesConfig
+
+tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8")
+quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+model = AutoModelForSeq2SeqLM.from_pretrained("google/switch-base-8", device_map="auto", quantization_config=quantization_config)
+
+input_text = "The capital of France is <extra_id_0>."
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(0)
+
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
 
-- [Translation task guide](../tasks/translation)
-- [Summarization task guide](../tasks/summarization)
 
 ## SwitchTransformersConfig
 

docs/source/en/model_doc/t5gemma.md

@@ -14,16 +14,25 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 
 -->
-
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
+</div>
 
 # T5Gemma
 
-T5Gemma (aka encoder-decoder Gemma) was proposed in a [research paper](https://arxiv.org/abs/2504.06225) by Google. It is a family of encoder-decoder large langauge models, developed by adapting pretrained decoder-only models into encoder-decoder. T5Gemma includes pretrained and instruction-tuned variants. The architecture is based on transformer encoder-decoder design following T5, with improvements from Gemma 2: GQA, RoPE, GeGLU activation, RMSNorm, and interleaved local/global attention.
+T5Gemma (aka encoder-decoder Gemma) was proposed in a [research paper](https://arxiv.org/abs/2504.06225) by Google. It is a family of encoder-decoder large language models, developed by adapting pretrained decoder-only models into encoder-decoder. T5Gemma includes pretrained and instruction-tuned variants. The architecture is based on transformer encoder-decoder design following T5, with improvements from Gemma 2: GQA, RoPE, GeGLU activation, RMSNorm, and interleaved local/global attention.
 
 T5Gemma has two groups of model sizes: 1) [Gemma 2](https://ai.google.dev/gemma/docs/core/model_card_2) sizes (2B-2B, 9B-2B, and 9B-9B), which are based on the offical Gemma 2 models (2B and 9B); and 2) [T5](https://arxiv.org/abs/1910.10683) sizes (Small, Base, Large, and XL), where are pretrained under the Gemma 2 framework following T5 configuration. In addition, we also provide a model at ML size (medium large, ~2B in total), which is in-between T5 Large and T5 XL.
 
 The pretrained varaints are trained with two objectives: prefix language modeling with knowledge distillation (PrefixLM) and UL2, separately. We release both variants for each model size. The instruction-turned varaints was post-trained with supervised fine-tuning and reinforcement learning.
 
+> [!TIP]
+> Click on the T5Gemma models in the right sidebar for more examples of how to apply T5Gemma to different language tasks.
+
 The example below demonstrates how to chat with the model with [`Pipeline`] or the [`AutoModel`] class, and from the command line.
 
 <hfoptions id="usage">
@@ -35,43 +44,52 @@ import torch
 from transformers import pipeline
 
 pipe = pipeline(
-    task="text2text-generation",
-    model="google/t5gemma-placeholder",
+    "text2text-generation",
+    model="google/t5gemma-2b-2b-prefixlm-it",
     torch_dtype=torch.bfloat16,
-    device="cuda",
+    device="cuda",  # replace with "mps" to run on a Mac device
 )
 
-pipe("Question: Why is the sky blue?\nAnswer:", max_new_tokens=50)
+messages = [
+    {"role": "user", "content": "Tell me an unknown interesting biology fact about the brain."},
+]
+prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+pipe(prompt, max_new_tokens=32)
 ```
 
 </hfoption>
 <hfoption id="AutoModel">
 
 ```python
-import torch
+# pip install accelerate
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+import torch
 
-tokenizer = AutoTokenizer.from_pretrained("google/t5gemma-placeholder")
+tokenizer = AutoTokenizer.from_pretrained("google/t5gemma-2b-2b-prefixlm-it")
 model = AutoModelForSeq2SeqLM.from_pretrained(
-    "google/t5gemma-placeholder",
+    "google/t5gemma-2b-2b-prefixlm-it",
+    device_map="auto",
     torch_dtype=torch.bfloat16,
-    device_map="auto"
 )
 
-input_text = "Question: Why is the sky blue?\nAnswer:"
-input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+messages = [
+    {"role": "user", "content": "Tell me an unknown interesting biology fact about the brain."},
+]
+input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True, add_generation_prompt=True).to("cuda")
 
 outputs = model.generate(**input_ids, max_new_tokens=32)
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-
+print(tokenizer.decode(outputs[0]))
 ```
 
 </hfoption>
 <hfoption id="transformers CLI">
 
 ```
-echo -e "Question: Why is the sky blue? Answer:" | transformers run --task text2text-generation --model google/t5gemma-placeholder --device 0
+echo -e "Write me a poem about Machine Learning. Answer:" | transformers run --task text2text-generation --model google/t5gemma-2b-2b-prefixlm --device 0
 ```
+</hfoption>
+</hfoptions>
 
 ## T5GemmaConfig
 

docs/source/en/model_doc/timesfm.md

@@ -37,6 +37,7 @@ The original code can be found [here](https://github.com/google-research/timesfm
 To use the model:
 
 ```python
+import numpy as np
 import torch
 from transformers import TimesFmModelForPrediction
 

docs/source/en/model_doc/vitpose.md

@@ -10,52 +10,39 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# ViTPose
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+  </div>
 </div>
 
-## Overview
+# ViTPose
 
-The ViTPose model was proposed in [ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation](https://huggingface.co/papers/2204.12484) by Yufei Xu, Jing Zhang, Qiming Zhang, Dacheng Tao. ViTPose employs a standard, non-hierarchical [Vision Transformer](vit) as backbone for the task of keypoint estimation. A simple decoder head is added on top to predict the heatmaps from a given image. Despite its simplicity, the model gets state-of-the-art results on the challenging MS COCO Keypoint Detection benchmark. The model was further improved in [ViTPose++: Vision Transformer for Generic Body Pose Estimation](https://huggingface.co/papers/2212.04246) where the authors employ
-a mixture-of-experts (MoE) module in the ViT backbone along with pre-training on more data, which further enhances the performance.
+[ViTPose](https://huggingface.co/papers/2204.12484) is a vision transformer-based model for keypoint (pose) estimation. It uses a simple, non-hierarchical [ViT](./vit) backbone and a lightweight decoder head. This architecture simplifies model design, takes advantage of transformer scalability, and can be adapted to different training strategies.
 
-The abstract from the paper is the following:
-
-*Although no specific domain knowledge is considered in the design, plain vision transformers have shown excellent performance in visual recognition tasks. However, little effort has been made to reveal the potential of such simple structures for pose estimation tasks. In this paper, we show the surprisingly good capabilities of plain vision transformers for pose estimation from various aspects, namely simplicity in model structure, scalability in model size, flexibility in training paradigm, and transferability of knowledge between models, through a simple baseline model called ViTPose. Specifically, ViTPose employs plain and non-hierarchical vision transformers as backbones to extract features for a given person instance and a lightweight decoder for pose estimation. It can be scaled up from 100M to 1B parameters by taking the advantages of the scalable model capacity and high parallelism of transformers, setting a new Pareto front between throughput and performance. Besides, ViTPose is very flexible regarding the attention type, input resolution, pre-training and finetuning strategy, as well as dealing with multiple pose tasks. We also empirically demonstrate that the knowledge of large ViTPose models can be easily transferred to small ones via a simple knowledge token. Experimental results show that our basic ViTPose model outperforms representative methods on the challenging MS COCO Keypoint Detection benchmark, while the largest model sets a new state-of-the-art.*
+[ViTPose++](https://huggingface.co/papers/2212.04246) improves on ViTPose by incorporating a mixture-of-experts (MoE) module in the backbone and using more diverse pretraining data.
 
 <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/vitpose-architecture.png"
 alt="drawing" width="600"/>
 
-<small> ViTPose architecture. Taken from the <a href="https://huggingface.co/papers/2204.12484">original paper.</a> </small>
+You can find all ViTPose and ViTPose++ checkpoints under the [ViTPose collection](https://huggingface.co/collections/usyd-community/vitpose-677fcfd0a0b2b5c8f79c4335).
 
-This model was contributed by [nielsr](https://huggingface.co/nielsr) and [sangbumchoi](https://github.com/SangbumChoi).
-The original code can be found [here](https://github.com/ViTAE-Transformer/ViTPose).
-
-## Usage Tips
-
-ViTPose is a so-called top-down keypoint detection model. This means that one first uses an object detector, like [RT-DETR](rt_detr.md), to detect people (or other instances) in an image. Next, ViTPose takes the cropped images as input and predicts the keypoints for each of them.
+The example below demonstrates pose estimation with the [`VitPoseForPoseEstimation`] class.
 
 ```py
 import torch
 import requests
 import numpy as np
-
+import supervision as sv
 from PIL import Image
-
 from transformers import AutoProcessor, RTDetrForObjectDetection, VitPoseForPoseEstimation
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-url = "http://images.cocodataset.org/val2017/000000000139.jpg"
+url = "https://www.fcbarcelona.com/fcbarcelona/photo/2021/01/31/3c55a19f-dfc1-4451-885e-afd14e890a11/mini_2021-01-31-BARCELONA-ATHLETIC-BILBAOI-30.JPG"
 image = Image.open(requests.get(url, stream=True).raw)
 
-# ------------------------------------------------------------------------
-# Stage 1. Detect humans on the image
-# ------------------------------------------------------------------------
-
-# You can choose any detector of your choice
+# Detect humans in the image
 person_image_processor = AutoProcessor.from_pretrained("PekingU/rtdetr_r50vd_coco_o365")
 person_model = RTDetrForObjectDetection.from_pretrained("PekingU/rtdetr_r50vd_coco_o365", device_map=device)
 
@@ -67,7 +54,7 @@ with torch.no_grad():
 results = person_image_processor.post_process_object_detection(
     outputs, target_sizes=torch.tensor([(image.height, image.width)]), threshold=0.3
 )
-result = results[0]  # take first image results
+result = results[0]
 
 # Human label refers 0 index in COCO dataset
 person_boxes = result["boxes"][result["labels"] == 0]
@@ -77,10 +64,7 @@ person_boxes = person_boxes.cpu().numpy()
 person_boxes[:, 2] = person_boxes[:, 2] - person_boxes[:, 0]
 person_boxes[:, 3] = person_boxes[:, 3] - person_boxes[:, 1]
 
-# ------------------------------------------------------------------------
-# Stage 2. Detect keypoints for each person found
-# ------------------------------------------------------------------------
-
+# Detect keypoints for each person found
 image_processor = AutoProcessor.from_pretrained("usyd-community/vitpose-base-simple")
 model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-base-simple", device_map=device)
 
@@ -90,54 +74,7 @@ with torch.no_grad():
     outputs = model(**inputs)
 
 pose_results = image_processor.post_process_pose_estimation(outputs, boxes=[person_boxes])
-image_pose_result = pose_results[0]  # results for first image
-```
-
-### ViTPose++ models
-
-The best [checkpoints](https://huggingface.co/collections/usyd-community/vitpose-677fcfd0a0b2b5c8f79c4335) are those of the [ViTPose++ paper](https://huggingface.co/papers/2212.04246). ViTPose++ models employ a so-called [Mixture-of-Experts (MoE)](https://huggingface.co/blog/moe) architecture for the ViT backbone, resulting in better performance.
-
-The ViTPose+ checkpoints use 6 experts, hence 6 different dataset indices can be passed. 
-An overview of the various dataset indices is provided below:
-
-- 0: [COCO validation 2017](https://cocodataset.org/#overview) dataset, using an object detector that gets 56 AP on the "person" class
-- 1: [AiC](https://github.com/fabbrimatteo/AiC-Dataset) dataset
-- 2: [MPII](https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/software-and-datasets/mpii-human-pose-dataset) dataset
-- 3: [AP-10K](https://github.com/AlexTheBad/AP-10K) dataset
-- 4: [APT-36K](https://github.com/pandorgan/APT-36K) dataset
-- 5: [COCO-WholeBody](https://github.com/jin-s13/COCO-WholeBody) dataset
-
-Pass the `dataset_index` argument in the forward of the model to indicate which experts to use for each example in the batch. Example usage is shown below:
-
-```python
-image_processor = AutoProcessor.from_pretrained("usyd-community/vitpose-plus-base")
-model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-plus-base", device=device)
-
-inputs = image_processor(image, boxes=[person_boxes], return_tensors="pt").to(device)
-
-dataset_index = torch.tensor([0], device=device) # must be a tensor of shape (batch_size,)
-
-with torch.no_grad():
-    outputs = model(**inputs, dataset_index=dataset_index)
-```
-
-The ViTPose+ checkpoints use 6 experts, hence 6 different dataset indices can be passed. 
-An overview of the various dataset indices is provided below:
-
-- 0: [COCO validation 2017](https://cocodataset.org/#overview) dataset, using an object detector that gets 56 AP on the "person" class
-- 1: [AiC](https://github.com/fabbrimatteo/AiC-Dataset) dataset
-- 2: [MPII](https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/software-and-datasets/mpii-human-pose-dataset) dataset
-- 3: [AP-10K](https://github.com/AlexTheBad/AP-10K) dataset
-- 4: [APT-36K](https://github.com/pandorgan/APT-36K) dataset
-- 5: [COCO-WholeBody](https://github.com/jin-s13/COCO-WholeBody) dataset
-
-
-### Visualization
-
-To visualize the various keypoints, one can either leverage the `supervision` [library](https://github.com/roboflow/supervision (requires `pip install supervision`):
-
-```python
-import supervision as sv
+image_pose_result = pose_results[0]
 
 xy = torch.stack([pose_result['keypoints'] for pose_result in image_pose_result]).cpu().numpy()
 scores = torch.stack([pose_result['scores'] for pose_result in image_pose_result]).cpu().numpy()
@@ -162,119 +99,192 @@ annotated_frame = vertex_annotator.annotate(
     scene=annotated_frame,
     key_points=key_points
 )
+annotated_frame
 ```
 
-Alternatively, one can also visualize the keypoints using [OpenCV](https://opencv.org/) (requires `pip install opencv-python`):
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/vitpose.png"/>
+</div>
 
-```python
-import math
-import cv2
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
 
-def draw_points(image, keypoints, scores, pose_keypoint_color, keypoint_score_threshold, radius, show_keypoint_weight):
-    if pose_keypoint_color is not None:
-        assert len(pose_keypoint_color) == len(keypoints)
-    for kid, (kpt, kpt_score) in enumerate(zip(keypoints, scores)):
-        x_coord, y_coord = int(kpt[0]), int(kpt[1])
-        if kpt_score > keypoint_score_threshold:
-            color = tuple(int(c) for c in pose_keypoint_color[kid])
-            if show_keypoint_weight:
-                cv2.circle(image, (int(x_coord), int(y_coord)), radius, color, -1)
-                transparency = max(0, min(1, kpt_score))
-                cv2.addWeighted(image, transparency, image, 1 - transparency, 0, dst=image)
-            else:
-                cv2.circle(image, (int(x_coord), int(y_coord)), radius, color, -1)
+The example below uses [torchao](../quantization/torchao) to only quantize the weights to int4.
 
-def draw_links(image, keypoints, scores, keypoint_edges, link_colors, keypoint_score_threshold, thickness, show_keypoint_weight, stick_width = 2):
-    height, width, _ = image.shape
-    if keypoint_edges is not None and link_colors is not None:
-        assert len(link_colors) == len(keypoint_edges)
-        for sk_id, sk in enumerate(keypoint_edges):
-            x1, y1, score1 = (int(keypoints[sk[0], 0]), int(keypoints[sk[0], 1]), scores[sk[0]])
-            x2, y2, score2 = (int(keypoints[sk[1], 0]), int(keypoints[sk[1], 1]), scores[sk[1]])
-            if (
-                x1 > 0
-                and x1 < width
-                and y1 > 0
-                and y1 < height
-                and x2 > 0
-                and x2 < width
-                and y2 > 0
-                and y2 < height
-                and score1 > keypoint_score_threshold
-                and score2 > keypoint_score_threshold
-            ):
-                color = tuple(int(c) for c in link_colors[sk_id])
+```py
+# pip install torchao
+import torch
+import requests
+import numpy as np
+from PIL import Image
+from transformers import AutoProcessor, RTDetrForObjectDetection, VitPoseForPoseEstimation, TorchAoConfig
+
+url = "https://www.fcbarcelona.com/fcbarcelona/photo/2021/01/31/3c55a19f-dfc1-4451-885e-afd14e890a11/mini_2021-01-31-BARCELONA-ATHLETIC-BILBAOI-30.JPG"
+image = Image.open(requests.get(url, stream=True).raw)
+
+person_image_processor = AutoProcessor.from_pretrained("PekingU/rtdetr_r50vd_coco_o365")
+person_model = RTDetrForObjectDetection.from_pretrained("PekingU/rtdetr_r50vd_coco_o365", device_map=device)
+
+inputs = person_image_processor(images=image, return_tensors="pt").to(device)
+
+with torch.no_grad():
+    outputs = person_model(**inputs)
+
+results = person_image_processor.post_process_object_detection(
+    outputs, target_sizes=torch.tensor([(image.height, image.width)]), threshold=0.3
+)
+result = results[0]
+
+person_boxes = result["boxes"][result["labels"] == 0]
+person_boxes = person_boxes.cpu().numpy()
+
+person_boxes[:, 2] = person_boxes[:, 2] - person_boxes[:, 0]
+person_boxes[:, 3] = person_boxes[:, 3] - person_boxes[:, 1]
+
+quantization_config = TorchAoConfig("int4_weight_only", group_size=128)
+
+image_processor = AutoProcessor.from_pretrained("usyd-community/vitpose-plus-huge")
+model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-plus-huge", device_map=device, quantization_config=quantization_config)
+
+inputs = image_processor(image, boxes=[person_boxes], return_tensors="pt").to(device)
+
+with torch.no_grad():
+    outputs = model(**inputs)
+
+pose_results = image_processor.post_process_pose_estimation(outputs, boxes=[person_boxes])
+image_pose_result = pose_results[0]
+```
+
+## Notes
+
+- Use [`AutoProcessor`] to automatically prepare bounding box and image inputs.
+- ViTPose is a top-down pose estimator. It uses a object detector to detect individuals first before keypoint prediction.
+- ViTPose++ has 6 different MoE expert heads (COCO validation `0`, AiC `1`, MPII `2`, AP-10K `3`, APT-36K `4`, COCO-WholeBody `5`) which supports 6 different datasets. Pass a specific value corresponding to the dataset to the `dataset_index` to indicate which expert to use.
+
+    ```py
+    from transformers import AutoProcessor, VitPoseForPoseEstimation
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    image_processor = AutoProcessor.from_pretrained("usyd-community/vitpose-plus-base")
+    model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-plus-base", device=device)
+
+    inputs = image_processor(image, boxes=[person_boxes], return_tensors="pt").to(device)
+    dataset_index = torch.tensor([0], device=device) # must be a tensor of shape (batch_size,)
+
+    with torch.no_grad():
+        outputs = model(**inputs, dataset_index=dataset_index)
+    ```
+
+- [OpenCV](https://opencv.org/) is an alternative option for visualizing the estimated pose.
+
+    ```py
+    # pip install opencv-python
+    import math
+    import cv2
+
+    def draw_points(image, keypoints, scores, pose_keypoint_color, keypoint_score_threshold, radius, show_keypoint_weight):
+        if pose_keypoint_color is not None:
+            assert len(pose_keypoint_color) == len(keypoints)
+        for kid, (kpt, kpt_score) in enumerate(zip(keypoints, scores)):
+            x_coord, y_coord = int(kpt[0]), int(kpt[1])
+            if kpt_score > keypoint_score_threshold:
+                color = tuple(int(c) for c in pose_keypoint_color[kid])
                 if show_keypoint_weight:
-                    X = (x1, x2)
-                    Y = (y1, y2)
-                    mean_x = np.mean(X)
-                    mean_y = np.mean(Y)
-                    length = ((Y[0] - Y[1]) ** 2 + (X[0] - X[1]) ** 2) ** 0.5
-                    angle = math.degrees(math.atan2(Y[0] - Y[1], X[0] - X[1]))
-                    polygon = cv2.ellipse2Poly(
-                        (int(mean_x), int(mean_y)), (int(length / 2), int(stick_width)), int(angle), 0, 360, 1
-                    )
-                    cv2.fillConvexPoly(image, polygon, color)
-                    transparency = max(0, min(1, 0.5 * (keypoints[sk[0], 2] + keypoints[sk[1], 2])))
+                    cv2.circle(image, (int(x_coord), int(y_coord)), radius, color, -1)
+                    transparency = max(0, min(1, kpt_score))
                     cv2.addWeighted(image, transparency, image, 1 - transparency, 0, dst=image)
                 else:
-                    cv2.line(image, (x1, y1), (x2, y2), color, thickness=thickness)
+                    cv2.circle(image, (int(x_coord), int(y_coord)), radius, color, -1)
 
+    def draw_links(image, keypoints, scores, keypoint_edges, link_colors, keypoint_score_threshold, thickness, show_keypoint_weight, stick_width = 2):
+        height, width, _ = image.shape
+        if keypoint_edges is not None and link_colors is not None:
+            assert len(link_colors) == len(keypoint_edges)
+            for sk_id, sk in enumerate(keypoint_edges):
+                x1, y1, score1 = (int(keypoints[sk[0], 0]), int(keypoints[sk[0], 1]), scores[sk[0]])
+                x2, y2, score2 = (int(keypoints[sk[1], 0]), int(keypoints[sk[1], 1]), scores[sk[1]])
+                if (
+                    x1 > 0
+                    and x1 < width
+                    and y1 > 0
+                    and y1 < height
+                    and x2 > 0
+                    and x2 < width
+                    and y2 > 0
+                    and y2 < height
+                    and score1 > keypoint_score_threshold
+                    and score2 > keypoint_score_threshold
+                ):
+                    color = tuple(int(c) for c in link_colors[sk_id])
+                    if show_keypoint_weight:
+                        X = (x1, x2)
+                        Y = (y1, y2)
+                        mean_x = np.mean(X)
+                        mean_y = np.mean(Y)
+                        length = ((Y[0] - Y[1]) ** 2 + (X[0] - X[1]) ** 2) ** 0.5
+                        angle = math.degrees(math.atan2(Y[0] - Y[1], X[0] - X[1]))
+                        polygon = cv2.ellipse2Poly(
+                            (int(mean_x), int(mean_y)), (int(length / 2), int(stick_width)), int(angle), 0, 360, 1
+                        )
+                        cv2.fillConvexPoly(image, polygon, color)
+                        transparency = max(0, min(1, 0.5 * (keypoints[sk[0], 2] + keypoints[sk[1], 2])))
+                        cv2.addWeighted(image, transparency, image, 1 - transparency, 0, dst=image)
+                    else:
+                        cv2.line(image, (x1, y1), (x2, y2), color, thickness=thickness)
 
-# Note: keypoint_edges and color palette are dataset-specific
-keypoint_edges = model.config.edges
+    # Note: keypoint_edges and color palette are dataset-specific
+    keypoint_edges = model.config.edges
 
-palette = np.array(
-    [
-        [255, 128, 0],
-        [255, 153, 51],
-        [255, 178, 102],
-        [230, 230, 0],
-        [255, 153, 255],
-        [153, 204, 255],
-        [255, 102, 255],
-        [255, 51, 255],
-        [102, 178, 255],
-        [51, 153, 255],
-        [255, 153, 153],
-        [255, 102, 102],
-        [255, 51, 51],
-        [153, 255, 153],
-        [102, 255, 102],
-        [51, 255, 51],
-        [0, 255, 0],
-        [0, 0, 255],
-        [255, 0, 0],
-        [255, 255, 255],
-    ]
-)
+    palette = np.array(
+        [
+            [255, 128, 0],
+            [255, 153, 51],
+            [255, 178, 102],
+            [230, 230, 0],
+            [255, 153, 255],
+            [153, 204, 255],
+            [255, 102, 255],
+            [255, 51, 255],
+            [102, 178, 255],
+            [51, 153, 255],
+            [255, 153, 153],
+            [255, 102, 102],
+            [255, 51, 51],
+            [153, 255, 153],
+            [102, 255, 102],
+            [51, 255, 51],
+            [0, 255, 0],
+            [0, 0, 255],
+            [255, 0, 0],
+            [255, 255, 255],
+        ]
+    )
 
-link_colors = palette[[0, 0, 0, 0, 7, 7, 7, 9, 9, 9, 9, 9, 16, 16, 16, 16, 16, 16, 16]]
-keypoint_colors = palette[[16, 16, 16, 16, 16, 9, 9, 9, 9, 9, 9, 0, 0, 0, 0, 0, 0]]
+    link_colors = palette[[0, 0, 0, 0, 7, 7, 7, 9, 9, 9, 9, 9, 16, 16, 16, 16, 16, 16, 16]]
+    keypoint_colors = palette[[16, 16, 16, 16, 16, 9, 9, 9, 9, 9, 9, 0, 0, 0, 0, 0, 0]]
 
-numpy_image = np.array(image)
+    numpy_image = np.array(image)
 
-for pose_result in image_pose_result:
-    scores = np.array(pose_result["scores"])
-    keypoints = np.array(pose_result["keypoints"])
+    for pose_result in image_pose_result:
+        scores = np.array(pose_result["scores"])
+        keypoints = np.array(pose_result["keypoints"])
 
-    # draw each point on image
-    draw_points(numpy_image, keypoints, scores, keypoint_colors, keypoint_score_threshold=0.3, radius=4, show_keypoint_weight=False)
+        # draw each point on image
+        draw_points(numpy_image, keypoints, scores, keypoint_colors, keypoint_score_threshold=0.3, radius=4, show_keypoint_weight=False)
 
-    # draw links
-    draw_links(numpy_image, keypoints, scores, keypoint_edges, link_colors, keypoint_score_threshold=0.3, thickness=1, show_keypoint_weight=False)
+        # draw links
+        draw_links(numpy_image, keypoints, scores, keypoint_edges, link_colors, keypoint_score_threshold=0.3, thickness=1, show_keypoint_weight=False)
 
-pose_image = Image.fromarray(numpy_image)
-pose_image
-```
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/vitpose-coco.jpg" alt="drawing" width="600"/>
+    pose_image = Image.fromarray(numpy_image)
+    pose_image
+    ```
 
 ## Resources
 
-A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ViTPose. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+Refer to resources below to learn more about using ViTPose.
 
-- A demo of ViTPose on images and video can be found [here](https://huggingface.co/spaces/hysts/ViTPose-transformers).
-- A notebook illustrating inference and visualization can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/ViTPose/Inference_with_ViTPose_for_human_pose_estimation.ipynb).
+- This [notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/ViTPose/Inference_with_ViTPose_for_body_pose_estimation.ipynb) demonstrates inference and visualization.
+- This [Space](https://huggingface.co/spaces/hysts/ViTPose-transformers) demonstrates ViTPose on images and video.
 
 ## VitPoseImageProcessor
 

docs/source/en/model_doc/voxtral.md

@@ -0,0 +1,300 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Voxtral
+
+Voxtral is an upgrade of [Ministral 3B and Mistral Small 3B](https://mistral.ai/news/ministraux), extending its language capabilities with audio input support. It is designed to handle tasks such as speech transcription, translation, and audio understanding.
+
+You can read more in Mistral's [realease blog post](https://mistral.ai/news/voxtral).
+
+The model is available in two checkpoints:
+- 3B: [mistralai/Voxtral-Mini-3B-2507](https://huggingface.co/mistralai/Voxtral-Mini-3B-2507)
+- 24B: [mistralai/Voxtral-Small-24B-2507](https://huggingface.co/mistralai/Voxtral-Small-24B-2507)
+
+## Key Features
+
+Voxtral builds on Ministral-3B by adding audio processing capabilities:
+
+- **Transcription mode**: Includes a dedicated mode for speech transcription. By default, Voxtral detects the spoken language and transcribes it accordingly.  
+- **Long-form context**: With a 32k token context window, Voxtral can process up to 30 minutes of audio for transcription or 40 minutes for broader audio understanding.  
+- **Integrated Q&A and summarization**: Supports querying audio directly and producing structured summaries without relying on separate ASR and language models.  
+- **Multilingual support**: Automatically detects language and performs well across several widely spoken languages, including English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian.  
+- **Function calling via voice**: Can trigger functions or workflows directly from spoken input based on detected user intent.  
+- **Text capabilities**: Maintains the strong text processing performance of its Ministral-3B foundation.
+
+## Usage
+
+Let's first load the model!
+```python
+from transformers import VoxtralForConditionalGeneration, AutoProcessor
+import torch
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+repo_id = "mistralai/Voxtral-Mini-3B-2507"
+
+processor = AutoProcessor.from_pretrained(repo_id)
+model = VoxtralForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch.bfloat16, device_map=device)
+```
+
+### Audio Instruct Mode
+
+The model supports audio-text instructions, including multi-turn and multi-audio interactions, all processed in batches.
+
+➡️ audio + text instruction
+```python
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "audio",
+                "url": "https://huggingface.co/datasets/eustlb/audio-samples/resolve/main/dude_where_is_my_car.wav",
+            },
+            {"type": "text", "text": "What can you tell me about this audio?"},
+        ],
+    }
+]
+
+inputs = processor.apply_chat_template(conversation)
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated response:")
+print("=" * 80)
+print(decoded_outputs[0])
+print("=" * 80)
+```
+
+➡️ multi-audio + text instruction 
+```python
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "audio",
+                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/mary_had_lamb.mp3",
+            },
+            {
+                "type": "audio",
+                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/winning_call.mp3",
+            },
+            {"type": "text", "text": "What sport and what nursery rhyme are referenced?"},
+        ],
+    }
+]
+
+inputs = processor.apply_chat_template(conversation)
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated response:")
+print("=" * 80)
+print(decoded_outputs[0])
+print("=" * 80)
+```
+
+➡️ multi-turn:
+```python
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "audio",
+                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/obama.mp3",
+            },
+            {
+                "type": "audio",
+                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3",
+            },
+            {"type": "text", "text": "Describe briefly what you can hear."},
+        ],
+    },
+    {
+        "role": "assistant",
+        "content": "The audio begins with the speaker delivering a farewell address in Chicago, reflecting on his eight years as president and expressing gratitude to the American people. The audio then transitions to a weather report, stating that it was 35 degrees in Barcelona the previous day, but the temperature would drop to minus 20 degrees the following day.",
+    },
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "audio",
+                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/dude_where_is_my_car.wav",
+            },
+            {"type": "text", "text": "Ok, now compare this new audio with the previous one."},
+        ],
+    },
+]
+
+inputs = processor.apply_chat_template(conversation)
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated response:")
+print("=" * 80)
+print(decoded_outputs[0])
+print("=" * 80)
+```
+
+➡️ text only:
+```python
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "text",
+                "text": "What if a cyber brain could possibly generate its own ghost, and create a soul all by itself?",
+            },
+        ],
+    }
+]
+
+inputs = processor.apply_chat_template(conversation)
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated response:")
+print("=" * 80)
+print(decoded_outputs[0])
+print("=" * 80)
+```
+
+➡️ audio only:
+```python
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "audio",
+                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/dude_where_is_my_car.wav",
+            },
+        ],
+    }
+]
+
+inputs = processor.apply_chat_template(conversation)
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated response:")
+print("=" * 80)
+print(decoded_outputs[0])
+print("=" * 80)
+```
+
+➡️ batched inference!
+```python
+conversations = [
+    [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "audio",
+                    "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/obama.mp3",
+                },
+                {
+                    "type": "audio",
+                    "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3",
+                },
+                {
+                    "type": "text",
+                    "text": "Who's speaking in the speach and what city's weather is being discussed?",
+                },
+            ],
+        }
+    ],
+    [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "audio",
+                    "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/winning_call.mp3",
+                },
+                {"type": "text", "text": "What can you tell me about this audio?"},
+            ],
+        }
+    ],
+]
+
+inputs = processor.apply_chat_template(conversations)
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated responses:")
+print("=" * 80)
+for decoded_output in decoded_outputs:
+    print(decoded_output)
+    print("=" * 80)
+```
+
+### Transcription Mode
+
+Use the model to transcribe audio (supports English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian)!
+
+```python
+inputs = processor.apply_transcrition_request(language="en", audio="https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/obama.mp3")
+inputs = inputs.to(device, dtype=torch.bfloat16)
+
+outputs = model.generate(**inputs, max_new_tokens=500)
+decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+print("\nGenerated responses:")
+print("=" * 80)
+for decoded_output in decoded_outputs:
+    print(decoded_output)
+    print("=" * 80)
+```
+
+This model was contributed by [Eustache Le Bihan](https://huggingface.co/eustlb).
+
+## VoxtralConfig
+
+[[autodoc]] VoxtralConfig
+
+## VoxtralEncoderConfig
+
+[[autodoc]] VoxtralEncoderConfig
+
+## VoxtralProcessor
+
+[[autodoc]] VoxtralProcessor
+
+## VoxtralEncoder
+
+[[autodoc]] VoxtralEncoder
+    - forward
+
+## VoxtralForConditionalGeneration
+
+[[autodoc]] VoxtralForConditionalGeneration
+    - forward

docs/source/en/model_doc/wav2vec2.md

@@ -172,9 +172,9 @@ Otherwise, [`~Wav2Vec2ProcessorWithLM.batch_decode`] performance will be slower
 >>> dataset = dataset.cast_column("audio", datasets.Audio(sampling_rate=16_000))
 
 
->>> def map_to_array(batch):
-...     batch["speech"] = batch["audio"]["array"]
-...     return batch
+>>> def map_to_array(example):
+...     example["speech"] = example["audio"]["array"]
+...     return example
 
 
 >>> # prepare speech data for batch inference

docs/source/en/optimizers.md

@@ -164,7 +164,7 @@ args = TrainingArguments(
     output_dir="./test-schedulefree",
     max_steps=1000,
     per_device_train_batch_size=4,
-+   optim="schedule_free_radamw,
++   optim="schedule_free_radamw",
 +   lr_scheduler_type="constant",
     gradient_checkpointing=True,
     logging_strategy="steps",
@@ -174,3 +174,29 @@ args = TrainingArguments(
     run_name="sfo",
 )
 ```
+
+## StableAdamW
+
+```bash
+pip install torch-optimi
+```
+
+[StableAdamW](https://arxiv.org/pdf/2304.13013) is a hybrid between AdamW and AdaFactor. It ports AdaFactor's update clipping into AdamW, which removes the need for gradient clipping. Otherwise, it behaves as a drop-in replacement for AdamW.
+
+> [!TIP]
+> If training on large batch sizes or still observing training loss spikes, consider reducing beta_2 between [0.95, 0.99].
+
+```diff
+args = TrainingArguments(
+    output_dir="./test-stable-adamw",
+    max_steps=1000,
+    per_device_train_batch_size=4,
++   optim="stable_adamw",
+    gradient_checkpointing=True,
+    logging_strategy="steps",
+    logging_steps=1,
+    learning_rate=2e-6,
+    save_strategy="no",
+    run_name="stable-adamw",
+)
+```

docs/source/en/perf_hardware.md

@@ -15,7 +15,7 @@ rendered properly in your Markdown viewer.
 
 # Build your own machine
 
-One of the most important consideration when building a machine for deep learning is the GPU choice. GPUs are the standard workhorse for deep learning owing to their tensor cores for performing very efficient matrix multiplication and high memory bandwidth. To train large models, you either need a more powerful GPU, multiple GPUs, or take advantage of techniques that offload some of the load to the CPU or NVMe.
+One of the most important considerations when building a machine for deep learning is the GPU choice. GPUs are the standard workhorse for deep learning owing to their tensor cores for performing very efficient matrix multiplication and high memory bandwidth. To train large models, you either need a more powerful GPU, multiple GPUs, or take advantage of techniques that offload some of the load to the CPU or NVMe.
 
 This guide provides some practical tips for setting up a GPU for deep learning. For a more detailed discussion and comparison of GPUs, take a look at the [Which GPU(s) to Get for Deep Learning](https://timdettmers.com/2023/01/30/which-gpu-for-deep-learning/) blog post.
 
@@ -25,11 +25,11 @@ High-end consumer GPUs may have two or three PCIe 8-pin power sockets, and you s
 
 Each PCIe 8-pin power cable should be connected to a 12V rail on the power supply unit (PSU) and can deliver up to 150W. Other GPUs may use a PCIe 12-pin connector which can deliver up to 500-600W. Lower-end GPUs may only use a PCIe 6-pin connector which supplies up to 75W.
 
-It is important the PSU has stable voltage otherwise it may not be able to supply the GPU with enough power to function properly during peak usage.
+It is important that the PSU maintains stable voltage; otherwise, it may fail to supply the GPU with enough power during peak usage.
 
 ## Cooling
 
-An overheated GPU throttles its performance and can even shutdown if it's too hot to prevent damage. Keeping the GPU temperature low, anywhere between 158 - 167F, is essential for delivering full performance and maintaining its lifespan. Once temperatures reach 183 - 194F, the GPU may begin to throttle performance.
+An overheated GPU throttles its performance and can even shutdown if it's too hot to prevent damage. Keeping the GPU temperature low, anywhere between 158–167°F, is essential for delivering full performance and maintaining its lifespan. Once temperatures reach 183 - 194°F, the GPU may begin to throttle performance.
 
 ## Multi-GPU connectivity
 

docs/source/en/perf_infer_gpu_one.md

@@ -177,10 +177,16 @@ There are three supported implementations available.
 
 SDPA is used by default for PyTorch v2.1.1. and greater when an implementation is available. You could explicitly enable SDPA by setting `attn_implementation="sdpa"` in [`~PreTrainedModel.from_pretrained`] though. Certain attention parameters, such as `head_mask` and `output_attentions=True`, are unsupported and returns a warning that Transformers will fall back to the (slower) eager implementation.
 
+Refer to the [AttentionInterface](./attention_interface) guide to learn how to change the attention implementation after loading a model.
+
 ```py
 from transformers import AutoModelForCausalLM
 
 model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B", device_map="auto", attn_implementation="sdpa")
+
+# Change the model's attention dynamically after loading it
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B", device_map="auto")
+model.set_attention_implementation("sdpa")
 ```
 
 SDPA selects the most performant implementation available, but you can also explicitly select an implementation with [torch.nn.attention.sdpa_kernel](https://pytorch.org/docs/master/backends.html#torch.backends.cuda.sdp_kernel) as a context manager. The example below shows how to enable the FlashAttention2 implementation with `enable_flash=True`.
@@ -234,7 +240,7 @@ FlashAttention2 support is currently limited to Instinct MI210, Instinct MI250 a
 </hfoption>
 </hfoptions>
 
-Enable FlashAttention2 by setting `attn_implementation="flash_attention_2"` in [`~PreTrainedModel.from_pretrained`]. FlashAttention2 is only supported for models with the fp16 or bf16 torch type. Make sure to cast your model to the appropriate data type first.
+Enable FlashAttention2 by setting `attn_implementation="flash_attention_2"` in [`~PreTrainedModel.from_pretrained`] or by setting `model.set_attention_implementation("flash_attention_2")` to dynamically update the [attention interface](./attention_interface). FlashAttention2 is only supported for models with the fp16 or bf16 torch type. Make sure to cast your model to the appropriate data type first.
 
 ```py
 from transformers import AutoModelForCausalLM

docs/source/en/perf_train_tpu_tf.md

@@ -1,355 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-
-⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
-rendered properly in your Markdown viewer.
-
--->
-
-# TPU
-
-TPU (Tensor Processing Unit) is a type of hardware designed to accelerate tensor computations for training and inference. TPUs are generally accessed through Google cloud services, but smaller TPUs are also available for free from [Google Colab](https://colab.research.google.com/notebooks/tpu.ipynb) or [Kaggle](https://www.kaggle.com/docs/tpu).
-
-This guide focuses on training a Keras model for sequence classification on a TPU from Google Colab. Make sure the TPU runtime is enabled by going to **Runtime > Change runtime type** and selecting a TPU.
-
-Run the command below to install the latest version of Transformers and [Datasets](https://huggingface.co/docs/datasets).
-
-```py
-!pip install --U transformers datasets
-```
-
-Create an instance of [tf.distribute.cluster_resolver.TPUClusterResolver](https://www.tensorflow.org/api_docs/python/tf/distribute/cluster_resolver/TPUClusterResolver), and then connect to the remote cluster and initialize the TPUs.
-
-```py
-import tensorflow as tf
-
-resolver = tf.distribute.cluster_resolver.TPUClusterResolver()
-tf.config.experimental_connect_to_cluster(resolver)
-tf.tpu.experimental.initialize_tpu_system(resolver)
-```
-
-There are various distribution strategies for running your model on multiple TPUs. The [tpu.distribute.TPUStrategy](https://www.tensorflow.org/api_docs/python/tf/distribute/TPUStrategy) offers synchronized distributed training.
-
-```py
-strategy = tf.distribute.TPUStrategy(resolver)
-```
-
-Load and tokenize a dataset - this example uses [CoLA](https://huggingface.co/datasets/nyu-mll/glue/viewer/cola) from the GLUE benchmark - and pad all samples to the maximum length so it is easier to load as an array and to avoid [XLA compilation issues](#xla).
-
-```py
-from transformers import AutoTokenizer
-from datasets import load_dataset
-import numpy as np
-
-dataset = load_dataset("glue", "cola")["train"]
-tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
-
-train_data = tokenizer(
-    dataset["sentence"],
-    padding="max_length",
-    truncation=True,
-    max_length=128,
-    return_tensors="np",
-)
-train_data = dict(train_data)
-train_labels = np.array(dataset["label"])
-```
-
-The model **must** be created inside [Strategy.scope](https://www.tensorflow.org/api_docs/python/tf/distribute/MirroredStrategy#scope) in order to replicate the model layers on each TPU device.
-
-```py
-from transformers import TFAutoModelForSequenceClassification
-
-with strategy.scope():
-    model = TFAutoModelForSequenceClassification.from_pretrained(model_checkpoint)
-    model.compile(optimizer="adam")
-```
-
-TPUs only accept [tf.data.Dataset](https://www.tensorflow.org/api_docs/python/tf/data/Dataset) inputs unlike the Keras [fit](https://keras.io/api/models/model_training_apis/#fit-method) method which accepts a broader range of inputs.
-
-```py
-BATCH_SIZE = 8 * strategy.num_replicas_in_sync
-
-tf_dataset = tf.data.Dataset.from_tensor_slices((train_data, train_labels))
-tf_dataset = tf_dataset.shuffle(len(tf_dataset))
-tf_dataset = tf_dataset.batch(BATCH_SIZE, drop_remainder=True)
-```
-
-Finally, call [fit](https://keras.io/api/models/model_training_apis/#fit-method) to start training.
-
-```py
-model.fit(tf_dataset)
-```
-
-## Large datasets
-
-The dataset created above pads every sample to the maximum length and loads the whole dataset into memory. This may not be possible if you're working with larger datasets. When training on large datasets, you may want to create a [tf.TFRecord](https://www.tensorflow.org/tutorials/load_data/tfrecord) or stream the data.
-
-### tf.TFRecord
-
-[tf.TFRecord](https://www.tensorflow.org/tutorials/load_data/tfrecord) is the standard [tf.data](https://www.tensorflow.org/guide/data) format for storing training data. For very large training jobs, it's worth preprocessing your data and storing it in the `tf.TFRecord` format and building a `tf.data` pipeline on top. Refer to the table below to help you decide whether `tf.TFRecord` is helpful for you.
-
-| pros | cons |
-|---|---|
-| works on all TPU instances | costs associated with cloud storage |
-| supports huge datasets and massive throughput | some data types (images) can take a lot of space to store |
-| suitable for training on entire TPU pods |  |
-| preprocessing is done in advance, maximizing training speed |  |
-
-Preprocess and tokenize the dataset before writing it to a `tf.TFRecord` to avoid writing every time the data is loaded.
-
-An exception is made for *train-time augmentations*, because augmentations applied after writing to a `tf.TFRecord` results in the same augmentation for each epoch. Instead, apply augmentations in the `tf.data` pipeline that loads the data.
-
-> [!TIP]
-> In practice, you probably won't be able to load the entire dataset in memory. Load a chunk of the dataset at a time and convert it to `TFRecord`, and repeat until the entire dataset is in the `TFRecord` format. Then you can use a list of all the files to create a `TFRecordDataset`. The example below demonstrates a single file for simplicity.
-
-```py
-tokenized_data = tokenizer(
-    dataset["sentence"],
-    padding="max_length",
-    truncation=True,
-    max_length=128,
-    return_tensors="np",
-)
-labels = dataset["label"]
-
-with tf.io.TFRecordWriter("dataset.tfrecords") as file_writer:
-    for i in range(len(labels)):
-        features = {
-            "input_ids": tf.train.Feature(
-                int64_list=tf.train.Int64List(value=tokenized_data["input_ids"][i])
-            ),
-            "attention_mask": tf.train.Feature(
-                int64_list=tf.train.Int64List(value=tokenized_data["attention_mask"][i])
-            ),
-            "labels": tf.train.Feature(
-                int64_list=tf.train.Int64List(value=[labels[i]])
-            ),
-        }
-        features = tf.train.Features(feature=features)
-        example = tf.train.Example(features=features)
-        record_bytes = example.SerializeToString()
-        file_writer.write(record_bytes)
-```
-
-Build a [TFRecordDataset](https://www.tensorflow.org/api_docs/python/tf/data/TFRecordDataset) using the saved filename to load it.
-
-```py
-def decode_fn(sample):
-    features = {
-        "input_ids": tf.io.FixedLenFeature((128,), dtype=tf.int64),
-        "attention_mask": tf.io.FixedLenFeature((128,), dtype=tf.int64),
-        "labels": tf.io.FixedLenFeature((1,), dtype=tf.int64),
-    }
-    return tf.io.parse_example(sample, features)
-
-# TFRecordDataset can handle gs:// paths
-tf_dataset = tf.data.TFRecordDataset(["gs://matt-tf-tpu-tutorial-datasets/cola/dataset.tfrecords"])
-tf_dataset = tf_dataset.map(decode_fn)
-tf_dataset = tf_dataset.shuffle(len(dataset)).batch(BATCH_SIZE, drop_remainder=True)
-tf_dataset = tf_dataset.apply(
-    tf.data.experimental.assert_cardinality(len(labels) // BATCH_SIZE)
-)
-```
-
-The dataset can now be passed to the [fit](https://keras.io/api/models/model_training_apis/#fit-method) method.
-
-```py
-model.fit(tf_dataset)
-```
-
-### Stream from raw data
-
-Data can be stored in its native format and preprocessed in a [tf.data](https://www.tensorflow.org/guide/data) pipeline as the data is loaded. This approach isn't supported for many models with complex tokenization schemes, but some models like BERT are supported because their tokenization can be compiled. Refer to the table below to help you decide whether this approach is helpful for you.
-
-| pros | cons |
-|---|---|
-| suitable for highly compressed big data in native format (images, audio) | requires writing a full preprocessing pipeline |
-| convenient if raw data is available in a public cloud bucket | complex preprocessing on-the-fly can hurt throughput |
-| works on all TPU instances if data is stored in Google Cloud | must place data in cloud storage if not already there |
-|  | not as suitable for text data because writing a tokenization pipeline is hard (use `TFRecord` for text) |
-
-The example below demonstrates streaming data for an image model.
-
-Load an image dataset and get a list of the underlying image file paths and labels.
-
-```py
-from datasets import load_dataset
-
-image_dataset = load_dataset("beans", split="train")
-filenames = image_dataset["image_file_path"]
-labels = image_dataset["labels"]
-```
-
-Convert the local filenames in the dataset into `gs://` paths in Google Cloud Storage.
-
-```py
-# strip everything but the category directory and filenames
-base_filenames = ['/'.join(filename.split('/')[-2:]) for filename in filenames]
-# prepend the Google Cloud base path to everything instead
-gs_paths = ["gs://matt-tf-tpu-tutorial-datasets/beans/"+filename for filename in base_filenames]
-
-# create tf_dataset
-tf_dataset = tf.data.Dataset.from_tensor_slices(
-    {"filename": gs_paths, "labels": labels}
-)
-tf_dataset = tf_dataset.shuffle(len(tf_dataset))
-```
-
-Transformers preprocessing classes like [`AutoImageProcessor`] are framework-agnostic and can't be compiled into a pipeline by `tf.data`. To get around this, get the normalization values (`mean` and `std`) from the [`AutoImageProcessor`] and use them in the `tf.data` pipeline.
-
-```py
-from transformers import AutoImageProcessor
-
-processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
-image_size = (processor.size["height"], processor.size["width"])
-image_mean = processor.image_mean
-image_std = processor.image_std
-```
-
-Use these normalization values to create a function to load and preprocess the images.
-
-```py
-BATCH_SIZE = 8 * strategy.num_replicas_in_sync
-
-def decode_fn(sample):
-    image_data = tf.io.read_file(sample["filename"])
-    image = tf.io.decode_jpeg(image_data, channels=3)
-    image = tf.image.resize(image, image_size)
-    array = tf.cast(image, tf.float32)
-    array /= 255.0
-    array = (array - image_mean) / image_std
-    array = tf.transpose(array, perm=[2, 0, 1])
-    return {"pixel_values": array, "labels": sample["labels"]}
-
-tf_dataset = tf_dataset.map(decode_fn)
-tf_dataset = tf_dataset.batch(BATCH_SIZE, drop_remainder=True)
-print(tf_dataset.element_spec)
-```
-
-The dataset can now be passed to the [fit](https://keras.io/api/models/model_training_apis/#fit-method) method.
-
-```py
-from transformers import TFAutoModelForImageClassification
-
-with strategy.scope():
-    model = TFAutoModelForImageClassification.from_pretrained(image_model_checkpoint)
-    model.compile(optimizer="adam")
-
-model.fit(tf_dataset)
-```
-
-### Stream with prepare_tf_dataset
-
-[`~TFPreTrainedModel.prepare_tf_dataset`] creates a `tf.data` pipeline that loads samples from [tf.data.Dataset](https://www.tensorflow.org/api_docs/python/tf/data/Dataset). The pipeline uses [tf.numpy_function]() or [`~datasets.Dataset.from_generator`], which can't be compiled by TensorFlow, to access the underlying `tf.data.Dataset`. It also won't work on a Colab TPU or TPU Nodes because the pipeline streams data from a local disk. Refer to the table below to help you decide whether this approach is helpful for you.
-
-| pros | cons |
-|---|---|
-| simple code | only works on TPU VM |
-| same approach on TPU/GPU | data must be available as a Hugging Face Dataset |
-| dataset doesn't have to fit in memory | data must fit on local storage |
-| supports variable padding | data loading may be a bottleneck on a big TPU pod slice |
-
-[`~TFPreTrainedModel.prepare_tf_dataset`] only works on [TPU VM](#tpu-types). Add the tokenizer output as columns in the dataset since the dataset is stored on disk, which means it can handle data larger than the available memory. Use [`~TFPreTrainedModel.prepare_tf_dataset`] to stream data from the dataset by wrapping it with a `tf.data` pipeline.
-
-```py
-def tokenize_function(examples):
-    return tokenizer(
-        examples["sentence"], padding="max_length", truncation=True, max_length=128
-    )
-# add the tokenizer output to the dataset as new columns
-dataset = dataset.map(tokenize_function)
-
-# prepare_tf_dataset() chooses columns that match the models input names
-tf_dataset = model.prepare_tf_dataset(
-    dataset, batch_size=BATCH_SIZE, shuffle=True, tokenizer=tokenizer
-)
-```
-
-The dataset can now be passed to the [fit](https://keras.io/api/models/model_training_apis/#fit-method) method.
-
-```py
-from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-
-with strategy.scope():
-    model = TFAutoModelForSequenceClassification.from_pretrained(model_checkpoint)
-    model.compile(optimizer="adam")
-
-model.fit(tf_dataset)
-```
-
-## TPU types
-
-There are two types of TPUs, a TPU Node and a TPU VM.
-
-A TPU Node indirectly accesses a remote TPU. It requires a separate VM to initialize your network and data pipeline, and then forwards it to the remote node. Google Colab TPUs are an example of a TPU Node. You can't use local data because the TPU is remotely located, and data must be stored in Google Cloud Storage where the data pipeline can access it.
-
-TPU VM are connected directly to the machine the TPU is located on, and they are generally easier to work with, especially when it comes to your data pipeline.
-
-> [!TIP]
-> We recommend avoiding TPU Nodes if possible because it is more difficult to debug than TPU VMs. TPU Nodes may also be unsupported in the future and become a legacy access method.
-
-A single TPU (v2-8, v3-8, v4-8) runs 8 replicas. TPUs can exist in **pods** which run hundreds or even thousands of replicas simultaneously. When you only use a portion of a pod, it is referred to as a **pod slice**. On Google Colab, you'll typically get a single v2-8 TPU.
-
-## XLA
-
-[XLA](https://openxla.org/xla) is a linear algebra compiler for high-performance execution and it is used by default to improve performance on TPUs.
-
-Before executing your code on a TPU, it's a good idea to try it first on a CPU or GPU because it is easier to debug. You can train for a few steps to make sure the model and data pipeline work as expected. Set `jit_compile=True` in the [compile](https://keras.io/api/models/model_training_apis/#compile-method) method to enable XLA compilation (but remember to remove this line of code before running on a TPU).
-
-The section below outlines three rules for making your code XLA-compatible. Transformers enforce the first two rules for models and loss functions by default, but don't forget about them if you're writing your own models and loss functions.
-
-### Data dependent conditionals
-
-Any `if` statements cannot depend on values inside a [tf.Tensor](https://www.tensorflow.org/api_docs/python/tf/Tensor). The code below can't be compiled by XLA.
-
-```py
-if tf.reduce_sum(tensor) > 10:
-    tensor = tensor / 2.0
-```
-
-To compile with XLA, use [tf.cond](https://www.tensorflow.org/api_docs/python/tf/cond) or remove the conditional and use indicator variables instead as shown below.
-
-```py
-sum_over_10 = tf.cast(tf.reduce_sum(tensor) > 10, tf.float32)
-tensor = tensor / (1.0 + sum_over_10)
-```
-
-### Data dependent shapes
-
-The shape of a [tf.Tensor](https://www.tensorflow.org/api_docs/python/tf/Tensor) cannot depend on their values. For example, [tf.unique](https://www.tensorflow.org/api_docs/python/tf/unique) can't be compiled because it returns a tensor containing an instance of each unique value in the input. The shape of this output depends on how repetitive the input [tf.Tensor](https://www.tensorflow.org/api_docs/python/tf/Tensor) is.
-
-This is an issue during **label masking**, where labels are set to a negative value to indicate they should be ignored when computing the loss. The code below can't be compiled by XLA because the shape of `masked_outputs` and `masked_labels` depend on how many positions are masked.
-
-```py
-label_mask = labels >= 0
-masked_outputs = outputs[label_mask]
-masked_labels = labels[label_mask]
-loss = compute_loss(masked_outputs, masked_labels)
-mean_loss = torch.mean(loss)
-```
-
-To compile with XLA, avoid the data-dependent shapes by computing the loss for every position and zeroing out the masked positions in both the numerator and denominator when calculating the mean. Convert `tf.bool` to `tf.float32` as an indicator variable to make your code XLA-compatible.
-
-```py
-label_mask = tf.cast(labels >= 0, tf.float32)
-loss = compute_loss(outputs, labels)
-loss = loss * label_mask
-mean_loss = tf.reduce_sum(loss) / tf.reduce_sum(label_mask)
-```
-
-### Recompile different input shapes
-
-XLA recompiles your model if input shapes are variable which create huge performance problems. It is especially common in text models because input texts have variable lengths after tokenization.
-
-> [!WARNING]
-> Execessive padding can also severely slow down training because requires more compute and memory to process.
-
-To avoid different shapes, use padding to pad all your inputs to the same length and use an `attention_mask`. Try padding batches of samples to a multiple of 32 or 64 tokens. Use the parameters `padding="max_length"`, `padding="longest"`, or `pad_to_multiple_of` to help with padding. This often increases the number of tokens by a small amount, but it significantly reduces the number of unique input shapes because every input shape is a multiple of 32 or 64. Fewer unique input shapes requires fewer recompilation.

docs/source/en/quantization/hqq.md

@@ -20,7 +20,7 @@ rendered properly in your Markdown viewer.
 
 HQQ further supports fine-tuning with [PEFT](https://huggingface.co/docs/peft) and is fully compatible with [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) for even faster inference and training.
 
-Install HQQ with the following command to get the latest version and to build its corresponding CUDA kernels.
+Install HQQ with the following command to get the latest version and to build its corresponding CUDA kernels if you are using a cuda device. It also support Intel XPU with pure pytorch implementation.
 
 ```bash
 pip install hqq
@@ -34,13 +34,14 @@ You can choose to either replace all the linear layers in a model with the same
 Quantize a model by creating a [`HqqConfig`] and specifying the `nbits` and `group_size` to replace for all the linear layers ([torch.nn.Linear](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html)) of the model.
 
 ``` py
+import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, HqqConfig
 
 quant_config = HqqConfig(nbits=8, group_size=64)
 model = transformers.AutoModelForCausalLM.from_pretrained(
     "meta-llama/Llama-3.1-8B", 
     torch_dtype=torch.float16, 
-    device_map="cuda", 
+    device_map="auto", 
     quantization_config=quant_config
 )
 ```
@@ -67,7 +68,7 @@ quant_config  = HqqConfig(dynamic_config={
 model = transformers.AutoModelForCausalLM.from_pretrained(
     "meta-llama/Llama-3.1-8B", 
     torch_dtype=torch.float16, 
-    device_map="cuda", 
+    device_map="auto", 
     quantization_config=quant_config
 )
 ```

docs/source/en/serving.md

@@ -16,49 +16,195 @@ rendered properly in your Markdown viewer.
 
 # Serving
 
-Transformer models can be served for inference with specialized libraries such as Text Generation Inference (TGI) and vLLM. These libraries are specifically designed to optimize performance with LLMs and include many unique optimization features that may not be included in Transformers.
+Transformer models can be efficiently deployed using libraries such as vLLM, Text Generation Inference (TGI), and others. These libraries are designed for production-grade user-facing services, and can scale to multiple servers and millions of concurrent users. Refer to [Transformers as Backend for Inference Servers](./transformers_as_backends) for usage examples.
 
-## TGI
+Apart from that you can also serve transformer models easily using the `transformers serve` CLI. This is ideal for experimentation purposes, or to run models locally for personal and private use.
 
-[TGI](https://huggingface.co/docs/text-generation-inference/index) can serve models that aren't [natively implemented](https://huggingface.co/docs/text-generation-inference/supported_models) by falling back on the Transformers implementation of the model. Some of TGIs high-performance features aren't available in the Transformers implementation, but other features like continuous batching and streaming are still supported.
+## Serve CLI
 
-> [!TIP]
-> Refer to the [Non-core model serving](https://huggingface.co/docs/text-generation-inference/basic_tutorials/non_core_models) guide for more details.
+> [!WARNING]
+> This section is experimental and subject to change in future versions
 
-Serve a Transformers implementation the same way you'd serve a TGI model.
+You can serve models of diverse modalities supported by `transformers` with the `transformers serve` CLI. It spawns a local server that offers compatibility with the OpenAI SDK, which is the _de facto_ standard for LLM conversations and other related tasks. This way, you can use the server from many third party applications, or test it using the `transformers chat` CLI ([docs](conversations.md#chat-cli)).
 
-```docker
-docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:latest --model-id gpt2
-```
+The server supports the following REST APIs:
+- `/v1/chat/completions`
+- `/v1/responses`
+- `/v1/audio/transcriptions`
+- `/v1/models`
 
-Add `--trust-remote_code` to the command to serve a custom Transformers model.
-
-```docker
-docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:latest --model-id <CUSTOM_MODEL_ID> --trust-remote-code
-```
-
-## vLLM
-
-[vLLM](https://docs.vllm.ai/en/latest/index.html) can also serve a Transformers implementation of a model if it isn't [natively implemented](https://docs.vllm.ai/en/latest/models/supported_models.html#list-of-text-only-language-models) in vLLM.
-
-Many features like quantization, LoRA adapters, and distributed inference and serving are supported for the Transformers implementation.
-
-> [!TIP]
-> Refer to the [Transformers fallback](https://docs.vllm.ai/en/latest/models/supported_models.html#transformers-fallback) section for more details.
-
-By default, vLLM serves the native implementation and if it doesn't exist, it falls back on the Transformers implementation. But you can also set `--model-impl transformers` to explicitly use the Transformers model implementation.
+To launch a server, simply use the `transformers serve` CLI command:
 
 ```shell
-vllm serve Qwen/Qwen2.5-1.5B-Instruct \
-    --task generate \
-    --model-impl transformers
+transformers serve
 ```
 
-Add the `trust-remote-code` parameter to enable loading a remote code model.
+The simplest way to interact with the server is through our `transformers chat` CLI
 
 ```shell
-vllm serve Qwen/Qwen2.5-1.5B-Instruct \
-    --task generate \
-    --model-impl transformers \
-    --trust-remote-code
-```
+transformers chat localhost:8000 --model-name-or-path Qwen/Qwen3-4B
+```
+
+or by sending an HTTP request with `cURL`, e.g.
+
+```shell
+curl -X POST http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{"messages": [{"role": "system", "content": "hello"}], "temperature": 0.9, "max_tokens": 1000, "stream": true, "model": "Qwen/Qwen2.5-0.5B-Instruct"}'
+```
+
+from which you'll receive multiple chunks in the Completions API format
+
+```shell
+data: {"object": "chat.completion.chunk", "id": "req_0", "created": 1751377863, "model": "Qwen/Qwen2.5-0.5B-Instruct", "system_fingerprint": "", "choices": [{"delta": {"role": "assistant", "content": "", "tool_call_id": null, "tool_calls": null}, "index": 0, "finish_reason": null, "logprobs": null}]}
+
+data: {"object": "chat.completion.chunk", "id": "req_0", "created": 1751377863, "model": "Qwen/Qwen2.5-0.5B-Instruct", "system_fingerprint": "", "choices": [{"delta": {"role": "assistant", "content": "", "tool_call_id": null, "tool_calls": null}, "index": 0, "finish_reason": null, "logprobs": null}]}
+
+(...)
+```
+
+The server is also an MCP client, so it can interact with MCP tools in agentic use cases. This, of course, requires the use of an LLM that is designed to use tools.
+
+> [!TIP]
+> At the moment, MCP tool usage in `transformers` is limited to the `qwen` family of models.
+
+<!-- TODO: example with a minimal python example, and explain that it is possible to pass a full generation config in the request -->
+
+
+### Usage example 1: chat with local requests (feat. Jan)
+
+This example shows how to use `transformers serve` as a local LLM provider for the [Jan](https://jan.ai/) app. Jan is a ChatGPT-alternative graphical interface, fully running on your machine. The requests to `transformers serve` come directly from the local app -- while this section focuses on Jan, you can extrapolate some instructions to other apps that make local requests.
+
+To connect `transformers serve` with Jan, you'll need to set up a new model provider ("Settings" > "Model Providers"). Click on "Add Provider", and set a new name. In your new model provider page, all you need to set is the "Base URL" to the following pattern:
+
+```shell
+http://[host]:[port]/v1
+```
+
+where `host` and `port` are the `transformers serve` CLI parameters (`localhost:8000` by default). After setting this up, you should be able to see some models in the "Models" section, hitting "Refresh". Make sure you add some text in the "API key" text field too -- this data is not actually used, but the field can't be empty. Your custom model provider page should look like this:
+
+<h3 align="center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_jan_model_providers.png"/>
+</h3>
+
+You are now ready to chat!
+
+> [!TIP]
+> You can add any `transformers`-compatible model to Jan through `transformers serve`. In the custom model provider you created, click on the "+" button in the "Models" section and add its Hub repository name, e.g. `Qwen/Qwen3-4B`.
+
+To conclude this example, let's look into a more advanced use-case. If you have a beefy machine to serve models with, but prefer using Jan on a different device, you need to add port forwarding. If you have `ssh` access from your Jan machine into your server, this can be accomplished by typing the following to your Jan machine's terminal
+
+```
+ssh -N -f -L 8000:localhost:8000 your_server_account@your_server_IP -p port_to_ssh_into_your_server
+```
+
+Port forwarding is not Jan-specific: you can use it to connect `transformers serve` running in a different machine with an app of your choice.
+
+
+### Usage example 2: chat with external requests (feat. Cursor)
+
+This example shows how to use `transformers serve` as a local LLM provider for [Cursor](https://cursor.com/), the popular IDE. Unlike in the previous example, requests to `transformers serve` will come from an external IP (Cursor's server IPs), which requires some additional setup. Furthermore, some of Cursor's requests require [CORS](https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/CORS), which is disabled by default for security reasons.
+
+To launch a server with CORS enabled, run
+
+```shell
+transformers serve --enable-cors
+```
+
+You'll also need to expose your server to external IPs. A potential solution is to use [`ngrok`](https://ngrok.com/), which has a permissive free tier. After setting up your `ngrok` account and authenticating on your server machine, you run
+
+```shell
+ngrok http [port]
+```
+
+where `port` is the port used by `transformers serve` (`8000` by default). On the terminal where you launched `ngrok`, you'll see an https address in the "Forwarding" row, as in the image below. This is the address to send requests to.
+
+<h3 align="center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_ngrok.png"/>
+</h3>
+
+You're now ready to set things up on the app side! In Cursor, while you can't set a new provider, you can change the endpoint for OpenAI requests in the model selection settings. First, navigate to "Settings" > "Cursor Settings", "Models" tab, and expand the "API Keys" collapsible. To set your `transformers serve` endpoint, follow this order:
+1. Unselect ALL models in the list above (e.g. `gpt4`, ...);
+2. Add and select the model you want to use (e.g. `Qwen/Qwen3-4B`)
+3. Add some random text to OpenAI API Key. This field won't be used, but it can’t be empty;
+4. Add the https address from `ngrok` to the "Override OpenAI Base URL" field, appending `/v1` to the address (i.e. `https://(...).ngrok-free.app/v1`);
+5. Hit "Verify".
+
+After you follow these steps, your "Models" tab should look like the image below. Your server should also have received a few requests from the verification step.
+
+<h3 align="center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_cursor.png"/>
+</h3>
+
+You are now ready to use your local model in Cursor! For instance, if you toggle the AI Pane, you can select the model you added and ask it questions about your local files.
+
+<h3 align="center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_cursor_chat.png"/>
+</h3>
+
+
+### Usage example 3: `tiny-agents` CLI and MCP Tools
+
+To showcase the use of MCP tools, let's see how to integrate the `transformers serve` server with the [`tiny-agents`](https://huggingface.co/blog/python-tiny-agents) CLI.
+
+> [!TIP]
+> Many Hugging Face Spaces can be used as MCP servers, as in this example. You can find all compatible Spaces [here](https://huggingface.co/spaces?filter=mcp-server).
+
+The first step to use MCP tools is to let the model know which tools are available. As an example, let's consider a `tiny-agents` configuration file with a reference to an [image generation MCP server](https://evalstate-flux1-schnell.hf.space/).
+
+```json
+{
+    "model": "Menlo/Jan-nano",
+    "endpointUrl": "http://localhost:8000",
+    "servers": [
+        {
+            "type": "sse",
+            "url": "https://evalstate-flux1-schnell.hf.space/gradio_api/mcp/sse"
+        }
+    ]
+}
+```
+
+You can then launch your `tiny-agents` chat interface with the following command.
+
+```bash
+tiny-agents run path/to/your/config.json
+```
+
+If you have `transformers serve` running in the background, you're ready to use MCP tools from a local model! For instance, here's the example of a chat session with `tiny-agents`:
+
+```bash
+Agent loaded with 1 tools:
+ • flux1_schnell_infer
+»  Generate an image of a cat on the moon
+<Tool req_0_tool_call>flux1_schnell_infer {"prompt": "a cat on the moon", "seed": 42, "randomize_seed": true, "width": 1024, "height": 1024, "num_inference_steps": 4}
+
+Tool req_0_tool_call
+[Binary Content: Image image/webp, 57732 bytes]
+The task is complete and the content accessible to the User
+Image URL: https://evalstate-flux1-schnell.hf.space/gradio_api/file=/tmp/gradio/3dbddc0e53b5a865ed56a4e3dbdd30f3f61cf3b8aabf1b456f43e5241bd968b8/image.webp
+380576952
+
+I have generated an image of a cat on the moon using the Flux 1 Schnell Image Generator. The image is 1024x1024 pixels and was created with 4 inference steps. Let me know if you would like to make any changes or need further assistance!
+```
+
+### Usage example 4: speech to text transcription (feat. Open WebUI)
+
+This guide shows how to do audio transcription for chat purposes, using `transformers serve` and [Open WebUI](https://openwebui.com/). This guide assumes you have Open WebUI installed on your machine and ready to run. Please refer to the examples above to use the text functionalities of `transformer serve` with Open WebUI -- the instructions are the same.
+
+To start, let's launch the server. Some of Open WebUI's requests require [CORS](https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/CORS), which is disabled by default for security reasons, so you need to enable it:
+
+```shell
+transformers serve --enable-cors
+```
+
+Before you can speak into Open WebUI, you need to update its settings to use your server for speech to text (STT) tasks. Launch Open WebUI, and navigate to the audio tab inside the admin settings. If you're using Open WebUI with the default ports, [this link (default)](http://localhost:3000/admin/settings/audio) or [this link (python deployment)](http://localhost:8080/admin/settings/audio) will take you there. Do the following changes there:
+1. Change the type of "Speech-to-Text Engine" to "OpenAI";
+2. Update the address to your server's address -- `http://localhost:8000/v1` by default;
+3. Type your model of choice into the "STT Model" field, e.g. `openai/whisper-large-v3` ([available models](https://huggingface.co/models?pipeline_tag=automatic-speech-recognition&sort=trending)).
+
+If you've done everything correctly, the audio tab should look like this
+
+<h3 align="center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_openwebui_stt_settings.png"/>
+</h3>
+
+You're now ready to speak! Open a new chat, utter a few words after hitting the microphone button, and you should see the corresponding text on the chat input after the model transcribes it.

docs/source/en/tf_xla.md

@@ -1,129 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
-
-⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
-rendered properly in your Markdown viewer.
-
--->
-
-# XLA
-
-[[open-in-colab]]
-
-[Accelerated Linear Algebra (XLA)](https://openxla.org/xla) is a linear algebra compiler that optimizes model runtime across different hardware and frameworks.
-
-This guide will look specifically at how to accelerate *TensorFlow* models with XLA.
-
-## TensorFlow
-
-XLA can potentially accelerate a TensorFlow model without making any source code changes. It is already packaged with the TensorFlow library, and it is triggered with `jit_compile` in any graph creating function such as [tf.function](https://www.tensorflow.org/api_docs/python/tf/function).
-
-If you're using Keras methods like [fit](https://keras.io/api/models/model_training_apis/#fit-method) and [predict](https://keras.io/api/models/model_training_apis/#predict-method), enable XLA by passing `jit_compile=True` to [compile](https://keras.io/api/models/model_training_apis/#compile-method).
-
-```py
-model.compile(jit_compile=True)
-```
-
-XLA can be used to accelerate any arbitrary [tf.function](https://www.tensorflow.org/api_docs/python/tf/function).
-
-Models with a TensorFlow implementation like [GPT2](./model_doc/gpt2), [T5](./model_doc/t5), [OPT](./model_doc/opt), and [Whisper](./model_doc/whisper) are XLA compatible. The speed up depends on a model, but in general, TensorFlow models in Transformers get a ~100x speed up.
-
-### Functions
-
-A typical forward pass in a TensorFlow model is shown below. To run a forward pass with XLA, wrap the model with [tf.function](https://www.tensorflow.org/api_docs/python/tf/function) and set `jit_compile=True`.
-
-```diff
-import tensorflow as tf
-
-model = tf.keras.Sequential(
-    [tf.keras.layers.Dense(10, input_shape=(10,), activation="relu"), tf.keras.layers.Dense(5, activation="softmax")]
-)
-# Generate random inputs for the model.
-batch_size = 16
-input_vector_dim = 10
-random_inputs = tf.random.normal((batch_size, input_vector_dim))
-
-# Run a forward pass.
-- _ = model(random_inputs)
-+ xla_fn = tf.function(model, jit_compile=True)
-+ _ = xla_fn(random_inputs)
-```
-
-The default `call` function of the model is used to compile the XLA graph. But if there's any other model function you want to compile with XLA, wrap them with [tf.function](https://www.tensorflow.org/api_docs/python/tf/function).
-
-```py
-my_xla_fn = tf.function(model.my_xla_fn, jit_compile=True)
-```
-
-### Text generation
-
-You could also compile other model functions with XLA. For example, enable XLA for text generation by wrapping [`~TFGenerationMixin.generate`] with [tf.function](https://www.tensorflow.org/api_docs/python/tf/function).
-
-```py
-import tensorflow as tf
-from transformers import AutoTokenizer, TFAutoModelForCausalLM
-# Will error if the minimal version of Transformers is not installed.
-from transformers.utils import check_min_version
-
-check_min_version("4.21.0")
-
-tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2", padding_side="left", pad_token="</s>")
-model = TFAutoModelForCausalLM.from_pretrained("openai-community/gpt2")
-input_string = ["TensorFlow is"]
-
-xla_generate = tf.function(model.generate, jit_compile=True)
-
-tokenized_input = tokenizer(input_string, return_tensors="tf")
-generated_tokens = xla_generate(**tokenized_input, num_beams=2)
-
-decoded_text = tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
-print(f"Generated -- {decoded_text}")
-"Generated -- TensorFlow is an open-source, open-source, distributed-source application framework for the"
-```
-
-## Tracing
-
-When executing an XLA-enabled function for the first time, it tries to infer the computation graph in a process known as *tracing*. This is a time-consuming step, but any consecutive calls to the function will be much faster because it won't have to trace the computation graph again.
-
-To ensure a function is only traced once, the inputs must have the same shape as when the graph was built. This usually isn't an issue for fixed input shapes like images, but it can be an issue for inputs with variable shapes like text.
-
-One way to handle this is to pad your text so it always has the same shape. Configure padding options such as [pad_to_multiple_of](https://hf.co/docs/transformers/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.pad.pad_to_multiple_of) in the tokenizer.
-
-```py
-import tensorflow as tf
-from transformers import AutoTokenizer, TFAutoModelForCausalLM
-
-tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2", padding_side="left", pad_token="</s>")
-model = TFAutoModelForCausalLM.from_pretrained("openai-community/gpt2")
-input_string = ["TensorFlow is"]
-
-xla_generate = tf.function(model.generate, jit_compile=True)
-
-# Call tokenizer with padding options.
-tokenized_input = tokenizer(input_string, pad_to_multiple_of=8, padding=True, return_tensors="tf")
-
-generated_tokens = xla_generate(**tokenized_input, num_beams=2)
-decoded_text = tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
-print(f"Generated -- {decoded_text}")
-```
-
-In addition to the input shape, any changes to the generation options at any point also triggers tracing.
-
-## Resources
-
-Learn more about XLA with the following resources.
-
-- A [notebook](https://colab.research.google.com/github/huggingface/blog/blob/main/notebooks/91_tf_xla_generate.ipynb) demonstrating XLA-compatible encoder-decoder and decoder-only text generation models.
-- The [Faster Text Generation with TensorFlow and XLA](https://hf.co/blog/tf-xla-generate) blog post compares benchmarks for XLA-compatible models and provides a friendly introduction to XLA in TensorFlow.
-- The [How Hugging Face improved Text Generation performance with XLA](https://blog.tensorflow.org/2022/11/how-hugging-face-improved-text-generation-performance-with-xla.html) blog post discusses the design philosophy behind adding XLA to TensorFlow models in Transformers.
-- The [Introduction to graphs and tf.function](https://www.tensorflow.org/guide/intro_to_graphs) guide.
-- The [Better performance with tf.function](https://www.tensorflow.org/guide/function) guide.
-- The [XLA](https://openxla.org/xla) documentation.

docs/source/en/tools.md

@@ -14,5 +14,9 @@ rendered properly in your Markdown viewer.
 
 -->
 
+# Tools
+
+(deprecated)
+
 > [!WARNING]
 > Agents and tools were spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. They were removed from `transformers` in v4.52.

docs/source/en/trainer.md

@@ -187,13 +187,13 @@ from torch import nn
 from transformers import Trainer
 
 class CustomTrainer(Trainer):
-    def compute_losss(self, model: nn.Module, inputs: dict[str, Union[torch.Tensor, Any]], return_outputs: bool = False num_items_in_batch: Optional[torch.Tensor] = None):
+    def compute_loss(self, model: nn.Module, inputs: dict[str, Union[torch.Tensor, Any]], return_outputs: bool = False num_items_in_batch: Optional[torch.Tensor] = None):
         labels = inputs.pop("labels")
         # forward pass
         outputs = model(**inputs)
         logits = outputs.get("logits")
         # compute custom loss for 3 labels with different weights
-        reduction = "mean" if num_items_in_batch is not None else "sum"
+        reduction = "sum" if num_items_in_batch is not None else "mean"
         loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 2.0, 3.0], device=model.device, reduction=reduction))
         loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
         if num_items_in_batch is not None:

docs/source/en/transformers_as_backend.md

@@ -0,0 +1,254 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Inference server backends
+
+Transformers' models are compatible with different inference servers like vLLM and SGLang. Instead of implementing a model for each inference server, you only need one model, which can be plugged into any inference server. It simplifies maintenance and makes it easy for users to use different inference servers for different use cases.
+
+With Transformers as a backend, you can also serve any model - including custom and Hub-hosted models - without waiting for native support.
+
+This guide shows how to use Transformers' models as a backend to some popular inference servers and how to build a model that supports all inference servers.
+
+## vLLM
+
+[vLLM](https://github.com/vllm-project/vllm) is a high-performance inference engine optimized for serving LLMs at scale. It supports many Transformers' models, including all decoder-only LLMs and several vision-language models (VLMs). VLMs currently support image inputs only, with video support planned.
+
+vLLM automatically selects the best backend, and if a model isn’t natively supported, it falls back to the Transformers model. To explicitly use a Transformers' model, set `model_impl="transformers"`.
+
+```python
+from vllm import LLM
+llm = LLM(model="meta-llama/Llama-3.2-1B", model_impl="transformers")
+```
+Add `--model-impl transformers` to `vllm serve` to launch a server with a Transformers' model.
+
+```bash
+vllm serve meta-llama/Llama-3.2-1B \
+    --task generate \
+    --model-impl transformers
+```
+
+Refer to the [vLLM docs](https://docs.vllm.ai/en/latest/models/transformers_backend.html) for more usage examples and tips on using a Transformers as the backend.
+
+
+## SGLang
+
+[SGLang](https://github.com/InternLM/sglang) is a high-performance, OpenAI-compatible server and runtime designed for chat-based LLMs. It offers fast inference, role-based conversation handling, and support for custom pipelines, making it great for building real-world LLM apps.
+
+SGLang automatically falls back to the Transformers backend if a model isn’t natively supported. To explicitly use a Transformers' model, set `impl="transformers"`.
+
+```python
+import sglang as sgl
+
+llm = sgl.Engine("meta-llama/Llama-3.2-1B-Instruct", impl="transformers")
+print(llm.generate(["The capital of France is"], {"max_new_tokens": 20})[0])
+```
+
+Add `impl transformers` to `sglang.launch_server` to launch a server with a Transformers' model.
+          
+      
+    
+    
+  
+
+
+```bash
+python3 -m sglang.launch_server \
+  --model-path kyutai/helium-1-preview-2b \
+  --impl transformers \
+  --host 0.0.0.0 \
+  --port 30000
+```
+
+Refer to the [SGLang docs](https://docs.sglang.ai/supported_models/transformers_fallback.html) for more usage examples and tips on using a Transformers as the backend.
+
+## TGI
+
+[TGI](https://huggingface.co/docs/text-generation-inference/index) can serve models that aren't [natively implemented](https://huggingface.co/docs/text-generation-inference/supported_models) by falling back on the Transformers implementation of the model. Some of TGIs high-performance features aren't available in the Transformers implementation, but other features like continuous batching and streaming are still supported.
+
+> [!TIP]
+> Refer to the [Non-core model serving](https://huggingface.co/docs/text-generation-inference/basic_tutorials/non_core_models) guide for more details.
+
+Serve a Transformers implementation the same way you'd serve a TGI model.
+
+```docker
+docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:latest --model-id gpt2
+```
+
+Add `--trust-remote_code` to the command to serve a custom Transformers model.
+
+```docker
+docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:latest --model-id <CUSTOM_MODEL_ID> --trust-remote-code
+```
+
+## Building a compatible model backend
+
+To ensure a model is compatible as a backend to any inference server, make sure it is compatible with Transformers and supports the [AttentionInterface](./attention_interface) class.
+
+1. A model must be Transformers-compatible following the model [contribution guidelines](./add_new_model) or the [custom model contribution guidelines](./custom_models). Make sure the model has a valid `config.json` in its directory and a valid `auto_map` field pointing to the model class in the config.
+
+2. A model's attentions needs to be configurable with the [AttentionInterface](./attention_interface) to allow custom and optimized attention functions. This is important for enabling the performance features of the different inference servers.
+   Use `ALL_ATTENTION_FUNCTIONS` when defining the attention layer and propagate `**kwargs**` from the base `MyModel` class to the attention layers. Set `_supports_attention_backend` to `True` in [`PreTrainedModel`]. Expand the code below for an example.
+
+<details>
+<summary>modeling_my_model.py</summary>
+
+```python
+
+from transformers import PreTrainedModel
+from torch import nn
+
+class MyAttention(nn.Module):
+
+    def forward(self, hidden_states, **kwargs):
+        ...
+        attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            **kwargs,
+        )
+        ...
+
+class MyModel(PreTrainedModel):
+    _supports_attention_backend = True
+```
+
+</details>
+
+3. This step is optional, but if you want to support tensor parallel and/or pipeline parallel features, add the following keys to the config.
+    * `base_model_tp_plan` enables [tensor parallelism](./perf_infer_gpu_multi) by mapping fully qualified layer name patterns to tensor parallel styles. Only the `"colwise"` and `"rowwise"` partitioning strategies are currently supported.
+    * `base_model_pp_plan` enables pipeline parallelism by mapping direct child layer names to tuples of lists of strings. The list in the first element of the tuple contains the names of the input arguments. The list in the last element of the tuple contains the names of the variables the layer outputs to in the modeling code.
+ 
+ Expand the code below for an example.
+
+<details>
+<summary>configuration_my_model.py</summary>
+
+```python
+
+from transformers import PretrainedConfig
+
+class MyConfig(PretrainedConfig):
+    base_model_tp_plan = {
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+```
+</details>
+
+### Multimodal models
+
+For multimodal models, you need to include a few more changes on top of the general recommendations. These rules ensure that your model integrates properly with multimodal data.
+
+1. A multimodal model requires a base `MyMultiModalModel` class to handle multimodal fusion without a language modeling head and a separate generative class that adds a head.
+
+    The base model needs to implement the `get_image_features()` method to accept image pixel values and return encoded outputs. These are later merged with the language embeddings and don't require any postprocessing. The shape of the returned features must match the number of input images. If a vision encoder returns variable-length outputs (patch-based), return a list of 2D tensors of size `(image_seq_len, image_dim)` for each image.
+
+Expand the code below for an example.
+
+<details>
+<summary>modeling_my_multimodal_model.py</summary>
+
+```python
+from transformers.generation import GenerationMixin
+
+class MyMultimodalModel(MyMultimodalPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.language_model = AutoModel.from_config(config.text_config)
+        self.vision_tower = AutoModel.from_config(config.vision_config)
+        self.multimodal_projection = nn.Linear(vision_dim, text_dim)
+    
+    def get_image_features(self, pixel_values):
+        return self.vision_tower(pixel_values).last_hidden_states
+    
+    def forward(self, input_ids, pixel_values, **kwargs):
+        # process your inputs
+        return MyModelOutputWithPast(
+            last_hidden_state=last_hidden_state,
+            image_hidden_states=image_features,
+            [...]
+        )
+
+class MyMultimodalModelForConditionalGeneration(MyMultimodalPreTrainedModel, GenerationMixin):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = MyMultimodalModel(config)
+        self.lm_head = nn.Linear(hidden_dim, vocab_size)
+```
+</details>
+
+
+2. A multimodal model config must be nested with the following fields.
+    * text_config: decoder language model config
+    * vision_config: vision encoder config
+    * image_token_id: ID of the image placeholder token used in the input to indicate image position
+
+3. A multimodal model's processing class must have the `self.image_token` and `self.image_token_ids` attributes. These are placeholder tokens used to indicate image positions in the input. The placeholder token is the same token used in the input prompt and to mask scatter image features.
+
+   The processing class also needs ` self._get_num_multimodal_tokens` method to compute the number of placeholder tokens needed for multimodal inputs with given sizes and to return a [`MultiModalData`] object. The placeholder for row and column tokens don't count as image placeholders. Only the tokens that are actually replaced by image features are computed.
+
+Finally, when `return_mm_token_type_ids=True`, the class has to return `mm_token_type_ids` to indicate whether each position is a text token (`0`) or image placeholder token (`1`). Each image's token type IDs must be contiguous with no breaks between consecutive ones.
+
+Expand the code below for an example.
+
+<details>
+<summary>processing_my_multimodal_model.py</summary>
+
+```python
+class MyMultimodalProcessor(ProcessorMixin):
+
+    def __call__(self, images=None, text=None, **kwargs):
+        if return_mm_token_type_ids:
+            mm_token_type_ids = np.zeros_like(input_ids)
+            mm_token_type_ids[input_ids == self.image_token_id] = 1
+            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
+        return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)
+
+    def _get_num_multimodal_tokens(self, image_sizes=None, **kwargs):
+        """
+        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
+        Args:
+            image_sizes (`list[list[int]]`, *optional*):
+                The input sizes formatted as (height, width) per each image.
+        Returns:
+            `MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided
+            input modalities, along with other useful data.
+        """
+        vision_data = {}
+        if image_sizes is not None:
+            num_image_tokens = [256] * len(image_sizes) # 256 placeholder tokens for each image always
+            num_image_patches = [1] * len(image_sizes) # no patching, thus each image is processed as a single base image
+            vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})
+        return MultiModalData(**vision_data)
+```
+</details>
+
+## Resources
+
+* Read the [Transformers backend integration in vLLM](https://blog.vllm.ai/2025/04/11/transformers-backend.html) blog post for more details about the Transformers backend in vLLM.
+* Read the [Transformers backend integration in SGLang](https://huggingface.co/blog/transformers-backend-sglang) blog post for more details about the Transformers backend in SGLang.

docs/source/it/perf_train_special.md

@@ -7,6 +7,7 @@ http://www.apache.org/licenses/LICENSE-2.0
 
 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
 an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
 
 ⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
 rendered properly in your Markdown viewer.

docs/source/ko/_toctree.yml

@@ -157,6 +157,8 @@
     title: (번역중) VPTQ
   - local: quantization/quanto
     title: Quanto
+  - local: quantization/quark
+    title: Quark
   - local: quantization/eetq
     title: EETQ
   - local: in_translation
@@ -200,6 +202,8 @@
       title: CPU로 추론하기
     - local: perf_infer_gpu_one
       title: 하나의 GPU를 활용한 추론
+    - local: perf_infer_gpu_multi
+      title: 다중 GPU를 활용한 추론
     title: 추론 최적화하기
   - local: big_models
     title: 대형 모델을 인스턴스화
@@ -225,7 +229,7 @@
 - sections:
   - local: philosophy
     title: 이념과 목표
-  - local: in_translation
+  - local: glossary
     title: (번역중) Glossary
   - local: task_summary
     title: 🤗 Transformers로 할 수 있는 작업

docs/source/ko/glossary.md

@@ -0,0 +1,454 @@
+<!--Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# 용어집(Glossary)
+
+이 용어집은 전반적인 머신러닝 및 🤗 Transformers 관련 용어를 정의하여 문서를 더 잘 이해하는 데 도움을 줍니다.
+
+## A
+
+### 어텐션 마스크 (attention mask)
+
+어텐션 마스크(attention mask)는 여러 시퀀스를 배치(batch)로 처리할 때 사용되는 선택적 인자입니다.
+
+<Youtube id="M6adb1j2jPI"/>
+
+이 인자는 모델에게 어떤 토큰에 주의를 기울여야 하는지, 그리고 어떤 토큰은 무시해야 하는지를 알려줍니다.
+
+예를 들어, 다음 두 개의 시퀀스가 있다고 가정해 봅시다:
+
+```python
+>>> from transformers import BertTokenizer
+
+>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
+
+>>> sequence_a = "This is a short sequence."
+>>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
+
+>>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
+>>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
+```
+
+인코딩된 버전들의 길이가 다릅니다:
+
+```python
+>>> len(encoded_sequence_a), len(encoded_sequence_b)
+(8, 19)
+```
+
+따라서 이 두 시퀀스를 그대로 하나의 텐서에 넣을 수는 없습니다. 첫 번째 시퀀스를 두 번째 길이에 맞춰 패딩 하거나, 반대로 두 번째 시퀀스를 첫 번째 길이에 맞춰 잘라내야 합니다.
+
+첫 번째 경우에는 ID 목록이 패딩 인덱스로 확장됩니다. 이렇게 패딩을 적용하려면 토크나이저에 리스트를 전달하고 다음과 같이 요청할 수 있습니다:
+
+```python
+>>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
+```
+
+첫 번째 문장 오른쪽에 0이 추가되어 두 번째 문장과 길이가 같아진 것을 볼 수 있습니다:
+
+```python
+>>> padded_sequences["input_ids"]
+[[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
+```
+
+이것은 PyTorch나 TensorFlow의 텐서로 변환될 수 있습니다. 어텐션 마스크는 모델이 패딩 된 인덱스를 참조하지 않도록 해당 위치를 나타내는 이진 텐서입니다. [`BertTokenizer`]의 경우, `1`은 어텐션이 필요한 값을 나타내고, `0`은 패딩 된 값을 나타냅니다. 이 어텐션 마스크는 토크나이저가 반환되는 딕셔너리의 "attention_mask" 키 아래에 포함되어 있습니다:
+
+```python
+>>> padded_sequences["attention_mask"]
+[[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
+```
+
+### 오토인코딩 모델 (autoencoding models)
+
+[인코더 모델](#encoder-models)과 [마스킹된 언어 모델링](#masked-language-modeling-mlm)을 참고하세요.
+
+### 자기회귀 모델 (autoregressive models)
+
+[인과적 언어 모델링](#causal-language-modeling)과 [디코더 모델](#decoder-models)을 참고하세요.
+
+## B
+
+### 백본 (backbone)
+
+백본(backbone)은 원시(hidden) 은닉 상태(hidden state) 또는 특징(feature)을 출력하는 네트워크(임베딩과 레이어)입니다. 일반적으로 이 백본은 해당 특징을 입력으로 받아 예측을 수행하는 [헤드](#head)와 연결됩니다. 예를 들어, [`ViTModel`]은 특정 헤드가 없는 백본입니다. 다른 모델들도[`VitModel`]을 백본으로 사용할 수 있으며, [DPT](model_doc/dpt)등이 그 예시입니다.
+
+## C
+
+### 인과적 언어 모델링 (causal language modeling)
+
+모델이 텍스트를 순서대로 읽으며 다음 단어를 예측해야 하는 사전 학습(pretraining) 작업입니다. 일반적으로 문장을 전체로 읽되, 모델 내부에서 특징 시점 이후의 토큰을 마스킹(masking)하여 다음 단어를 예측하게 됩니다.
+
+### 채널 (channel)
+
+컬러 이미지는 빨간색(R), 초록색(G), 파란색(B)의 세 채널 값을 조합하여 구성되며, 흑백 이미지는 단일 채널만을 가집니다. 🤗 Transformers에서는 이미지 텐서의 채널이 첫 번째 또는 마지막 차원에 위치할 수 있습니다:[`n_channels`, `height`, `width`] 또는 [`height`, `width`, `n_channels`]와 같은 형식입니다.
+
+### 연결 시간분류(connectionist temporal classification, CTC) 
+
+입력과 출력의 정렬 상태를 정확히 몰라도 모델이 학습할 수 있도록 돕는 알고리즘입니다. CTC는 주어진 입력에 대해 가능한 모든 출력의 확률 분포를 계산하고, 그중 가장 가능성이 높은 출력을 선택합니다. CTC는 말하는 속도의 차이 등 여러 이유로 음성과 텍스트가 항상 정확하게 일치하지 않기 때문에 음성 인식 작업에서 자주 사용됩니다.
+
+### 컨볼루션 (convolution)
+
+신경망에서 사용되는 레이어의 한 종류로, 입력 행렬에 대해 더 작은 행렬(커널 또는 필터)을 원소별로 곱한 뒤 그 값을 합산해 새로운 행렬을 만드는 연산입니다. 이 연산을 컨볼루션 연산이라고 하며, 입력 행렬 전체에 걸쳐 반복적으로 수행됩니다. 각 연산은 입력 행렬의 서로 다른 구간에 적용됩니다. 컨볼루션 신경망(CNN)은 컴퓨터 비전 분야에서 널리 사용됩니다.
+
+## D
+
+### 데이터 병렬화 (DataParallel)
+
+여러 개의 GPU에서 훈련을 수행할 때 사용하는 병렬화 기법으로, 동일한 모델 구성이 여러 번 복제되며 각 인스턴스는 서로 다른 데이터 조각을 받습니다. 모든 인스턴스는 병렬로 처리를 수행하며, 각 훈련 단계가 끝난 후 결과를 동기화합니다.
+
+DataParallel 방식에 대해 더 알아보려면 [여기](perf_train_gpu_many#dataparallel-vs-distributeddataparallel)를 참고하세요.
+
+### 디코더 입력 ID (decoder input IDs)
+
+이 입력은 인코더-디코더 모델에 특화된 것으로, 디코더에 전달될 input ID 들을 포함합니다. 이러한 입력은 번역이나 요약과 같은 시퀀스-투-시퀀스(sequence-to-sequence) 작업에 사용되며, 일반적으로 모델마다 고유한 방식으로 구성됩니다.
+
+대부분의 인코더-디코더 모델(BART, T5 등)은 `labels`로부터 자동으로 `decoder_input_ids`를 생성합니다. 이러한 모델에서는 학습 시 `labels`를 전달하는 것이 일반적으로 권장됩니다.
+
+시퀀스-투-시퀀스 학습에서 각 모델이 이러한 input ID를 어떻게 처리하는지는 모델 문서를 참고하시기를 바랍니다.
+
+### 디코더 모델 (decoder models)
+
+자기회귀 모델(Autoregressive models)이라고도 불리는 디코더 모델은 인과 언어 모델링(causal language modeling)이라 불리는 사전 학습 작업을 수행합니다. 이 작업에서는 모델이 텍스트를 순서대로 읽고 다음 단어를 예측해야 합니다. 일반적으로 문장의 전체를 읽되, 특정 시점 이후의 토큰은 마스크로 가려 예측하게 합니다.
+
+<Youtube id="d_ixlCubqQw"/>
+
+### 딥러닝 (deep learning)
+
+여러 층의 신경망(neural network)을 사용하는 머신러닝 알고리즘입니다.
+
+## E
+
+### 인코더 모델 (encoder models)
+
+자동 인코딩 모델(Autoencoding models)이라고도 불리는 인코더 모델은 텍스트나 이미지와 같은 입력을 받아 임베딩이라 불리는 압축된 수치 표현으로 반환합니다. 일반적으로 인코더 모델은 입력 시퀀스의 일부를 마스킹하고 더 의미 있는 표현을 생성하도록 학습하는 [masked language modeling](#masked-language-modeling-mlm)과 같은 기술을 사용하여 사전 학습됩니다.
+
+<Youtube id="H39Z_720T5s"/>
+
+## F
+
+### 특징 추출 (feature extraction)
+
+머신러닝 알고리즘이 더 효과적으로 학습할 수 있도록, 원시 데이터를 선택하고 변환하여 더 유용한 특징(feature) 집합으로 만드는 과정입니다. 예를 들어, 원시 텍스트를 워드 임베딩으로 변환하거나 이미지나 비디오 데이터에서 윤곽선이나 형태와 같은 중요한 특징을 추출하는 것이 있습니다.
+
+### 피드 포워드 청킹 (feed forward chunking)
+
+트랜스포머의 각 residual attention Block에서는 self-Attention Layer 다음에 보통 두 개의 Feed Forward Layer가 이어집니다. 이 Feed Forward Layers의 중간 임베딩 크기는 종종 모델의 히든 사이즈(hidden size)보다 큽니다(예:
+`google-bert/bert-base-uncased` 모델의 경우).
+
+입력 크기가 `[batch_size, sequence_length]`일 경우, 중간 Feed Forward 임베딩
+`[batch_size, sequence_length, config.intermediate_size]`을 저장하는 데 필요한 메모리는 전체 메모리 사용량의 큰 부분을 차지할 수 있습니다.
+[Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 논문의 저자들은 이 연산이 `sequence_length` 차원에 대해 독립적이기 때문에,토큰마다 Feed Forward Layer의 출력 임베딩을 각 토큰별로 `[batch_size, config.hidden_size]`을 개별적으로 계산한 뒤, 이를 이어 붙여 `[batch_size, sequence_length, config.hidden_size]` 형태로 만들 수 있습니다.`n = sequence_length`. 이 방식은 계산 시간은 늘어나지만, 메모리 사용량은 줄어들게 됩니다.
+
+[`apply_chunking_to_forward`] 함수를 사용하는 모델의 경우, `chunk_size`는 병렬로 계산되는 출력 임베딩의 개수를 정의하며, 이는 메모리 사용량과 계산 시간 간의 트레이드오프를 결정합니다.
+`chunk_size`가 0으로 설정되면, 피드 포워드 청킹(Feed Forward Chunking)은 수행되지 않습니다.
+
+### 파인튜닝 모델 (finetuned models)
+
+파인튜닝(Finetuning)은 전이 학습(transfer learning)의 한 형태로, 사전 학습된 (pretrained) 모델을 사용하여 가중치를 고정(freeze)하고, 출력층을 새롭게 추가된 [모델 헤드](#head)로 교체한 뒤, 해당 모델 헤드를 목표 데이터셋에 맞게 학습시키는 방식입니다.
+
+자세한 내용은 [Fine-tune a pretrained model](https://huggingface.co/docs/transformers/training) 튜토리얼을 참고하시고, 🤗 Transformers를 사용해 모델을 파인 튜닝하는 방법도 함께 확인해 보세요.
+
+## H
+
+### 헤드 (head)
+
+모델 헤드(model head)란 신경망의 마지막 층을 의미하며, 이 층은 이전 층에서 나온 히든 상태(hidden states)를 받아 다른 차원으로 변환합니다. 각 작업(task)에 따라 서로 다른 모델 헤드가 사용됩니다. 예를 들어:
+
+  * [`GPT2ForSequenceClassification`]은 기본 [`GPT2Model`] 위에 시퀀스 분류를 위한 선형계층(linear layer)을 추가한 모델 헤드입니다.
+  * [`ViTForImageClassification`]은 이미지 분류를 위한 모델 헤드로, 기본 [`ViTModel`] 위에 `CLS` 토큰의 마지막 히든 상태에 선형 계층(linear layer)을 추가한 구조입니다.
+  * [`Wav2Vec2ForCTC`]는 기본 [`Wav2Vec2Model`] 위에 [CTC](#connectionist-temporal-classification-ctc)를 적용한 언어 모델링 헤드입니다.
+
+## I
+
+### 이미지 패치 (image patch)
+
+비전 기반 Transformer 모델은 이미지를 작은 패치로 분할한 후, 각 패치를 선형 임베딩하여 시퀀스로 모델에 입력합니다. 모델의 구성 파일에서 `patch_size`(또는 해상도)를 확인할 수 있습니다.
+
+### 인퍼런스 (inference)
+
+인퍼런스는 학습이 완료된 모델에 새로운 데이터를 입력하여 예측을 수행하는 과정입니다. 🤗 Transformer에서 인퍼런스를 수행하는 방법은 [Pipeline for inference](https://huggingface.co/docs/transformers/pipeline_tutorial) 튜토리얼을 참고하세요.
+
+### 입력 ID (input IDs)
+
+입력 ID는 종종 모델에 입력으로 전달해야 하는 유일한 필수 파라미터입니다. 이들은 토큰의 인덱스로, 모델이 입력으로 사용할 시퀀스를 구성하는 토큰들의 숫자 표현입니다.
+
+<Youtube id="VFp38yj8h3A"/>
+
+토크나이저마다 작동 방식은 다르지만, 기본 메커니즘은 동일합니다. 다음은 [WordPiece](https://arxiv.org/pdf/1609.08144.pdf) 토크나이저인 BERT 토크나이저를 사용한 예시입니다:
+
+```python
+>>> from transformers import BertTokenizer
+
+>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
+
+>>> sequence = "A Titan RTX has 24GB of VRAM"
+```
+
+토크나이저는 시퀀스를 토크나이저의 토큰 목록에 있는 항목으로 분리합니다.
+
+```python
+>>> tokenized_sequence = tokenizer.tokenize(sequence)
+```
+
+토큰은 단어이거나 서브 워드(subword)입니다. 예를 들어, "VRAM"은 모델의 어휘 사전에 없는 단어이기 때문에 "V", "RA", "M"으로 나뉘었습니다. 이 토큰들이 개별 단어가 아니라 같은 단어의 일부임을 나타내기 위해 "RA"와 "M" 앞에 더블 해시(`##`)가 추가 됩니다.
+
+```python
+>>> print(tokenized_sequence)
+['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
+```
+
+이러한 토큰들은 모델이 이해할 수 있는 ID로 변환될 수 있습니다. 이 과정은 문장을 바로 토크나이저에 입력함으로써 수행되며, 성능 최적화를 위해 [🤗 Tokenizers](https://github.com/huggingface/tokenizers)의 Rust 구현을 활용합니다.
+
+```python
+>>> inputs = tokenizer(sequence)
+```
+
+토크나이저는 해당 모델이 올바르게 작동하는 데 필요한 모든 인자를 포함한 딕셔너리를 반환합니다. 토큰 인덱스는 `input_ids`라는 키에 저장됩니다.
+
+```python
+>>> encoded_sequence = inputs["input_ids"]
+>>> print(encoded_sequence)
+[101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
+```
+
+토크나이저는 (연결된 모델이 이를 사용하는 경우) 자동으로 "특수 토큰"을 추가합니다. 이들은 모델이 특정 상황에서 사용하는 특별한 ID입니다.
+
+이전의 ID 시퀀스를 디코딩하면,
+
+```python
+>>> decoded_sequence = tokenizer.decode(encoded_sequence)
+```
+
+우리는 다음과 같은 결과를 보게 될 것입니다.
+
+```python
+>>> print(decoded_sequence)
+[CLS] A Titan RTX has 24GB of VRAM [SEP]
+```
+
+이는 [`BertModel`]이 입력값을 기대하는 방식이기 때문입니다.
+
+## L
+
+### 레이블 (labels)
+
+레이블은 모델이 손실(loss)을 직접 계산할 수 있도록 전달되는 선택적 인자입니다. 이 레이블은 모델이 예측해야 할 정답 값을 의미하며, 모델은 예측값과 이 정답(label) 사이의 차이를 표준 손실 함수를 이용해 계산하게 됩니다.
+
+이 레이블(label)의 형태는 모델 헤드(model head)의 종류에 따라 달라집니다. 예를 들어:
+
+- 시퀀스 분류 모델([`BertForSequenceClassification`] 등)의 경우, 모델은
+  `(batch_size)` 차원의 텐서를 입력으로 받으며, 배치의 각 값은 전체 시퀀스에 대한 예상 레이블을 나타냅니다.
+- 토큰 분류 모델([`BertForTokenClassification`] 등)의 경우, 모델은 `(batch_size, seq_length)` 차원의 텐서를 입력으로 받으며, 각 값은 개별 토큰에 대한 예상 레이블을 나타냅니다.
+- 마스킹 언어 모델([`BertForMaskedLM`])의 경우, 모델은 `(batch_size,seq_length)` 차원의 텐서를 입력으로 받으며, 각 값은 개별 토큰에 대한 예상 레이블을 나타냅니다. 레이블은 마스킹 된 토큰의 토큰 ID이며, 나머지 토큰에 대해서는 무시할 값을 사용합니다(일반적으로 -100).
+- 시퀀스 투 시퀀스 작업([`BartForConditionalGeneration`], [`MBartForConditionalGeneration`]등)의 경우, 모델은 `(batch_size, tgt_seq_length)` 차원의 텐서를 입력으로 받으며, 각 값은 입력 시퀀스에 대응하는 타겟 시퀀스를 나타냅니다. 학습 중에는 BART와 T5가 적절한 `decoder_input_ids`와 디코더 attention 마스크를 내부적으로 생성하므로, 일반적으로 따로 제공할 필요가 없습니다. 단, 이는 Encoder-Decoder 프레임워크를 직접 활용하는 모델에는 적용되지 않습니다.
+- 이미지 분류 모델([`ViTForImageClassification`] 등)의 경우, 모델은 `(batch_size)` 차원의 텐서를 입력으로 받으며, 배치의 각 값은 개별 이미지에 대한 예상 레이블을 나타냅니다.
+- 시멘틱 세그멘테이션 모델([`SegformerForSemanticSegmentation`] 등)의 경우, 모델은 `(batch_size, height, width)` 차원의 텐서를 입력으로 받으며, 배치의 각 값은 개별 픽셀에 대한 예상 레이블을 나타냅니다.
+- 객체 탐지 모델([`DetrForObjectDetection`] 등)의 경우, 모델은 `class_labels`와 `boxes` 키를 포함하는 딕셔너리들의 리스트를 입력으로 받습니다. 배치의 각 값은 개별 이미지에 대한 예상 클래스 레이블과 바운딩 박스 정보를 나타냅니다.
+- 자동 음성 인식 모델([`Wav2Vec2ForCTC`] 등)의 경우 모델은 `(batch_size,target_length)` 차원의 텐서를 입력으로 받으며, 각 값은 개별 토큰에 대한 예상 레이블을 나타냅니다.
+  
+<Tip>
+
+모델마다 요구하는 레이블 형식이 다를 수 있으므로, 각 모델의 문서를 확인하여 해당 모델에 맞는 레이블 형식을 반드시 확인하세요!
+
+</Tip>
+
+기본 모델([`BertModel`] 등)은 레이블을 입력으로 받지 않습니다. 이러한 모델은 단순히 특징(feature)을 출력하는 기본 트랜스포머 모델이기 때문입니다.
+
+### 대규모 언어 모델 (LLM)
+
+대규모 데이터로 학습된 트랜스포머 언어 모델(GPT-3, BLOOM, OPT 등)을 지칭하는 일반적인 용어입니다. 이러한 모델은 학습할 수 있는 파라미터(parameter)의 수가 매우 많으며, 예를 들어 GPT-3는 약 1,750억 개의 파라미터를 가지고 있습니다.
+
+## M
+
+### 마스킹된 언어 모델링 (MLM)
+
+사전 학습 단계 중 하나로, 모델은 일부 토큰이 무작위로 마스킹 된 손상된 문장을 입력받고, 원래의 문장을 예측해야 합니다.
+
+### 멀티모달 (multimodal)
+
+텍스트와 이미지와 같은 다른 형태의 입력을 함께 사용하는 작업입니다.
+
+## N
+
+### 자연어 생성 (NLG)
+
+텍스트를 생성하는 모든 작업을 의미합니다. (예: [Write With Transformers](https://transformer.huggingface.co/), 번역 등).
+
+### 자연어 처리 (NLP)
+
+텍스트를 다루는 작업 전반을 지칭하는 일반적인 용어입니다.
+
+### 자연어 이해 (NLU)
+
+텍스트에 담긴 의미를 이해하는 모든 작업을 포함합니다. (예: 전체 문서 분류, 개별 단어 분류 등).
+
+## P
+
+### 파이프라인 (pipeline)
+
+🤗 Transformers에서 파이프라인은 데이터를 전처리하고 변환한 후, 모델을 통해 예측값을 반환하는 일련의 단계를 순차적으로 수행하는 추상화된 개념입니다. 파이프라인에 포함될 수 있는 단계로는 데이터 전처리, 특징 추출(feature extraction), 정규화(normalization) 등이 있습니다.   
+
+자세한 내용은 [Pipelines for inference](https://huggingface.co/docs/transformers/pipeline_tutorial) 문서를 참고하세요.
+
+### 파이프라인 병렬화 (PP)
+
+모델을 수직 방향(레이어 단위)으로 여러 GPU에 분할하여 병렬로 처리하는 병렬화 기법입니다. 각 GPU는 모델의 하나 또는 여러 개의 레이어만을 담당하며, 전체 파이프라인의 서로 다른 단계를 병렬로 처리하게 됩니다. 또한 각 GPU는 배치(batch)의 일부 작은 조각만 처리합니다. Pipeline Parallel 방식에 대해 더 알아보려면 [이 문서](perf_train_gpu_many#from-naive-model-parallelism-to-pipeline-parallelism)를 참고하세요.
+
+### 픽셀 값 (pixel values)
+
+이미지를 수치상으로 표현한 텐서로, 모델에 입력으로 전달됩니다. 이 텐서는 이미지 프로세서를 통해 생성되면, 값은 [`batch_size`, `num_channels`, `height`, `width`] 형태의 차원을 가집니다.
+
+### 풀링 (pooling)
+
+행렬의 특정 차원에서 최댓값이나 평균값을 취하여 더 작은 행렬로 줄이는 연산입니다. 풀링 계층은 주로 합성곱 계층 사이에 위치하여 특징 표현을 다운샘플링 하는 데 사용됩니다.
+
+### 포지션 ID (position IDs)
+
+RNN 모델과 달리 트랜스포머는 각 토큰의 위치 정보를 내부적으로 가지고 있지 않습니다. 따라서 모델은 `position_ids`를 사용하여 각 토큰이 시퀀스 내에서 어느 위치에 있는지를 인식합니다. 이 값은 선택적인 파라미터입니다. 모델에 `position_ids`를 전달하지 않으면, 절대 위치 임베딩 방식으로 자동 생성됩니다. 절대 위치 임베딩은 `[0, config.max_position_embeddings - 1]` 범위 내에서 선택됩니다. 일부 모델은 사인파 형태의 위치 임베딩(sinusoidal position embeddings) 또는 상대 위치 임베딩(relative position embeddings)과 같은 다른 유형의 위치 임베딩을 사용하기도 합니다.
+
+### 전처리 (preprocessing)
+
+머신러닝 모델이 쉽게 처리할 수 있도록 가공되지 않은 데이터를 정제하는 작업입니다. 예를 들어, 텍스트는 일반적으로 토큰화(tokenization) 과정을 거칩니다. 다른 입력 유형에 대한 전처리 방식이 궁금하다면 [Preprocess](https://huggingface.co/docs/transformers/preprocessing) 튜토리얼을 참고해 보세요.
+
+### 사전 학습된 모델 (pretrained model)
+
+일부 데이터(예: 위키피디아 전체)로 사전 학습(pretraining)된 모델입니다. 사전 학습은 자기 지도 학습(self-supervised learning)의 목표를 포함하며, 예를 들어 문장을 읽고 다음 단어를 예측하거나 ([causal language modeling](#causal-language-modeling)) 참고, 일부 단어를 마스킹하고 이를 예측하는 방식([masked language modeling](#masked-language-modeling-mlm))이 있습니다.
+
+음성 및 비전 모델은 고유의 사전 학습 목표를 가지고 있습니다. 예를 들어, Wav2Vec2는 음성 표현 중 "진짜"를 "가짜" 중에서 구분하는 대조 학습(contrastive learning) 방식으로 사전 학습된 음성 모델입니다. 반면, BEiT는 이미지 패치 중 일부를 마스킹하고 이를 예측하는 마스킹 이미지 모델링 방식으로 사전 학습된 비전 모델입니다. 이는 마스킹 언어 모델링과 유사한 방식입니다.
+
+## R
+
+### 순환 신경망 (RNN)
+
+텍스트와 같은 시퀀스 데이터를 처리하기 위해 레이어에 반복 구조(루프)를 사용하는 신경망 모델의 한 종류입니다.
+
+### 표현학습 (representation learning)
+
+머신러닝의 하위 분야로, 원시 데이터로부터 의미 있는 표현을 학습하는 데 중점을 둡니다. 대표적인 기법으로는 단어 임베딩, 오토인코더(autoencoder), 생성적 적대 신경망(GAN) 등이 있습니다.
+
+## S
+
+### 샘플링 속도 (sampling rate)
+
+샘플링 속도는 1초에 추출하는 (오디오 신호) 샘플의 개수를 헤르츠(Hz) 단위로 나타낸 측정값입니다. 이는 음성처럼 연속적인 신호를 디지털화하여 이산적인 형태로 만드는 결과입니다.
+
+### 셀프 어텐션 (self-attention)
+
+입력의 각 요소가 다른 어떤 요소에 주목해야 하는지를 스스로 판단하는 메커니즘입니다. 이는 모델이 문장에서 특정 단어만을 보는 것이 아니라, 다른 단어들과의 관계를 고려하여 어떤 정보에 더 집중해야 할지를 학습하게 합니다.
+
+### 자기지도 학습 (self-supervised learning) 
+
+레이블이 없는 데이터로부터 모델이 스스로 학습 목표를 정의하여 학습하는 머신러닝 기법의 한 종류입니다. [비지도 학습](#unsupervised-learning)이나 [지도 학습](#supervised-learning)과 달리, 학습 과정 자체는 감독 방식 되지만, 라벨이 명시적으로 주어지는 것은 아닙니다.
+
+예시로는 [마스크 언어 모델링](#masked-language-modeling-mlm)이 있으며, 이는 문장의 일부 토큰을 제거한 상태로 모델에 입력하고, 모델이 해당 토큰을 예측하도록 학습하는 방식입니다.
+
+### 준지도 학습 (semi-supervised learning)
+
+소량의 라벨이 달린 데이터와 대량의 라벨이 없는 데이터를 함께 사용하여 모델의 정확도를 높이는 머신러닝 훈련 기법의 넓은 범주입니다. 이는 [지도 학습](#supervised-learning)이나 [비지도 학습](#unsupervised-learning)과는 다른 방식입니다.
+
+준지도 학습 기법의 예로는 "자기 학습(self-training)"이 있습니다. 이 방식은 먼저 라벨이 있는 데이터로 모델을 학습시키고, 그 모델을 사용해 라벨이 없는 데이터에 대한 예측을 수행합니다. 모델이 가장 높은 확신을 가지고 예측한 라벨이 없는 데이터 일부를 라벨이 있는 데이터로 추가하고, 이를 통해 모델을 다시 학습시킵니다.
+
+### 시퀀스 투 시퀀스 (seq2seq)
+
+입력으로부터 새로운 시퀀스를 생성하는 모델입니다. 예를 들어 번역 모델이나 요약 모델이 이에 해당하며, 대표적인 예로는 [Bart](model_doc/bart)나[T5](model_doc/t5) 모델이 있습니다.
+
+### 분할 DDP (Sharded DDP)
+
+[ZeRO](#zero-redundancy-optimizer-zero) 개념을 기반으로 다양한 구현에서 사용되는 다른 이름으로 불립니다.
+
+### 스트라이드 (stride)
+
+[convolution](#convolution) 또는 [pooling](#pooling)에서 스트라이드(stride)는 커널이 행렬 위를 이동하는 간격을 의미합니다. 스트라이드가 1이면 커널이 한 픽셀씩 이동하고, 2이면 두 픽셀씩 이동합니다.
+
+### 지도학습 (supervised learning)
+
+정답이 포함된 라벨링된 데이터를 직접 사용하여 모델의 성능을 개선하는 학습 방식입니다. 학습 중인 모델에 데이터를 입력하고, 예측 결과를 정답과 비교하여 오차를 계산합니다. 모델은 이 오차를 기반으로 가중치를 업데이트하며, 이러한 과정을 반복하여 성능을 최적화합니다.
+
+## T
+
+### 텐서 병렬화 (TP)
+
+여러 GPU에서 훈련하기 위한 병렬화 기법으로, 각 텐서를 여러 덩어리(chunk)로 나눕니다. 따라서 전체 텐서가 단일 GPU에 상주하는 대신, 텐서의 각 조각(shard)이 지정된 GPU에 상주하게 됩니다. 이 조각들은 각각 다른 GPU에서 개별적으로 병렬 처리되며, 처리 단계가 끝날 때 결과가 동기화됩니다. 이러한 분할이 수평 방향으로 일어나기 때문에, 이는 때때로 수평적 병렬화라고 불립니다. Tensor Parallelism에 대해 더 알아보려면 [여기](perf_train_gpu_many#tensor-parallelism)를 참고하세요.
+
+### 토큰 (token)
+
+일반적인 단어 단위이지만, 때에 따라 서브 워드(자주 사용되지 않는 단어는 서브 워드로 분리됨)나 문장 부호도 포함될 수 있는 문장의 구성 요소입니다.
+
+### 토큰 타입 ID (token type IDs)
+
+일부 모델은 문장 쌍 분류나 질의 응답 작업을 수행하는 데 사용됩니다.
+
+<Youtube id="0u3ioSwev3s"/>
+
+이러한 작업에서는 두 개의 서로 다른 시퀀스를 하나의 "input_ids" 항목으로 결합해야 하며, 일반적으로 `[CLS]` 분류용 및 `[SEP]` 구분용과 같은 특수 토큰을 사용하여 처리합니다. 예를 들어, BERT 모델은 두 개의 시퀀스를 다음과 같은 방식으로 구성합니다:
+
+```python
+>>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
+```
+
+두 개의 시퀀스를 `tokenizer`에 리스트가 아닌 개별 인자로 전달하면, 토크나이저가 자동으로 이러한 문장을 생성해 줍니다. 예시는 다음과 같습니다:
+
+```python
+>>> from transformers import BertTokenizer
+
+>>> tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
+>>> sequence_a = "HuggingFace is based in NYC"
+>>> sequence_b = "Where is HuggingFace based?"
+
+>>> encoded_dict = tokenizer(sequence_a, sequence_b)
+>>> decoded = tokenizer.decode(encoded_dict["input_ids"])
+```
+
+결과는 아래와 같습니다:
+
+```python
+>>> print(decoded)
+[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
+```
+
+이 코드는 일부 모델이 두 개의 시퀀스를 어떻게 구분하는지 이해하는 데 충분합니다. 그러나 BERT와 같은 다른 모델은 토큰 타입 ID(또는 세그먼트 ID)를 추가로 사용합니다. 이 ID는 0과 1로 구성된 이진 마스크로, 두 시퀀스를 구분하는 역할을 합니다.
+
+토크나이저는 이 마스크를 "token_type_id" 항목으로 반환합니다:
+
+```python
+>>> encoded_dict["token_type_ids"]
+[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
+```
+
+질문에 사용되는 첫 번째 시퀀스인 "context"는 모든 토큰이 `0`으로 표시됩니다. 반면 두 번째 시퀀스인 "question"은 모든 토큰이 `1`로 표시됩니다.
+
+일부 모델(예: [`XLNetModel`])은 `2`로 표시되는 추가 토큰을 사용하기도 합니다.
+
+### 전이학습 (transfer learning)
+
+사전 학습된(pretrained) 모델을 가져와 특정 작업에 맞는 데이터셋에 대해 추가 학습하는 기술입니다. 모델을 처음부터 학습시키는 대신, 기존 모델이 학습한 지식을 출발점으로 삼아 더욱 빠르게 학습할 수 있습니다. 이를 통해 학습 속도를 높이고 필요한 데이터양도 줄일 수 있습니다.
+
+### 트랜스포머 (transformer)
+
+셀프 어텐션 메커니즘을 기반으로 한 딥러닝 모델 아키텍처입니다.
+
+## U
+
+### 비지도 학습 (unsupervised learning)
+
+정답(레이블)이 포함되지 않은 데이터를 이용해 모델을 학습시키는 방식입니다. 비지도 학습은 데이터 분포의 통계적 특성을 활용해 유용한 패턴을 찾아냅니다.
+
+## Z
+
+### Zero Redundancy Optimizer (ZeRO)
+
+[TensorParallel](#tensor-parallelism-tp)과 유사하게 텐서를 샤딩(sharding)하는 병렬 처리 기법이지만, 순전파(forward)나 역전파(backward) 계산 시점에 전체 텐서를 다시 복원한다는 점에서 차이가 있습니다. 따라서 모델 자체를 수정할 필요가 없습니다. 이 방법은 GPU 메모리가 부족할 경우 이를 보완하기 위한 다양한 오프로딩 (offloading) 기법도 지원합니다. 
+ZeRO에 대해 더 알아보려면 [이 문서](perf_train_gpu_many#zero-data-parallelism)를 참고하세요.

docs/source/ko/internal/generation_utils.md

@@ -345,12 +345,6 @@ generation_output[:2]
 [[autodoc]] Cache
     - update
 
-[[autodoc]] CacheConfig
-    - update
-
-[[autodoc]] QuantizedCacheConfig
-    - validate
-
 [[autodoc]] DynamicCache
     - update
     - get_seq_length

docs/source/ko/perf_infer_gpu_multi.md

@@ -0,0 +1,311 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# 분산 추론[[distributed-inference]]
+
+모델이 단일 GPU에 올라가지 않는 경우, [텐서 병렬 처리](./perf_train_gpu_many#tensor-parallelism)를 사용한 분산 추론이 도움이 될 수 있습니다. 텐서 병렬화는 모델을 여러 가속기(CUDA GPU, Intel XPU 등)에 분할하여 행렬 곱셈과 같은 계산을 병렬화합니다. 이를 통해 더 큰 모델을 메모리에 올릴 수 있으며, 각 가속기가 텐서의 일부를 처리하므로 추론 속도가 향상됩니다.
+
+그러나 텐서 병렬화는 통신 오버헤드를 발생시키므로, 빠른 노드 내 통신을 활용할 수 있는 다중 가속기 환경에서 사용하는 것이 가장 효과적입니다. 다중 노드 학습 환경에서는 사용 사례에 따라 파이프라인 병렬화나 데이터 병렬화를 사용하는 것이 더 효율적일 수 있습니다.
+
+> [!TIP]
+> 텐서 병렬화에 대해 더 자세히 알아보려면 [Ultra-Scale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=tensor_parallelism)의 텐서 병렬화 섹션을 참조하세요.
+
+아래 목록에서 텐서 병렬 처리를 기본적으로 지원하는 모델을 확인할 수 있습니다. 새로운 모델에 대한 지원을 추가하려면 GitHub 이슈나 풀 리퀘스트를 열어주세요.
+
+<details>
+<summary>지원되는 모델 보기</summary>
+
+* [Cohere](./model_doc/cohere) 및 [Cohere 2](./model_doc/cohere2)
+* [Gemma](./model_doc/gemma) 및 [Gemma 2](./model_doc/gemma2)
+* [GLM](./model_doc/glm)
+* [Granite](./model_doc/granite)
+* [Llama](./model_doc/llama)
+* [Mistral](./model_doc/mistral)
+* [Mixtral](./model_doc/mixtral)
+* [OLMo](./model_doc/olmo) 및 [OLMo2](./model_doc/olmo2)
+* [Phi](./model_doc/phi) 및 [Phi-3](./model_doc/phi3)
+* [Qwen2](./model_doc/qwen2), [Qwen2Moe](./model_doc/qwen2_moe), 및 [Qwen2-VL](./model_doc/qwen2_5_vl)
+* [Starcoder2](./model_doc/starcoder2)
+
+</details>
+
+이 가이드는 Transformers에서 다양한 분할 전략을 사용하여 텐서 병렬화를 활성화하는 방법을 설명합니다.
+
+## 모델 분할[[partitioning-a-model]]
+
+Transformers는 `tp_plan`매개변수를 활용할 수 있는 모델에 대해 텐서 병렬 처리를 지원합니다. 모델 분할 방식은 두 가지가 있습니다.
+
+- `auto` 텐서 병렬화 계획은 사전 정의된 구성을 기반으로 모델(위에 언급된 지원 모델)을 자동으로 분할합니다.
+- 사용자 지정 분할 계획을 직접 정의하여 [~PreTrainedModel.from_pretrained] 메소드의 `tp_plan` 매개변수로 전달할 수 있습니다.
+
+<hfoptions id="sharding">
+<hfoption id="auto plan">
+
+```py
+import os
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct" # 모든 가능한 전략을 시각화하기에 더 좋음
+model_id = "meta-llama/Meta-Llama-3-8B-Instruct"  # 적은 수의 GPU에 더 좋음
+
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, tp_plan="auto")
+print(model._tp_plan)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")
+prompt = "Can I help"
+inputs = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)
+
+# 분산 실행
+outputs = model(inputs)
+```
+
+위의 추론 스크립트를 GPU당 4개 프로세스로 [torchrun](https://pytorch.org/docs/stable/elastic/run.html)에서 실행하세요.
+
+```bash
+torchrun --nproc-per-node 4 demo.py
+```
+
+</hfoption>
+<hfoption id="manual plan">
+
+각 레이어에 대한 텐서 병렬 계획을 `tp_plan`에 정의한 후 [`~PreTrainedModel.from_pretrained`]에 전달하세요. 아래 예시는 열 및 행 분할을 조합하여 사용합니다. 지원되는 다른 분할 전략은 [분할 전략](#partitioning-strategies) 섹션을 참고하세요.
+
+> [!WARNING]
+> 사용자 지정 분할 계획을 수동으로 지정하려면 모델 아키텍처와 분할 전략이 함께 상호 작용하는 방식에 대한 충분한 이해가 필요합니다. 분할 전략을 잘못 설정하면 모델이 매우 느려지거나, 오류가 발생하거나, 부정확한 결과를 낼 수 있습니다. 자세히 알아보려면 [Ultra-Scale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=tensor_parallelism)을 참고하세요.
+
+```py
+from transformers import AutoModelForCausalLM
+
+tp_plan = {
+    "model.layers.*.self_attn.q_proj": "colwise",
+    "model.layers.*.self_attn.k_proj": "colwise",
+    "model.layers.*.self_attn.v_proj": "colwise",
+    "model.layers.*.self_attn.o_proj": "rowwise",
+    ...
+}
+
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, tp_plan=tp_plan)
+print(model._tp_plan)
+```
+
+</hfoption>
+</hfoptions>
+
+## 분할 전략[[partitioning-strategies]]
+
+모든 분할 전략은 문자열을 전략 구현에 매핑하는 [`ParallelInterface`] 클래스에서 정의됩니다. 모든 전략은 [`~PreTrainedModel.from_pretrained`]의 `tp_plan`을 통해 설정되므로 이 클래스와 직접 상호 작용할 필요는 없지만, 어떤 전략을 사용할 수 있는지 확인할 때 유용합니다.
+
+```py
+class ParallelInterface(MutableMapping):
+    """
+    허용된 어텐션 함수를 추적하는 딕셔너리 같은 객체입니다. `register()` 호출로 새로운 어텐션 함수를 쉽게 추가할 수 있습니다. 
+    모델이 기존 어텐션 함수(예: `sdpa`)를 로컬에서 덮어쓰려면 `modeling_<model>.py` 내부에서 이 클래스의 새 인스턴스를 선언하고 
+    해당 인스턴스에서 선언해야 합니다.
+    """
+    _global_mapping = {
+        "colwise": ColwiseParallel(),
+        "rowwise": RowwiseParallel(),
+        "colwise_rep": ColwiseParallel(output_layouts=Replicate()),
+        "rowwise_rep": RowwiseParallel(input_layouts=Replicate()),
+        "local_colwise": ColwiseParallel(use_dtensor=False),
+        "local_rowwise": RowwiseParallel(use_dtensor=False),
+        "local": IsolatedParallel(),
+        "gather": GatherParallel(),
+        "local_packed_rowwise": PackedRowwiseParallel(use_dtensor=False),
+        "sequence_parallel": SequenceParallel(),
+        "replicate": ReplicateParallel(),
+    }
+```
+
+각 전략에 대해 자세히 알아보려면 아래 표를 참고하세요.
+
+| 전략 | 설명 |
+|---|---|
+| `ColwiseParallel` | 가중치와 편향의 열 방향 분할. |
+| `RowwiseParallel` | 가중치와 편향의 행 방향 분할. `nn.Embedding` 모듈 분할도 지원. |
+| `SequenceParallel` | `LayerNorm`과 `Dropout` 레이어를 지원하는 시퀀스 병렬 구현. [RMSNorm](https://github.com/facebookresearch/llama/blob/main/llama/model.py#L34)의 Python 구현도 지원. |
+| `PackedColwiseParallel` | 패킹된 가중치를 지원하는 `ColwiseParallel`의 변형(예: `up_proj`와 `gate_proj`를 함께 패킹). 자세한 내용은 [코드](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/tensor_parallel.py#L79-#L108)를 참조하세요. |
+| `PackedRowwiseParallel` | 패킹된 가중치를 지원하는 `RowwiseParallel`의 변형([코드](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/tensor_parallel.py#L79-#L108) 참조). |
+| `GatherParallel` | 기기 간 모듈의 출력을 수집. |
+| `IsolatedParallel` | Mixture-of-Experts(MoE) 레이어의 전문가에 사용되어 다른 기기로부터 모듈을 격리. |
+| `ReplicateParallel` | 부분적으로 분할된 모델로 인해 `torch.distributed` API가 중단되는 것을 방지하기 위해 모든 기기에 모듈을 복제. |
+
+### 패킹된 전략[[packed-strategies]]
+
+가중치 패킹은 여러 선형 레이어를 하나의 더 큰 레이어로 합치는 기법입니다. 패킹된 전략인 `PackedColwiseParallel`과 `PackedRowwiseParallel`은 패킹된 가중치를 분할하는 데 사용됩니다. 기본적인 `ColwiseParallel`이나 `RowwiseParallel`은 패킹된 가중치를 올바르게 분할하지 못합니다.
+
+아래 예시는 `up_proj`와 `gate_proj`를 단일 `gate_up_proj` 모듈로 패킹하고 `gate_up_proj`를 분할하기 위해 `PackedRowwiseParallel` 전략이 필요합니다.
+
+```python
+class Llama4TextExperts(nn.Module):
+    ...
+    self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_size, 2 * self.expert_dim))
+```
+
+배치 행렬 곱셈을 `forward` 패스에서 사용하여 `gate_up_proj` 모듈의 출력을 계산할 수 있습니다.
+
+```python
+def forward(self, hidden_states):
+    ...
+    gate_up = torch.bmm(hidden_states, self.gate_up_proj) # gate_up_proj 모듈의 출력 계산
+    gate, up = gate_up.chunk(2, dim=-1) # 출력을 gate와 up으로 분할
+```
+
+> [!TIP]
+> `Packed*`를 사용해야 하는 이유에 대한 시각적 표현은 [이 주석](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/tensor_parallel.py#L79-#L108)을 참고하세요.
+
+### 로컬 전략[[local-strategies]]
+
+로컬 전략(`local_colwise`, `local_rowwise`, `local_packed_rowwise`)은 [torch.chunk](https://docs.pytorch.org/docs/stable/generated/torch.chunk.html)와 같은 일부 연산에서 지원되지 않기 때문에 [DTensor](https://docs.pytorch.org/docs/stable/distributed.tensor.html)를 사용하지 않습니다. 대신 로컬 전략은 기본 [torch.Tensor](https://docs.pytorch.org/docs/stable/tensors.html)를 사용하고 일부 분산 로직을 수동으로 수행합니다.
+
+<!--
+Readd this when I get the exact error message
+> [!TIP]
+> 사용자 정의 분할 전략을 사용하는데 `... is not supported` 오류로 작동하지 않는 경우, `local*` 전략을 사용해서 더 잘 작동하는지 시도해보세요.
+-->
+
+## 사용자 정의 분할 전략[[custom-partitioning-strategies]]
+
+사용자 정의 분할 전략은 [`TensorParallelLayer`](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/tensor_parallel.py)를 상속하고 `partition_tensor`, `_prepare_input_fn`, `_prepare_output_fn`을 구현해야 합니다.
+
+그런 다음 `tp_plan`에서 해당 전략을 지정했을 때 디스패칭 로직이 찾을 수 있도록 `ParallelInterface` 매핑에 등록해야 합니다.
+
+아래 예시는 이 워크플로우로 `ColwiseParallel`을 구현하는 방법을 보여줍니다.
+
+1. `TensorParallelLayer`를 상속합니다. `__init__` 메소드에서 입력 및 출력 텐서가 기기에 어떻게 배치되어야 하는지 설명하는 `input_layouts`과 `output_layouts`을 정의합니다. `desired_input_layouts` 속성은 입력이 기기에 어떻게 배치*되어야만* 하는지를 명시하는 데 사용됩니다.
+
+    ```python
+    class ColwiseParallel(TensorParallelLayer):
+        def __init__(
+            self,
+            *,
+            input_layouts: Optional[Placement] = None, # 이전 레이어에서 오는 입력 레이아웃
+            output_layouts: Optional[Placement] = None, # 달성하고자 하는 출력 레이아웃
+            use_local_output: bool = True, # 로컬 출력 사용 여부
+            use_dtensor=True, # DTensor 사용 여부
+        ):
+            self.input_layouts = (input_layouts or Replicate(),) # 이전 레이어에서 오는 입력 분할
+            self.output_layouts = (output_layouts or Shard(-1),) # 원하는 출력 분할
+            self.desired_input_layouts = (Replicate(),) # 원하는 입력 분할, 입력은 GPU 간에 복제되어야 함
+            self.use_local_output = use_local_output
+            self.use_dtensor = use_dtensor
+    ```
+
+2. `partition_tensor`, `_prepare_input_fn`, `_prepare_output_fn` 메서드를 구현합니다.
+
+    `partition_tensor` 메소드는 텐서를 분할하고 분할된 텐서로 `empty_param`을 채웁니다. 유틸리티 함수 `get_tensor_shard`를 사용하여 주어진 랭크에 대한 원본 매개변수의 올바른 분할을 얻고, 패킹된 가중치에 대해서는 `get_packed_weights`를 사용하세요.
+
+    ```python
+    def partition_tensor(
+        self,
+        param, # 매개변수의 전체 텐서
+        empty_param, # 매개변수의 빈 텐서, 분할된 텐서로 채워짐
+        param_type, # 매개변수 유형, `bias` 또는 `weight`
+        param_casting_dtype, # 매개변수를 캐스팅할 유형
+        to_contiguous, # 텐서를 연속적인 메모리 레이아웃으로 변환할지 여부
+        rank, # 현재 기기의 랭크
+        device_mesh, # 기기 메시
+    ) -> nn.Parameter: # 분할된 매개변수 반환
+        ...
+    ```
+
+    `_prepare_input_fn`과 `_prepare_output_fn` 메소드는 [사전 포워드](https://docs.pytorch.org/docs/stable/generated/torch.nn.modules.module.register_module_forward_pre_hook.html) 및 [포워드](https://docs.pytorch.org/docs/stable/generated/torch.nn.modules.module.register_module_forward_hook.html) 훅에서 사용됩니다. `__init__`에서 지정된 대로 입력과 출력을 원하는 레이아웃으로 재분배합니다.
+
+    ```python
+    def _prepare_input_fn(input_layouts, desired_input_layouts, mod, inputs, device_mesh):
+        ...
+        # 사용자 정의 로직 수행, DTensor로 캐스팅 등.
+        ...
+        return inputs.redistribute(placements=desired_input_layouts, device_mesh=device_mesh)
+    def _prepare_output_fn(output_layouts, use_local_output, mod, outputs, device_mesh):
+        ...
+        # 사용자 정의 로직 수행, DTensor로 캐스팅 등.
+        ...
+        return outputs.redistribute(placements=output_layouts, device_mesh=device_mesh)
+    ```
+
+3. `tp_plan`과 함께 사용할 수 있도록 전략을 [`ParallelInterface`]에 등록합니다.
+
+    ```python
+    from transformers.integrations.tensor_parallel import ParallelInterface
+
+    ParallelInterface.register_strategy("colwise_custom", ColwiseParallel)
+    tp_plan = {
+        "model.layers.*.self_attn.q_proj": "colwise_custom",
+        ...
+    }
+    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, tp_plan=tp_plan)
+    ```
+
+## 벤치마크[[benchmarks]]
+
+텐서 병렬화는 특히 큰 배치 크기나 긴 시퀀스를 가진 입력에 대한 추론 속도를 크게 향상시킬 수 있습니다.
+
+시퀀스 길이가 512인 [Llama](./model_doc/llama)에서 단일 포워드 패스에 대한 예상 속도 향상 수치는 아래 차트를 참조하세요.
+
+<div style="text-align: center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Meta-Llama-3-8B-Instruct%2C%20seqlen%20%3D%20512%2C%20python%2C%20w_%20compile.png">
+</div>
+
+## 설계 구현[[design-implementation]]
+
+Transformers 텐서 병렬화 구현은 프레임워크에 구애받지 않지만, 구체적인 구현을 위해서는 [DeviceMesh](https://docs.pytorch.org/tutorials/recipes/distributed_device_mesh.html)와 [torch.distributed](https://docs.pytorch.org/tutorials/beginner/dist_overview.html)의 [DTensor](https://docs.pytorch.org/docs/stable/distributed.tensor.html)에 의존하여 간단하고 확장 가능한 인터페이스를 제공합니다.
+
+### DeviceMesh[[devicemesh]]
+
+`DeviceMesh`를 함께 통신하는 기기들의 다차원 그리드로 상상해보세요. 병렬 처리 전략마다 각기 다른 통신 패턴이 필요하므로, 여러 하위 메시를 가진 `DeviceMesh`를 만들 수 있습니다.
+
+```python
+from torch.distributed.device_mesh import init_device_mesh
+
+# 4개 GPU의 1D 메시 생성
+device_mesh = init_device_mesh("cuda", (4,), mesh_dim_names=["tp"])
+```
+
+`torch.distributed`에서 정의된 대부분의 병렬화 전략은 메시 자체나 하위 메시에 적용할 수 있으며, 자동으로 통신 패턴을 처리합니다.
+
+### DTensor[[dtensor]]
+
+`DTensor`(분산 텐서)는 일반적인 텐서 연산 위에 분산 로직을 처리하는 텐서 하위 클래스입니다. 텐서 병렬화의 대부분의 모델 가중치는 `DTensor` 형태로 저장됩니다.
+
+DTensor의 가장 중요한 부분은 `placement` 속성입니다. 이는 PyTorch에게 텐서가 `DeviceMesh`의 기기에 어떻게 배치되는지 알려주기 때문입니다. `placement` 속성은 다음 값을 가질 수 있습니다.
+
+- `Shard(dimension)` - `DTensor`가 구성된 `DeviceMesh`에서 주어진 차원에 걸쳐 어떻게 분할되는지 나타냅니다. 아래 예시는 열 방향 분할을 위해 다양한 차원에 걸쳐 가중치를 분할하는 방법을 보여줍니다.
+
+    ```python
+    weight = ...
+    weight = DTensor.from_local(weight, device_mesh["tp"], placements=[Shard(0)]) # 첫 번째(열 방향) 차원에 걸쳐 분할
+    bias = ...
+    bias = DTensor.from_local(bias, device_mesh["tp"], placements=[Shard(-1)]) # 유일한 차원에 걸쳐 분할
+    ```
+
+    이 예시는 행 방향 분할을 위해 여러 차원에 걸쳐 가중치를 분할하는 방법을 보여줍니다.
+
+    ```python
+    weight = ...
+    weight = DTensor.from_local(weight, device_mesh["tp"], placements=[Shard(1)]) # 두 번째(행 방향) 차원에 걸쳐 분할
+    bias = ...
+    bias = DTensor.from_local(bias, device_mesh["tp"], placements=[Replicate()]) # 모든 GPU에 편향 복제
+    ```
+
+- `Replicate()` - `DTensor`가 `DeviceMesh`에 걸쳐 복제됨을 나타냅니다. 각 기기에 텐서의 전체 사본만 생성합니다.
+
+    ```py
+    bias = ...
+    bias = DTensor.from_local(bias, device_mesh["tp"], placements=[Replicate()]) # 모든 GPU에 편향 복제
+    ```
+
+- `Partial()` - 텐서가 감소 연산을 기다리고 있는 상태임을 나타냅니다 (일반적으로 Transformers에서의 사용 사례와는 직접적인 관련이 적습니다).

docs/source/ko/quantization/quark.md

@@ -0,0 +1,85 @@
+<!--Copyright 2025 Advanced Micro Devices, Inc. and The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Quark[[quark]]
+
+[Quark](https://quark.docs.amd.com/latest/)는 특정 데이터 타입, 알고리즘, 하드웨어에 구애받지 않도록 설계된 딥러닝 양자화 툴킷입니다. Quark에서는 다양한 전처리 전략, 알고리즘, 데이터 타입을 조합하여 사용할 수 있습니다.
+
+🤗 Transformers를 통해 통합된 PyTorch 지원은 주로 AMD CPU 및 GPU를 대상으로 하며, 주로 평가 목적으로 사용됩니다. 예를 들어, [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)를 🤗 Transformers 백엔드와 함께 사용하여 Quark로 양자화된 다양한 모델을 원활하게 평가할 수 있습니다.
+
+Quark에 관심이 있는 사용자는 [문서](https://quark.docs.amd.com/latest/)를 참고하여 모델 양자화를 시작하고 지원되는 오픈 소스 라이브러리에서 사용할 수 있습니다!
+
+Quark는 자체 체크포인트/[설정 포맷](https://huggingface.co/amd/Llama-3.1-8B-Instruct-FP8-KV-Quark-test/blob/main/config.json#L26)를 가지고 있지만, 다른 양자화/런타임 구현체 ([AutoAWQ](https://huggingface.co/docs/transformers/quantization/awq), [네이티브 fp8](https://huggingface.co/docs/transformers/quantization/finegrained_fp8))와 호환되는 직렬화 레이아웃으로 모델을 생성하는 것도 지원합니다.
+
+Transformer에서 Quark 양자화 모델을 로드하려면 먼저 라이브러리를 설치해야 합니다:
+
+```bash
+pip install amd-quark
+```
+
+## 지원 매트릭스[[Support matrix]]
+
+Quark를 통해 양자화된 모델은 함께 조합할 수 있는 광범위한 기능을 지원합니다. 구성에 관계없이 모든 양자화된 모델은 `PretrainedModel.from_pretrained`를 통해 원활하게 다시 로드할 수 있습니다.
+
+아래 표는 Quark에서 지원하는 몇 가지 기능을 보여줍니다:
+
+| **기능**                        | **Quark에서 지원하는 항목**                                                                             |   |
+|---------------------------------|-----------------------------------------------------------------------------------------------------------|---|
+| 데이터 타입                     | int8, int4, int2, bfloat16, float16, fp8_e5m2, fp8_e4m3, fp6_e3m2, fp6_e2m3, fp4, OCP MX, MX6, MX9, bfp16 |   |
+| 양자화 전 모델 변환 | SmoothQuant, QuaRot, SpinQuant, AWQ                                                                       |   |
+| 양자화 알고리즘                 | GPTQ                                                                                                      |   |
+| 지원 연산자                     | ``nn.Linear``, ``nn.Conv2d``, ``nn.ConvTranspose2d``, ``nn.Embedding``, ``nn.EmbeddingBag``               |   |
+| 세분성(Granularity)             | per-tensor, per-channel, per-block, per-layer, per-layer type                                             |   |
+| KV 캐시                         | fp8                                                                                                       |   |
+| 활성화 캘리브레이션             | MinMax / Percentile / MSE                                                                                 |   |
+| 양자화 전략                     | weight-only, static, dynamic, with or without output quantization                                         |   |
+
+## Hugging Face Hub의 모델[[Models on Hugging Face Hub]]
+
+Quark 네이티브 직렬화를 사용하는 공개 모델은 https://huggingface.co/models?other=quark 에서 찾을 수 있습니다.
+
+Quark는 [`quant_method="fp8"`을 이용하는 모델](https://huggingface.co/models?other=fp8)과 [`quant_method="awq"`을 사용하는 모델](https://huggingface.co/models?other=awq)도 지원하지만, Transformers는 이러한 모델을 [AutoAWQ](https://huggingface.co/docs/transformers/quantization/awq)를 통해 불러오거나 
+[🤗 Transformers의 네이티브 fp8 지원](https://huggingface.co/docs/transformers/quantization/finegrained_fp8)을 사용합니다.
+
+## Transformers에서 Quark모델 사용하기[[Using Quark models in Transformers]]
+
+다음은 Transformers에서 Quark 모델을 불러오는 방법의 예시입니다:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "EmbeddedLLM/Llama-3.1-8B-Instruct-w_fp8_per_channel_sym"
+model = AutoModelForCausalLM.from_pretrained(model_id)
+model = model.to("cuda")
+
+print(model.model.layers[0].self_attn.q_proj)
+# QParamsLinear(
+#   (weight_quantizer): ScaledRealQuantizer()
+#   (input_quantizer): ScaledRealQuantizer()
+#   (output_quantizer): ScaledRealQuantizer()
+# )
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+inp = tokenizer("Where is a good place to cycle around Tokyo?", return_tensors="pt")
+inp = inp.to("cuda")
+
+res = model.generate(**inp, min_new_tokens=50, max_new_tokens=100)
+
+print(tokenizer.batch_decode(res)[0])
+# <|begin_of_text|>Where is a good place to cycle around Tokyo? There are several places in Tokyo that are suitable for cycling, depending on your skill level and interests. Here are a few suggestions:
+# 1. Yoyogi Park: This park is a popular spot for cycling and has a wide, flat path that's perfect for beginners. You can also visit the Meiji Shrine, a famous Shinto shrine located in the park.
+# 2. Imperial Palace East Garden: This beautiful garden has a large, flat path that's perfect for cycling. You can also visit the
+```

examples/flax/question-answering/run_qa.py

@@ -545,7 +545,7 @@ def main():
 
     # region Tokenizer check: this script requires a fast tokenizer.
     if not isinstance(tokenizer, PreTrainedTokenizerFast):
-        raise ValueError(
+        raise TypeError(
             "This example script only works for models that have a fast tokenizer. Check out the big table of models at"
             " https://huggingface.co/transformers/index.html#supported-frameworks to find the model types that meet"
             " this requirement"

examples/modular-transformers/configuration_duplicated_method.py

@@ -0,0 +1,216 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from examples/modular-transformers/modular_duplicated_method.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_duplicated_method.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+
+from ...configuration_utils import PretrainedConfig
+from ...modeling_rope_utils import rope_config_validation
+
+
+class DuplicatedMethodConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DuplicatedMethodModel`]. It is used to instantiate an DuplicatedMethod
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DuplicatedMethod-7B.
+    e.g. [meta-duplicated_method/DuplicatedMethod-2-7b-hf](https://huggingface.co/meta-duplicated_method/DuplicatedMethod-2-7b-hf)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the DuplicatedMethod model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DuplicatedMethodModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. DuplicatedMethod 1 supports up to 2048 tokens,
+            DuplicatedMethod 2 up to 4096, CodeLlama up to 16384.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
+            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
+            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'duplicated_method3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'duplicated_method3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'duplicated_method3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'duplicated_method3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
+        head_dim (`int`, *optional*):
+            The attention head dimension. If None, it will default to hidden_size // num_attention_heads
+
+    ```python
+    >>> from transformers import DuplicatedMethodModel, DuplicatedMethodConfig
+
+    >>> # Initializing a DuplicatedMethod duplicated_method-7b style configuration
+    >>> configuration = DuplicatedMethodConfig()
+
+    >>> # Initializing a model from the duplicated_method-7b style configuration
+    >>> model = DuplicatedMethodModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "duplicated_method"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    # Default tensor parallel plan for base model `DuplicatedMethodModel`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        head_dim=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, copy it it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        return 45
+
+    @vocab_size.setter
+    def vocab_size(self, value):
+        self.vocab_size = value

examples/modular-transformers/configuration_my_new_model.py

@@ -125,8 +125,6 @@ class MyNewModelConfig(PretrainedConfig):
     >>> # Accessing the model configuration
     >>> configuration = model.config
     ```
-        new_param (`int`, *optional*, defaults to `False`):
-            A fun new parameter
     """
 
     model_type = "my_new_model"

examples/modular-transformers/modeling_dummy.py

@@ -1,446 +0,0 @@
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-#           This file was automatically generated from examples/modular-transformers/modular_dummy.py.
-#               Do NOT edit this file manually as any edits will be overwritten by the generation of
-#             the file from the modular. If any change should be done, please apply the change to the
-#                          modular_dummy.py file directly. One of our CI enforces this.
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-from typing import Callable, Optional
-
-import torch
-from torch import nn
-
-from ...activations import ACT2FN
-from ...cache_utils import Cache, DynamicCache
-from ...integrations import use_kernel_forward_from_hub
-from ...masking_utils import create_causal_mask
-from ...modeling_flash_attention_utils import FlashAttentionKwargs
-from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutputWithPast
-from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
-from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
-from ...processing_utils import Unpack
-from ...utils import auto_docstring, can_return_tuple, logging
-from .configuration_dummy import DummyConfig
-
-
-logger = logging.get_logger(__name__)
-
-
-@use_kernel_forward_from_hub("RMSNorm")
-class DummyRMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        """
-        DummyRMSNorm is equivalent to T5LayerNorm
-        """
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-    def extra_repr(self):
-        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
-
-
-class DummyRotaryEmbedding(nn.Module):
-    def __init__(self, config: DummyConfig, device=None):
-        super().__init__()
-        # BC: "rope_type" was originally "type"
-        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
-            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
-        else:
-            self.rope_type = "default"
-        self.max_seq_len_cached = config.max_position_embeddings
-        self.original_max_seq_len = config.max_position_embeddings
-
-        self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
-
-        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = self.inv_freq
-
-    @torch.no_grad()
-    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
-    def forward(self, x, position_ids):
-        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
-        position_ids_expanded = position_ids[:, None, :].float()
-
-        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
-            emb = torch.cat((freqs, freqs), dim=-1)
-            cos = emb.cos() * self.attention_scaling
-            sin = emb.sin() * self.attention_scaling
-
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
-
-
-class DummyMLP(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.hidden_size = config.hidden_size
-        self.intermediate_size = config.intermediate_size
-        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
-        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
-        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
-        self.act_fn = ACT2FN[config.hidden_act]
-
-    def forward(self, x):
-        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
-        return down_proj
-
-
-def rotate_half(x):
-    """Rotates half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 4]
-    x2 = x[..., x.shape[-1] // 4 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
-    """Applies Rotary Position Embedding to the query and key tensors.
-
-    Args:
-        q (`torch.Tensor`): The query tensor.
-        k (`torch.Tensor`): The key tensor.
-        cos (`torch.Tensor`): The cosine part of the rotary embedding.
-        sin (`torch.Tensor`): The sine part of the rotary embedding.
-        position_ids (`torch.Tensor`, *optional*):
-            Deprecated and unused.
-        unsqueeze_dim (`int`, *optional*, defaults to 1):
-            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
-            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
-            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
-            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
-            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
-            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
-    Returns:
-        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
-    """
-    cos = cos.unsqueeze(unsqueeze_dim)
-    sin = sin.unsqueeze(unsqueeze_dim)
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-    return q_embed, k_embed
-
-
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
-    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
-
-def eager_attention_forward(
-    module: nn.Module,
-    query: torch.Tensor,
-    key: torch.Tensor,
-    value: torch.Tensor,
-    attention_mask: Optional[torch.Tensor],
-    scaling: float,
-    dropout: float = 0.0,
-    **kwargs,
-):
-    key_states = repeat_kv(key, module.num_key_value_groups)
-    value_states = repeat_kv(value, module.num_key_value_groups)
-
-    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
-    if attention_mask is not None:
-        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
-        attn_weights = attn_weights + causal_mask
-
-    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
-    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
-    attn_output = torch.matmul(attn_weights, value_states)
-    attn_output = attn_output.transpose(1, 2).contiguous()
-
-    return attn_output, attn_weights
-
-
-class DummyAttention(nn.Module):
-    """Multi-headed attention from 'Attention Is All You Need' paper"""
-
-    def __init__(self, config: DummyConfig, layer_idx: int):
-        super().__init__()
-        self.config = config
-        self.layer_idx = layer_idx
-        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
-        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
-        self.scaling = self.head_dim**-0.5
-        self.attention_dropout = config.attention_dropout
-        self.is_causal = True
-
-        self.q_proj = nn.Linear(
-            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.k_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.v_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.o_proj = nn.Linear(
-            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
-        )
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: Optional[torch.Tensor],
-        past_key_value: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
-        input_shape = hidden_states.shape[:-1]
-        hidden_shape = (*input_shape, -1, self.head_dim)
-
-        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
-        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
-        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
-
-        cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        if past_key_value is not None:
-            # sin and cos are specific to RoPE models; cache_position needed for the static cache
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
-
-        attn_output, attn_weights = attention_interface(
-            self,
-            query_states,
-            key_states,
-            value_states,
-            attention_mask,
-            dropout=0.0 if not self.training else self.attention_dropout,
-            scaling=self.scaling,
-            **kwargs,
-        )
-
-        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
-        attn_output = self.o_proj(attn_output)
-        return attn_output, attn_weights
-
-
-class DummyDecoderLayer(GradientCheckpointingLayer):
-    def __init__(self, config: DummyConfig, layer_idx: int):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-
-        self.self_attn = DummyAttention(config=config, layer_idx=layer_idx)
-
-        self.mlp = DummyMLP(config)
-        self.input_layernorm = DummyRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = DummyRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Cache] = None,
-        output_attentions: Optional[bool] = False,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
-        residual = hidden_states
-        hidden_states = self.input_layernorm(hidden_states)
-
-        # Self Attention
-        hidden_states, self_attn_weights = self.self_attn(
-            hidden_states=hidden_states,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_value=past_key_value,
-            output_attentions=output_attentions,
-            use_cache=use_cache,
-            cache_position=cache_position,
-            position_embeddings=position_embeddings,
-            **kwargs,
-        )
-        hidden_states = residual + hidden_states
-
-        # Fully Connected
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
-        outputs = (hidden_states,)
-        if output_attentions:
-            outputs += (self_attn_weights,)
-
-        return outputs
-
-
-@auto_docstring
-class DummyPreTrainedModel(PreTrainedModel):
-    config_class = DummyConfig
-    base_model_prefix = "model"
-    supports_gradient_checkpointing = True
-    _no_split_modules = ["DummyDecoderLayer"]
-    _skip_keys_device_placement = ["past_key_values"]
-    _supports_flash_attn_2 = True
-    _supports_sdpa = True
-    _supports_flex_attn = True
-    _supports_cache_class = True
-    _supports_quantized_cache = True
-    _supports_static_cache = True
-    _supports_attention_backend = True
-
-    def _init_weights(self, module):
-        std = self.config.initializer_range
-        if isinstance(module, nn.Linear):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.bias is not None:
-                module.bias.data.zero_()
-        elif isinstance(module, nn.Embedding):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.padding_idx is not None:
-                module.weight.data[module.padding_idx].zero_()
-        elif isinstance(module, DummyRMSNorm):
-            module.weight.data.fill_(1.0)
-
-
-@auto_docstring
-class DummyModel(DummyPreTrainedModel):
-    def __init__(self, config: DummyConfig):
-        super().__init__(config)
-        self.padding_idx = config.pad_token_id
-        self.vocab_size = config.vocab_size
-
-        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
-        self.layers = nn.ModuleList(
-            [DummyDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
-        )
-        self.norm = DummyRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.rotary_emb = DummyRotaryEmbedding(config=config)
-        self.gradient_checkpointing = False
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.embed_tokens = value
-
-    @can_return_tuple
-    @auto_docstring
-    def forward(
-        self,
-        input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
-    ) -> BaseModelOutputWithPast:
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
-
-        if (input_ids is None) ^ (inputs_embeds is not None):
-            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
-
-        if self.gradient_checkpointing and self.training and use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
-            )
-            use_cache = False
-
-        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
-        if not isinstance(past_key_values, (type(None), Cache)):
-            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
-
-        if inputs_embeds is None:
-            inputs_embeds = self.embed_tokens(input_ids)
-
-        if use_cache and past_key_values is None:
-            past_key_values = DynamicCache()
-
-        if cache_position is None:
-            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
-            cache_position = torch.arange(
-                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
-            )
-
-        if position_ids is None:
-            position_ids = cache_position.unsqueeze(0)
-
-        causal_mask = create_causal_mask(
-            config=self.config,
-            input_embeds=inputs_embeds,
-            attention_mask=attention_mask,
-            cache_position=cache_position,
-            past_key_values=past_key_values,
-        )
-
-        hidden_states = inputs_embeds
-
-        # create position embeddings to be shared across the decoder layers
-        position_embeddings = self.rotary_emb(hidden_states, position_ids)
-
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-
-        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-
-            layer_outputs = decoder_layer(
-                hidden_states,
-                attention_mask=causal_mask,
-                position_ids=position_ids,
-                past_key_value=past_key_values,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
-                cache_position=cache_position,
-                position_embeddings=position_embeddings,
-                **flash_attn_kwargs,
-            )
-
-            hidden_states = layer_outputs[0]
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        hidden_states = self.norm(hidden_states)
-
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        return BaseModelOutputWithPast(
-            last_hidden_state=hidden_states,
-            past_key_values=past_key_values if use_cache else None,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
-        )

examples/modular-transformers/modeling_global_indexing.py

@@ -0,0 +1,169 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from examples/modular-transformers/modular_global_indexing.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_global_indexing.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+from typing import Callable, Optional
+
+import torch
+from torch import nn
+
+from transformers.modeling_utils import AttentionInterface
+
+from ...cache_utils import Cache
+from ...processing_utils import Unpack
+from ...utils import TransformersKwargs
+from .configuration_global_indexing import GlobalIndexingConfig
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def custom_flex(x, **kwargs):
+    """Dummy function."""
+    return x
+
+
+ALL_ATTENTION_FUNCTIONS = AttentionInterface()
+# This indexing statement and associated function should be exported correctly!
+ALL_ATTENTION_FUNCTIONS["flex_attention"] = custom_flex
+
+
+class GlobalIndexingAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: GlobalIndexingConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights

examples/modular-transformers/modeling_multimodal2.py

@@ -289,7 +289,6 @@ class Multimodal2VisionEncoder(nn.Module):
         self.layers = nn.ModuleList([Multimodal2VisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
         self.gradient_checkpointing = False
 
-    @can_return_tuple
     def forward(
         self,
         inputs_embeds,
@@ -455,7 +454,6 @@ class Multimodal2VisionTransformer(nn.Module):
         self.encoder = Multimodal2VisionEncoder(config)
         self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
 
-    @can_return_tuple
     @auto_docstring
     def forward(
         self,
@@ -499,7 +497,7 @@ class Multimodal2VisionPreTrainedModel(PreTrainedModel):
     base_model_prefix = "multimodal2_vision"
     supports_gradient_checkpointing = True
     _supports_sdpa = True
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_flex_attn = True
     _supports_attention_backend = True
 

examples/modular-transformers/modeling_my_new_model2.py

@@ -12,13 +12,13 @@ from torch import nn
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...masking_utils import create_causal_mask
-from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, SequenceClassifierOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import auto_docstring, can_return_tuple, logging
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
+from ...utils.generic import check_model_inputs
 from .configuration_my_new_model2 import MyNewModel2Config
 
 
@@ -65,7 +65,7 @@ class MyNewModel2RotaryEmbedding(nn.Module):
     def __init__(self, config: MyNewModel2Config, device=None):
         super().__init__()
         # BC: "rope_type" was originally "type"
-        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
             self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
         else:
             self.rope_type = "default"
@@ -149,7 +149,7 @@ def eager_attention_forward(
     attention_mask: Optional[torch.Tensor],
     scaling: float,
     dropout: float = 0.0,
-    **kwargs,
+    **kwargs: Unpack[TransformersKwargs],
 ):
     key_states = repeat_kv(key, module.num_key_value_groups)
     value_states = repeat_kv(value, module.num_key_value_groups)
@@ -200,8 +200,8 @@ class MyNewModel2Attention(nn.Module):
         attention_mask: Optional[torch.Tensor],
         past_key_value: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -254,22 +254,19 @@ class MyNewModel2DecoderLayer(GradientCheckpointingLayer):
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
         past_key_value: Optional[Cache] = None,
-        output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
         position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor]:
         residual = hidden_states
         hidden_states = self.input_layernorm(hidden_states)
-
         # Self Attention
-        hidden_states, self_attn_weights = self.self_attn(
+        hidden_states, _ = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
             past_key_value=past_key_value,
-            output_attentions=output_attentions,
             use_cache=use_cache,
             cache_position=cache_position,
             position_embeddings=position_embeddings,
@@ -282,12 +279,7 @@ class MyNewModel2DecoderLayer(GradientCheckpointingLayer):
         hidden_states = self.post_attention_layernorm(hidden_states)
         hidden_states = self.mlp(hidden_states)
         hidden_states = residual + hidden_states
-
-        outputs = (hidden_states,)
-        if output_attentions:
-            outputs += (self_attn_weights,)
-
-        return outputs
+        return hidden_states
 
 
 @auto_docstring
@@ -297,13 +289,17 @@ class MyNewModel2PreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["MyNewModel2DecoderLayer"]
     _skip_keys_device_placement = ["past_key_values"]
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_cache_class = True
     _supports_quantized_cache = True
     _supports_static_cache = True
     _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": MyNewModel2DecoderLayer,
+        "attentions": MyNewModel2Attention,
+    }
 
     def _init_weights(self, module):
         std = self.config.initializer_range
@@ -337,13 +333,7 @@ class MyNewModel2Model(MyNewModel2PreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    def get_input_embeddings(self):
-        return self.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.embed_tokens = value
-
-    @can_return_tuple
+    @check_model_inputs
     @auto_docstring
     def forward(
         self,
@@ -353,26 +343,12 @@ class MyNewModel2Model(MyNewModel2PreTrainedModel):
         past_key_values: Optional[Cache] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
+        **kwargs: Unpack[TransformersKwargs],
     ) -> BaseModelOutputWithPast:
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
-
         if (input_ids is None) ^ (inputs_embeds is not None):
             raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 
-        if self.gradient_checkpointing and self.training and use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
-            )
-            use_cache = False
-
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
 
@@ -394,6 +370,7 @@ class MyNewModel2Model(MyNewModel2PreTrainedModel):
             attention_mask=attention_mask,
             cache_position=cache_position,
             past_key_values=past_key_values,
+            position_ids=position_ids,
         )
 
         # embed positions
@@ -408,42 +385,21 @@ class MyNewModel2Model(MyNewModel2PreTrainedModel):
         normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
         hidden_states = hidden_states * normalizer
 
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-
         for decoder_layer in self.layers[: self.config.num_hidden_layers]:
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-
-            layer_outputs = decoder_layer(
+            hidden_states = decoder_layer(
                 hidden_states,
                 attention_mask=causal_mask,
                 position_ids=position_ids,
                 past_key_value=past_key_values,
-                output_attentions=output_attentions,
                 use_cache=use_cache,
                 cache_position=cache_position,
                 position_embeddings=position_embeddings,
                 **kwargs,
             )
-
-            hidden_states = layer_outputs[0]
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
         hidden_states = self.norm(hidden_states)
-
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
         return BaseModelOutputWithPast(
             last_hidden_state=hidden_states,
             past_key_values=past_key_values if use_cache else None,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
         )
 
 
@@ -471,12 +427,6 @@ class MyNewModel2ForSequenceClassification(MyNewModel2PreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    def get_input_embeddings(self):
-        return self.model.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.model.embed_tokens = value
-
     @can_return_tuple
     @auto_docstring
     def forward(
@@ -488,8 +438,7 @@ class MyNewModel2ForSequenceClassification(MyNewModel2PreTrainedModel):
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
+        **kwargs: Unpack[TransformersKwargs],
     ) -> SequenceClassifierOutputWithPast:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
@@ -505,8 +454,7 @@ class MyNewModel2ForSequenceClassification(MyNewModel2PreTrainedModel):
             past_key_values=past_key_values,
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
+            **kwargs,
         )
         hidden_states = transformer_outputs.last_hidden_state
         logits = self.score(hidden_states)

examples/modular-transformers/modeling_new_task_model.py

@@ -95,7 +95,7 @@ class NewTaskModelPreTrainedModel(PreTrainedModel):
     _supports_cache_class = True
     _supports_quantized_cache = True
     _supports_static_cache = True
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_attention_backend = True
@@ -118,6 +118,8 @@ class NewTaskModelPreTrainedModel(PreTrainedModel):
 )
 class NewTaskModelModel(NewTaskModelPreTrainedModel):
     _checkpoint_conversion_mapping = {"language_model.model": "language_model"}
+    # we are filtering the logits/labels so we shouldn't divide the loss based on num_items_in_batch
+    accepts_loss_kwargs = False
 
     def __init__(self, config: NewTaskModelConfig):
         super().__init__(config)
@@ -313,9 +315,11 @@ class NewTaskModelModel(NewTaskModelPreTrainedModel):
                 special_image_mask = inputs_embeds == self.get_input_embeddings()(
                     torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
                 )
+                special_image_mask = special_image_mask.all(-1)
             else:
-                special_image_mask = (input_ids == self.config.image_token_id).unsqueeze(-1)
-                special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+                special_image_mask = input_ids == self.config.image_token_id
+
+            special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
 
             if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
                 image_tokens_in_text = (special_image_mask).sum(dim=1).sum(dim=0)[0]
@@ -385,12 +389,6 @@ class NewTaskModelForNewTask(NewTaskModelPreTrainedModel, GenerationMixin):
     def set_input_embeddings(self, value):
         self.model.set_input_embeddings(value)
 
-    def get_output_embeddings(self):
-        return self.lm_head
-
-    def set_output_embeddings(self, new_embeddings):
-        self.lm_head = new_embeddings
-
     def set_decoder(self, decoder):
         self.model.set_decoder(decoder)
 
@@ -433,32 +431,6 @@ class NewTaskModelForNewTask(NewTaskModelPreTrainedModel, GenerationMixin):
         num_logits_to_keep: int = 0,
     ) -> Union[tuple, NewTaskModelCausalLMOutputWithPast]:
         r"""
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-            config.text_config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.text_config.vocab_size]`.
-
-        Example:
-
-        ```python
-        >>> from PIL import Image
-        >>> import requests
-        >>> from transformers import AutoProcessor, NewTaskModelForNewTask
-
-        >>> model = NewTaskModelForNewTask.from_pretrained("google/new_task_model2-3b-mix-224")
-        >>> processor = AutoProcessor.from_pretrained("google/new_task_model2-3b-mix-224")
-
-        >>> prompt = "Where is the cat standing?"
-        >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
-        >>> image = Image.open(requests.get(url, stream=True).raw)
-
-        >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
-
-        >>> # Generate
-        >>> generate_ids = model.generate(**inputs,)
-        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-        "Where is the cat standing?\nsnow"
-        ```
         Returns:
         """
         vlm_outputs = super().forward(

examples/modular-transformers/modeling_super.py

@@ -14,12 +14,12 @@ from transformers.modeling_outputs import CausalLMOutputWithPast
 from ...activations import ACT2FN
 from ...cache_utils import Cache
 from ...integrations import use_kernel_forward_from_hub
-from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import auto_docstring, can_return_tuple
+from ...utils import TransformersKwargs, auto_docstring
+from ...utils.generic import check_model_inputs
 from .configuration_super import SuperConfig
 
 
@@ -48,7 +48,7 @@ class SuperRotaryEmbedding(nn.Module):
     def __init__(self, config: SuperConfig, device=None):
         super().__init__()
         # BC: "rope_type" was originally "type"
-        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
             self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
         else:
             self.rope_type = "default"
@@ -148,7 +148,7 @@ def eager_attention_forward(
     attention_mask: Optional[torch.Tensor],
     scaling: float,
     dropout: float = 0.0,
-    **kwargs,
+    **kwargs: Unpack[TransformersKwargs],
 ):
     key_states = repeat_kv(key, module.num_key_value_groups)
     value_states = repeat_kv(value, module.num_key_value_groups)
@@ -199,8 +199,8 @@ class SuperAttention(nn.Module):
         attention_mask: Optional[torch.Tensor],
         past_key_value: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -253,22 +253,19 @@ class SuperDecoderLayer(GradientCheckpointingLayer):
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
         past_key_value: Optional[Cache] = None,
-        output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
         position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor]:
         residual = hidden_states
         hidden_states = self.input_layernorm(hidden_states)
-
         # Self Attention
-        hidden_states, self_attn_weights = self.self_attn(
+        hidden_states, _ = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
             past_key_value=past_key_value,
-            output_attentions=output_attentions,
             use_cache=use_cache,
             cache_position=cache_position,
             position_embeddings=position_embeddings,
@@ -281,12 +278,7 @@ class SuperDecoderLayer(GradientCheckpointingLayer):
         hidden_states = self.post_attention_layernorm(hidden_states)
         hidden_states = self.mlp(hidden_states)
         hidden_states = residual + hidden_states
-
-        outputs = (hidden_states,)
-        if output_attentions:
-            outputs += (self_attn_weights,)
-
-        return outputs
+        return hidden_states
 
 
 @auto_docstring
@@ -296,13 +288,17 @@ class SuperPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["SuperDecoderLayer"]
     _skip_keys_device_placement = ["past_key_values"]
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_cache_class = True
     _supports_quantized_cache = True
     _supports_static_cache = True
     _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": SuperDecoderLayer,
+        "attentions": SuperAttention,
+    }
 
     def _init_weights(self, module):
         std = self.config.initializer_range
@@ -336,13 +332,7 @@ class SuperModel(SuperPreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    def get_input_embeddings(self):
-        return self.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.embed_tokens = value
-
-    @can_return_tuple
+    @check_model_inputs
     @auto_docstring
     def forward(
         self,

examples/modular-transformers/modeling_switch_function.py

@@ -11,9 +11,9 @@ import torch
 from torch import nn
 
 from ...cache_utils import Cache
-from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
 from ...processing_utils import Unpack
+from ...utils import TransformersKwargs
 from .configuration_switch_function import SwitchFunctionConfig
 
 
@@ -72,7 +72,7 @@ def eager_attention_forward(
     attention_mask: Optional[torch.Tensor],
     scaling: float,
     dropout: float = 0.0,
-    **kwargs,
+    **kwargs: Unpack[TransformersKwargs],
 ):
     key_states = repeat_kv(key, module.num_key_value_groups)
     value_states = repeat_kv(value, module.num_key_value_groups)
@@ -123,8 +123,8 @@ class SwitchFunctionAttention(nn.Module):
         attention_mask: Optional[torch.Tensor],
         past_key_value: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 

examples/modular-transformers/modular_dummy.py

@@ -1,15 +0,0 @@
-import torch
-
-from transformers.models.llama.modeling_llama import LlamaModel
-
-
-def rotate_half(x):
-    """Rotates half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 4]
-    x2 = x[..., x.shape[-1] // 4 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-# example where we need some deps and some functions
-class DummyModel(LlamaModel):
-    pass