Initial stubs for P/D scheduling changes

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

.buildkite/lm-eval-harness/configs/Qwen1.5-MoE-W4A16-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16 -b auto -l 1319 -f 5 -t 1
+model_name: "nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.31
+  - name: "exact_match,flexible-extract"
+    value: 0.47
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/models-small.txt

@@ -4,7 +4,7 @@ Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
-Minitron-4B-Base-FP8.yaml
+Qwen1.5-MoE-W4A16-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-FP8W8.yaml
 Meta-Llama-3-8B-QQQ.yaml

.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh

@@ -10,15 +10,24 @@ set -x
 set -o pipefail
 
 check_gpus() {
-  # check the number of GPUs and GPU type.
-  declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  if command -v nvidia-smi; then
+    # check the number of GPUs and GPU type.
+    declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  elif command -v amd-smi; then
+    declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
+  fi
+
   if [[ $gpu_count -gt 0 ]]; then
     echo "GPU found."
   else
     echo "Need at least 1 GPU to run benchmarking."
     exit 1
   fi
-  declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  if command -v nvidia-smi; then
+    declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  elif command -v amd-smi; then
+    declare -g gpu_type=$(amd-smi static -g 0 -a | grep 'MARKET_NAME' | awk '{print $2}')
+  fi
   echo "GPU type is $gpu_type"
 }
 
@@ -90,9 +99,15 @@ kill_gpu_processes() {
 
 
   # wait until GPU memory usage smaller than 1GB
-  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
-    sleep 1
-  done
+  if command -v nvidia-smi; then
+    while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
+      sleep 1
+    done
+  elif command -v amd-smi; then
+    while [ "$(amd-smi metric -g 0 | grep 'USED_VRAM' | awk '{print $2}')" -ge 1000 ]; do
+      sleep 1
+    done
+  fi
 
   # remove vllm config file
   rm -rf ~/.config/vllm

.buildkite/nightly-benchmarks/tests/serving-tests.json

@@ -63,10 +63,12 @@
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "disable_log_requests": "", 
             "tensor_parallel_size": 4,
-            "swap_space": 16, 
-            "speculative_model": "turboderp/Qwama-0.5B-Instruct",
-            "num_speculative_tokens": 4,
-            "speculative_draft_tensor_parallel_size": 1
+            "swap_space": 16,
+            "speculative_config": {
+                "model": "turboderp/Qwama-0.5B-Instruct",
+                "num_speculative_tokens": 4,
+                "draft_tensor_parallel_size": 1
+            }
         },
         "client_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",

.buildkite/release-pipeline.yaml

@@ -3,10 +3,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -14,10 +14,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -31,10 +31,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -48,7 +48,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
   - label: "Build and publish TPU release image"
@@ -57,7 +57,7 @@ steps:
     agents:
       queue: tpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f Dockerfile.tpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
       - "docker push vllm/vllm-tpu:nightly"
       - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
     plugins:
@@ -82,7 +82,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain -f Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -105,19 +105,33 @@ fi
 if [[ $commands == *" entrypoints/openai "* ]]; then
   commands=${commands//" entrypoints/openai "/" entrypoints/openai \
   --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_chat.py \
   --ignore=entrypoints/openai/test_shutdown.py \
   --ignore=entrypoints/openai/test_completion.py \
   --ignore=entrypoints/openai/test_sleep.py \
   --ignore=entrypoints/openai/test_models.py \
+  --ignore=entrypoints/openai/test_lora_adapters.py \
+  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+  --ignore=entrypoints/openai/test_root_path.py \
+  --ignore=entrypoints/openai/test_tokenization.py \
   --ignore=entrypoints/openai/test_prompt_validation.py "}
 fi
 
 #ignore certain Entrypoints/llm tests
-if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
-  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+if [[ $commands == *" entrypoints/llm "* ]]; then
+  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
+  --ignore=entrypoints/llm/test_chat.py \
+  --ignore=entrypoints/llm/test_accuracy.py \
+  --ignore=entrypoints/llm/test_init.py \
+  --ignore=entrypoints/llm/test_generate_multiple_loras.py \
+  --ignore=entrypoints/llm/test_prompt_validation.py "}
 fi
 
+#Obsolete currently
+##ignore certain Entrypoints/llm tests
+#if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+#  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+#fi
+
 # --ignore=entrypoints/openai/test_encoder_decoder.py \
 # --ignore=entrypoints/openai/test_embedding.py \
 # --ignore=entrypoints/openai/test_oot_registration.py
@@ -134,9 +148,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     # assign shard-id for each shard
     commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
     echo "Shard ${GPU} commands:$commands_gpu"
+    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
     docker run \
-        --device /dev/kfd --device /dev/dri \
-        --network host \
+        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+        --network=host \
         --shm-size=16gb \
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
@@ -163,9 +178,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     fi
   done
 else
+  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-          --device /dev/kfd --device /dev/dri \
-          --network host \
+          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+          --network=host \
           --shm-size=16gb \
           --rm \
           -e HIP_VISIBLE_DEVICES=0 \

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+# This script build the CPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Setup cleanup
+remove_docker_container() { podman rm -f cpu-test-ubi9-ppc || true; podman system prune -f; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Try building the docker image
+podman build -t cpu-test-ubi9-ppc -f docker/Dockerfile.ppc64le .
+
+# Run the image
+podman run -itd --entrypoint /bin/bash -v /tmp/:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN --name cpu-test-ubi9-ppc cpu-test-ubi9-ppc
+
+function cpu_tests() {
+
+  # offline inference
+  podman exec cpu-test-ubi9-ppc bash -c "
+    set -e
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
+
+  # Run basic model test
+  podman exec cpu-test-ubi9-ppc bash -c "
+    set -e
+    pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
+    pip install sentence-transformers datamodel_code_generator
+    pytest -v -s tests/models/embedding/language/test_cls_models.py::test_classification_models[float-jason9693/Qwen2.5-1.5B-apeach]
+    pytest -v -s tests/models/embedding/language/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model"
+}
+
+# All of CPU tests are expected to be finished less than 40 mins.
+export -f cpu_tests
+timeout 40m bash -c cpu_tests
+

.buildkite/scripts/hardware_ci/run-cpu-test-s390x.sh

@@ -10,5 +10,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-docker build -t cpu-test -f Dockerfile.ppc64le .
-
+docker build -t cpu-test -f docker/Dockerfile.s390x .

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -8,15 +8,19 @@ set -ex
 CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
-# Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 -f Dockerfile.cpu .
-
 # Setup cleanup
-remove_docker_container() { set -e; docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; }
+remove_docker_container() { 
+    set -e; 
+    docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; 
+    docker image rm cpu-test-"$BUILDKITE_BUILD_NUMBER" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 || true; 
+}
 trap remove_docker_container EXIT
 remove_docker_container
 
+# Try building the docker image
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$BUILDKITE_BUILD_NUMBER" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+
 # Run the image, setting --shm-size=4g for tensor parallel.
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
  --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
@@ -36,8 +40,8 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pip install -r vllm/requirements/test.txt
-    pip install -r vllm/requirements/cpu.txt
+    pytest -v -s tests/kernels/test_cache.py -m cpu_model
+    pytest -v -s tests/kernels/test_mla_decode_cpu.py -m cpu_model
     pytest -v -s tests/models/decoder_only/language -m cpu_model
     pytest -v -s tests/models/embedding/language -m cpu_model
     pytest -v -s tests/models/encoder_decoder/language -m cpu_model

.buildkite/scripts/hardware_ci/run-gh200-test.sh

@@ -9,6 +9,7 @@ python3 use_existing_torch.py
 
 # Try building the docker image
 DOCKER_BUILDKIT=1 docker build . \
+  --file docker/Dockerfile \
   --target vllm-openai \
   --platform "linux/arm64" \
   -t gh200-test \

.buildkite/scripts/hardware_ci/run-hpu-test.sh

@@ -5,7 +5,7 @@
 set -ex
 
 # Try building the docker image
-docker build -t hpu-test-env -f Dockerfile.hpu .
+docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can

.buildkite/scripts/hardware_ci/run-neuron-test.sh

@@ -35,7 +35,7 @@ else
     date "+%s" > /tmp/neuron-docker-build-timestamp
 fi
 
-docker build -t "${image_name}" -f Dockerfile.neuron .
+docker build -t "${image_name}" -f docker/Dockerfile.neuron .
 
 # Setup cleanup
 remove_docker_container() {

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -1,9 +1,9 @@
 #!/bin/bash
 
-set -e
+set -xue
 
 # Build the docker image.
-docker build -f Dockerfile.tpu -t vllm-tpu .
+docker build -f docker/Dockerfile.tpu -t vllm-tpu .
 
 # Set up cleanup.
 remove_docker_container() { docker rm -f tpu-test || true; }
@@ -17,12 +17,16 @@ source /etc/environment
 docker run --privileged --net host --shm-size=16G -it \
     -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
     vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install pytest \
+    && python3 -m pip install pytest tpu-info \
     && python3 -m pip install lm_eval[api]==0.4.4 \
     && export VLLM_USE_V1=1 \
     && export VLLM_XLA_CHECK_RECOMPILATION=1 \
+    && echo HARDWARE \
+    && tpu-info \
+    && echo TEST_0 \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_perf.py \
     && echo TEST_1 \
-    && pytest /workspace/vllm/tests/tpu/test_compilation.py \
+    && pytest -v -s /workspace/vllm/tests/tpu/test_compilation.py \
     && echo TEST_2 \
     && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
     && echo TEST_3 \
@@ -30,9 +34,16 @@ docker run --privileged --net host --shm-size=16G -it \
     && echo TEST_4 \
     && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
     && echo TEST_5 \
-    && python3 /workspace/vllm/examples/offline_inference/tpu.py" \
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py \
+    && echo TEST_6 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py \
+    && echo TEST_7 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py \
+    && echo TEST_8 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py \
+    && echo TEST_9 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py" \
 
 
 # TODO: This test fails because it uses RANDOM_SEED sampling
 # && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
-

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -8,7 +8,7 @@ image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 
 # Try building the docker image
-docker build -t ${image_name} -f Dockerfile.xpu .
+docker build -t ${image_name} -f docker/Dockerfile.xpu .
 
 # Setup cleanup
 remove_docker_container() { 

.buildkite/scripts/run-benchmarks.sh

@@ -5,8 +5,8 @@
 set -ex
 set -o pipefail
 
-# cd into parent directory of this file
-cd "$(dirname "${BASH_SOURCE[0]}")/.."
+# cd 2 levels into the working directory
+cd "$(dirname "${BASH_SOURCE[0]}")/../.."
 
 (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 

.buildkite/scripts/run-multi-node-test.sh

@@ -3,7 +3,7 @@
 set -euox pipefail
 
 if [[ $# -lt 4 ]]; then
-    echo "Usage: .buildkite/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
+    echo "Usage: .buildkite/scripts/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
     exit 1
 fi
 

.buildkite/test-pipeline.yaml

@@ -104,7 +104,7 @@ steps:
 - label: Entrypoints Test # 40min
   working_dir: "/vllm-workspace/tests"
   fast_check: true
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
@@ -118,7 +118,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_openai_schema.py
   - pytest -v -s entrypoints/test_chat_utils.py
   - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
@@ -135,12 +135,14 @@ steps:
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
+  - tests/v1/test_async_llm_dp.py
   commands:
   # test with tp=2 and external_dp=2
   - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
@@ -148,11 +150,12 @@ steps:
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
-  - VLLM_ENABLE_V1_MULTIPROCESSING=0 python3 rlhf.py
-  - VLLM_ENABLE_V1_MULTIPROCESSING=0 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - python3 rlhf.py
+  - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
 - label: Metrics, Tracing Test # 10min
+  mirror_hardwares: [amd]
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -160,18 +163,13 @@ steps:
   - tests/tracing
   commands:
   - pytest -v -s metrics
-  - "pip install \
-      'opentelemetry-sdk>=1.26.0,<1.27.0' \
-      'opentelemetry-api>=1.26.0,<1.27.0' \
-      'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
-      'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'"
   - pytest -v -s tracing
 
 ##### fast check tests  #####
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -202,7 +200,6 @@ steps:
   commands:
     # split the test to avoid interference
     - pytest -v -s v1/core
-    - pytest -v -s v1/entrypoints
     - pytest -v -s v1/engine
     - pytest -v -s v1/entrypoints
     - pytest -v -s v1/sample
@@ -283,13 +280,21 @@ steps:
     - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py  --ignore=lora/test_transfomers_model.py
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
   parallelism: 4
 
+- label: PyTorch Compilation Unit Tests
+  source_file_dependencies:
+    - vllm/
+    - tests/compile
+  commands:
+    - pytest -v -s compile/test_pass_manager.py
+    - pytest -v -s compile/test_fusion.py
+
 - label: PyTorch Fullgraph Smoke Test # 9min
   source_file_dependencies:
   - vllm/
@@ -299,7 +304,6 @@ steps:
   # these tests need to be separated, cannot combine
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
-  - pytest -v -s compile/test_pass_manager.py
 
 - label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
@@ -309,7 +313,7 @@ steps:
   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Test %N # 1h each
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   source_file_dependencies:
   - csrc/
   - vllm/attention
@@ -319,7 +323,7 @@ steps:
   parallelism: 4
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -335,7 +339,14 @@ steps:
   source_file_dependencies:
   - benchmarks/
   commands:
-  - bash run-benchmarks.sh
+  - bash scripts/run-benchmarks.sh
+
+- label: Benchmarks CLI Test # 10min
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+  - pytest -v -s benchmarks/
 
 - label: Quantization Test # 33min
   source_file_dependencies:
@@ -370,12 +381,14 @@ steps:
 
 - label: OpenAI-Compatible Tool Use # 20 min
   fast_check: false
-  mirror_hardwares: [ amd ]
+  #mirror_hardwares: [ amd ]
   source_file_dependencies:
     - vllm/
     - tests/tool_use
+    - tests/mistral_tool_use
   commands:
     - pytest -v -s tool_use
+    - pytest -v -s mistral_tool_use
 
 #####  models test  #####
 
@@ -387,7 +400,9 @@ steps:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
     # V1 Test: https://github.com/vllm-project/vllm/issues/14531
-    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'llama4'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'plamo2'
 
 - label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]
@@ -397,6 +412,8 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/language -m core_model
 
@@ -408,6 +425,8 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/language -m 'not core_model'
 
@@ -424,11 +443,12 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal
     - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
+    - pytest -v -s models/decoder_only/vision_language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/vision_language -m core_model
     - pytest -v -s models/encoder_decoder/audio_language -m core_model
     - pytest -v -s models/encoder_decoder/language -m core_model
     - pytest -v -s models/encoder_decoder/vision_language -m core_model
+    - pytest -v -s models/decoder_only/vision_language/test_interleaved.py
 
 - label: Multi-Modal Models Test (Extended) 1 # 48m
   optional: true
@@ -442,10 +462,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
     - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
-    # HACK - run phi3v tests separately to sidestep this transformers bug
-    # https://github.com/huggingface/transformers/issues/34307
-    - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
+    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/vision_language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
@@ -461,6 +478,7 @@ steps:
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
+  mirror_hardwares: [amd]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -472,6 +490,7 @@ steps:
 #####  multi gpus test  #####
 
 - label: Distributed Comm Ops Test # 7min
+  mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -514,8 +533,11 @@ steps:
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
+  - tests/v1/test_async_llm_dp.py
+  - vllm/v1/engine/
   commands:
-  - VLLM_ENABLE_V1_MULTIPROCESSING=0 pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
@@ -530,6 +552,7 @@ steps:
   # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
 
 - label: Plugin Tests (2 GPUs) # 40min
   working_dir: "/vllm-workspace/tests"
@@ -592,14 +615,10 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    # This test runs llama 13B, so it is required to run on 4 GPUs.
-    - pytest -v -s -x lora/test_long_context.py
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
-    - pytest -v -s -x lora/test_minicpmv_tp.py
-    - pytest -v -s -x lora/test_transfomers_model.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

.github/ISSUE_TEMPLATE/200-installation.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/300-usage.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/400-bug-report.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/600-new-model.yml

@@ -9,7 +9,7 @@ body:
     value: >
       #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).
 
-      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/adding_model.html first to understand how to add a new model.
+      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/index.html first to understand how to add a new model.
 - type: textarea
   attributes:
     label: The model to consider.

.github/ISSUE_TEMPLATE/700-performance-discussion.yml

@@ -35,7 +35,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/PULL_REQUEST_TEMPLATE.md

@@ -3,4 +3,4 @@ FILL IN THE PR DESCRIPTION HERE
 FIX #xxxx (*link existing issues this PR will resolve*)
 
 <!--- pyml disable-next-line no-emphasis-as-heading -->
-**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>**
+**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>** (anything written below this line will be removed by GitHub Actions)

.github/mergify.yml

@@ -19,7 +19,7 @@ pull_request_rules:
       - files~=\.buildkite/
       - files~=^cmake/
       - files=CMakeLists.txt
-      - files~=^Dockerfile
+      - files~=^docker/Dockerfile
       - files~=^requirements.*\.txt
       - files=setup.py
   actions:
@@ -88,6 +88,36 @@ pull_request_rules:
       add:
         - v1
 
+- name: label-tpu
+  description: Automatically apply tpu label
+  # Keep this list in sync with `label-tpu-remove` conditions
+  conditions:
+    - or:
+      - files~=tpu.py
+      - files~=_tpu
+      - files~=tpu_
+      - files~=/tpu/
+      - files~=pallas
+  actions:
+    label:
+      add:
+        - tpu
+
+- name: label-tpu-remove
+  description: Automatically remove tpu label
+  # Keep this list in sync with `label-tpu` conditions
+  conditions:
+    - and:
+      - -files~=tpu.py
+      - -files~=_tpu
+      - -files~=tpu_
+      - -files~=/tpu/
+      - -files~=pallas
+  actions:
+    label:
+      remove:
+        - tpu
+
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
       - conflict

.github/workflows/lint-and-deploy.yaml

@@ -50,7 +50,7 @@ jobs:
         uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0
 
       - name: Build the Docker image vllm cpu
-        run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .
+        run: docker buildx build -f docker/Dockerfile.cpu -t vllm-cpu-env .
 
       - name: Configuration of docker images, network and namespace for the kind cluster
         run: |

.gitignore

@@ -203,3 +203,6 @@ benchmarks/**/*.json
 # Linting
 actionlint
 shellcheck*/
+
+# Ingore moe/marlin_moe gen code
+csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -1,3 +1,6 @@
+default_install_hook_types:
+  - pre-commit
+  - commit-msg
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
@@ -8,7 +11,6 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
-    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.9.3
   hooks:
@@ -119,6 +121,12 @@ repos:
     language: system
     always_run: true
     pass_filenames: false
+  - id: update-dockerfile-graph
+    name: Update Dockerfile dependency graph
+    entry: tools/update-dockerfile-graph.sh
+    language: script
+    files: ^docker/Dockerfile$
+    pass_filenames: false
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -34,7 +34,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -44,7 +44,7 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
 #
 # Note: the CUDA torch version is derived from pyproject.toml and various
 # requirements.txt files and should be kept consistent.  The ROCm torch
-# versions are derived from Dockerfile.rocm
+# versions are derived from docker/Dockerfile.rocm
 #
 set(TORCH_SUPPORTED_VERSION_CUDA "2.6.0")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
@@ -230,10 +230,12 @@ set(VLLM_EXT_SRC
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
+  "csrc/attention/merge_attn_states.cu"
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
+  "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
@@ -241,6 +243,7 @@ set(VLLM_EXT_SRC
   "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
+  "csrc/custom_all_reduce.cu"
   "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -282,7 +285,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
-    "csrc/custom_all_reduce.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
@@ -461,6 +463,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(FP4_ARCHS)
   endif()
 
+  #
+  # CUTLASS MoE kernels
+
+  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
+  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
+  # to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
+             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
+    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper.")
+    else()
+      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
   #
   # Machete kernels
 
@@ -580,21 +609,51 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
   cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
-    set(MARLIN_MOE_SRC
-        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.cu"
-        "csrc/moe/marlin_moe_ops.cu")
 
+    #
+    # For the Marlin MOE kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MOE_MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
+    file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
+          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE moe_marlin_generation_result
+        OUTPUT_VARIABLE moe_marlin_generation_output
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+      )
+
+      if (NOT moe_marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin MOE generation failed."
+                            " Result: \"${moe_marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
+      else()
+        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
+        message(STATUS "Marlin MOE generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
     set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_MOE_SRC}"
+      SRCS "${MOE_WNAA16_MARLIN_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
 
-    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_SRC}")
+    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
+
     message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
   else()
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"

Dockerfile.cpu

@@ -1,69 +0,0 @@
-# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
-
-FROM ubuntu:22.04 AS cpu-test-1
-
-ENV CCACHE_DIR=/root/.cache/ccache
-
-ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
-
-RUN --mount=type=cache,target=/var/cache/apt \
-    apt-get update -y \
-    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
-    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
-    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
-
-# https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
-# intel-openmp provides additional performance improvement vs. openmp
-# tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install intel-openmp==2025.0.1
-
-ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so"
-
-RUN echo 'ulimit -c 0' >> ~/.bashrc
-
-RUN pip install intel_extension_for_pytorch==2.6.0
-
-WORKDIR /workspace
-
-ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
-ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
-    pip install --upgrade pip && \
-    pip install -r requirements/build.txt
-
-FROM cpu-test-1 AS build
-
-WORKDIR /workspace/vllm
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements/common.txt,target=requirements/common.txt \
-    --mount=type=bind,src=requirements/cpu.txt,target=requirements/cpu.txt \
-    pip install -v -r requirements/cpu.txt
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
-ARG VLLM_CPU_DISABLE_AVX512
-ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
-    pip install dist/*.whl && \
-    rm -rf dist
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -e tests/vllm_test_utils
-
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.ppc64le

@@ -1,37 +0,0 @@
-FROM mambaorg/micromamba
-ARG MAMBA_DOCKERFILE_ACTIVATE=1
-USER root
-
-ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"
-
-RUN apt-get update -y && apt-get install -y git wget kmod curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 libssl-dev 
-
-# Some packages in requirements/cpu are installed here
-# IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
-# Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 rust && micromamba clean --all --yes
-
-COPY ./ /workspace/vllm
-
-WORKDIR /workspace/vllm
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
-
-RUN --mount=type=cache,target=/root/.cache/pip  \
-    RUSTFLAGS='-L /opt/conda/lib' pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
-        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        -r requirements/cpu.txt \
-        xformers uvloop==0.20.0
-
-RUN --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py install
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]

README.md

@@ -10,19 +10,14 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---
 
-[2025/03] We are collaborating with Ollama to host an [Inference Night](https://lu.ma/vllm-ollama) at Y Combinator in San Francisco on Thursday, March 27, at 6 PM. Discuss all things inference local or data center!
-
-[2025/04] We're hosting our first-ever *vLLM Asia Developer Day* in Singapore on *April 3rd*! This is a full-day event (9 AM - 9 PM SGT) in partnership with SGInnovate, AMD, and Embedded LLM. Meet the vLLM team and learn about LLM inference for RL, MI300X, and more! [Register Now](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)
-
----
-
 *Latest News* 🔥
-
+- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
+- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
 - [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
 - [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
@@ -103,7 +98,7 @@ Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
 ## Contributing
 
 We welcome and value any contributions and collaborations.
-Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
+Please check out [Contributing to vLLM](https://docs.vllm.ai/en/stable/contributing/overview.html) for how to get involved.
 
 ## Sponsors
 
@@ -126,6 +121,7 @@ Compute Resources:
 - Databricks
 - DeepInfra
 - Google Cloud
+- Intel
 - Lambda Lab
 - Nebius
 - Novita AI

benchmarks/README.md

@@ -41,29 +41,39 @@ become available.
       <td><code>synthetic</code></td>
     </tr>
     <tr>
-      <td><strong>HuggingFace</strong></td>
-      <td style="text-align: center;">🟡</td>
-      <td style="text-align: center;">🟡</td>
-      <td>Specify your dataset path on HuggingFace</td>
+      <td><strong>HuggingFace-VisionArena</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/VisionArena-Chat</code></td>
     </tr>
     <tr>
-      <td><strong>VisionArena</strong></td>
+      <td><strong>HuggingFace-InstructCoder</strong></td>
       <td style="text-align: center;">✅</td>
       <td style="text-align: center;">✅</td>
-      <td><code>lmarena-ai/vision-arena-bench-v0.1</code> (a HuggingFace dataset)</td>
+      <td><code>likaixin/InstructCoder</code></td>
+    </tr>
+      <tr>
+      <td><strong>HuggingFace-AIMO</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>AI-MO/aimo-validation-aime</code> , <code>AI-MO/NuminaMath-1.5</code>, <code>AI-MO/NuminaMath-CoT</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Other</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
     </tr>
   </tbody>
 </table>
 
 ✅: supported
 
+🟡: Partial support
+
 🚧: to be supported
 
-🟡: Partial support. Currently, HuggingFaceDataset only supports dataset formats
-similar to `lmms-lab/LLaVA-OneVision-Data` and `Aeala/ShareGPT_Vicuna_unfiltered`.
-If you need support for other dataset formats, please consider contributing.
-
-**Note**: VisionArena’s `dataset-name` should be set to `hf`
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
 
 ---
 ## Example - Online Benchmark
@@ -71,8 +81,7 @@ If you need support for other dataset formats, please consider contributing.
 First start serving your model
 
 ```bash
-MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
-vllm serve ${MODEL_NAME} --disable-log-requests
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
 ```
 
 Then run the benchmarking script
@@ -80,12 +89,13 @@ Then run the benchmarking script
 ```bash
 # download dataset
 # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
-NUM_PROMPTS=10
-BACKEND="vllm"
-DATASET_NAME="sharegpt"
-DATASET_PATH="<your data path>/ShareGPT_V3_unfiltered_cleaned_split.json"
-python3 vllm/benchmarks/benchmark_serving.py --backend ${BACKEND} --model ${MODEL_NAME} --endpoint /v1/completions --dataset-name ${DATASET_NAME} --dataset-path ${DATASET_PATH} --num-prompts ${NUM_PROMPTS}
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --num-prompts 10
 ```
 
 If successful, you will see the following output
@@ -122,88 +132,105 @@ vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 ```
 
 ```bash
-MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
-NUM_PROMPTS=10
-BACKEND="openai-chat"
-DATASET_NAME="hf"
-DATASET_PATH="lmarena-ai/vision-arena-bench-v0.1"
-DATASET_SPLIT='train'
-
 python3 vllm/benchmarks/benchmark_serving.py \
-  --backend "${BACKEND}" \
-  --model "${MODEL_NAME}" \
-  --endpoint "/v1/chat/completions" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --hf-split "${DATASET_SPLIT}" \
-  --num-prompts "${NUM_PROMPTS}"
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 1000
 ```
 
-### HuggingFaceDataset Examples
+### InstructCoder Benchmark with Speculative Decoding
 
-Currently, HuggingFaceDataset only supports dataset formats
-similar to `lmms-lab/LLaVA-OneVision-Data` and `Aeala/ShareGPT_Vicuna_unfiltered`. If you need support for other dataset
-formats, please consider contributing.
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-model "[ngram]" \
+    --ngram_prompt_lookup_min 2 \
+    --ngram-prompt-lookup-max 5 \
+    --num_speculative_tokens 5
+```
+
+``` bash
+python3 benchmarks/benchmark_serving.py \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name hf \
+    --dataset-path likaixin/InstructCoder \
+    --num-prompts 2048
+```
+
+### Other HuggingFaceDataset Examples
 
 ```bash
-# need a model with vision capability here
 vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 ```
 
 **`lmms-lab/LLaVA-OneVision-Data`**
 
 ```bash
-MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
-NUM_PROMPTS=10
-BACKEND="openai-chat"
-DATASET_NAME="hf"
-DATASET_PATH="lmms-lab/LLaVA-OneVision-Data"
-DATASET_SPLIT='train'
-DATASET_SUBSET='chart2text(cauldron)'
 python3 vllm/benchmarks/benchmark_serving.py \
-  --backend "${BACKEND}" \
-  --model "${MODEL_NAME}" \
-  --endpoint "/v1/chat/completions" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --hf-split "${DATASET_SPLIT}" \
-  --num-prompts "${NUM_PROMPTS}" \
-  --hf-subset "${DATASET_SUBSET}"
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
 ```
 
 **`Aeala/ShareGPT_Vicuna_unfiltered`**
 
 ```bash
-MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
-NUM_PROMPTS=10
-BACKEND="openai-chat"
-DATASET_NAME="hf"
-DATASET_PATH="Aeala/ShareGPT_Vicuna_unfiltered"
-DATASET_SPLIT='train'
 python3 vllm/benchmarks/benchmark_serving.py \
-  --backend "${BACKEND}" \
-  --model "${MODEL_NAME}" \
-  --endpoint "/v1/chat/completions" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --hf-split "${DATASET_SPLIT}" \
-  --num-prompts "${NUM_PROMPTS}" \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path AI-MO/aimo-validation-aime \
+    --num-prompts 10 \
+    --seed 42
+```
+
+### Running With Sampling Parameters
+
+When using OpenAI-compatible backends such as `vllm`, optional sampling
+parameters can be specified. Example client command:
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --top-k 10 \
+  --top-p 0.9 \
+  --temperature 0.5 \
+  --num-prompts 10
 ```
 
 ---
 ## Example - Offline Throughput Benchmark
 
 ```bash
-MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
-NUM_PROMPTS=10
-DATASET_NAME="sonnet"
-DATASET_PATH="vllm/benchmarks/sonnet.txt"
-
 python3 vllm/benchmarks/benchmark_throughput.py \
-  --model "${MODEL_NAME}" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --num-prompts "${NUM_PROMPTS}"
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset-name sonnet \
+  --dataset-path vllm/benchmarks/sonnet.txt \
+  --num-prompts 10
 ```
 
 If successful, you will see the following output
@@ -217,19 +244,13 @@ Total num output tokens:  1500
 ### VisionArena Benchmark for Vision Language Models
 
 ``` bash
-MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
-NUM_PROMPTS=10
-DATASET_NAME="hf"
-DATASET_PATH="lmarena-ai/vision-arena-bench-v0.1"
-DATASET_SPLIT="train"
-
 python3 vllm/benchmarks/benchmark_throughput.py \
-  --model "${MODEL_NAME}" \
-  --backend "vllm-chat" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --num-prompts "${NUM_PROMPTS}" \
-  --hf-split "${DATASET_SPLIT}"
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --num-prompts 1000 \
+  --hf-split train
 ```
 
 The `num prompt tokens` now includes image token counts
@@ -240,29 +261,83 @@ Total num prompt tokens:  14527
 Total num output tokens:  1280
 ```
 
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_USE_V1=1 \
+python3 vllm/benchmarks/benchmark_throughput.py \
+    --dataset-name=hf \
+    --dataset-path=likaixin/InstructCoder \
+    --model=meta-llama/Meta-Llama-3-8B-Instruct \
+    --input-len=1000 \
+    --output-len=100 \
+    --num-prompts=2048 \
+    --async-engine \
+    --speculative-model="[ngram]" \
+    --ngram_prompt_lookup_min=2 \
+    --ngram-prompt-lookup-max=5 \
+    --num_speculative_tokens=5
+```
+
+```
+Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
+Total num prompt tokens:  261136
+Total num output tokens:  204800
+```
+
+### Other HuggingFaceDataset Examples
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+```bash
+python3 benchmarks/benchmark_throughput.py \
+  --model Qwen/QwQ-32B \
+  --backend vllm \
+  --dataset-name hf \
+  --dataset-path AI-MO/aimo-validation-aime \
+  --hf-split train \
+  --num-prompts 10
+```
+
 ### Benchmark with LoRA Adapters
 
 ``` bash
 # download dataset
 # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-MODEL_NAME="meta-llama/Llama-2-7b-hf"
-BACKEND="vllm"
-DATASET_NAME="sharegpt"
-DATASET_PATH="<your data path>/ShareGPT_V3_unfiltered_cleaned_split.json"
-NUM_PROMPTS=10
-MAX_LORAS=2
-MAX_LORA_RANK=8
-ENABLE_LORA="--enable-lora"
-LORA_PATH="yard1/llama-2-7b-sql-lora-test"
-
 python3 vllm/benchmarks/benchmark_throughput.py \
-  --model "${MODEL_NAME}" \
-  --backend "${BACKEND}" \
-  --dataset_path "${DATASET_PATH}" \
-  --dataset_name "${DATASET_NAME}" \
-  --num-prompts "${NUM_PROMPTS}" \
-  --max-loras "${MAX_LORAS}" \
-  --max-lora-rank "${MAX_LORA_RANK}" \
-  ${ENABLE_LORA} \
-  --lora-path "${LORA_PATH}"
+  --model meta-llama/Llama-2-7b-hf \
+  --backend vllm \
+  --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --dataset_name sharegpt \
+  --num-prompts 10 \
+  --max-loras 2 \
+  --max-lora-rank 8 \
+  --enable-lora \
+  --lora-path yard1/llama-2-7b-sql-lora-test
   ```

benchmarks/backend_request_func.py

@@ -219,7 +219,15 @@ async def async_request_deepspeed_mii(
                 if response.status == 200:
                     parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
-                    output.generated_text = parsed_resp["text"][0]
+                    if "choices" in parsed_resp:
+                        output.generated_text = parsed_resp["choices"][0][
+                            "text"]
+                    elif "text" in parsed_resp:
+                        output.generated_text = parsed_resp["text"][0]
+                    else:
+                        output.error = ("Unexpected response format: "
+                                        "neither 'choices' nor 'text' found")
+                        output.success = False
                     output.success = True
                 else:
                     output.error = response.reason or ""
@@ -489,3 +497,9 @@ ASYNC_REQUEST_FUNCS = {
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
 }
+
+OPENAI_COMPATIBLE_BACKENDS = [
+    k for k, v in ASYNC_REQUEST_FUNCS.items()
+    if v in (async_request_openai_completions,
+             async_request_openai_chat_completions)
+]

benchmarks/benchmark_dataset.py

@@ -23,7 +23,8 @@ from abc import ABC, abstractmethod
 from collections.abc import Mapping
 from dataclasses import dataclass
 from functools import cache
-from typing import Any, Optional, Union
+from io import BytesIO
+from typing import Any, Callable, Optional, Union
 
 import numpy as np
 import pandas as pd
@@ -239,21 +240,24 @@ def process_image(image: Any) -> Mapping[str, Any]:
     """
     Process a single image input and return a multimedia content dictionary.
 
-    For a PIL.Image.Image input:
-      - Converts the image to RGB.
-      - Saves the image as a JPEG in-memory.
-      - Encodes the JPEG data as a base64 string.
-      - Returns a dictionary with the image as a base64 data URL.
+    Supports three input types:
 
-    For a string input:
-      - Treats the string as a URL or file path.
-      - Prepends "file://" if the string doesn't start with "http://" or
-        "file://".
-      - Returns a dictionary with the image URL.
+    1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key
+       containing raw image data.  - Loads the bytes as a PIL.Image.Image.
+
+    2. PIL.Image.Image input: - Converts the image to RGB.  - Saves the image as
+       a JPEG in memory.  - Encodes the JPEG data as a base64 string.  - Returns
+       a dictionary with the image as a base64 data URL.
+
+    3. String input: - Treats the string as a URL or local file path.  -
+       Prepends "file://" if the string doesn't start with "http://" or
+       "file://".  - Returns a dictionary with the image URL.
 
     Raises:
-      ValueError: If the input is neither a PIL.Image.Image nor a string.
+        ValueError: If the input is not a supported type.
     """
+    if isinstance(image, dict) and 'bytes' in image:
+        image = Image.open(BytesIO(image['bytes']))
     if isinstance(image, Image.Image):
         image = image.convert("RGB")
         with io.BytesIO() as image_data:
@@ -272,8 +276,8 @@ def process_image(image: Any) -> Mapping[str, Any]:
             ("http://", "file://")) else f"file://{image}")
         return {"type": "image_url", "image_url": {"url": image_url}}
 
-    raise ValueError(
-        f"Invalid image input {image}. Must be a PIL.Image.Image or str.")
+    raise ValueError(f"Invalid image input {image}. Must be a PIL.Image.Image"
+                     " or str or dictionary with raw image bytes.")
 
 
 # -----------------------------------------------------------------------------
@@ -284,7 +288,7 @@ def process_image(image: Any) -> Mapping[str, Any]:
 class RandomDataset(BenchmarkDataset):
     # Default values copied from benchmark_serving.py for the random dataset.
     DEFAULT_PREFIX_LEN = 0
-    DEFAULT_RANGE_RATIO = 1.0
+    DEFAULT_RANGE_RATIO = 0.0
     DEFAULT_INPUT_LEN = 1024
     DEFAULT_OUTPUT_LEN = 128
 
@@ -304,19 +308,32 @@ class RandomDataset(BenchmarkDataset):
         output_len: int = DEFAULT_OUTPUT_LEN,
         **kwargs,
     ) -> list[SampleRequest]:
+        # Enforce range_ratio < 1
+        assert range_ratio < 1.0, (
+            "random_range_ratio must be < 1.0 to ensure a valid sampling range"
+        )
+
         vocab_size = tokenizer.vocab_size
 
         prefix_token_ids = (np.random.randint(
             0, vocab_size, size=prefix_len).tolist() if prefix_len > 0 else [])
 
-        input_low = int(input_len * range_ratio)
-        output_low = int(output_len * range_ratio)
+        # New sampling logic: [X * (1 - b), X * (1 + b)]
+        input_low = int(input_len * (1 - range_ratio))
+        input_high = int(input_len * (1 + range_ratio))
+        output_low = int(output_len * (1 - range_ratio))
+        output_high = int(output_len * (1 + range_ratio))
+
+        # Add logging for debugging
+        logger.info("Sampling input_len from [%s, %s]", input_low, input_high)
+        logger.info("Sampling output_len from [%s, %s]", output_low,
+                    output_high)
 
         input_lens = np.random.randint(input_low,
-                                       input_len + 1,
+                                       input_high + 1,
                                        size=num_requests)
         output_lens = np.random.randint(output_low,
-                                        output_len + 1,
+                                        output_high + 1,
                                         size=num_requests)
         offsets = np.random.randint(0, vocab_size, size=num_requests)
 
@@ -468,11 +485,11 @@ class SonnetDataset(BenchmarkDataset):
 
         # Determine how many poem lines to use.
         num_input_lines = round((input_len - base_offset) / avg_len)
-        num_prefix_lines = round((prefix_len - base_offset) / avg_len)
+        num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0)
         prefix_lines = self.data[:num_prefix_lines]
 
         samples = []
-        for _ in range(num_requests):
+        while len(samples) < num_requests:
             extra_lines = random.choices(self.data,
                                          k=num_input_lines - num_prefix_lines)
             prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
@@ -480,13 +497,14 @@ class SonnetDataset(BenchmarkDataset):
             prompt_formatted = tokenizer.apply_chat_template(
                 msg, add_generation_prompt=True, tokenize=False)
             prompt_len = len(tokenizer(prompt_formatted).input_ids)
-            samples.append(
-                SampleRequest(
-                    prompt=prompt_formatted
-                    if return_prompt_formatted else prompt,
-                    prompt_len=prompt_len,
-                    expected_output_len=output_len,
-                ))
+            if prompt_len <= input_len:
+                samples.append(
+                    SampleRequest(
+                        prompt=prompt_formatted
+                        if return_prompt_formatted else prompt,
+                        prompt_len=prompt_len,
+                        expected_output_len=output_len,
+                    ))
         return samples
 
 
@@ -562,48 +580,47 @@ class BurstGPTDataset(BenchmarkDataset):
 
 
 # -----------------------------------------------------------------------------
-# HuggingFace Dataset Implementation
+# HuggingFace Dataset Base Implementation
 # -----------------------------------------------------------------------------
-
-
 class HuggingFaceDataset(BenchmarkDataset):
-    """
-    Dataset class for processing a HuggingFace dataset with conversation data
-    and optional images.
-    """
+    """Base class for datasets hosted on HuggingFace."""
+
+    SUPPORTED_DATASET_PATHS: Union[set[str], dict[str, Callable]] = set()
 
     def __init__(
         self,
+        dataset_path: str,
         dataset_split: str,
         dataset_subset: Optional[str] = None,
         **kwargs,
     ) -> None:
-        super().__init__(**kwargs)
+        super().__init__(dataset_path=dataset_path, **kwargs)
+
         self.dataset_split = dataset_split
         self.dataset_subset = dataset_subset
-
         self.load_data()
 
     def load_data(self) -> None:
-        if not self.dataset_path:
-            raise ValueError("dataset_path must be provided for loading data.")
-
+        """Load data from HuggingFace datasets."""
         self.data = load_dataset(
             self.dataset_path,
             name=self.dataset_subset,
             split=self.dataset_split,
             streaming=True,
         )
-        if self.data.features is None or "conversations" \
-            not in self.data.features:
-            raise ValueError(
-                "HuggingFaceDataset currently only supports datasets with "
-                "a 'conversations' column like lmms-lab/LLaVA-OneVision-Data. "
-                "Please consider contributing if you would like to add "
-                "support for additional dataset formats.")
-        # Shuffle and filter examples with at least 2 conversations.
-        self.data = self.data.shuffle(seed=self.random_seed).filter(
-            lambda x: len(x["conversations"]) >= 2)
+        self.data = self.data.shuffle(seed=self.random_seed)
+
+
+# -----------------------------------------------------------------------------
+# Conversation Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class ConversationDataset(HuggingFaceDataset):
+    """Dataset for conversation data with multimodal support."""
+    SUPPORTED_DATASET_PATHS = {
+        'lmms-lab/LLaVA-OneVision-Data', 'Aeala/ShareGPT_Vicuna_unfiltered'
+    }
 
     def sample(self,
                tokenizer: PreTrainedTokenizerBase,
@@ -611,10 +628,13 @@ class HuggingFaceDataset(BenchmarkDataset):
                output_len: Optional[int] = None,
                enable_multimodal_chat: bool = False,
                **kwargs) -> list:
+        # Filter examples with at least 2 conversations
+        filtered_data = self.data.filter(
+            lambda x: len(x["conversations"]) >= 2)
         sampled_requests = []
         dynamic_output = output_len is None
 
-        for item in self.data:
+        for item in filtered_data:
             if len(sampled_requests) >= num_requests:
                 break
             conv = item["conversations"]
@@ -659,29 +679,12 @@ class VisionArenaDataset(HuggingFaceDataset):
     """
 
     DEFAULT_OUTPUT_LEN = 128
-    VISION_ARENA_DATASET_PATH = "lmarena-ai/vision-arena-bench-v0.1"
-
-    def __init__(
-        self,
-        **kwargs,
-    ) -> None:
-        super().__init__(**kwargs)
-        if self.dataset_path != self.VISION_ARENA_DATASET_PATH:
-            raise ValueError(f"Only support Vision Arena dataset.\
-                    This data path {self.dataset_path} is not valid.")
-        if self.dataset_subset is None and self.dataset_split != "train":
-            raise ValueError("Dataset split must be 'train'.")
-
-        self.load_data()
-
-    def load_data(self) -> None:
-        dataset = load_dataset(
-            self.dataset_path,
-            name=self.dataset_subset,
-            split=self.dataset_split,
-            streaming=True,
-        )
-        self.data = dataset.shuffle(seed=self.random_seed)
+    SUPPORTED_DATASET_PATHS = {
+        "lmarena-ai/VisionArena-Chat":
+        lambda x: x["conversation"][0][0]["content"],
+        "lmarena-ai/vision-arena-bench-v0.1":
+        lambda x: x["turns"][0][0]["content"]
+    }
 
     def sample(
         self,
@@ -697,7 +700,11 @@ class VisionArenaDataset(HuggingFaceDataset):
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
-            prompt = item["turns"][0][0]["content"]
+            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path)
+            if parser_fn is None:
+                raise ValueError(
+                    f"Unsupported dataset path: {self.dataset_path}")
+            prompt = parser_fn(item)
             mm_content = process_image(item["images"][0])
             prompt_len = len(tokenizer(prompt).input_ids)
             if enable_multimodal_chat:
@@ -715,3 +722,96 @@ class VisionArenaDataset(HuggingFaceDataset):
                 ))
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# Instruct Coder Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class InstructCoderDataset(HuggingFaceDataset):
+    """
+    InstructCoder Dataset.
+    https://huggingface.co/datasets/likaixin/InstructCoder
+
+    InstructCoder is the dataset designed for general code editing.  It consists
+    of 114,239 instruction-input-output triplets, and covers multiple distinct
+    code editing scenario.
+    """
+
+    DEFAULT_OUTPUT_LEN = 200  # this is the average default output length
+    SUPPORTED_DATASET_PATHS = {
+        "likaixin/InstructCoder",
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               enable_multimodal_chat: bool = False,
+               **kwargs) -> list:
+        output_len = (output_len
+                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = f"{item['instruction']}:\n{item['input']}"
+            prompt_len = len(tokenizer(prompt).input_ids)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# AIMO Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class AIMODataset(HuggingFaceDataset):
+    """
+    Dataset class for processing a AIMO dataset with reasoning questions.
+    """
+    SUPPORTED_DATASET_PATHS = {
+        "AI-MO/aimo-validation-aime", "AI-MO/NuminaMath-1.5",
+        "AI-MO/NuminaMath-CoT"
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               **kwargs) -> list:
+        sampled_requests = []
+        dynamic_output = output_len is None
+
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt, completion = item['problem'], item["solution"]
+
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            completion_len = len(completion_ids)
+            output_len = completion_len if dynamic_output else output_len
+            assert isinstance(output_len, int) and output_len > 0
+            if dynamic_output and not is_valid_sequence(prompt_len,
+                                                        completion_len,
+                                                        max_prompt_len=2048,
+                                                        max_total_len=32000):
+                continue
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=None,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests

benchmarks/benchmark_serving.py

@@ -7,9 +7,6 @@ On the server side, run one of the following commands:
         --swap-space 16 \
         --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
     python benchmarks/benchmark_serving.py \
         --backend <backend> \
@@ -37,7 +34,8 @@ from datetime import datetime
 from typing import Any, Optional
 
 import numpy as np
-from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
+from backend_request_func import (ASYNC_REQUEST_FUNCS,
+                                  OPENAI_COMPATIBLE_BACKENDS, RequestFuncInput,
                                   RequestFuncOutput)
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
@@ -52,9 +50,11 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
-from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
-                               RandomDataset, SampleRequest, ShareGPTDataset,
-                               SonnetDataset, VisionArenaDataset)
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, HuggingFaceDataset,
+                               InstructCoderDataset, RandomDataset,
+                               SampleRequest, ShareGPTDataset, SonnetDataset,
+                               VisionArenaDataset)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
@@ -156,7 +156,7 @@ def calculate_metrics(
         if outputs[i].success:
             output_len = outputs[i].output_tokens
 
-            if output_len is None:
+            if not output_len:
                 # We use the tokenizer to count the number of output tokens
                 # for some serving backends instead of looking at
                 # len(outputs[i].itl) since multiple output tokens may be
@@ -261,6 +261,7 @@ async def benchmark(
     goodput_config_dict: dict[str, float],
     max_concurrency: Optional[int],
     lora_modules: Optional[Iterable[str]],
+    extra_body: Optional[dict],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -288,6 +289,7 @@ async def benchmark(
         logprobs=logprobs,
         multi_modal_content=test_mm_content,
         ignore_eos=ignore_eos,
+        extra_body=extra_body,
     )
 
     test_output = await request_func(request_func_input=test_input)
@@ -314,7 +316,8 @@ async def benchmark(
                                          output_len=test_output_len,
                                          logprobs=logprobs,
                                          multi_modal_content=test_mm_content,
-                                         ignore_eos=ignore_eos)
+                                         ignore_eos=ignore_eos,
+                                         extra_body=extra_body)
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
             print("Profiler started")
@@ -364,7 +367,8 @@ async def benchmark(
                                               output_len=output_len,
                                               logprobs=logprobs,
                                               multi_modal_content=mm_content,
-                                              ignore_eos=ignore_eos)
+                                              ignore_eos=ignore_eos,
+                                              extra_body=extra_body)
         tasks.append(
             asyncio.create_task(
                 limited_request_func(request_func_input=request_func_input,
@@ -586,19 +590,39 @@ def main(args: argparse.Namespace):
                                             return_prompt_formatted=True)
 
     elif args.dataset_name == "hf":
-        # Choose between VisionArenaDataset
-        # and HuggingFaceDataset based on provided parameters.
-        dataset_class = (VisionArenaDataset if args.dataset_path
-                         == VisionArenaDataset.VISION_ARENA_DATASET_PATH
-                         and args.hf_subset is None else HuggingFaceDataset)
+        # all following datasets are implemented from the
+        # HuggingFaceDataset base class
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = VisionArenaDataset
+            args.hf_split = "train"
+            args.hf_subset = None
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = InstructCoderDataset
+            args.hf_split = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = ConversationDataset
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = AIMODataset
+            args.hf_split = "train"
+        else:
+            supported_datasets = set([
+                dataset_name for cls in HuggingFaceDataset.__subclasses__()
+                for dataset_name in cls.SUPPORTED_DATASET_PATHS
+            ])
+            raise ValueError(
+                f"Unsupported dataset path: {args.dataset_path}. "
+                "Huggingface dataset only supports dataset_path"
+                f" from one of following: {supported_datasets}. "
+                "Please consider contributing if you would "
+                "like to add support for additional dataset formats.")
         input_requests = dataset_class(
             dataset_path=args.dataset_path,
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
+            random_seed=args.seed,
         ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
-            random_seed=args.seed,
             output_len=args.hf_output_len,
         )
 
@@ -633,6 +657,26 @@ def main(args: argparse.Namespace):
             raise ValueError(f"Unknown dataset: {args.dataset_name}") from err
     goodput_config_dict = check_goodput_args(args)
 
+    # Collect the sampling parameters.
+    sampling_params = {
+        k: v
+        for k, v in {
+            "top_p": args.top_p,
+            "top_k": args.top_k,
+            "min_p": args.min_p,
+            "temperature": args.temperature
+        }.items() if v is not None
+    }
+
+    # Sampling parameters are only supported by openai-compatible backend.
+    if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS:
+        raise ValueError(
+            "Sampling parameters are only supported by openai-compatible "
+            "backends.")
+
+    if "temperature" not in sampling_params:
+        sampling_params["temperature"] = 0.0  # Default to greedy decoding.
+
     # Avoid GC processing "static" data - reduce pause times.
     gc.collect()
     gc.freeze()
@@ -659,6 +703,7 @@ def main(args: argparse.Namespace):
             goodput_config_dict=goodput_config_dict,
             max_concurrency=args.max_concurrency,
             lora_modules=args.lora_modules,
+            extra_body=sampling_params,
         ))
 
     # Save config and results to json
@@ -876,7 +921,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -884,7 +929,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -951,18 +996,23 @@ if __name__ == "__main__":
     random_group.add_argument(
         "--random-range-ratio",
         type=float,
-        default=1.0,
-        help="Range of sampled ratio of input/output length, "
-        "used only for random sampling.",
+        default=0.0,
+        help="Range ratio for sampling input/output length, "
+        "used only for random sampling. Must be in the range [0, 1) to define "
+        "a symmetric sampling range"
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
     )
     random_group.add_argument(
         "--random-prefix-len",
         type=int,
         default=0,
-        help="Number of fixed prefix tokens before random "
-        " context. The length range of context in a random "
-        " request is [random-prefix-len, "
-        " random-prefix-len + random-prefix-len * random-range-ratio).")
+        help=("Number of fixed prefix tokens before the random context "
+              "in a request. "
+              "The total input length is the sum of `random-prefix-len` and "
+              "a random "
+              "context length sampled from [input_len * (1 - range_ratio), "
+              "input_len * (1 + range_ratio)]."),
+    )
 
     hf_group = parser.add_argument_group("hf dataset options")
     hf_group.add_argument("--hf-subset",
@@ -981,6 +1031,33 @@ if __name__ == "__main__":
         "from the sampled HF dataset.",
     )
 
+    sampling_group = parser.add_argument_group("sampling parameters")
+    sampling_group.add_argument(
+        "--top-p",
+        type=float,
+        default=None,
+        help="Top-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--top-k",
+        type=int,
+        default=None,
+        help="Top-k sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--min-p",
+        type=float,
+        default=None,
+        help="Min-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--temperature",
+        type=float,
+        default=None,
+        help="Temperature sampling parameter. Only has effect on "
+        "openai-compatible backends. If not specified, default to greedy "
+        "decoding (i.e. temperature==0.0).")
+
     parser.add_argument(
         '--tokenizer-mode',
         type=str,

benchmarks/benchmark_serving_structured_output.py

@@ -5,16 +5,13 @@ On the server side, run one of the following commands:
     (vLLM OpenAI API server)
     vllm serve <your_model> --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
     python benchmarks/benchmark_serving_structured_output.py \
         --backend <backend> \
         --model <your_model> \
         --dataset json \
         --structured-output-ratio 1.0 \
-        --structured-output-backend xgrammar \
+        --structured-output-backend auto \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -133,10 +130,11 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
                         "description":
                         "An unique optional field to avoid cached schemas"
                     }
+        else:
+            json_schemas = [schema] * args.num_prompts
 
         def gen_prompt(index: int):
-            schema = json_schemas[index % len(json_schemas)]
-            return f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
+            return f"Generate an example of a user profile given the following schema: {json.dumps(get_schema(index))}"  # noqa: E501
 
         def get_schema(index: int):
             return json_schemas[index % len(json_schemas)]
@@ -966,7 +964,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -974,7 +972,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -999,12 +997,14 @@ if __name__ == "__main__":
                         type=float,
                         default=1.0,
                         help="Ratio of Structured Outputs requests")
-    parser.add_argument(
-        "--structured-output-backend",
-        type=str,
-        choices=["outlines", "lm-format-enforcer", "xgrammar", "guidance"],
-        default="xgrammar",
-        help="Backend to use for structured outputs")
+    parser.add_argument("--structured-output-backend",
+                        type=str,
+                        choices=[
+                            "outlines", "lm-format-enforcer", "xgrammar",
+                            "guidance", "auto"
+                        ],
+                        default="auto",
+                        help="Backend to use for structured outputs")
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -11,7 +11,8 @@ from typing import Any, Optional, Union
 
 import torch
 import uvloop
-from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, InstructCoderDataset,
                                RandomDataset, SampleRequest, ShareGPTDataset,
                                SonnetDataset, VisionArenaDataset)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
@@ -212,14 +213,17 @@ def run_hf(
     max_prompt_len = 0
     max_output_len = 0
     for i in range(len(requests)):
-        prompt, prompt_len, output_len = requests[i]
+        prompt = requests[i].prompt
+        prompt_len = requests[i].prompt_len
+        output_len = requests[i].expected_output_len
         # Add the prompt to the batch.
         batch.append(prompt)
         max_prompt_len = max(max_prompt_len, prompt_len)
         max_output_len = max(max_output_len, output_len)
         if len(batch) < max_batch_size and i != len(requests) - 1:
             # Check if we can add more requests to the batch.
-            _, next_prompt_len, next_output_len = requests[i + 1]
+            next_prompt_len = requests[i + 1].prompt_len
+            next_output_len = requests[i + 1].expected_output_len
             if (max(max_prompt_len, next_prompt_len) +
                     max(max_output_len, next_output_len)) <= 2048:
                 # We can add more requests to the batch.
@@ -300,6 +304,7 @@ def get_requests(args, tokenizer):
         "input_len": args.input_len,
         "output_len": args.output_len,
     }
+
     if args.dataset_path is None or args.dataset_name == "random":
         sample_kwargs["range_ratio"] = args.random_range_ratio
         sample_kwargs["prefix_len"] = args.prefix_len
@@ -317,18 +322,23 @@ def get_requests(args, tokenizer):
     elif args.dataset_name == "burstgpt":
         dataset_cls = BurstGPTDataset
     elif args.dataset_name == "hf":
-        if args.backend != "vllm-chat":
-            raise ValueError(
-                "hf datasets only are supported by vllm-chat backend")
-        # Choose between VisionArenaDataset and HuggingFaceDataset based on
-        # provided parameters.
-        dataset_cls = (VisionArenaDataset if args.dataset_path
-                       == VisionArenaDataset.VISION_ARENA_DATASET_PATH
-                       and args.hf_subset is None else HuggingFaceDataset)
-        common_kwargs['dataset_subset'] = args.hf_subset
-        common_kwargs['dataset_split'] = args.hf_split
-        sample_kwargs["enable_multimodal_chat"] = True
-
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = VisionArenaDataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = InstructCoderDataset
+            common_kwargs['dataset_split'] = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = ConversationDataset
+            common_kwargs['dataset_subset'] = args.hf_subset
+            common_kwargs['dataset_split'] = args.hf_split
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = AIMODataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
     else:
         raise ValueError(f"Unknown dataset name: {args.dataset_name}")
     # Remove None values
@@ -462,9 +472,17 @@ def validate_args(args):
         warnings.warn("--hf-subset and --hf-split will be ignored \
                 since --dataset-name is not 'hf'.",
                       stacklevel=2)
-    elif args.dataset_name == "hf" and args.backend != "vllm-chat":
-        raise ValueError(
-            "When --dataset-name is 'hf', backend must be 'vllm-chat'")
+    elif args.dataset_name == "hf":
+        if args.dataset_path in (
+                VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
+                | ConversationDataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm-chat", f"{args.dataset_path} needs to use vllm-chat as the backend."  #noqa: E501
+        elif args.dataset_path in (InstructCoderDataset.SUPPORTED_DATASET_PATHS
+                                   | AIMODataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm", f"{args.dataset_path} needs to use vllm as the backend."  #noqa: E501
+        else:
+            raise ValueError(
+                f"{args.dataset_path} is not supported by hf dataset.")
 
     # --random-range-ratio: only used when dataset_name is 'random'
     if args.dataset_name != 'random' and args.random_range_ratio is not None:
@@ -576,18 +594,30 @@ if __name__ == "__main__":
         default=None,
         help="Path to the lora adapters to use. This can be an absolute path, "
         "a relative path, or a Hugging Face model identifier.")
-    parser.add_argument("--prefix-len",
-                        type=int,
-                        default=None,
-                        help="Number of prefix tokens per request."
-                        "This is for the RandomDataset and SonnetDataset")
+    parser.add_argument(
+        "--prefix-len",
+        type=int,
+        default=None,
+        help=f"Number of prefix tokens to be used in RandomDataset "
+        "and SonnetDataset. For RandomDataset, the total input "
+        "length is the sum of prefix-len (default: "
+        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
+        "sampled from [input_len * (1 - range_ratio), "
+        "input_len * (1 + range_ratio)]. For SonnetDataset, "
+        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
+        "controls how much of the input is fixed lines versus "
+        "random lines, but the total input length remains approximately "
+        "input_len tokens.")
     # random dataset
     parser.add_argument(
         "--random-range-ratio",
         type=float,
         default=None,
-        help="Range of sampled ratio of input/output length, "
-        "used only for RandomDataSet.",
+        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
+        "for sampling input/output length, "
+        "used only for RandomDataset. Must be in the range [0, 1) to "
+        "define a symmetric sampling range "
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
     )
 
     # hf dtaset

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -0,0 +1,340 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import torch.utils.benchmark as benchmark
+from benchmark_shapes import WEIGHT_SHAPES_MOE
+
+from vllm import _custom_ops as ops
+from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.fused_moe import (cutlass_moe_fp8,
+                                                            fused_experts,
+                                                            fused_topk)
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = [
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1", "nm-testing/deepseekv2-lite",
+    "ibm-granite/granite-3.0-1b-a400m", "ibm-granite/granite-3.0-3b-a800m"
+]
+DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+PER_ACT_TOKEN_OPTS = [False]
+PER_OUT_CH_OPTS = [False]
+
+
+def to_fp8(tensor: torch.Tensor):
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def bench_run(results: list[benchmark.Measurement], model: str,
+              num_experts: int, topk: int, per_act_token: bool,
+              per_out_ch: bool, mkn: tuple[int, int, int]):
+    label = "Quant Matmul"
+
+    sub_label = (
+        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, "
+        "MKN=({})".format(model, num_experts, topk, per_act_token, per_out_ch,
+                          mkn))
+
+    print(f"Testing: {sub_label}")
+
+    (m, k, n) = mkn
+
+    dtype = torch.half
+
+    a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
+    w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10
+    w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10
+
+    _, a_scale = ops.scaled_fp8_quant(a)
+
+    w1_q = torch.empty((num_experts, 2 * n, k),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w2_q = torch.empty((num_experts, k, n),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w1_scale = torch.empty((num_experts, 1, 1),
+                           device="cuda",
+                           dtype=torch.float32)
+    w2_scale = torch.empty((num_experts, 1, 1),
+                           device="cuda",
+                           dtype=torch.float32)
+
+    ab_strides1 = torch.full((num_experts, ),
+                             k,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides1 = torch.full((num_experts, ),
+                            2 * n,
+                            device="cuda",
+                            dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts, ),
+                             n,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides2 = torch.full((num_experts, ),
+                            k,
+                            device="cuda",
+                            dtype=torch.int64)
+
+    for expert in range(num_experts):
+        w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert])
+        w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert])
+    w1_q_notransp = w1_q.clone()
+    w2_q_notransp = w2_q.clone()
+    w1_q = w1_q.transpose(1, 2)
+    w2_q = w2_q.transpose(1, 2)
+
+    score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
+
+    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+
+    def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
+                       topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                       w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                       a_scale: torch.Tensor, num_repeats: int):
+        for _ in range(num_repeats):
+            fused_experts(a,
+                          w1,
+                          w2,
+                          topk_weights,
+                          topk_ids,
+                          use_fp8_w8a8=True,
+                          w1_scale=w1_scale,
+                          w2_scale=w2_scale,
+                          a1_scale=a_scale)
+
+    def run_cutlass_moe(a: torch.Tensor, a_scale: torch.Tensor,
+                        w1: torch.Tensor, w2: torch.Tensor,
+                        w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                        topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                        ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
+                        ab_strides2: torch.Tensor, c_strides2: torch.Tensor,
+                        num_repeats: int):
+        for _ in range(num_repeats):
+            cutlass_moe_fp8(a,
+                            w1,
+                            w2,
+                            w1_scale,
+                            w2_scale,
+                            topk_weights,
+                            topk_ids,
+                            ab_strides1,
+                            c_strides1,
+                            ab_strides2,
+                            c_strides2,
+                            a1_scale=a_scale)
+
+    def run_cutlass_from_graph(
+            a: torch.Tensor, a_scale: torch.Tensor, w1_q: torch.Tensor,
+            w2_q: torch.Tensor, w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+            topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+            ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
+            ab_strides2: torch.Tensor, c_strides2: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return cutlass_moe_fp8(a,
+                                   w1_q,
+                                   w2_q,
+                                   w1_scale,
+                                   w2_scale,
+                                   topk_weights,
+                                   topk_ids,
+                                   ab_strides1,
+                                   c_strides1,
+                                   ab_strides2,
+                                   c_strides2,
+                                   a1_scale=a_scale)
+
+    def run_triton_from_graph(a: torch.Tensor, w1: torch.Tensor,
+                              w2: torch.Tensor, topk_weights: torch.Tensor,
+                              topk_ids: torch.Tensor, w1_scale: torch.Tensor,
+                              w2_scale: torch.Tensor, a_scale: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return fused_experts(a,
+                                 w1,
+                                 w2,
+                                 topk_weights,
+                                 topk_ids,
+                                 use_fp8_w8a8=True,
+                                 w1_scale=w1_scale,
+                                 w2_scale=w2_scale,
+                                 a1_scale=a_scale)
+
+    def replay_graph(graph, num_repeats):
+        for _ in range(num_repeats):
+            graph.replay()
+        torch.cuda.synchronize()
+
+    cutlass_stream = torch.cuda.Stream()
+    cutlass_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
+        run_cutlass_from_graph(a, a_scale, w1_q, w2_q, w1_scale, w2_scale,
+                               topk_weights, topk_ids, ab_strides1, c_strides1,
+                               ab_strides2, c_strides2)
+    torch.cuda.synchronize()
+
+    triton_stream = torch.cuda.Stream()
+    triton_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(triton_graph, stream=triton_stream):
+        run_triton_from_graph(a, w1_q_notransp, w2_q_notransp, topk_weights,
+                              topk_ids, w1_scale, w2_scale, a_scale)
+    torch.cuda.synchronize()
+
+    min_run_time = 5
+    num_warmup = 5
+    num_runs = 25
+
+    globals = {
+        # Baseline params
+        "w1": w1,
+        "w2": w2,
+        "score": score,
+        "topk": topk,
+        "w1_q_notransp": w1_q_notransp,
+        "w2_q_notransp": w2_q_notransp,
+        # Cutlass params
+        "a_scale": a_scale,
+        "w1_q": w1_q,
+        "w2_q": w2_q,
+        "w1_scale": w1_scale,
+        "w2_scale": w2_scale,
+        "ab_strides1": ab_strides1,
+        "c_strides1": c_strides1,
+        "ab_strides2": ab_strides2,
+        "c_strides2": c_strides2,
+        # cuda graph params
+        "cutlass_graph": cutlass_graph,
+        "triton_graph": triton_graph,
+        # Gen params
+        "a": a,
+        "topk_weights": topk_weights,
+        "topk_ids": topk_ids,
+        "num_runs": num_runs,
+        # Kernels
+        "run_triton_moe": run_triton_moe,
+        "run_cutlass_moe": run_cutlass_moe,
+        "replay_graph": replay_graph,
+    }
+
+    # Warmup
+    run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids,
+                   w1_scale, w2_scale, a_scale, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(triton_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(triton_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights,
+                    topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2,
+                    num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(cutlass_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(cutlass_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+
+def main(args):
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    results: list[benchmark.Measurement] = []
+
+    for model in args.models:
+        for tp in args.tp_sizes:
+            for layer in WEIGHT_SHAPES_MOE[model]:
+                num_experts = layer[0]
+                topk = layer[1]
+                size_k = layer[2]
+                size_n = layer[3] // tp
+
+                if len(args.limit_k) > 0 and size_k not in args.limit_k:
+                    continue
+
+                if len(args.limit_n) > 0 and size_n not in args.limit_n:
+                    continue
+
+                for per_act_token in PER_ACT_TOKEN_OPTS:
+                    for per_out_ch in PER_OUT_CH_OPTS:
+                        for size_m in DEFAULT_BATCH_SIZES:
+                            mkn = (size_m, size_k, size_n)
+                            bench_run(results, model, num_experts, topk,
+                                      per_act_token, per_out_ch, mkn)
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark Marlin across specified models/shapes/batches")
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES_MOE.keys(),
+    )
+    parser.add_argument("--tp-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_TP_SIZES)
+    parser.add_argument("--batch-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-per-act-token",
+                        nargs="+",
+                        type=int,
+                        default=[])
+    parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_moe.py

@@ -30,19 +30,18 @@ class BenchmarkConfig(TypedDict):
     num_stages: int
 
 
-def benchmark_config(
-    config: BenchmarkConfig,
-    num_tokens: int,
-    num_experts: int,
-    shard_intermediate_size: int,
-    hidden_size: int,
-    topk: int,
-    dtype: torch.dtype,
-    use_fp8_w8a8: bool,
-    use_int8_w8a16: bool,
-    num_iters: int = 100,
-    block_quant_shape: List[int] = None,
-) -> float:
+def benchmark_config(config: BenchmarkConfig,
+                     num_tokens: int,
+                     num_experts: int,
+                     shard_intermediate_size: int,
+                     hidden_size: int,
+                     topk: int,
+                     dtype: torch.dtype,
+                     use_fp8_w8a8: bool,
+                     use_int8_w8a16: bool,
+                     num_iters: int = 100,
+                     block_quant_shape: List[int] = None,
+                     use_deep_gemm: bool = False) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
     if use_int8_w8a16:
@@ -115,22 +114,41 @@ def benchmark_config(
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
         with override_config(config):
-            fused_moe(
-                x,
-                w1,
-                w2,
-                input_gating,
-                topk,
-                renormalize=True,
-                inplace=True,
-                use_fp8_w8a8=use_fp8_w8a8,
-                use_int8_w8a16=use_int8_w8a16,
-                w1_scale=w1_scale,
-                w2_scale=w2_scale,
-                a1_scale=a1_scale,
-                a2_scale=a2_scale,
-                block_shape=block_quant_shape,
-            )
+            if use_deep_gemm:
+                topk_weights, topk_ids = fused_topk(x, input_gating, topk,
+                                                    False)
+                return fused_experts(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                    allow_deep_gemm=True,
+                )
+            else:
+                fused_moe(
+                    x,
+                    w1,
+                    w2,
+                    input_gating,
+                    topk,
+                    renormalize=True,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    use_int8_w8a16=use_int8_w8a16,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                )
 
     # JIT compilation & warmup
     run()
@@ -366,6 +384,7 @@ class BenchmarkWorker:
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
         block_quant_shape: List[int] = None,
+        use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
@@ -396,7 +415,8 @@ class BenchmarkWorker:
                                        use_fp8_w8a8,
                                        use_int8_w8a16,
                                        num_iters=100,
-                                       block_quant_shape=block_quant_shape)
+                                       block_quant_shape=block_quant_shape,
+                                       use_deep_gemm=use_deep_gemm)
         return config, kernel_time
 
     def tune(
@@ -411,6 +431,7 @@ class BenchmarkWorker:
         use_int8_w8a16: bool,
         search_space: list[dict[str, int]],
         block_quant_shape: list[int],
+        use_deep_gemm: bool,
     ) -> dict[str, int]:
         best_config = None
         best_time = float("inf")
@@ -436,7 +457,8 @@ class BenchmarkWorker:
                         use_fp8_w8a8,
                         use_int8_w8a16,
                         num_iters=20,
-                        block_quant_shape=block_quant_shape)
+                        block_quant_shape=block_quant_shape,
+                        use_deep_gemm=use_deep_gemm)
                 except triton.runtime.autotuner.OutOfResources:
                     # Some configurations may be invalid and fail to compile.
                     continue
@@ -531,6 +553,9 @@ def main(args: argparse.Namespace):
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
+        if not hasattr(config, "hidden_size"):
+            # Support for llama4
+            config = config.text_config
         # Default: Mixtral.
         E = config.num_local_experts
         topk = config.num_experts_per_tok
@@ -550,6 +575,8 @@ def main(args: argparse.Namespace):
     else:
         batch_sizes = [args.batch_size]
 
+    use_deep_gemm = bool(args.use_deep_gemm)
+
     ray.init()
     num_gpus = int(ray.available_resources()["GPU"])
     workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
@@ -572,10 +599,10 @@ def main(args: argparse.Namespace):
 
         start = time.time()
         configs = _distribute(
-            "tune",
-            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, search_space, block_quant_shape)
-             for batch_size in batch_sizes])
+            "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
+                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space,
+                      block_quant_shape, use_deep_gemm)
+                     for batch_size in batch_sizes])
         best_configs = {
             M: sort_config(config)
             for M, config in zip(batch_sizes, configs)
@@ -589,7 +616,7 @@ def main(args: argparse.Namespace):
         outputs = _distribute(
             "benchmark",
             [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, block_quant_shape)
+              use_fp8_w8a8, use_int8_w8a16, block_quant_shape, use_deep_gemm)
              for batch_size in batch_sizes])
 
         for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
@@ -611,6 +638,7 @@ if __name__ == "__main__":
                         type=str,
                         choices=["auto", "fp8_w8a8", "int8_w8a16"],
                         default="auto")
+    parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)
     parser.add_argument("--tune", action="store_true")

benchmarks/kernels/benchmark_paged_attention.py

@@ -7,10 +7,13 @@ from typing import Optional
 import torch
 
 from vllm import _custom_ops as ops
+from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
                         create_kv_caches_with_random)
 
+logger = init_logger(__name__)
+
 NUM_BLOCKS = 128 * 1024
 PARTITION_SIZE = 512
 PARTITION_SIZE_ROCM = 256
@@ -193,6 +196,9 @@ def main(
 
 
 if __name__ == '__main__':
+    logger.warning("This script benchmarks the paged attention kernel. "
+                   "By default this is no longer used in vLLM inference.")
+
     parser = FlexibleArgumentParser(
         description="Benchmark the paged attention kernel.")
     parser.add_argument("--version",

benchmarks/kernels/benchmark_shapes.py

@@ -75,3 +75,19 @@ WEIGHT_SHAPES = {
         [7168, 8192],
     ],
 }
+
+WEIGHT_SHAPES_MOE = {
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
+        [8, 2, 4096, 28672],
+        [8, 2, 14336, 4096],
+    ],
+    "nm-testing/deepseekv2-lite": [
+        [64, 6, 2048, 1408],
+    ],
+    "ibm-granite/granite-3.0-1b-a400m": [
+        [32, 8, 1024, 1024],
+    ],
+    "ibm-granite/granite-3.0-3b-a800m": [
+        [40, 8, 1024, 1536],
+    ],
+}

benchmarks/launch_tgi_server.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-PORT=8000
-MODEL=$1
-TOKENS=$2
-
-docker run -e "HF_TOKEN=$HF_TOKEN" --gpus all --shm-size 1g -p $PORT:80 \
-           -v "$PWD/data:/data" \
-           ghcr.io/huggingface/text-generation-inference:2.2.0 \
-           --model-id "$MODEL" \
-           --sharded false  \
-           --max-input-length 1024 \
-           --max-total-tokens 2048 \
-           --max-best-of 5 \
-           --max-concurrent-requests 5000 \
-           --max-batch-total-tokens "$TOKENS"

cmake/cpu_extension.cmake

@@ -33,8 +33,6 @@ endif()
 
 if(MACOSX_FOUND)
     list(APPEND CXX_COMPILE_FLAGS
-        "-Xpreprocessor"
-        "-fopenmp"
         "-DVLLM_CPU_EXTENSION")
 else()
     list(APPEND CXX_COMPILE_FLAGS
@@ -190,12 +188,14 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cache.cpp"
     "csrc/cpu/utils.cpp"
     "csrc/cpu/layernorm.cpp"
+    "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
     "csrc/cpu/torch_bindings.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
         "csrc/cpu/quant.cpp"
+        "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
 endif()
 

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG dc9d410b3e2d6534a4c70724c2515f4def670a22
+          GIT_TAG 0a721daebe4fa7149f06ecf3d3eabeb6dcd0f1fa
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

csrc/attention/merge_attn_states.cu

@@ -0,0 +1,178 @@
+#include <optional>
+#include <torch/all.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <algorithm>
+
+#include "attention_dtypes.h"
+#include "attention_utils.cuh"
+
+namespace vllm {
+
+// Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+// can be used to combine partial attention results (in the split-KV case)
+template <typename scalar_t, const uint NUM_THREADS>
+__global__ void merge_attn_states_kernel(
+    scalar_t* output, float* output_lse, const scalar_t* prefix_output,
+    const float* prefix_lse, const scalar_t* suffix_output,
+    const float* suffix_lse, const uint num_tokens, const uint num_heads,
+    const uint head_size) {
+  using pack_128b_t = uint4;
+  const uint pack_size = 16 / sizeof(scalar_t);
+  const uint threads_per_head = head_size / pack_size;
+
+  const uint global_idx = blockIdx.x * NUM_THREADS + threadIdx.x;
+  const uint token_head_threads = num_tokens * num_heads * threads_per_head;
+
+  if (global_idx >= token_head_threads) return;
+
+  // global_idx -> token_idx + head_idx + pack_idx
+  const uint token_head_idx = global_idx / threads_per_head;
+  const uint pack_idx = global_idx % threads_per_head;
+
+  const uint token_idx = token_head_idx / num_heads;
+  const uint head_idx = token_head_idx % num_heads;
+
+  const uint pack_offset = pack_idx * pack_size;  // (0~15)*8, etc.
+  const uint head_offset =
+      token_idx * num_heads * head_size + head_idx * head_size;
+  const scalar_t* prefix_head_ptr = prefix_output + head_offset;
+  const scalar_t* suffix_head_ptr = suffix_output + head_offset;
+  scalar_t* output_head_ptr = output + head_offset;
+
+  float p_lse = prefix_lse[head_idx * num_tokens + token_idx];
+  float s_lse = suffix_lse[head_idx * num_tokens + token_idx];
+  p_lse = std::isinf(p_lse) ? -std::numeric_limits<float>::infinity() : p_lse;
+  s_lse = std::isinf(s_lse) ? -std::numeric_limits<float>::infinity() : s_lse;
+
+  const float max_lse = fmaxf(p_lse, s_lse);
+  p_lse = p_lse - max_lse;
+  s_lse = s_lse - max_lse;
+  const float p_se = expf(p_lse);
+  const float s_se = expf(s_lse);
+  const float out_se = p_se + s_se;
+  const float p_scale = p_se / out_se;
+  const float s_scale = s_se / out_se;
+
+  if (pack_offset < head_size) {
+    // Pack 128b load
+    pack_128b_t p_out_pack = reinterpret_cast<const pack_128b_t*>(
+        prefix_head_ptr)[pack_offset / pack_size];
+    pack_128b_t s_out_pack = reinterpret_cast<const pack_128b_t*>(
+        suffix_head_ptr)[pack_offset / pack_size];
+    pack_128b_t o_out_pack;
+
+#pragma unroll
+    for (uint i = 0; i < pack_size; ++i) {
+      // Always use float for FMA to keep high precision.
+      // half(uint16_t), bfloat16, float -> float.
+      const float p_out_f =
+          vllm::to_float(reinterpret_cast<const scalar_t*>(&p_out_pack)[i]);
+      const float s_out_f =
+          vllm::to_float(reinterpret_cast<const scalar_t*>(&s_out_pack)[i]);
+      // fma: a * b + c = p_out_f * p_scale + (s_out_f * s_scale)
+      const float o_out_f = p_out_f * p_scale + (s_out_f * s_scale);
+      // float -> half(uint16_t), bfloat16, float.
+      vllm::from_float(reinterpret_cast<scalar_t*>(&o_out_pack)[i], o_out_f);
+    }
+
+    // Pack 128b storage
+    reinterpret_cast<pack_128b_t*>(output_head_ptr)[pack_offset / pack_size] =
+        o_out_pack;
+  }
+  // We only need to write to output_lse once per head.
+  if (output_lse != nullptr && pack_idx == 0) {
+    float out_lse = logf(out_se) + max_lse;
+    output_lse[head_idx * num_tokens + token_idx] = out_lse;
+  }
+}
+
+}  // namespace vllm
+
+// The following macro is used to dispatch the conversion function based on
+// the output data type. The FN is a macro that calls a function with
+// template<typename scalar_t>.
+#define DISPATCH_BY_SCALAR_DTYPE(scalar_dtype, fn)                      \
+  {                                                                     \
+    if (scalar_dtype == at::ScalarType::Float) {                        \
+      fn(float);                                                        \
+    } else if (scalar_dtype == at::ScalarType::Half) {                  \
+      fn(uint16_t);                                                     \
+    } else if (scalar_dtype == at::ScalarType::BFloat16) {              \
+      fn(__nv_bfloat16);                                                \
+    } else {                                                            \
+      TORCH_CHECK(false, "Unsupported data type of O: ", scalar_dtype); \
+    }                                                                   \
+  }
+
+#define LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS)                     \
+  {                                                                         \
+    vllm::merge_attn_states_kernel<scalar_t, NUM_THREADS>                   \
+        <<<grid, block, 0, stream>>>(                                       \
+            reinterpret_cast<scalar_t*>(output.data_ptr()), output_lse_ptr, \
+            reinterpret_cast<scalar_t*>(prefix_output.data_ptr()),          \
+            reinterpret_cast<float*>(prefix_lse.data_ptr()),                \
+            reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),          \
+            reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,    \
+            num_heads, head_size);                                          \
+  }
+
+/*@brief Merges the attention states from prefix and suffix
+ * into the output tensor. NUM_TOKENS: n, NUM_HEADS: h, HEAD_SIZE: d
+ *
+ * @param output [n,h,d] The output tensor to store the merged attention states.
+ * @param output_lse [h,d] Optional tensor to store the log-sum-exp values.
+ * @param prefix_output [n,h,d] The prefix attention states.
+ * @param prefix_lse [h,n] The log-sum-exp values for the prefix attention
+ * states.
+ * @param suffix_output [n,h,d] The suffix attention states.
+ * @param suffix_lse [h,n] The log-sum-exp values for the suffix attention
+ * states.
+ */
+template <typename scalar_t>
+void merge_attn_states_launcher(torch::Tensor& output,
+                                std::optional<torch::Tensor> output_lse,
+                                const torch::Tensor& prefix_output,
+                                const torch::Tensor& prefix_lse,
+                                const torch::Tensor& suffix_output,
+                                const torch::Tensor& suffix_lse) {
+  constexpr uint NUM_THREADS = 128;
+  const uint num_tokens = output.size(0);
+  const uint num_heads = output.size(1);
+  const uint head_size = output.size(2);
+  const uint pack_size = 16 / sizeof(scalar_t);
+  TORCH_CHECK(head_size % pack_size == 0,
+              "headsize must be multiple of pack_size:", pack_size);
+  float* output_lse_ptr = nullptr;
+  if (output_lse.has_value()) {
+    output_lse_ptr = output_lse.value().data_ptr<float>();
+  }
+  // Process one pack elements per thread. for float, the
+  // pack_size is 4 for half/bf16, the pack_size is 8.
+  const uint threads_per_head = head_size / pack_size;
+  const uint total_threads = num_tokens * num_heads * threads_per_head;
+
+  dim3 block(NUM_THREADS);
+  dim3 grid((total_threads + NUM_THREADS - 1) / NUM_THREADS);
+
+  const c10::cuda::OptionalCUDAGuard device_guard(prefix_output.device());
+  auto stream = at::cuda::getCurrentCUDAStream();
+
+  LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS);
+}
+
+#define CALL_MERGE_ATTN_STATES_LAUNCHER(scalar_t)                           \
+  {                                                                         \
+    merge_attn_states_launcher<scalar_t>(output, output_lse, prefix_output, \
+                                         prefix_lse, suffix_output,         \
+                                         suffix_lse);                       \
+  }
+
+void merge_attn_states(torch::Tensor& output,
+                       std::optional<torch::Tensor> output_lse,
+                       const torch::Tensor& prefix_output,
+                       const torch::Tensor& prefix_lse,
+                       const torch::Tensor& suffix_output,
+                       const torch::Tensor& suffix_lse) {
+  DISPATCH_BY_SCALAR_DTYPE(output.dtype(), CALL_MERGE_ATTN_STATES_LAUNCHER);
+}

csrc/cpu/cache.cpp

@@ -88,6 +88,48 @@ void reshape_and_cache_cpu_impl(
 }
 };  // namespace
 
+template <typename scalar_t>
+void concat_and_cache_mla_cpu_impl(
+    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
+    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
+    scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
+                                      // + pe_dim)]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int num_tokens,                      //
+    const int block_stride,                    //
+    const int entry_stride,                    //
+    const int kv_c_stride,                     //
+    const int k_pe_stride,                     //
+    const int kv_lora_rank,                    //
+    const int pe_dim,                          //
+    const int block_size                       //
+) {
+#pragma omp parallel for
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    const int64_t slot_idx = slot_mapping[token_idx];
+    // NOTE: slot_idx can be -1 if the token is padded
+    if (slot_idx < 0) {
+      continue;
+    }
+    const int64_t block_idx = slot_idx / block_size;
+    const int64_t block_offset = slot_idx % block_size;
+
+    auto copy = [&](const scalar_t* __restrict__ src,
+                    scalar_t* __restrict__ dst, int src_stride, int dst_stride,
+                    int size, int offset) {
+      for (int i = 0; i < size; i++) {
+        const int64_t src_idx = token_idx * src_stride + i;
+        const int64_t dst_idx =
+            block_idx * block_stride + block_offset * entry_stride + i + offset;
+        dst[dst_idx] = src[src_idx];
+      }
+    };
+
+    copy(kv_c, kv_cache, kv_c_stride, block_stride, kv_lora_rank, 0);
+    copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
+  }
+}
+
 // Note: the key_caches and value_caches vectors are constant but
 // not the Tensors they contain. The vectors need to be const refs
 // in order to satisfy pytorch's C++ operator registration code.
@@ -134,6 +176,38 @@ void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
   });
 }
 
+void concat_and_cache_mla(
+    torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
+    torch::Tensor& k_pe,          // [num_tokens, pe_dim]
+    torch::Tensor& kv_cache,      // [num_blocks, block_size, (kv_lora_rank +
+                                  // pe_dim)]
+    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    const std::string& kv_cache_dtype, torch::Tensor& scale) {
+  int num_tokens = slot_mapping.size(0);
+  int kv_lora_rank = kv_c.size(1);
+  int pe_dim = k_pe.size(1);
+  int block_size = kv_cache.size(1);
+
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+  TORCH_CHECK(kv_cache_dtype != "fp8");
+
+  int kv_c_stride = kv_c.stride(0);
+  int k_pe_stride = k_pe.stride(0);
+  int block_stride = kv_cache.stride(0);
+  int entry_stride = kv_cache.stride(1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      kv_c.scalar_type(), "concat_and_cache_mla_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(concat_and_cache_mla_cpu_impl)
+        concat_and_cache_mla_cpu_impl<scalar_t>(
+            kv_c.data_ptr<scalar_t>(), k_pe.data_ptr<scalar_t>(),
+            kv_cache.data_ptr<scalar_t>(), slot_mapping.data_ptr<int64_t>(),
+            num_tokens, block_stride, entry_stride, kv_c_stride, k_pe_stride,
+            kv_lora_rank, pe_dim, block_size);
+        CPU_KERNEL_GUARD_OUT(concat_and_cache_mla_cpu_impl)
+      });
+}
+
 void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
                  const torch::Tensor& block_mapping) {
   TORCH_CHECK(false, "swap_blocks is unsupported on CPU.")

csrc/cpu/cpu_types_x86.hpp

@@ -78,9 +78,14 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
 
   __m256i reg;
 
+  // normal load
   explicit FP16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
+  // non-temproal load
+  explicit FP16Vec16(bool, void* ptr)
+      : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
+
   explicit FP16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
@@ -110,9 +115,14 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   __m256i reg;
 
+  // normal load
   explicit BF16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
+  // non-temproal load
+  explicit BF16Vec16(bool, void* ptr)
+      : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
+
   explicit BF16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
@@ -130,6 +140,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
 
   __m512i reg;
 
+  explicit BF16Vec32() : reg(_mm512_setzero_si512()) {}
+
   explicit BF16Vec32(const void* ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
 
   explicit BF16Vec32(__m512i data) : reg(data) {}
@@ -311,8 +323,13 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
 
+  // normal load
   explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {}
 
+  // non-temproal load
+  explicit FP32Vec16(bool, void* ptr)
+      : reg((__m512)_mm512_stream_load_si512(ptr)) {}
+
   explicit FP32Vec16(__m512 data) : reg(data) {}
 
   explicit FP32Vec16(const FP32Vec4& data)
@@ -545,6 +562,33 @@ struct INT8Vec16 : public Vec<INT8Vec16> {
     _mm_mask_storeu_epi8(ptr, mask, reg);
   }
 };
+
+struct INT8Vec64 : public Vec<INT8Vec64> {
+  constexpr static int VEC_ELEM_NUM = 64;
+  union AliasReg {
+    __m512i reg;
+    int8_t values[VEC_ELEM_NUM];
+  };
+
+  __m512i reg;
+
+  // normal load
+  explicit INT8Vec64(void* ptr) : reg(_mm512_loadu_epi8(ptr)) {}
+
+  // non-temproal load
+  explicit INT8Vec64(bool, void* ptr) : reg(_mm512_stream_load_si512(ptr)) {}
+
+  void save(void* ptr) const { _mm512_storeu_epi8(ptr, reg); }
+
+  void save(int8_t* ptr, const int elem_num) const {
+    constexpr uint64_t M = 0xFFFFFFFFFFFFFFFF;
+    __mmask64 mask = _cvtu64_mask64(M >> (64 - elem_num));
+    _mm512_mask_storeu_epi8(ptr, mask, reg);
+  }
+
+  // non-temproal save
+  void nt_save(int8_t* ptr) { _mm512_stream_si512((__m512i*)ptr, reg); }
+};
 #endif
 
 template <typename T>
@@ -655,6 +699,22 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
 
 inline void prefetch(const void* addr) { _mm_prefetch(addr, _MM_HINT_T1); }
 
+#ifdef __AVX512F__
+inline void non_temporal_save(FP16Vec16& vec, void* ptr) {
+  _mm256_stream_si256((__m256i*)ptr, vec.reg);
+}
+inline void non_temporal_save(BF16Vec32& vec, void* ptr) {
+  _mm512_stream_si512((__m512i*)ptr, vec.reg);
+}
+inline void non_temporal_save(BF16Vec16& vec, void* ptr) {
+  _mm256_stream_si256((__m256i*)ptr, vec.reg);
+}
+inline void non_temporal_save(FP32Vec16& vec, void* ptr) {
+  _mm512_stream_ps((float*)ptr, vec.reg);
+}
+#endif
+
+inline void mem_barrier() { _mm_mfence(); }
 };  // namespace vec_op
 
 #endif

csrc/cpu/mla_decode.cpp

@@ -0,0 +1,393 @@
+#include "cpu_types.hpp"
+#include <float.h>
+
+namespace {
+template <typename scalar_t>
+struct KernelVecType {
+  using qk_load_vec_type = void;
+  using qk_vec_type = void;
+  using v_load_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using qk_load_vec_type = vec_op::FP32Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power and s390x architecture-specific vector types
+  using qk_load_vec_type = vec_op::FP32Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures, including x86
+  using qk_load_vec_type = vec_op::FP16Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP16Vec16;
+#endif
+};
+
+#ifdef __AVX512BF16__
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using qk_load_vec_type = vec_op::BF16Vec32;
+  using qk_vec_type = vec_op::BF16Vec32;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#elif defined(__aarch64__) && !defined(ARM_BF16_SUPPORT)
+// pass
+#else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using qk_load_vec_type = vec_op::BF16Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#endif
+
+template <int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE, int HEAD_UNROLL,
+          typename qk_vec_type>
+void mla_decode_block_head(
+    const qk_vec_type* __restrict__ q_vecs,          // [HEAD_UNROLL, head_dim]
+    const qk_vec_type* __restrict__ k_vecs,          // [block_size, head_dim]
+    const vec_op::FP32Vec16* __restrict v_vecs_f32,  // [block_size, v_head_dim]
+    float* __restrict__ acc_out,  // [HEAD_UNROLL, v_head_dim]
+    float* __restrict__ acc_lse,  // [HEAD_UNROLL]
+    const float scale, const int num_tokens) {
+  using f32_vec_type = vec_op::FP32Vec16;
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+  constexpr int V_NUM_ELEM = f32_vec_type::VEC_ELEM_NUM;
+
+  float logits[BLOCK_SIZE][HEAD_UNROLL] = {};  // initialize to zeros
+  float max_val[HEAD_UNROLL];
+  std::fill(max_val, max_val + HEAD_UNROLL, -FLT_MAX);
+
+  f32_vec_type acc_vec[BLOCK_SIZE][HEAD_UNROLL];
+  for (int i = 0; i < HEAD_DIM; i += QK_NUM_ELEM) {
+    // load to registers
+    qk_vec_type q_vec[HEAD_UNROLL];
+
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+      q_vec[unroll] =
+          qk_vec_type{q_vecs[(i + unroll * HEAD_DIM) / QK_NUM_ELEM]};
+
+    for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+      qk_vec_type k_vec(k_vecs[(block_offset * HEAD_DIM + i) / QK_NUM_ELEM]);
+
+#pragma unroll
+      for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+        vec_op::fma(acc_vec[block_offset][unroll], q_vec[unroll], k_vec);
+    }
+  }
+
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      const float acc = acc_vec[block_offset][unroll].reduce_sum() * scale;
+      logits[block_offset][unroll] = acc;
+      max_val[unroll] = std::max(max_val[unroll], acc);
+    }
+  }
+
+  float sum_exp[HEAD_UNROLL] = {};
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      const float val =
+          std::exp(logits[block_offset][unroll] - max_val[unroll]);
+      logits[block_offset][unroll] = val;
+      sum_exp[unroll] += val;
+    }
+  }
+
+  f32_vec_type this_out[V_HEAD_DIM / V_NUM_ELEM][HEAD_UNROLL];
+
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+    // load to registers
+    f32_vec_type scale_[HEAD_UNROLL];
+
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+      scale_[unroll] =
+          f32_vec_type{logits[block_offset][unroll] / sum_exp[unroll]};
+
+    for (int i = 0; i < V_HEAD_DIM; i += V_NUM_ELEM) {
+      f32_vec_type v_vec(
+          v_vecs_f32[(block_offset * HEAD_DIM + i) / V_NUM_ELEM]);
+
+#pragma unroll
+      for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+        vec_op::fma(this_out[i / V_NUM_ELEM][unroll], v_vec, scale_[unroll]);
+    }
+  }
+
+  // merge attention state
+  // section 2.2 in https://arxiv.org/pdf/2501.01005
+  f32_vec_type prev_scale[HEAD_UNROLL];
+  f32_vec_type curr_scale[HEAD_UNROLL];
+
+#pragma unroll
+  for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+    const float prev_lse = acc_lse[unroll];
+    const float curr_lse = std::log(sum_exp[unroll]) +
+                           max_val[unroll];  // add back max_val to get true lse
+    // softmax trick
+    const float max_lse = std::max(prev_lse, curr_lse);
+    const float prev_sum_exp = std::exp(prev_lse - max_lse);
+    const float curr_sum_exp = std::exp(curr_lse - max_lse);
+
+    const float new_sum_exp = prev_sum_exp + curr_sum_exp;
+    acc_lse[unroll] = std::log(new_sum_exp) + max_lse;
+
+    prev_scale[unroll] = f32_vec_type{prev_sum_exp / new_sum_exp};
+    curr_scale[unroll] = f32_vec_type{curr_sum_exp / new_sum_exp};
+  }
+
+  for (int i = 0; i < V_HEAD_DIM; i += V_NUM_ELEM) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      f32_vec_type o_vec(acc_out + i + V_HEAD_DIM * unroll);
+      o_vec = o_vec * prev_scale[unroll] +
+              this_out[i / V_NUM_ELEM][unroll] * curr_scale[unroll];
+      o_vec.save(acc_out + i + V_HEAD_DIM * unroll);
+    }
+  }
+
+  q_vecs += HEAD_DIM / QK_NUM_ELEM * HEAD_UNROLL;
+  acc_out += V_HEAD_DIM * HEAD_UNROLL;
+}
+
+template <typename scalar_t, int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE,
+          typename qk_vec_type>
+void mla_decode_block(
+    const qk_vec_type* __restrict__ q_vecs,  // [num_heads, head_dim]
+    const scalar_t* __restrict__ kv_cache,   // [block_size, head_dim]
+    float* __restrict__ acc_out,             // [num_heads, v_head_dim]
+    float* __restrict__ acc_lse,             // [num_heads]
+    const int num_heads, const float scale, const int num_tokens) {
+  using qk_load_vec_type = typename KernelVecType<scalar_t>::qk_load_vec_type;
+  static_assert(
+      std::is_same<qk_vec_type,
+                   typename KernelVecType<scalar_t>::qk_vec_type>::value);
+  using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+  using f32_vec_type = vec_op::FP32Vec16;
+  static_assert(qk_load_vec_type::VEC_ELEM_NUM == qk_vec_type::VEC_ELEM_NUM);
+  static_assert(v_load_vec_type::VEC_ELEM_NUM == f32_vec_type::VEC_ELEM_NUM);
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+  constexpr int V_NUM_ELEM = v_load_vec_type::VEC_ELEM_NUM;
+
+  const qk_vec_type* k_vecs;
+  const f32_vec_type* v_vecs_f32;
+  float* kv_cache_f32 = nullptr;
+
+  if constexpr (!std::is_same<scalar_t, float>::value) {
+    // convert KV cache block to FP32 to reuse it across query heads and
+    // attn @ V computation, since FP16/BF16->FP32 is expensive.
+    // TODO: move malloc outside of this fn to reuse across iterations.
+    const int nbytes = BLOCK_SIZE * HEAD_DIM * sizeof(float);
+    kv_cache_f32 = static_cast<float*>(std::aligned_alloc(64, nbytes));
+
+    for (int block_offset = 0; block_offset < num_tokens; ++block_offset)
+      for (int i = 0; i < HEAD_DIM; i += V_NUM_ELEM) {
+        v_load_vec_type kv_load_vec(kv_cache + block_offset * HEAD_DIM + i);
+        f32_vec_type kv_vec_f32(kv_load_vec);
+        kv_vec_f32.save(kv_cache_f32 + block_offset * HEAD_DIM + i);
+      }
+
+    if constexpr (std::is_same<qk_load_vec_type, qk_vec_type>::value) {
+      // for AVX512_BF16, Q @ K.T uses BF16 for K (no conversion)
+      // NOTE: in this case, we only need to convert the V section to FP32.
+      // But for simplicity, we will convert the whole KV block to FP32.
+      k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache);
+    } else {
+      k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache_f32);
+    }
+
+    // attn @ V always use FP32 for V, since attn is FP32.
+    v_vecs_f32 = reinterpret_cast<const f32_vec_type*>(kv_cache_f32);
+
+  } else {
+    // KV cache is FP32. don't need to do anything.
+    k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache);
+    v_vecs_f32 = reinterpret_cast<const f32_vec_type*>(kv_cache);
+  }
+
+  // compute 2 heads at the same time to improve ILP and
+  // take advantage of register cache for K and V.
+  constexpr int HEAD_UNROLL = 2;
+  for (int iter = 0; iter < num_heads / HEAD_UNROLL; ++iter) {
+    mla_decode_block_head<HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE, HEAD_UNROLL>(
+        q_vecs, k_vecs, v_vecs_f32, acc_out, acc_lse, scale, num_tokens);
+
+    q_vecs += HEAD_UNROLL * HEAD_DIM / QK_NUM_ELEM;
+    acc_out += HEAD_UNROLL * V_HEAD_DIM;
+    acc_lse += HEAD_UNROLL;
+  }
+
+  // take care of the remaining heads
+  for (int iter = 0; iter < num_heads % HEAD_UNROLL; ++iter) {
+    mla_decode_block_head<HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE, 1>(
+        q_vecs, k_vecs, v_vecs_f32, acc_out, acc_lse, scale, num_tokens);
+
+    q_vecs += HEAD_DIM / QK_NUM_ELEM;
+    acc_out += V_HEAD_DIM;
+    acc_lse += 1;
+  }
+
+  if (kv_cache_f32 != nullptr) {
+    std::free(kv_cache_f32);
+  }
+}
+}  // namespace
+
+template <typename scalar_t, int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE>
+void mla_decode_kvcache_cpu_impl(
+    scalar_t* __restrict__ out,             // [num_seqs, num_heads, v_head_dim]
+    const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_dim]
+    const scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size,
+                                            // head_dim]
+    const int num_heads, const float scale,
+    const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_blocks_per_seq, const int o_stride, const int q_stride,
+    const int kv_stride, const int num_seqs) {
+  using qk_load_vec_type = typename KernelVecType<scalar_t>::qk_load_vec_type;
+  using qk_vec_type = typename KernelVecType<scalar_t>::qk_vec_type;
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+
+  // shared across threads
+  const int max_threads = omp_get_max_threads();
+  const int acc_out_nbytes =
+      max_threads * num_heads * V_HEAD_DIM * sizeof(float);
+  float* acc_out = static_cast<float*>(std::aligned_alloc(64, acc_out_nbytes));
+  std::vector<float> acc_lse(max_threads * num_heads);
+
+  // allocate memory to pre-convert query to FP32 later
+  float* q_f32;
+  constexpr bool PRE_CONVERT_QUERY =
+      !std::is_same<scalar_t, float>::value &&
+      std::is_same<qk_vec_type, vec_op::FP32Vec16>::value;
+  if constexpr (PRE_CONVERT_QUERY) {
+    const int q_f32_nbytes = num_heads * HEAD_DIM * sizeof(float);
+    q_f32 = static_cast<float*>(std::aligned_alloc(64, q_f32_nbytes));
+  }
+
+#pragma omp parallel
+  {
+    const int num_threads = omp_get_num_threads();
+    const int thread_id = omp_get_thread_num();
+    float* __restrict__ acc_out_thread =
+        acc_out + thread_id * num_heads * V_HEAD_DIM;
+    float* __restrict__ acc_lse_thread = acc_lse.data() + thread_id * num_heads;
+
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      // reset accumulator
+      std::fill(acc_out_thread, acc_out_thread + num_heads * V_HEAD_DIM, 0.0f);
+      std::fill(acc_lse_thread, acc_lse_thread + num_heads, -FLT_MAX);
+
+      const int seq_len = seq_lens[seq_idx];
+      const int block_num = (seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+      const int last_block_size = seq_len - (block_num - 1) * BLOCK_SIZE;
+
+      const qk_vec_type* q_vecs;
+      if constexpr (PRE_CONVERT_QUERY) {
+// pre-convert query to FP32 since FP16/BF16->FP32 is slow.
+#pragma omp for
+        for (int i = 0; i < num_heads * HEAD_DIM; i += QK_NUM_ELEM) {
+          qk_load_vec_type q_load_vec(q + seq_idx * q_stride + i);
+          qk_vec_type q_vec(q_load_vec);
+          q_vec.save(q_f32 + i);
+        }
+        q_vecs = reinterpret_cast<const qk_vec_type*>(q_f32);
+      } else {
+        q_vecs = reinterpret_cast<const qk_vec_type*>(q + seq_idx * q_stride);
+      }
+
+#pragma omp for
+      for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+        const int physical_block_idx =
+            block_tables[seq_idx * max_num_blocks_per_seq + block_idx];
+        const int num_tokens =
+            block_idx < block_num - 1 ? BLOCK_SIZE : last_block_size;
+
+        mla_decode_block<scalar_t, HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE>(
+            q_vecs, kv_cache + physical_block_idx * kv_stride, acc_out_thread,
+            acc_lse_thread, num_heads, scale, num_tokens);
+      }
+
+// merge attention states across threads
+// section 2.2 in https://arxiv.org/pdf/2501.01005
+// each thread is responsible for 1 head
+#pragma omp for
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        float* acc_lse_head = acc_lse.data() + head_idx;
+        float* acc_out_head = acc_out + head_idx * V_HEAD_DIM;
+
+        float max_val = -FLT_MAX;
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          max_val = std::max(max_val, acc_lse_head[thread_id_ * num_heads]);
+        }
+
+        float sum_exp = 0.0f;
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          float val = std::exp(acc_lse_head[thread_id_ * num_heads] - max_val);
+          acc_lse_head[thread_id_ * num_heads] = val;
+          sum_exp += val;
+        }
+
+        float inv_sum = 1.0f / sum_exp;
+        float out_head[V_HEAD_DIM] = {};
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          float scale_ = acc_lse_head[thread_id_ * num_heads] * inv_sum;
+          for (int i = 0; i < V_HEAD_DIM; ++i) {
+            out_head[i] +=
+                acc_out_head[thread_id_ * num_heads * V_HEAD_DIM + i] * scale_;
+          }
+        }
+
+        for (int i = 0; i < V_HEAD_DIM; ++i) {
+          vec_op::storeFP32(out_head[i], out + seq_idx * o_stride +
+                                             head_idx * V_HEAD_DIM + i);
+        }
+      }
+    }
+  }
+  if (PRE_CONVERT_QUERY) {
+    std::free(q_f32);
+  }
+  std::free(acc_out);
+}
+
+void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
+                        torch::Tensor& kv_cache, double scale,
+                        torch::Tensor& block_tables, torch::Tensor& seq_lens) {
+  const int num_seqs = query.size(0);
+  const int num_heads = query.size(1);
+  const int head_dim = query.size(2);
+  const int block_size = kv_cache.size(1);
+  const int v_head_dim = out.size(2);
+
+  const int max_num_blocks_per_seq = block_tables.size(1);
+  const int o_stride = out.stride(0);
+  const int q_stride = query.stride(0);
+  const int kv_stride = kv_cache.stride(0);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      query.scalar_type(), "mla_decode_kvcache_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(mla_decode_kvcache_cpu_impl)
+        if (head_dim == 576 && v_head_dim == 512 && block_size == 16)
+          mla_decode_kvcache_cpu_impl<scalar_t, 576, 512, 16>(
+              out.data_ptr<scalar_t>(), query.data_ptr<scalar_t>(),
+              kv_cache.data_ptr<scalar_t>(), num_heads, scale,
+              block_tables.data_ptr<int>(), seq_lens.data_ptr<int>(),
+              max_num_blocks_per_seq, o_stride, q_stride, kv_stride, num_seqs);
+        else
+          TORCH_CHECK(false, "Unsupported block size: ", block_size);
+        CPU_KERNEL_GUARD_OUT(mla_decode_kvcache_cpu_impl)
+      });
+}

csrc/cpu/shm.cpp

@@ -0,0 +1,781 @@
+#include "cpu/cpu_types.hpp"
+
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+namespace {
+#define MAX_SHM_RANK_NUM 8
+#define MAX_THREAD_NUM 12
+#define PER_THREAD_SHM_BUFFER_BYTES (4 * 1024 * 1024)
+#define MIN_THREAD_PROCESS_SIZE (8 * 1024)
+#define MAX_P2P_SEND_TENSOR_NUM 8
+
+template <typename scalar_t>
+struct KernelVecType {
+  using scalar_vec_t = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using scalar_vec_t = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using scalar_vec_t = vec_op::BF16Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+  using scalar_vec_t = vec_op::FP16Vec16;
+};
+
+enum class ThreadSHMStat : char { THREAD_READY = 0, SHM_DATA_READY, DONE };
+
+struct ThreadSHMContext {
+  volatile ThreadSHMStat thread_stats[MAX_SHM_RANK_NUM];
+  int thread_id;
+  int thread_num;
+  int rank;
+  int group_size;
+  size_t _spinning_count;
+  int swizzled_ranks[MAX_SHM_RANK_NUM];
+  void* thread_shm_ptrs[MAX_SHM_RANK_NUM];
+  ThreadSHMContext* shm_contexts[MAX_SHM_RANK_NUM];
+
+  ThreadSHMContext(const int thread_id, const int thread_num, const int rank,
+                   const int group_size, void* thread_shm_ptr)
+      : thread_id(thread_id),
+        thread_num(thread_num),
+        rank(rank),
+        group_size(group_size),
+        _spinning_count(0) {
+    static_assert(sizeof(ThreadSHMContext) % 64 == 0);
+    TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
+    TORCH_CHECK((size_t)this % 64 == 0);
+    TORCH_CHECK((size_t)thread_shm_ptr % 64 == 0);
+    for (int i = 0; i < MAX_SHM_RANK_NUM; ++i) {
+      shm_contexts[i] = nullptr;
+      thread_shm_ptrs[i] = nullptr;
+      swizzled_ranks[i] = (i + rank) % group_size;
+      thread_stats[i] = ThreadSHMStat::DONE;
+    }
+    set_context(rank, this, thread_shm_ptr);
+  }
+
+  void set_context(int rank, ThreadSHMContext* ptr, void* thread_shm_ptr) {
+    TORCH_CHECK(rank < MAX_SHM_RANK_NUM);
+    TORCH_CHECK(ptr);
+    TORCH_CHECK(thread_shm_ptr);
+    TORCH_CHECK_EQ(ptr->thread_num, thread_num);
+    TORCH_CHECK_EQ(ptr->thread_id, thread_id);
+    shm_contexts[rank] = ptr;
+    thread_shm_ptrs[rank] = thread_shm_ptr;
+  }
+
+  template <typename T>
+  T* get_thread_shm_ptr(int rank) {
+    return reinterpret_cast<T*>(thread_shm_ptrs[rank]);
+  }
+
+  int get_swizzled_rank(int idx) { return swizzled_ranks[idx]; }
+
+  void wait_for_all(ThreadSHMStat prev_stat) {
+    for (int idx = 0; idx < group_size; ++idx) {
+      int rank = get_swizzled_rank(idx);
+      while (thread_stats[rank] == prev_stat) {
+        ++_spinning_count;
+        _mm_pause();
+      }
+    }
+    vec_op::mem_barrier();
+  }
+
+  void wait_for_one(int rank, ThreadSHMStat prev_stat) {
+    while (thread_stats[rank] == prev_stat) {
+      ++_spinning_count;
+      _mm_pause();
+    }
+    vec_op::mem_barrier();
+  }
+
+  void set_thread_stat(ThreadSHMStat stat) {
+    for (int idx = 0; idx < group_size; ++idx) {
+      int rank = get_swizzled_rank(idx);
+      shm_contexts[rank]->thread_stats[this->rank] = stat;
+    }
+  }
+
+  void set_thread_stat(int target_rank, ThreadSHMStat stat) {
+    for (int idx = 0; idx < group_size; ++idx) {
+      int rank = get_swizzled_rank(idx);
+      shm_contexts[rank]->thread_stats[target_rank] = stat;
+    }
+  }
+
+  // barrier for all ranks in the group, used for all2all ops
+  // DONE -> THREAD_READY -> SHM_DATA_READY -> DONE -> ...
+  void barrier(ThreadSHMStat next_stat) {
+    if (next_stat == ThreadSHMStat::THREAD_READY) {
+      set_thread_stat(ThreadSHMStat::THREAD_READY);
+      wait_for_all(ThreadSHMStat::DONE);
+    } else if (next_stat == ThreadSHMStat::SHM_DATA_READY) {
+      set_thread_stat(ThreadSHMStat::SHM_DATA_READY);
+      wait_for_all(ThreadSHMStat::THREAD_READY);
+    } else if (next_stat == ThreadSHMStat::DONE) {
+      set_thread_stat(ThreadSHMStat::DONE);
+      wait_for_all(ThreadSHMStat::SHM_DATA_READY);
+    } else {
+      TORCH_CHECK(false, "Invalid next_stat to barrier.");
+    }
+  }
+
+  std::string to_string() const {
+    std::stringstream ss;
+    ss << "SHMContext:";
+    ss << "\nrank: " << rank;
+    ss << "\ngroup_size: " << group_size;
+    ss << "\nthread_num: " << thread_num;
+    ss << "\nthread_id: " << thread_id;
+
+    ss << "\nshm_ctx_stat_loop_seq: [";
+    for (int i = 0; i < group_size; ++i) {
+      ss << swizzled_ranks[i] << ", ";
+    }
+    ss << "]";
+
+    ss << "\nshm_contexts: [";
+    for (int i = 0; i < group_size; ++i) {
+      if (shm_contexts[i]) {
+        ss << shm_contexts[i]->rank << ", ";
+      }
+    }
+    ss << "]";
+
+    return ss.str();
+  }
+};
+
+class SHMManager {
+ public:
+  explicit SHMManager(const std::string& name, const int rank,
+                      const int group_size)
+      : _rank(rank),
+        _group_size(group_size),
+        _thread_num(std::min(torch::get_num_threads(), MAX_THREAD_NUM)),
+        _shm_names({""}),
+        _shared_mem_ptrs({nullptr}),
+        _shm_ctx(nullptr) {
+    _shm_names[rank] = get_shm_name(name, rank);
+    _shared_mem_ptrs[rank] = init_shm(rank);
+    _shm_ctx = reinterpret_cast<ThreadSHMContext*>(_shared_mem_ptrs[rank]);
+
+    for (int i = 0; i < _thread_num; ++i) {
+      ThreadSHMContext* ctx = new (_shm_ctx + i)
+          ThreadSHMContext(i, _thread_num, _rank, _group_size,
+                           compute_thread_shm_ptr(_shm_ctx, i));
+    }
+  }
+
+  void join(const std::string& name) {
+    for (int rank_idx = 0; rank_idx < _group_size; ++rank_idx) {
+      if (rank_idx != _rank) {
+        TORCH_CHECK(_shm_names[rank_idx].empty());
+        TORCH_CHECK(_shared_mem_ptrs[rank_idx] == nullptr);
+        _shm_names[rank_idx] = get_shm_name(name, rank_idx);
+        _shared_mem_ptrs[rank_idx] = init_shm(rank_idx);
+        ThreadSHMContext* target_ctx =
+            reinterpret_cast<ThreadSHMContext*>(_shared_mem_ptrs[rank_idx]);
+        for (int thread_idx = 0; thread_idx < _thread_num; ++thread_idx) {
+          _shm_ctx[thread_idx].set_context(
+              rank_idx, target_ctx + thread_idx,
+              compute_thread_shm_ptr(target_ctx, thread_idx));
+        }
+      }
+    }
+  }
+
+  ~SHMManager() { destroy_shm(); }
+
+  ThreadSHMContext* get_shm_ctx() const { return _shm_ctx; }
+
+  static std::string get_shm_name(const std::string& name, int rank) {
+    return name + "_" + std::to_string(rank);
+  }
+
+  static int64_t create_singleton_instance(const std::string& name,
+                                           const int group_size,
+                                           const int rank) {
+    std::lock_guard<std::mutex> guard(SingletonInstancesLock);
+    SingletonInstances.emplace_back(
+        std::make_unique<SHMManager>(name, rank, group_size));
+    return static_cast<int64_t>(SingletonInstances.size() - 1);
+  }
+
+  static SHMManager* get_singleton_instance(int64_t handle) {
+    return SingletonInstances[handle].get();
+  }
+
+ protected:
+  static std::vector<std::unique_ptr<SHMManager>> SingletonInstances;
+  static std::mutex SingletonInstancesLock;
+
+ private:
+  static size_t round_to_alignment(size_t num) {
+    return ((num + 63) / 64) * 64;
+  }
+
+  int8_t* compute_thread_shm_ptr(ThreadSHMContext* ctx, int thread_id) {
+    int8_t* thread_shm_ptr =
+        reinterpret_cast<int8_t*>(ctx) +
+        round_to_alignment(_thread_num * sizeof(ThreadSHMContext));
+    return thread_shm_ptr +
+           thread_id * round_to_alignment(PER_THREAD_SHM_BUFFER_BYTES);
+  }
+
+  size_t compute_shm_size() {
+    const size_t rounded_rank_buffer_size =
+        round_to_alignment(PER_THREAD_SHM_BUFFER_BYTES) * _thread_num;
+    const size_t rounded_thread_shm_ctx_size =
+        round_to_alignment(_thread_num * sizeof(ThreadSHMContext));
+    const size_t shm_size =
+        rounded_thread_shm_ctx_size + rounded_rank_buffer_size;
+    return shm_size;
+  }
+
+  void* init_shm(int target_rank) {
+    const std::string& shm_name = _shm_names[target_rank];
+    const int local_rank = _rank;
+    const size_t shm_size = compute_shm_size();
+
+    int fd = -1;
+    if (local_rank == target_rank) {
+      fd = shm_open(shm_name.c_str(), O_CREAT | O_EXCL | O_RDWR,
+                    S_IRUSR | S_IWUSR);
+
+      if (fd == -1)
+        TORCH_CHECK(false, "create shm in SHMManager failed. errno: " +
+                               std::to_string(errno));
+
+      if (ftruncate(fd, shm_size) == -1)
+        TORCH_CHECK(false, "ftruncate in SHMManager failed. errno: " +
+                               std::to_string(errno));
+    } else {
+      fd = shm_open(shm_name.c_str(), O_RDWR, S_IRUSR | S_IWUSR);
+
+      if (fd == -1)
+        TORCH_CHECK(false, "open shm in SHMManager failed. errno: " +
+                               std::to_string(errno));
+    }
+
+    void* shm_ptr = mmap(nullptr, shm_size, PROT_READ | PROT_WRITE,
+                         MAP_SHARED | MAP_POPULATE, fd, 0);
+
+    if (shm_ptr == MAP_FAILED) {
+      TORCH_CHECK(false,
+                  "mmap in SHMManager failed. errno: " + std::to_string(errno));
+    }
+
+    if (close(fd) != 0) {
+      TORCH_CHECK(
+          false, "close in SHMManager failed. errno: " + std::to_string(errno));
+    }
+
+    TORCH_CHECK((size_t)shm_ptr % 64 == 0);
+
+    return shm_ptr;
+  }
+
+  void destroy_shm() {
+    std::stringstream ss;
+    ss << "local rank " << _rank << ": [";
+    for (int thread_id = 0; thread_id < _thread_num; ++thread_id) {
+      ss << _shm_ctx[thread_id]._spinning_count << ", ";
+    }
+    ss << "]\n";
+
+    for (int i = 0; i < MAX_SHM_RANK_NUM; ++i) {
+      if (_shared_mem_ptrs[i] != nullptr) {
+        munmap(_shared_mem_ptrs[i], compute_shm_size());
+      }
+
+      if (!_shm_names[i].empty()) {
+        shm_unlink(_shm_names[i].c_str());
+      }
+    }
+  }
+
+  int _rank;
+  int _group_size;
+  int _thread_num;
+  std::array<std::string, MAX_SHM_RANK_NUM> _shm_names;
+  std::array<void*, MAX_SHM_RANK_NUM> _shared_mem_ptrs;
+  ThreadSHMContext* _shm_ctx;
+};
+
+namespace shm_cc_ops {
+template <typename scalar_t, typename F>
+void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
+  int thread_num = ctx->thread_num;
+  int64_t total_bytes = elem_num * sizeof(scalar_t);
+  int64_t total_units_num =
+      (total_bytes + MIN_THREAD_PROCESS_SIZE - 1) / MIN_THREAD_PROCESS_SIZE;
+  int64_t per_thread_units_num =
+      (total_units_num + thread_num - 1) / thread_num;
+  int64_t per_unit_elem_num = MIN_THREAD_PROCESS_SIZE / sizeof(scalar_t);
+  int64_t max_per_thread_iteration_elem_num =
+      PER_THREAD_SHM_BUFFER_BYTES / sizeof(scalar_t);
+  int64_t per_thread_elem_num = per_unit_elem_num * per_thread_units_num;
+
+#pragma omp parallel for schedule(static, 1)
+  for (int i = 0; i < thread_num; ++i) {
+    int64_t offset = i * per_thread_elem_num;
+    int64_t end = std::min(elem_num, offset + per_thread_elem_num);
+    int64_t curr_elem_num =
+        std::min(max_per_thread_iteration_elem_num, end - offset);
+    ThreadSHMContext* thread_ctx = ctx + i;
+
+    while (curr_elem_num > 0) {
+      inner_func(thread_ctx, offset, curr_elem_num);
+
+      offset += max_per_thread_iteration_elem_num;
+      curr_elem_num = std::min(max_per_thread_iteration_elem_num, end - offset);
+    }
+  }
+}
+};  // namespace shm_cc_ops
+
+namespace shm_cc_ops {
+
+void memcpy_from_shm(void* dst, void* src, const int64_t bytes) {
+  const int64_t aligned_bytes = ((bytes >> 6) << 6);  // 64 bytes aligned
+  int64_t i = 0;
+#pragma GCC unroll 4
+  for (; i < aligned_bytes; i += 64) {
+    vec_op::INT8Vec64 data(
+        true, (int8_t*)src + i);  // stream loading shm to avoid caching
+    data.save((int8_t*)dst + i);
+  }
+  if (aligned_bytes < bytes) {
+    vec_op::INT8Vec64 data(true, (int8_t*)src + aligned_bytes);
+    data.save((int8_t*)dst + aligned_bytes, bytes - aligned_bytes);
+  }
+}
+
+void memcpy_to_shm(void* dst, void* src, const int64_t bytes) {
+#pragma GCC unroll 4
+  for (int64_t i = 0; i < bytes; i += 64) {
+    vec_op::INT8Vec64 data((int8_t*)src + i);
+    data.nt_save((int8_t*)dst + i);
+  }
+}
+
+void memcpy(void* dst, void* src, const int64_t bytes) {
+  const int64_t aligned_bytes = ((bytes >> 6) << 6);  // 64 bytes aligned
+  int64_t i = 0;
+#pragma GCC unroll 4
+  for (; i < aligned_bytes; i += 64) {
+    vec_op::INT8Vec64 data((int8_t*)src + i);
+    data.save((int8_t*)dst + i);
+  }
+  if (aligned_bytes < bytes) {
+    vec_op::INT8Vec64 data((int8_t*)src + aligned_bytes);
+    data.save((int8_t*)dst + aligned_bytes, bytes - aligned_bytes);
+  }
+}
+
+template <typename scalar_t, int RANKS>
+void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
+                         size_t elem_num) {
+  CPU_KERNEL_GUARD_IN(all_reduce_sum_impl)
+  using vec_t = typename KernelVecType<scalar_t>::scalar_vec_t;
+  constexpr int64_t vec_elem_num = vec_t::get_elem_num();
+  const int worldsize = ctx->group_size;
+
+  shm_cc_ops::shm_cc_loop<scalar_t>(
+      ctx, elem_num,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        int rank = thread_ctx->rank;
+        scalar_t* thread_shm_ptr =
+            thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
+        scalar_t* thread_data_ptr = data + data_offset;
+        int64_t thread_data_elem_num = data_elem_num * sizeof(scalar_t);
+
+        scalar_t* remote_data_ptrs[RANKS - 1];
+        vec_op::unroll_loop<int, RANKS - 1>([&](int idx) {
+          remote_data_ptrs[idx] = thread_ctx->get_thread_shm_ptr<scalar_t>(
+              thread_ctx->get_swizzled_rank(idx + 1));
+        });
+
+        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+
+        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, thread_data_ptr,
+                                  thread_data_elem_num);
+
+        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+
+        int64_t aligned_data_elem_num =
+            (data_elem_num / vec_elem_num) * vec_elem_num;
+        int64_t i = 0;
+#pragma GCC unroll 4
+        for (; i < aligned_data_elem_num; i += vec_elem_num) {
+          vec_t local_data(thread_data_ptr + i);  // load from cache
+          vec_op::FP32Vec16 local_data_fp32(local_data);
+          vec_op::unroll_loop<int, RANKS - 1>([&](int idx) {
+            vec_t remote_data(
+                true, remote_data_ptrs[idx] + i);  // stream load from shm
+            vec_op::FP32Vec16 remote_data_fp32(remote_data);
+            local_data_fp32 = local_data_fp32 + remote_data_fp32;  // sum reduce
+          });
+          vec_t reduced_data(local_data_fp32);
+          reduced_data.save(thread_data_ptr + i);
+        }
+
+        if (i < data_elem_num) {
+          vec_t local_data(thread_data_ptr + i);  // load from cache
+          vec_op::FP32Vec16 local_data_fp32(local_data);
+          vec_op::unroll_loop<int, RANKS - 1>([&](int idx) {
+            vec_t remote_data(
+                true, remote_data_ptrs[idx] + i);  // stream load from shm
+            vec_op::FP32Vec16 remote_data_fp32(remote_data);
+            local_data_fp32 = local_data_fp32 + remote_data_fp32;  // sum reduce
+          });
+          vec_t reduced_data(local_data_fp32);
+          reduced_data.save(thread_data_ptr + i,
+                            data_elem_num - aligned_data_elem_num);
+        }
+
+        thread_ctx->barrier(ThreadSHMStat::DONE);
+      });
+
+  return;
+}
+};  // namespace shm_cc_ops
+
+std::vector<std::unique_ptr<SHMManager>> SHMManager::SingletonInstances = {};
+std::mutex SHMManager::SingletonInstancesLock = {};
+
+template <typename scalar_t>
+void shm_allreduce_sum(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num) {
+  switch (ctx->group_size) {
+    case 2:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 2>(ctx, data, elem_num);
+      break;
+    case 3:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 3>(ctx, data, elem_num);
+      break;
+    case 4:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 4>(ctx, data, elem_num);
+      break;
+    case 8:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 8>(ctx, data, elem_num);
+      break;
+    default:
+      TORCH_CHECK(false,
+                  "Invalid world size: " + std::to_string(ctx->group_size));
+  }
+}
+
+template <typename scalar_t>
+void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
+                     scalar_t** outputs, const int dst) {
+  CPU_KERNEL_GUARD_IN(shm_gather_impl)
+  const int worldsize = ctx->group_size;
+  TORCH_CHECK_LT(dst, worldsize);
+  shm_cc_ops::shm_cc_loop<scalar_t>(
+      ctx, elem_num,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        int rank = thread_ctx->rank;
+        scalar_t* thread_shm_ptr =
+            thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
+
+        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+
+        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, data + data_offset,
+                                  data_elem_num * sizeof(scalar_t));
+
+        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+
+        if (rank == dst) {
+          shm_cc_ops::memcpy(outputs[rank] + data_offset, data + data_offset,
+                             data_elem_num * sizeof(scalar_t));
+          for (int i = 1; i < worldsize; ++i) {
+            int src_rank = thread_ctx->get_swizzled_rank(i);
+            scalar_t* src_ptr =
+                thread_ctx->get_thread_shm_ptr<scalar_t>(src_rank);  // shm
+            scalar_t* dst_ptr = outputs[src_rank] + data_offset;
+            shm_cc_ops::memcpy_from_shm(dst_ptr, src_ptr,
+                                        data_elem_num * sizeof(scalar_t));
+          }
+        }
+
+        thread_ctx->barrier(ThreadSHMStat::DONE);
+      });
+
+  return;
+}
+
+struct MemPiece {
+  void* ptr;
+  int64_t size;
+
+  template <typename T>
+  T* data_ptr() {
+    return reinterpret_cast<T*>(ptr);
+  }
+};
+
+struct TensorListMeta {
+  int64_t tensor_bytes[MAX_P2P_SEND_TENSOR_NUM];
+  torch::ScalarType tensor_types[MAX_P2P_SEND_TENSOR_NUM];
+  int64_t tensor_num;
+  int64_t total_bytes;
+
+  TensorListMeta() : tensor_num(0), total_bytes(0) {
+    static_assert(sizeof(TensorListMeta) % 64 == 0);
+    static_assert(sizeof(TensorListMeta) <
+                  MIN_THREAD_PROCESS_SIZE);  // To ensure the metadata always
+                                             // hold by the thread 0
+    for (int i = 0; i < MAX_P2P_SEND_TENSOR_NUM; ++i) {
+      tensor_bytes[i] = 0;
+      tensor_ptrs[i] = nullptr;
+      tensor_types[i] = torch::ScalarType::Undefined;
+    }
+  }
+
+  // For send and recv
+  void bind_tensor_list(std::vector<torch::Tensor>& tensor_list) {
+    TORCH_CHECK(tensor_types[0] == torch::ScalarType::Undefined,
+                "Re-bind TensorListMeta is not allowed.")
+    TORCH_CHECK_LE(tensor_list.size(), MAX_P2P_SEND_TENSOR_NUM);
+    tensor_num = tensor_list.size();
+    int64_t bytes_sum = 0;
+    for (int i = 0; i < tensor_list.size(); ++i) {
+      torch::Tensor& t = tensor_list[i];
+      TORCH_CHECK(t.is_contiguous());
+      tensor_bytes[i] = t.nbytes();
+      tensor_types[i] = t.scalar_type();
+      tensor_ptrs[i] = t.data_ptr();
+      bytes_sum += t.nbytes();
+    }
+    total_bytes = bytes_sum;
+  }
+
+  // For recv
+  std::vector<torch::Tensor> generate_tensor_list() {
+    std::vector<torch::Tensor> tensor_list;
+    tensor_list.reserve(tensor_num);
+
+    for (int i = 0; i < tensor_num; ++i) {
+      int64_t bytes = tensor_bytes[i];
+      auto type = tensor_types[i];
+      int64_t elem_bytes = torch::elementSize(type);
+
+      TORCH_CHECK_EQ(bytes % elem_bytes, 0);
+      int64_t elem_num = bytes / elem_bytes;
+      auto options = torch::TensorOptions().dtype(type).device(torch::kCPU);
+      tensor_list.emplace_back(torch::empty({elem_num}, options));
+    }
+    return tensor_list;
+  }
+
+  MemPiece get_data(int64_t offset) {
+    for (int i = 0; i < tensor_num; ++i) {
+      if (offset < tensor_bytes[i]) {
+        return {reinterpret_cast<int8_t*>(tensor_ptrs[i]) + offset,
+                tensor_bytes[i] - offset};
+      }
+      offset -= tensor_bytes[i];
+    }
+    return {nullptr, 0};
+  }
+
+ private:
+  void* tensor_ptrs[MAX_P2P_SEND_TENSOR_NUM];
+  int8_t _padding[40];
+};
+
+void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
+                               const std::vector<torch::Tensor>& tensor_list) {
+  CPU_KERNEL_GUARD_IN(shm_send_tensor_list_impl)
+  std::vector<torch::Tensor> tensor_list_with_metadata;
+  tensor_list_with_metadata.reserve(1 + tensor_list.size());
+
+  auto options = torch::TensorOptions().dtype(torch::kInt8).device(torch::kCPU);
+  tensor_list_with_metadata.emplace_back(
+      torch::empty({sizeof(TensorListMeta)}, options));
+  tensor_list_with_metadata.insert(tensor_list_with_metadata.end(),
+                                   tensor_list.begin(), tensor_list.end());
+
+  torch::Tensor& metadata_tensor = tensor_list_with_metadata[0];
+  TORCH_CHECK_EQ(metadata_tensor.nbytes(), sizeof(TensorListMeta));
+
+  TensorListMeta* metadata = new (metadata_tensor.data_ptr()) TensorListMeta();
+  metadata->bind_tensor_list(tensor_list_with_metadata);
+
+  shm_cc_ops::shm_cc_loop<int8_t>(
+      ctx, metadata->total_bytes,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        int rank = thread_ctx->rank;
+        // Wait until the receiver set the stat to DONE
+        thread_ctx->wait_for_one(rank, ThreadSHMStat::SHM_DATA_READY);
+
+        int64_t curr_shm_offset = 0;
+        while (curr_shm_offset < data_elem_num) {
+          MemPiece frag = metadata->get_data(data_offset + curr_shm_offset);
+          frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
+          shm_cc_ops::memcpy(
+              thread_ctx->get_thread_shm_ptr<int8_t>(rank) + curr_shm_offset,
+              frag.ptr, frag.size);
+          curr_shm_offset += frag.size;
+        }
+
+        thread_ctx->set_thread_stat(rank, ThreadSHMStat::SHM_DATA_READY);
+      });
+}
+
+std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
+                                                     int64_t src) {
+  CPU_KERNEL_GUARD_IN(shm_recv_tensor_list_impl)
+  auto options = torch::TensorOptions().dtype(torch::kInt8).device(torch::kCPU);
+  torch::Tensor metadata_tensor =
+      torch::empty({sizeof(TensorListMeta)}, options);
+
+  // Wait until the sender set the stat of the thread 0 to SHM_DATA_READY
+  ctx->wait_for_one(src, ThreadSHMStat::DONE);
+  shm_cc_ops::memcpy(metadata_tensor.data_ptr(),
+                     ctx->get_thread_shm_ptr<void>(src),
+                     sizeof(TensorListMeta));
+  TensorListMeta* src_metadata =
+      reinterpret_cast<TensorListMeta*>(metadata_tensor.data_ptr());
+  std::vector<torch::Tensor> tensor_list_with_metadata =
+      src_metadata->generate_tensor_list();
+
+  TensorListMeta metadata;
+  metadata.bind_tensor_list(tensor_list_with_metadata);
+  TORCH_CHECK_EQ(metadata.tensor_num, src_metadata->tensor_num);
+  TORCH_CHECK_EQ(metadata.total_bytes, src_metadata->total_bytes);
+
+  shm_cc_ops::shm_cc_loop<int8_t>(
+      ctx, metadata.total_bytes,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        // Wait until the sender set the stat to SHM_DATA_READY
+        thread_ctx->wait_for_one(src, ThreadSHMStat::DONE);
+        int64_t curr_shm_offset = 0;
+        while (curr_shm_offset < data_elem_num) {
+          MemPiece frag = metadata.get_data(data_offset + curr_shm_offset);
+          frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
+          shm_cc_ops::memcpy(
+              frag.ptr,
+              thread_ctx->get_thread_shm_ptr<int8_t>(src) + curr_shm_offset,
+              frag.size);
+          curr_shm_offset += frag.size;
+        }
+
+        thread_ctx->set_thread_stat(src, ThreadSHMStat::DONE);
+      });
+
+  std::vector<torch::Tensor> tensor_list;
+  tensor_list.reserve(metadata.tensor_num - 1);
+  tensor_list.insert(tensor_list.begin(), tensor_list_with_metadata.begin() + 1,
+                     tensor_list_with_metadata.end());
+
+  return tensor_list;
+}
+}  // namespace
+
+void shm_gather(int64_t handle, torch::Tensor& data,
+                const std::optional<std::vector<torch::Tensor>>& outputs,
+                int64_t dst) {
+  TORCH_CHECK(data.is_contiguous())
+  VLLM_DISPATCH_FLOATING_TYPES(data.scalar_type(), "shm_gather_impl", [&] {
+    CPU_KERNEL_GUARD_IN(shm_gather_impl)
+
+    if (outputs.has_value()) {
+      TORCH_CHECK_LE(outputs->size(), MAX_SHM_RANK_NUM);
+      scalar_t* output_ptrs[MAX_SHM_RANK_NUM] = {nullptr};
+      for (int i = 0; i < outputs->size(); ++i) {
+        output_ptrs[i] = outputs->at(i).data_ptr<scalar_t>();
+      }
+      shm_gather_impl(SHMManager::get_singleton_instance(handle)->get_shm_ctx(),
+                      data.data_ptr<scalar_t>(), data.numel(), output_ptrs,
+                      dst);
+    } else {
+      shm_gather_impl(SHMManager::get_singleton_instance(handle)->get_shm_ctx(),
+                      data.data_ptr<scalar_t>(), data.numel(), (scalar_t**)(0),
+                      dst);
+    }
+
+    CPU_KERNEL_GUARD_OUT(shm_gather_impl)
+  });
+}
+
+void shm_all_gather(int64_t handle, const torch::Tensor& data,
+                    torch::Tensor& output) {
+  TORCH_CHECK(data.is_contiguous())
+  TORCH_CHECK(output.is_contiguous())
+
+  const int64_t input_elem_num = data.numel();
+  const int64_t output_elem_num = output.numel();
+  TORCH_CHECK_EQ(output_elem_num % input_elem_num, 0);
+  const int world_size = output_elem_num / input_elem_num;
+
+  VLLM_DISPATCH_FLOATING_TYPES(data.scalar_type(), "shm_all_gather_impl", [&] {
+    CPU_KERNEL_GUARD_IN(shm_all_gather_impl)
+    auto ctx = SHMManager::get_singleton_instance(handle)->get_shm_ctx();
+    TORCH_CHECK_EQ(ctx->group_size, world_size);
+
+    scalar_t* output_ptrs[MAX_SHM_RANK_NUM] = {nullptr};
+    for (int i = 0; i < world_size; ++i) {
+      output_ptrs[i] = output.data_ptr<scalar_t>() + i * input_elem_num;
+    }
+    shm_gather_impl(ctx, data.data_ptr<scalar_t>(), data.numel(), output_ptrs,
+                    ctx->rank);
+    CPU_KERNEL_GUARD_OUT(shm_all_gather_impl)
+  });
+}
+
+void shm_allreduce(int64_t handle, torch::Tensor& data) {
+  TORCH_CHECK(data.is_contiguous())
+  VLLM_DISPATCH_FLOATING_TYPES(data.scalar_type(), "shm_allreduce_sum", [&] {
+    CPU_KERNEL_GUARD_IN(shm_allreduce_sum)
+    shm_allreduce_sum(SHMManager::get_singleton_instance(handle)->get_shm_ctx(),
+                      data.data_ptr<scalar_t>(), data.numel());
+    CPU_KERNEL_GUARD_OUT(shm_allreduce_sum)
+  });
+}
+
+void shm_send_tensor_list(int64_t handle,
+                          const std::vector<torch::Tensor>& tensor_list,
+                          int64_t dst) {
+  CPU_KERNEL_GUARD_IN(shm_send_tensor_list)
+  shm_send_tensor_list_impl(
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), tensor_list);
+  CPU_KERNEL_GUARD_OUT(shm_send_tensor_list)
+}
+
+std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src) {
+  CPU_KERNEL_GUARD_IN(shm_recv_tensor_list)
+  auto tensor_list = shm_recv_tensor_list_impl(
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), src);
+  CPU_KERNEL_GUARD_OUT(shm_recv_tensor_list)
+  return tensor_list;
+}
+
+int64_t init_shm_manager(const std::string& name, const int64_t group_size,
+                         const int64_t rank) {
+  return SHMManager::create_singleton_instance(name, group_size, rank);
+}
+
+std::string join_shm_manager(int64_t handle, const std::string& name) {
+  auto shm_manager = SHMManager::get_singleton_instance(handle);
+  TORCH_CHECK(shm_manager);
+  shm_manager->join(name);
+  return shm_manager->get_shm_ctx()->to_string();
+}

csrc/cpu/torch_bindings.cpp

@@ -18,6 +18,30 @@ void int8_scaled_mm_azp(torch::Tensor& c, const torch::Tensor& a,
                         const std::optional<torch::Tensor>& azp,
                         const std::optional<torch::Tensor>& bias);
 
+void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
+                        torch::Tensor& kv_cache, double scale,
+                        torch::Tensor& block_tables, torch::Tensor& seq_lens);
+
+int64_t init_shm_manager(const std::string& name, const int64_t group_size,
+                         const int64_t rank);
+
+std::string join_shm_manager(int64_t handle, const std::string& name);
+
+void shm_allreduce(int64_t handle, torch::Tensor& data);
+
+void shm_gather(int64_t handle, torch::Tensor& data,
+                const std::optional<std::vector<torch::Tensor>>& outputs,
+                int64_t dst);
+
+void shm_all_gather(int64_t handle, const torch::Tensor& data,
+                    torch::Tensor& output);
+
+void shm_send_tensor_list(int64_t handle,
+                          const std::vector<torch::Tensor>& tensor_list,
+                          int64_t dst);
+
+std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src);
+
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
@@ -127,6 +151,29 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "                  Tensor? azp, Tensor? bias) -> ()");
   ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
 #endif
+
+// SHM CCL
+#ifdef __AVX512F__
+  ops.def("init_shm_manager(str name, int group_size, int rank) -> int",
+          &init_shm_manager);
+  ops.def("join_shm_manager(int handle, str name) -> str", &join_shm_manager);
+  ops.def("shm_allreduce(int handle, Tensor! data) -> ()");
+  ops.impl("shm_allreduce", torch::kCPU, &shm_allreduce);
+  ops.def(
+      "shm_gather(int handle, Tensor data, Tensor[](a!)? outputs, int dst) -> "
+      "()");
+  ops.impl("shm_gather", torch::kCPU, &shm_gather);
+  ops.def(
+      "shm_all_gather(int handle, Tensor data, Tensor! output) -> "
+      "()");
+  ops.impl("shm_all_gather", torch::kCPU, &shm_all_gather);
+  ops.def(
+      "shm_send_tensor_list(int handle, Tensor[](a) tensor_list, int dst) -> "
+      "()");
+  ops.impl("shm_send_tensor_list", torch::kCPU, &shm_send_tensor_list);
+  ops.def("shm_recv_tensor_list(int handle, int src) -> Tensor[](a)",
+          &shm_recv_tensor_list);
+#endif
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
@@ -150,6 +197,14 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
       "                  str kv_cache_dtype,"
       "                  Tensor k_scale, Tensor v_scale) -> ()");
   cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache);
+
+  cache_ops.def(
+      "concat_and_cache_mla(Tensor kv_c, Tensor k_pe,"
+      "                     Tensor! kv_cache,"
+      "                     Tensor slot_mapping,"
+      "                     str kv_cache_dtype,"
+      "                     Tensor scale) -> ()");
+  cache_ops.impl("concat_and_cache_mla", torch::kCPU, &concat_and_cache_mla);
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {
@@ -157,4 +212,12 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {
   utils.def("init_cpu_threads_env(str cpu_ids) -> str", &init_cpu_threads_env);
 }
 
+TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cpu), cpu_ops) {
+  cpu_ops.def(
+      "mla_decode_kvcache("
+      "   Tensor! out, Tensor query, Tensor kv_cache,"
+      "   float scale, Tensor block_tables, Tensor seq_lens) -> ()");
+  cpu_ops.impl("mla_decode_kvcache", torch::kCPU, &mla_decode_kvcache);
+}
+
 REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

csrc/cpu/utils.cpp

@@ -4,6 +4,11 @@
   #include <string>
   #include <sched.h>
 #endif
+#if __GLIBC__ == 2 && __GLIBC_MINOR__ < 30
+  #include <unistd.h>
+  #include <sys/syscall.h>
+  #define gettid() syscall(SYS_gettid)
+#endif
 
 #include "cpu_types.hpp"
 
@@ -18,7 +23,7 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
 
 #ifndef VLLM_NUMA_DISABLED
 std::string init_cpu_threads_env(const std::string& cpu_ids) {
-  bitmask* omp_cpu_mask = numa_parse_cpustring(cpu_ids.c_str());
+  bitmask* omp_cpu_mask = numa_parse_cpustring_all(cpu_ids.c_str());
   TORCH_CHECK(omp_cpu_mask->size > 0);
   std::vector<int> omp_cpu_ids;
   omp_cpu_ids.reserve(omp_cpu_mask->size);

csrc/cuda_view.cu

@@ -0,0 +1,39 @@
+#include <torch/all.h>
+#include <torch/cuda.h>
+#include <cuda_runtime.h>
+
+// This function assumes that `cpu_tensor` is a CPU tensor allocated with pinned
+// memory, and that UVA (Unified Virtual Addressing) is enabled.
+torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
+  TORCH_CHECK(cpu_tensor.device().is_cpu(), "Input tensor must be on CPU");
+
+  // Get raw host pointer from CPU tensor
+  void* host_ptr = cpu_tensor.data_ptr();
+
+  // Get a device pointer corresponding to the pinned host memory
+  void* device_ptr = nullptr;
+  cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+  TORCH_CHECK(err == cudaSuccess,
+              "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+
+  // We'll use the same sizes, strides, and dtype as the CPU tensor.
+  // TODO: check if layout is respected.
+  auto sizes = cpu_tensor.sizes();
+  auto strides = cpu_tensor.strides();
+  auto options = cpu_tensor.options().device(torch::kCUDA);
+
+  // from_blob signature: from_blob(void *data, IntArrayRef sizes, ..., Deleter,
+  // const TensorOptions &) Provide a no-op deleter. The CPU tensor holds the
+  // memory, so we don't free it here.
+  auto deleter = [](void*) {
+    // no-op, since the memory is owned by the original CPU tensor
+  };
+
+  torch::Tensor cuda_tensor =
+      torch::from_blob(device_ptr, sizes, strides, deleter, options);
+
+  TORCH_CHECK(cuda_tensor.device().is_cuda(),
+              "Resulting tensor is not on CUDA device");
+
+  return cuda_tensor;
+}

csrc/custom_all_reduce.cu

@@ -12,7 +12,7 @@ static_assert(sizeof(void*) == sizeof(fptr_t));
 
 fptr_t init_custom_ar(const std::vector<fptr_t>& fake_ipc_ptrs,
                       torch::Tensor& rank_data, int64_t rank,
-                      bool full_nvlink) {
+                      bool fully_connected) {
   int world_size = fake_ipc_ptrs.size();
   if (world_size > 8)
     throw std::invalid_argument("world size > 8 is not supported");
@@ -27,7 +27,7 @@ fptr_t init_custom_ar(const std::vector<fptr_t>& fake_ipc_ptrs,
   }
   return (fptr_t) new vllm::CustomAllreduce(ipc_ptrs, rank_data.data_ptr(),
                                             rank_data.numel(), rank, world_size,
-                                            full_nvlink);
+                                            fully_connected);
 }
 
 /**
@@ -142,3 +142,48 @@ void register_graph_buffers(fptr_t _fa,
   bytes.reserve(handles.size());
   fa->register_graph_buffers(bytes, offsets);
 }
+
+std::tuple<fptr_t, torch::Tensor> allocate_shared_buffer_and_handle(
+    int64_t size) {
+  auto device_index = c10::cuda::current_device();
+  at::DeviceGuard device_guard(at::Device(at::DeviceType::CUDA, device_index));
+  void* buffer;
+  cudaStreamCaptureMode mode = cudaStreamCaptureModeRelaxed;
+  auto stream = c10::cuda::getCurrentCUDAStream().stream();
+  AT_CUDA_CHECK(cudaThreadExchangeStreamCaptureMode(&mode));
+
+  // Allocate buffer
+#if defined(USE_ROCM)
+  // data buffers need to be "uncached" for signal on MI200
+  AT_CUDA_CHECK(
+      hipExtMallocWithFlags((void**)&buffer, size, hipDeviceMallocUncached));
+#else
+  AT_CUDA_CHECK(cudaMalloc((void**)&buffer, size));
+#endif
+  AT_CUDA_CHECK(cudaMemsetAsync(buffer, 0, size, stream));
+  AT_CUDA_CHECK(cudaStreamSynchronize(stream));
+  AT_CUDA_CHECK(cudaThreadExchangeStreamCaptureMode(&mode));
+
+  // Create IPC memhandle for the allocated buffer.
+  // Will use it in open_mem_handle.
+  auto options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(torch::kCPU);
+  auto handle =
+      torch::empty({static_cast<int64_t>(sizeof(cudaIpcMemHandle_t))}, options);
+  AT_CUDA_CHECK(
+      cudaIpcGetMemHandle((cudaIpcMemHandle_t*)handle.data_ptr(), buffer));
+
+  return std::make_tuple(reinterpret_cast<fptr_t>(buffer), handle);
+}
+
+fptr_t open_mem_handle(torch::Tensor& mem_handle) {
+  void* ipc_ptr;
+  AT_CUDA_CHECK(cudaIpcOpenMemHandle(
+      (void**)&ipc_ptr, *((const cudaIpcMemHandle_t*)mem_handle.data_ptr()),
+      cudaIpcMemLazyEnablePeerAccess));
+  return reinterpret_cast<fptr_t>(ipc_ptr);
+}
+
+void free_shared_buffer(fptr_t buffer) {
+  AT_CUDA_CHECK(cudaFree(reinterpret_cast<void*>(buffer)));
+}

csrc/custom_all_reduce.cuh

@@ -5,6 +5,10 @@
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 
+#if defined(USE_ROCM)
+typedef __hip_bfloat16 nv_bfloat16;
+#endif
+
 #include <iostream>
 #include <array>
 #include <limits>
@@ -12,6 +16,7 @@
 #include <unordered_map>
 #include <vector>
 
+namespace vllm {
 #define CUDACHECK(cmd)                                              \
   do {                                                              \
     cudaError_t e = cmd;                                            \
@@ -22,24 +27,37 @@
     }                                                               \
   } while (0)
 
-namespace vllm {
-
+// Maximal number of blocks in allreduce kernel.
 constexpr int kMaxBlocks = 36;
+
+// Default number of blocks in allreduce kernel.
+#ifndef USE_ROCM
+const int defaultBlockLimit = 36;
+CUpointer_attribute rangeStartAddrAttr = CU_POINTER_ATTRIBUTE_RANGE_START_ADDR;
+#else
+const int defaultBlockLimit = 16;
+hipPointer_attribute rangeStartAddrAttr =
+    HIP_POINTER_ATTRIBUTE_RANGE_START_ADDR;
+#endif
+
 // Counter may overflow, but it's fine since unsigned int overflow is
 // well-defined behavior.
 using FlagType = uint32_t;
+
+// Two sets of peer counters are needed for two syncs: starting and ending an
+// operation. The reason is that it's possible for peer GPU block to arrive at
+// the second sync point while the current GPU block haven't passed the first
+// sync point. Thus, peer GPU may write counter+1 while current GPU is busy
+// waiting for counter. We use alternating counter array to avoid this
+// possibility.
 struct Signal {
-  alignas(128) FlagType self_counter[kMaxBlocks][8];
-  // Two sets of peer counters are needed for two syncs. The reason is that
-  // it's possible for peer GPU block to arrive at the second sync point while
-  // the current GPU block haven't passed the first sync point. Thus, peer GPU
-  // may write counter+1 while current GPU is busy waiting for counter. We use
-  // alternating counter array to avoid this possibility.
-  alignas(128) FlagType peer_counter[2][kMaxBlocks][8];
+  alignas(128) FlagType start[kMaxBlocks][8];
+  alignas(128) FlagType end[kMaxBlocks][8];
+  alignas(128) FlagType _flag[kMaxBlocks];  // incremental flags for each rank
 };
 
 struct __align__(16) RankData {
-  const void* __restrict__ ptrs[8];
+  const void* ptrs[8];
 };
 
 struct __align__(16) RankSignals {
@@ -134,27 +152,29 @@ DINLINE O downcast(array_t<float, O::size> val) {
   }
 }
 
+#if !defined(USE_ROCM)
+
 static DINLINE void st_flag_release(FlagType* flag_addr, FlagType flag) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
   asm volatile("st.release.sys.global.u32 [%1], %0;" ::"r"(flag),
                "l"(flag_addr));
-#else
+  #else
   asm volatile("membar.sys; st.volatile.global.u32 [%1], %0;" ::"r"(flag),
                "l"(flag_addr));
-#endif
+  #endif
 }
 
 static DINLINE FlagType ld_flag_acquire(FlagType* flag_addr) {
   FlagType flag;
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
   asm volatile("ld.acquire.sys.global.u32 %0, [%1];"
                : "=r"(flag)
                : "l"(flag_addr));
-#else
+  #else
   asm volatile("ld.volatile.global.u32 %0, [%1]; membar.gl;"
                : "=r"(flag)
                : "l"(flag_addr));
-#endif
+  #endif
   return flag;
 }
 
@@ -170,37 +190,99 @@ static DINLINE FlagType ld_flag_volatile(FlagType* flag_addr) {
   return flag;
 }
 
-// is_start: whether this is the very first synchronization barrier.
-// need_fence: whether a memory fence is needed. If true, a release-acquire
-// semantic is used to enforce memory access order before and after this
-// barrier.
-template <int ngpus, bool is_start, bool need_fence = false>
-DINLINE void multi_gpu_barrier(const RankSignals& sg, Signal* self_sg,
-                               int rank) {
-  if constexpr (!is_start) __syncthreads();
-  static_assert(
-      !(is_start && need_fence));  // Start barrier shouldn't need fence.
+// This function is meant to be used as the first synchronization in the all
+// reduce kernel. Thus, it doesn't need to make any visibility guarantees for
+// prior memory accesses. Note: volatile writes will not be reordered against
+// other volatile writes.
+template <int ngpus>
+DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
+                              int rank) {
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
   if (threadIdx.x < ngpus) {
-    // Increment the counter. Technically we only need one counter, but we use
-    // multiple per block to eliminate the need to share the counter via smem.
-    auto val = self_sg->self_counter[blockIdx.x][threadIdx.x] += 1;
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->start[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->start[blockIdx.x][threadIdx.x];
+    // Write the expected counter value to peer and wait for correct value
+    // from peer.
+    st_flag_volatile(peer_counter_ptr, flag);
+    while (ld_flag_volatile(self_counter_ptr) != flag);
+  }
+  __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+// This function is meant to be used as the second or the final
+// synchronization barrier in the all reduce kernel. If it's the final
+// synchronization barrier, we don't need to make any visibility guarantees
+// for prior memory accesses.
+template <int ngpus, bool final_sync = false>
+DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->end[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->end[blockIdx.x][threadIdx.x];
     // Write the expected counter value to peer and wait for correct value from
     // peer.
-    auto peer_counter_ptr =
-        &sg.signals[threadIdx.x]->peer_counter[val % 2][blockIdx.x][rank];
-    auto self_counter_ptr =
-        &self_sg->peer_counter[val % 2][blockIdx.x][threadIdx.x];
-    if constexpr (need_fence) {
-      st_flag_release(peer_counter_ptr, val);
-      while (ld_flag_acquire(self_counter_ptr) != val);
+    if constexpr (!final_sync) {
+      st_flag_release(peer_counter_ptr, flag);
+      while (ld_flag_acquire(self_counter_ptr) != flag);
     } else {
-      st_flag_volatile(peer_counter_ptr, val);
-      while (ld_flag_volatile(self_counter_ptr) != val);
+      st_flag_volatile(peer_counter_ptr, flag);
+      while (ld_flag_volatile(self_counter_ptr) != flag);
     }
   }
-  if constexpr (is_start || need_fence) __syncthreads();
+  if constexpr (!final_sync) __syncthreads();
+
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
 }
 
+#else
+
+template <int ngpus>
+DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
+                              int rank) {
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->start[blockIdx.x][rank],
+                            flag, __ATOMIC_RELAXED, __MEMORY_SCOPE_SYSTEM);
+    // wait until we got true from all ranks
+    while (__scoped_atomic_load_n(&self_sg->start[blockIdx.x][threadIdx.x],
+                                  __ATOMIC_RELAXED,
+                                  __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+template <int ngpus, bool final_sync = false>
+DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->end[blockIdx.x][rank],
+                            flag,
+                            final_sync ? __ATOMIC_RELAXED : __ATOMIC_RELEASE,
+                            __MEMORY_SCOPE_SYSTEM);
+    // wait until we got true from all ranks
+    while (
+        __scoped_atomic_load_n(&self_sg->end[blockIdx.x][threadIdx.x],
+                               final_sync ? __ATOMIC_RELAXED : __ATOMIC_ACQUIRE,
+                               __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  if constexpr (!final_sync) __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+#endif
+
 template <typename P, int ngpus, typename A>
 DINLINE P packed_reduce(const P* ptrs[], int idx) {
   A tmp = upcast(ptrs[0][idx]);
@@ -220,13 +302,13 @@ __global__ void __launch_bounds__(512, 1)
   // note: we don't reorder the address so the accumulation order is the same
   // for all ranks, ensuring bitwise identical results
   auto dp = *_dp;
-  multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
+  barrier_at_start<ngpus>(sg, self_sg, rank);
   // do the actual reduction
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
     ((P*)result)[idx] = packed_reduce<P, ngpus, A>((const P**)&dp.ptrs[0], idx);
   }
-  multi_gpu_barrier<ngpus, false>(sg, self_sg, rank);
+  barrier_at_end<ngpus, true>(sg, self_sg, rank);
 }
 
 template <typename P>
@@ -255,18 +337,20 @@ __global__ void __launch_bounds__(512, 1)
     tmps[i] = get_tmp_buf<P>(sg.signals[target]);
   }
   auto tmp_out = tmps[0];
-  multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
+  barrier_at_start<ngpus>(sg, self_sg, rank);
+
   // stage 1: reduce scatter
   for (int idx = start + tid; idx < end; idx += stride) {
     tmp_out[idx - start] = packed_reduce<P, ngpus, A>(ptrs, idx);
   }
-  multi_gpu_barrier<ngpus, false, true>(sg, self_sg, rank);
+  barrier_at_end<ngpus>(sg, self_sg, rank);
 
   // stage 2: allgather. Note: it's important to match the tid between
   // the two stages, because visibility across devices is only guaranteed
   // between threads that have the same tid. If thread i computes the sum of
-  // start + i in the first stage, then thread i also gathers start + i from all
-  // ranks.
+  // start + i in the first stage, then thread i also gathers start + i from
+  // all ranks.
+
   for (int idx = tid; idx < largest_part; idx += stride) {
 #pragma unroll
     for (int i = 0; i < ngpus; i++) {
@@ -287,21 +371,22 @@ class CustomAllreduce {
  public:
   int rank_;
   int world_size_;
-  bool full_nvlink_;
+  // Full NVLink or xGMI connection between GPUs.
+  bool fully_connected_;
 
   RankSignals sg_;
-  // Stores an map from a pointer to its peer pointters from all ranks.
+  // Stores a map from a pointer to its peer pointers from all ranks.
   std::unordered_map<void*, RankData*> buffers_;
   Signal* self_sg_;
 
   // Stores rank data from all ranks. This is mainly for cuda graph purposes.
   // For cuda graph to work, all kernel arguments must be fixed during graph
-  // capture time. However, the peer pointers are not known during graph capture
-  // time. Therefore, during capture, we increment the rank data pointer and use
-  // that as the argument to the kernel. The kernel arguments are stored in
-  // graph_unreg_buffers_. The actual peer pointers will be filled in at the
-  // memory pointed to by the pointers in graph_unreg_buffers_ when
-  // the IPC handles are exchanged between ranks.
+  // capture time. However, the peer pointers are not known during graph
+  // capture time. Therefore, during capture, we increment the rank data
+  // pointer and use that as the argument to the kernel. The kernel arguments
+  // are stored in graph_unreg_buffers_. The actual peer pointers will be
+  // filled in at the memory pointed to by the pointers in
+  // graph_unreg_buffers_ when the IPC handles are exchanged between ranks.
   //
   // The overall process looks like this:
   // 1. Graph capture.
@@ -319,17 +404,18 @@ class CustomAllreduce {
    * Signals are an array of ipc-enabled buffers from all ranks.
    * For each of the buffer, the layout is as follows:
    * | -- sizeof(Signal) -- | ------ a few MB ----- |
-   * The first section is for allreduce synchronization, and the second section
-   * is for storing the intermediate results required by some allreduce algos.
+   * The first section is for allreduce synchronization, and the second
+   * section is for storing the intermediate results required by some
+   * allreduce algos.
    *
    * Note: this class does not own any device memory. Any required buffers
    * are passed in from the constructor.
    */
   CustomAllreduce(Signal** signals, void* rank_data, size_t rank_data_sz,
-                  int rank, int world_size, bool full_nvlink = true)
+                  int rank, int world_size, bool fully_connected = true)
       : rank_(rank),
         world_size_(world_size),
-        full_nvlink_(full_nvlink),
+        fully_connected_(fully_connected),
         self_sg_(signals[rank]),
         d_rank_data_base_(reinterpret_cast<RankData*>(rank_data)),
         d_rank_data_end_(d_rank_data_base_ + rank_data_sz / sizeof(RankData)) {
@@ -361,8 +447,7 @@ class CustomAllreduce {
       void* base_ptr;
       // note: must share the base address of each allocation, or we get wrong
       // address
-      if (cuPointerGetAttribute(&base_ptr,
-                                CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,
+      if (cuPointerGetAttribute(&base_ptr, rangeStartAddrAttr,
                                 (CUdeviceptr)ptr) != CUDA_SUCCESS)
         throw std::runtime_error("failed to get pointer attr");
       CUDACHECK(cudaIpcGetMemHandle(
@@ -396,11 +481,11 @@ class CustomAllreduce {
 
   // Note: when registering graph buffers, we intentionally choose to not
   // deduplicate the addresses. That means if the allocator reuses some
-  // addresses, they will be registered again. This is to account for the remote
-  // possibility of different allocation patterns between ranks. For example,
-  // rank 1 may get the same input address for the second allreduce, but rank 2
-  // got a different address. IPC handles have internal reference counting
-  // mechanism so overhead should be small.
+  // addresses, they will be registered again. This is to account for the
+  // remote possibility of different allocation patterns between ranks. For
+  // example, rank 1 may get the same input address for the second allreduce,
+  // but rank 2 got a different address. IPC handles have internal reference
+  // counting mechanism so overhead should be small.
   void register_graph_buffers(
       const std::vector<std::string>& handles,
       const std::vector<std::vector<int64_t>>& offsets) {
@@ -431,15 +516,15 @@ class CustomAllreduce {
   /**
    * Performs allreduce, assuming input has already been registered.
    *
-   * Block and grid default configs are results after careful grid search. Using
-   * 36 blocks give the best or close to the best runtime on the devices I
-   * tried: A100, A10, A30, T4, V100. You'll notice that NCCL kernels also only
-   * take a small amount of SMs. Not quite sure the underlying reason, but my
-   * guess is that too many SMs will cause contention on NVLink bus.
+   * Block and grid default configs are results after careful grid search.
+   * Using 36 blocks give the best or close to the best runtime on the devices
+   * I tried: A100, A10, A30, T4, V100. You'll notice that NCCL kernels also
+   * only take a small amount of SMs. Not quite sure the underlying reason,
+   * but my guess is that too many SMs will cause contention on NVLink bus.
    */
   template <typename T>
   void allreduce(cudaStream_t stream, T* input, T* output, int size,
-                 int threads = 512, int block_limit = 36) {
+                 int threads = 512, int block_limit = defaultBlockLimit) {
     auto d = packed_t<T>::P::size;
     if (size % d != 0)
       throw std::runtime_error(
@@ -473,13 +558,11 @@ class CustomAllreduce {
 #define KL(ngpus, name)                                                       \
   name<T, ngpus><<<blocks, threads, 0, stream>>>(ptrs, sg_, self_sg_, output, \
                                                  rank_, size);
-    // TODO(hanzhi713): Threshold is different for A100 and H100.
-    // Add per device threshold.
 #define REDUCE_CASE(ngpus)                            \
   case ngpus: {                                       \
     if (world_size_ == 2) {                           \
       KL(ngpus, cross_device_reduce_1stage);          \
-    } else if (full_nvlink_) {                        \
+    } else if (fully_connected_) {                    \
       if ((world_size_ <= 4 && bytes < 512 * 1024) || \
           (world_size_ <= 8 && bytes < 256 * 1024)) { \
         KL(ngpus, cross_device_reduce_1stage);        \
@@ -497,7 +580,8 @@ class CustomAllreduce {
       REDUCE_CASE(8)
       default:
         throw std::runtime_error(
-            "custom allreduce only supports num gpus in (2,4,6,8). Actual num "
+            "custom allreduce only supports num gpus in (2,4,6,8). Actual "
+            "num "
             "gpus = " +
             std::to_string(world_size_));
     }
@@ -511,10 +595,11 @@ class CustomAllreduce {
     }
   }
 };
+
 /**
- * To inspect PTX/SASS, copy paste this header file to compiler explorer and add
- a template instantiation:
+ * To inspect PTX/SASS, copy paste this header file to compiler explorer and
+ add a template instantiation:
  * template void vllm::CustomAllreduce::allreduce<half>(cudaStream_t, half *,
  half *, int, int, int);
 */
-}  // namespace vllm
+}  // namespace vllm

csrc/custom_all_reduce_test.cu

@@ -1,9 +1,9 @@
 /**
  * This is a standalone test for custom allreduce.
  * To compile, make sure you have MPI and NCCL installed in your system.
- * export MPI_HOME=xxx
+ * export MPI_HOME=XXX
  * nvcc -O2 -arch=native -std=c++17 custom_all_reduce_test.cu -o
- * custom_all_reduce_test -lnccl -I${MPI_HOME} -lmpi
+ * custom_all_reduce_test -lnccl -I${MPI_HOME}/include -lmpi
  *
  * Warning: this C++ test is not designed to be very readable and was used
  * during the rapid prototyping process.
@@ -22,7 +22,15 @@
 #include "cuda_profiler_api.h"
 #include "custom_all_reduce.cuh"
 #include "mpi.h"
-#include "nccl.h"
+#ifdef USE_ROCM
+  #include <hip/hip_bf16.h>
+typedef __hip_bfloat16 nv_bfloat16;
+  #include "rccl/rccl.h"
+  #include "custom_all_reduce_hip.cuh"
+#else
+  #include "nccl.h"
+  #include "custom_all_reduce.cuh"
+#endif
 
 #define MPICHECK(cmd)                                                  \
   do {                                                                 \
@@ -43,16 +51,29 @@
     }                                                               \
   } while (0)
 
+#ifdef USE_ROCM
 __global__ void dummy_kernel() {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  for (int i = 0; i < 100; i++) {
+    uint64_t start = wall_clock64();
+    uint64_t cycles_elapsed;
+    do {
+      cycles_elapsed = wall_clock64() - start;
+    } while (cycles_elapsed < 100);
+  }
   for (int i = 0; i < 100; i++) __nanosleep(1000000);  // 100ms
+}
 #else
+__global__ void dummy_kernel() {
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  for (int i = 0; i < 100; i++) __nanosleep(1000000);  // 100ms
+  #else
   for (int i = 0; i < 100; i++) {
     long long int start = clock64();
     while (clock64() - start < 150000000);  // approximately 98.4ms on P40
   }
-#endif
+  #endif
 }
+#endif
 
 template <typename T>
 __global__ void set_data(T* data, int size, int myRank) {
@@ -121,8 +142,14 @@ void run(int myRank, int nRanks, ncclComm_t& comm, int threads, int block_limit,
    * registration, they are allocated and registered together in the test for
    * convenience.
    */
+#ifdef USE_ROCM
+  CUDACHECK(hipExtMallocWithFlags(
+      (void**)&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Signal),
+      hipDeviceMallocUncached));
+#else
   CUDACHECK(
       cudaMalloc(&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Signal)));
+#endif
   CUDACHECK(
       cudaMemset(buffer, 0, 2 * data_size * sizeof(T) + sizeof(vllm::Signal)));
   CUDACHECK(cudaMalloc(&self_data_copy, data_size * sizeof(T)));
@@ -311,13 +338,18 @@ int main(int argc, char** argv) {
 
   bool performance_test = true;
   cudaProfilerStart();
-  // Uncomment to scan through different block size configs.
-  // for (int threads : {256, 512, 1024}) {
-  //   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
-  //     run<half>(myRank, nRanks, comm, threads, block_limit, 1024 * 1024,
-  //     performance_test);
-  //   }
-  // }
+// Uncomment to scan through different block size configs.
+// for (int threads : {256, 512, 1024}) {
+//   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
+//     run<half>(myRank, nRanks, comm, threads, block_limit, 1024 * 1024,
+//     performance_test);
+//   }
+// }
+#ifdef USE_ROCM
+  const int block_limit = 16;
+#else
+  const int block_limit = 36;
+#endif
   // Scan through different sizes to test performance.
   for (int sz = 512; sz <= (8 << 20); sz *= 2) {
     run<half>(myRank, nRanks, comm, 512, 36, sz + 8 * 47, performance_test);
@@ -326,4 +358,4 @@ int main(int argc, char** argv) {
   cudaProfilerStop();
   MPICHECK(MPI_Finalize());
   return EXIT_SUCCESS;
-}
+}

csrc/cutlass_extensions/common.hpp

@@ -48,4 +48,14 @@ struct enable_sm90_or_later : Kernel {
     Kernel::operator()(std::forward<Args>(args)...);
 #endif
   }
-};
+};
+
+template <typename Kernel>
+struct enable_sm90_only : Kernel {
+  template <typename... Args>
+  CUTLASS_DEVICE void operator()(Args&&... args) {
+#if defined __CUDA_ARCH__ && __CUDA_ARCH__ == 900
+    Kernel::operator()(std::forward<Args>(args)...);
+#endif
+  }
+};

csrc/cutlass_extensions/epilogue/broadcast_load_epilogue_array_c3x.hpp

@@ -0,0 +1,457 @@
+/***************************************************************************************************
+ * Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights
+ *reserved. SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice,
+ *this list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ *ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+ *LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ *CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ *SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ *INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ *CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ *ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ *POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+//
+// This file is a modified excerpt of
+// include/cutlass/epilogue/fusion/sm90_visitor_load_tma_warpspecialized.hpp
+// from https://github.com/NVIDIA/cutlass v3.5.0
+// It has been modified to support either row/column or scalar broadcasting
+// where the tensor being loaded from is always passed in via a device pointer.
+// This lets one compiled kernel handle all cases of per-tensor or
+// per-channel/per-token quantization.
+//
+// This interface also allows the scales to be passed in as tensors that
+// consistently reside on the device, which avoids an issue with a previous
+// implementation where scalars needed to be on the CPU since they
+// were passed in via float values. This created a potential performance hazard
+// if scales were initially on the device, and caused torch.compile graphs
+// breaks when moving scales to the CPU.
+//
+#pragma once
+
+// Turn off clang-format for the entire file to keep it close to upstream
+// clang-format off
+
+#include "cutlass/cutlass.h"
+#include "cutlass/arch/barrier.h"
+
+#include "cute/tensor.hpp"
+#include "cutlass/epilogue/fusion/sm90_visitor_tma_warpspecialized.hpp"
+
+namespace cutlass::epilogue::fusion {
+
+using namespace cute;
+using namespace detail;
+
+// Row vector broadcast
+template<
+  int Stages,
+  class CtaTileShapeMNK,
+  class Element,
+  class StrideMNL = Stride<_0,_1,_0>,
+  int Alignment = 128 / sizeof_bits_v<Element>
+>
+struct Sm90RowOrScalarBroadcastArray {
+  static_assert(Stages == 0, "Row broadcast doesn't support smem usage");
+  static_assert(is_static_v<decltype(take<0,2>(StrideMNL{}))>); // batch stride can be dynamic or static
+  static_assert(take<0,2>(StrideMNL{}) == Stride<_0,_1>{});
+
+  struct SharedStorage { 
+    array_aligned<Element, size<1>(CtaTileShapeMNK{})> smem;
+  };
+
+  // This struct has been modified to have a bool indicating that ptr_row is a 
+  // scalar that must be broadcast, instead of containing a scalar that is 
+  // valid if ptr_row is null.
+  struct Arguments {
+    const Element* const* ptr_row_array = nullptr;
+    bool row_broadcast = true;
+    StrideMNL dRow = {};
+  };
+
+  using Params = Arguments;
+
+  template <class ProblemShape>
+  static constexpr Params
+  to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
+    return args;
+  }
+
+  template <class ProblemShape>
+  static bool
+  can_implement(ProblemShape const& problem_shape, Arguments const& args) {
+    return true;
+  }
+
+  template <class ProblemShape>
+  static size_t
+  get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
+    return 0;
+  }
+
+  template <class ProblemShape>
+  static cutlass::Status
+  initialize_workspace(ProblemShape const& problem_shape, Arguments const& args, void* workspace, cudaStream_t stream,
+    CudaHostAdapter* cuda_adapter = nullptr) {
+    return cutlass::Status::kSuccess;
+  }
+
+  CUTLASS_HOST_DEVICE
+  Sm90RowOrScalarBroadcastArray() { }
+
+  CUTLASS_HOST_DEVICE
+  Sm90RowOrScalarBroadcastArray(Params const& params, SharedStorage const& shared_storage)
+      : params(params)
+      , smem(const_cast<Element*>(shared_storage.smem.data())) { }
+
+  Params params;
+  Element *smem = nullptr;
+
+  CUTLASS_DEVICE bool
+  is_producer_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_C_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_zero() const {
+    return (!params.row_broadcast && *(params.ptr_row_array[group]) == Element(0));
+  }
+
+  template <class... Args>
+  CUTLASS_DEVICE auto
+  get_producer_load_callbacks(ProducerLoadArgs<Args...> const& args) {
+    return EmptyProducerLoadCallbacks{};
+  }
+
+  template <class GS_GTensor, class GS_STensor, class GS_CTensor, class Tiled_G2S, class SR_STensor, class SR_RTensor, class CTensor, class ThrResidue, class ThrNum>
+  struct ConsumerStoreCallbacks : EmptyConsumerStoreCallbacks {
+    CUTLASS_DEVICE
+    ConsumerStoreCallbacks(
+        GS_GTensor tGS_gRow_, GS_STensor tGS_sRow_, 
+        GS_CTensor tGS_cRow_, Tiled_G2S tiled_g2s_, 
+        SR_STensor tSR_sRow_, SR_RTensor tSR_rRow_,
+        CTensor tCcRow_, ThrResidue residue_tCcRow_, ThrNum thr_num_,
+        int group, Params const& params_)
+      : tGS_gRow(tGS_gRow_)
+      , tGS_sRow(tGS_sRow_)
+      , tGS_cRow(tGS_cRow_)
+      , tiled_G2S(tiled_g2s_)
+      , tSR_sRow(tSR_sRow_)
+      , tSR_rRow(tSR_rRow_)
+      , tCcRow(tCcRow_)
+      , residue_tCcRow(residue_tCcRow_)
+      , group(group)
+      , params(params_) {}
+
+    GS_GTensor tGS_gRow;                                                         // (CPY,CPY_M,CPY_N)
+    GS_STensor tGS_sRow;                                                         // (CPY,CPY_M,CPY_N)
+    GS_CTensor tGS_cRow;                                                         // (CPY,CPY_M,CPY_N)
+    Tiled_G2S tiled_G2S;
+
+    SR_STensor tSR_sRow;                                                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    SR_RTensor tSR_rRow;                                                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N) 
+  
+    CTensor tCcRow;                                                              // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    ThrResidue residue_tCcRow;                                                   // (m, n)
+    ThrNum thr_num;
+    int group;
+    Params const& params;
+
+    CUTLASS_DEVICE void
+    begin() {
+      if (!params.row_broadcast) {
+        fill(tSR_rRow, *(params.ptr_row_array[group]));
+        return;
+      }
+
+      auto synchronize = [&] () { cutlass::arch::NamedBarrier::sync(thr_num, cutlass::arch::ReservedNamedBarriers::EpilogueBarrier); };
+      Tensor tGS_gRow_flt = filter_zeros(tGS_gRow);
+      Tensor tGS_sRow_flt = filter_zeros(tGS_sRow);
+      Tensor tGS_cRow_flt = make_tensor(tGS_cRow.data(), make_layout(tGS_gRow_flt.shape(), tGS_cRow.stride()));
+
+      for (int i = 0; i < size(tGS_gRow_flt); ++i) {
+        if (get<1>(tGS_cRow_flt(i)) >= size<1>(CtaTileShapeMNK{})) {
+          continue; // OOB of SMEM, 
+        }
+        if (elem_less(tGS_cRow_flt(i), make_coord(get<0>(residue_tCcRow), get<1>(residue_tCcRow)))) {
+          tGS_sRow_flt(i) = tGS_gRow_flt(i);
+        }
+        else {
+          tGS_sRow_flt(i) = Element(0); // Set to Zero when OOB so LDS could be issue without any preds.
+        }
+      }
+      synchronize();
+    }
+
+    CUTLASS_DEVICE void
+    begin_loop(int epi_m, int epi_n) {
+      if (epi_m == 0) { // Assumes M-major subtile loop
+        if (!params.row_broadcast) return; // Do not issue LDS when row is scalar 
+        Tensor tSR_sRow_flt = filter_zeros(tSR_sRow(_,_,_,epi_m,epi_n));
+        Tensor tSR_rRow_flt = filter_zeros(tSR_rRow);
+        copy(tSR_sRow_flt, tSR_rRow_flt);
+      }
+    }
+
+    template <typename ElementAccumulator, int FragmentSize>
+    CUTLASS_DEVICE Array<Element, FragmentSize>
+    visit(Array<ElementAccumulator, FragmentSize> const& frg_acc, int epi_v, int epi_m, int epi_n) {
+      Array<Element, FragmentSize> frg_row;
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < FragmentSize; ++i) {
+        frg_row[i] = tSR_rRow(epi_v * FragmentSize + i);
+      }
+
+      return frg_row;
+    }
+  };
+
+  template <
+    bool ReferenceSrc, // do register tensors reference the src or dst layout of the tiled copy
+    class... Args
+  >
+  CUTLASS_DEVICE auto
+  get_consumer_store_callbacks(ConsumerStoreArgs<Args...> const& args) {
+    auto [M, N, K, L] = args.problem_shape_mnkl;
+    auto [m, n, k, l] = args.tile_coord_mnkl;
+    using ThreadCount = decltype(size(args.tiled_copy));
+
+    Tensor mRow = make_tensor(make_gmem_ptr(params.ptr_row_array[l]), make_shape(M,N,1), params.dRow);
+    Tensor gRow = local_tile(mRow(_,_,l), take<0,2>(args.tile_shape_mnk), make_coord(m, n));          // (CTA_M, CTA_N)
+    Tensor sRow = make_tensor(make_smem_ptr(smem), 
+        make_shape(size<0>(CtaTileShapeMNK{}), size<1>(CtaTileShapeMNK{})), make_shape(_0{}, _1{}));  // (CTA_M, CTA_N)
+    //// G2S: Gmem to Smem
+    auto tiled_g2s = make_tiled_copy(Copy_Atom<DefaultCopy, Element>{},
+                                     Layout< Shape<_1, ThreadCount>, 
+                                            Stride<_0,          _1>>{}, 
+                                     Layout<_1>{});   
+    auto thr_g2s = tiled_g2s.get_slice(args.thread_idx);
+    Tensor tGS_gRow = thr_g2s.partition_S(gRow);
+    Tensor tGS_sRow = thr_g2s.partition_D(sRow);
+
+    //// G2S: Coord 
+    auto cRow = make_identity_tensor(make_shape(size<0>(CtaTileShapeMNK{}), size<1>(CtaTileShapeMNK{})));
+    Tensor tGS_cRow = thr_g2s.partition_S(cRow);
+
+    //// S2R: Smem to Reg
+    Tensor tSR_sRow = sm90_partition_for_epilogue<ReferenceSrc>(sRow, args.epi_tile, args.tiled_copy, args.thread_idx);
+    Tensor tSR_rRow = make_tensor_like(take<0,3>(tSR_sRow));                                           // (CPY,CPY_M,CPY_N)
+
+    return ConsumerStoreCallbacks<decltype(tGS_gRow), decltype(tGS_sRow), decltype(tGS_cRow), decltype(tiled_g2s), decltype(tSR_sRow), decltype(tSR_rRow), decltype(args.tCcD), decltype(args.residue_cD), ThreadCount>(
+      tGS_gRow, 
+      tGS_sRow, 
+      tGS_cRow, tiled_g2s, 
+      tSR_sRow, 
+      tSR_rRow, 
+      args.tCcD, 
+      args.residue_cD,
+      ThreadCount{}, 
+      l,
+      params);
+  }
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Column vector broadcast
+template<
+  int Stages,
+  class CtaTileShapeMNK,
+  class Element,
+  class StrideMNL = Stride<_1,_0,_0>,
+  int Alignment = 128 / sizeof_bits_v<Element>
+>
+struct Sm90ColOrScalarBroadcastArray {
+  static_assert(Stages == 0, "Column broadcast doesn't support smem usage yet");
+  static_assert(Alignment * sizeof_bits_v<Element> % 128 == 0, "sub-16B alignment not supported yet");
+  static_assert(
+    (cute::is_same_v<StrideMNL, Stride<_1,_0, _0>>) || // col vector broadcast, e.g. per-row alpha/bias
+    (cute::is_same_v<StrideMNL, Stride<_1,_0,int>>));  // batched col vector broadcast, e.g. batched per-row bias
+
+  // Accumulator distributes col elements evenly amongst threads so we can just directly load from gmem
+  struct SharedStorage { };
+
+  // This struct has been modified to have a bool indicating that ptr_col is a 
+  // scalar that must be broadcast, instead of containing a scalar that is 
+  // valid if ptr_col is null.
+  struct Arguments {
+    const Element* const* ptr_col_array = nullptr;
+    bool col_broadcast = true;
+    StrideMNL dCol = {};
+  };
+
+  using Params = Arguments;
+
+  template <class ProblemShape>
+  static constexpr Params
+  to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
+    return args;
+  }
+
+  template <class ProblemShape>
+  static bool
+  can_implement(ProblemShape const& problem_shape, Arguments const& args) {
+    return true;
+  }
+
+  template <class ProblemShape>
+  static size_t
+  get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
+    return 0;
+  }
+
+  template <class ProblemShape>
+  static cutlass::Status
+  initialize_workspace(ProblemShape const& problem_shape, Arguments const& args, void* workspace, cudaStream_t stream,
+    CudaHostAdapter* cuda_adapter = nullptr) {
+    return cutlass::Status::kSuccess;
+  }
+
+  CUTLASS_DEVICE bool
+  is_producer_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_C_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_zero() const {
+    return (!params.col_broadcast && *(params.ptr_col_array[group]) == Element(0));
+  }
+
+  CUTLASS_HOST_DEVICE
+  Sm90ColOrScalarBroadcastArray() { }
+
+  CUTLASS_HOST_DEVICE
+  Sm90ColOrScalarBroadcastArray(Params const& params, SharedStorage const& shared_storage)
+      : params(params) { }
+
+  Params params;
+
+  template <class... Args>
+  CUTLASS_DEVICE auto
+  get_producer_load_callbacks(ProducerLoadArgs<Args...> const& args) {
+    return EmptyProducerLoadCallbacks{};
+  }
+
+  template<class GTensor, class RTensor, class CTensor, class ProblemShape>
+  struct ConsumerStoreCallbacks : EmptyConsumerStoreCallbacks {
+    CUTLASS_DEVICE
+    ConsumerStoreCallbacks(
+      GTensor&& tCgCol,
+      RTensor&& tCrCol,
+      CTensor&& tCcCol,
+      ProblemShape problem_shape,
+      int group,
+      Params const& params
+    ): 
+      tCgCol(cute::forward<GTensor>(tCgCol)),
+      tCrCol(cute::forward<RTensor>(tCrCol)),
+      tCcCol(cute::forward<CTensor>(tCcCol)),
+      m(get<0>(problem_shape)),
+      group(group),
+      params(params) {}
+
+    GTensor tCgCol;                                                                    // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    RTensor tCrCol;
+    CTensor tCcCol;                                                                    // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    Params const& params;
+    int m;
+    int group;
+
+    CUTLASS_DEVICE void
+    begin() {
+      Tensor pred = make_tensor<bool>(shape(tCgCol));
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < size(pred); ++i) {
+        pred(i) = get<0>(tCcCol(i)) < m;
+      }
+
+      if (!params.col_broadcast) {
+        fill(tCrCol, *(params.ptr_col_array[group]));
+        return;
+      }
+
+      // Filter so we don't issue redundant copies over stride-0 modes
+      // (only works if 0-strides are in same location, which is by construction)
+      copy_if(pred, filter(tCgCol), filter(tCrCol));
+    }
+
+    template <typename ElementAccumulator, int FragmentSize>
+    CUTLASS_DEVICE Array<Element, FragmentSize>
+    visit(Array<ElementAccumulator, FragmentSize> const& frg_acc, int epi_v, int epi_m, int epi_n) {
+      Array<Element, FragmentSize> frg_col;
+      Tensor tCrCol_mn = tCrCol(_,_,_,epi_m,epi_n);
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < FragmentSize; ++i) {
+        frg_col[i] = tCrCol_mn(epi_v * FragmentSize + i);
+      }
+
+      return frg_col;
+    }
+
+  };
+
+  template <
+    bool ReferenceSrc, // do register tensors reference the src or dst layout of the tiled copy
+    class... Args
+  >
+  CUTLASS_DEVICE auto
+  get_consumer_store_callbacks(ConsumerStoreArgs<Args...> const& args) {
+
+    auto [M, N, K, L] = args.problem_shape_mnkl;
+    auto [m, n, k, l] = args.tile_coord_mnkl;
+
+    Tensor mCol = make_tensor(make_gmem_ptr(params.ptr_col_array[l]), make_shape(M,N,1), params.dCol);
+    Tensor tCgCol = sm90_partition_for_epilogue<ReferenceSrc>(                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+      mCol, args.tile_shape_mnk, args.tile_coord_mnkl, args.epi_tile, args.tiled_copy, args.thread_idx);
+    Tensor tCrCol = make_tensor_like(tCgCol);                                          // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+
+    // Generate an identity tensor matching the shape of the global tensor and 
+    //  partition the same way, this will be used to generate the predicate
+    //  tensor for loading
+    Tensor cCol = make_identity_tensor(mCol.shape());
+    Tensor tCcCol = sm90_partition_for_epilogue<ReferenceSrc>(                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+      cCol, args.tile_shape_mnk, args.tile_coord_mnkl, args.epi_tile, args.tiled_copy, args.thread_idx);
+
+    return ConsumerStoreCallbacks(
+      cute::move(tCgCol), 
+      cute::move(tCrCol), 
+      cute::move(tCcCol), 
+      args.problem_shape_mnkl,
+      l,
+      params
+    );
+  }
+};
+
+}

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "cutlass_extensions/epilogue/broadcast_load_epilogue_c3x.hpp"
+#include "cutlass_extensions/epilogue/broadcast_load_epilogue_array_c3x.hpp"
 
 /*
    This file defines custom epilogues for fusing channel scales, token scales,
@@ -69,6 +70,16 @@ struct ScaledEpilogueBase {
       0 /*Stages*/, TileShape, T, T, Stride<Int<0>, Int<1>, Int<0>>,
       128 / sizeof_bits_v<T>, EnableNullPtr>;
 
+  template <typename T>
+  using ColOrScalarLoadArray =
+      cutlass::epilogue::fusion::Sm90ColOrScalarBroadcastArray<
+          0 /*Stages*/, TileShape, T, Stride<Int<1>, Int<0>, Int<0>>>;
+
+  template <typename T>
+  using RowOrScalarLoadArray =
+      cutlass::epilogue::fusion::Sm90RowOrScalarBroadcastArray<
+          0 /*Stages*/, TileShape, T, Stride<Int<0>, Int<1>, Int<0>>>;
+
   // This utility function constructs the arguments for the load descriptors
   // from a tensor. It can handle both row and column, as well as row/column or
   // scalar cases.
@@ -96,6 +107,14 @@ struct ScaledEpilogueBase {
                   std::is_same_v<Descriptor, RowLoad<T, true>>);
     return Arguments{data_ptr};
   }
+
+  template <typename Descriptor, typename T>
+  static auto args_from_tensor(const T* const* data_ptr, bool do_broadcast) {
+    using Arguments = typename Descriptor::Arguments;
+    static_assert(std::is_same_v<Descriptor, ColOrScalarLoadArray<T>> ||
+                  std::is_same_v<Descriptor, RowOrScalarLoadArray<T>>);
+    return Arguments{data_ptr, do_broadcast};
+  }
 };
 
 /*
@@ -381,4 +400,51 @@ struct ScaledEpilogueBiasAzpToken
   }
 };
 
+/*
+    This epilogue works like ScaledEpilogue, but ScaleA and ScaleB are pointers
+    to arrays containing different scales used in group gemm. The number of
+   pointers in ScaleA and the number of pointers in ScaleB are equal to the
+   group size.
+*/
+template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
+struct ScaledEpilogueArray
+    : private ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor> {
+ private:
+  using SUPER = ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::template ColOrScalarLoadArray<float>;
+  using ScaleB = typename SUPER::template RowOrScalarLoadArray<float>;
+
+  using Compute0 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiplies, float, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using EVTCompute0 =
+      cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
+
+  using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiplies, ElementD, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+ public:
+  using EVTCompute =
+      cutlass::epilogue::fusion::Sm90EVT<Compute1, ScaleA, EVTCompute0>;
+  using ArgumentType = typename EVTCompute::Arguments;
+
+  using ScaleAArray = typename SUPER::template ColOrScalarLoadArray<float>;
+  using ScaleBArray = typename SUPER::template RowOrScalarLoadArray<float>;
+
+  static ArgumentType prepare_args(float const* const* a_scales_ptr,
+                                   float const* const* b_scales_ptr,
+                                   bool a_col_broadcast, bool b_row_broadcast) {
+    auto a_args = SUPER::template args_from_tensor<ScaleAArray, float>(
+        a_scales_ptr, a_col_broadcast);
+    auto b_args = SUPER::template args_from_tensor<ScaleBArray, float>(
+        b_scales_ptr, b_row_broadcast);
+
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, {}};
+  }
+};
+
 };  // namespace vllm::c3x

csrc/mamba/causal_conv1d/causal_conv1d.cu

@@ -422,7 +422,7 @@ void causal_conv1d_fwd_kernel(ConvParamsBase params) {
         int final_state_position =  ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize);
         // in case the final state is separated between the last "smem_exchange" and 
         // and the one before it (chunk = n_chunks - 1 and chunk = n_chunks - 2), 
-        // (which occurs when `final_state_position` is a non-positivie index)
+        // (which occurs when `final_state_position` is a non-positive index)
         // we load the correct data from smem_exchange from both chunks, the last chunk iteration and the one before it
         if (conv_states != nullptr && final_state_position < 0 && seqlen > kWidth){
             input_t vals_load[kNElts] = {0};

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+import glob
+import itertools
+import os
+import subprocess
+
+import jinja2
+
+FILE_HEAD = """
+// auto generated by generate.py
+// clang-format off
+
+#include "kernel.h"
+#include "marlin_template.h"
+
+namespace MARLIN_NAMESPACE_NAME {
+""".strip()
+
+TEMPLATE = ("template __global__ void Marlin<"
+            "{{scalar_t}}, "
+            "{{w_type_id}}, "
+            "{{threads}}, "
+            "{{thread_m_blocks}}, "
+            "{{thread_n_blocks}}, "
+            "{{thread_k_blocks}}, "
+            "{{'true' if m_block_size_8 else 'false'}}, "
+            "{{stages}}, "
+            "{{'true' if has_act_order else 'false'}}, "
+            "{{'true' if has_zp else 'false'}}, "
+            "{{group_blocks}}, "
+            "{{'true' if is_zp_float else 'false'}}>"
+            "( MARLIN_KERNEL_PARAMS );")
+
+# int8 with zero point case (vllm::kU8) is also supported,
+# we don't add it to reduce wheel size.
+SCALAR_TYPES = ["vllm::kU4", "vllm::kU4B8", "vllm::kU8B128"]
+THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128)]
+
+THREAD_M_BLOCKS = [0.5, 1, 2, 3, 4]
+# group_blocks:
+#   = 0 : act order case
+#   = -1 : channelwise quantization
+#   > 0 : group_size=16*group_blocks
+GROUP_BLOCKS = [0, -1, 2, 4, 8]
+DTYPES = ["fp16", "bf16"]
+
+
+def remove_old_kernels():
+    for filename in glob.glob(os.path.dirname(__file__) + "/kernel_*.cu"):
+        subprocess.call(["rm", "-f", filename])
+
+
+def generate_new_kernels():
+    for scalar_type, dtype in itertools.product(SCALAR_TYPES, DTYPES):
+        has_zp = "B" not in scalar_type
+        all_template_str_list = []
+
+        for group_blocks, m_blocks, thread_configs in itertools.product(
+                GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS):
+
+            has_act_order = group_blocks == 0
+            if has_zp and has_act_order:
+                continue
+            if thread_configs[2] == 256:
+                if m_blocks <= 1 and thread_configs[0] != 128:
+                    continue
+                if m_blocks > 1 and thread_configs[0] != 64:
+                    continue
+
+            k_blocks = thread_configs[0] // 16
+            n_blocks = thread_configs[1] // 16
+            threads = thread_configs[2]
+
+            c_dtype = "half" if dtype == "fp16" else "nv_bfloat16"
+
+            template_str = jinja2.Template(TEMPLATE).render(
+                scalar_t=c_dtype,
+                w_type_id=scalar_type + ".id()",
+                threads=threads,
+                thread_m_blocks=max(m_blocks, 1),
+                thread_n_blocks=n_blocks,
+                thread_k_blocks=k_blocks,
+                m_block_size_8=m_blocks == 0.5,
+                stages="pipe_stages",
+                has_act_order=has_act_order,
+                has_zp=has_zp,
+                group_blocks=group_blocks,
+                is_zp_float=False,
+            )
+
+            all_template_str_list.append(template_str)
+
+        file_content = FILE_HEAD + "\n\n"
+        file_content += "\n\n".join(all_template_str_list) + "\n\n}\n"
+        filename = f"kernel_{dtype}_{scalar_type[6:].lower()}.cu"
+
+        with open(os.path.join(os.path.dirname(__file__), filename), "w") as f:
+            f.write(file_content)
+
+
+if __name__ == "__main__":
+    remove_old_kernels()
+    generate_new_kernels()

csrc/moe/marlin_moe_wna16/kernel.h

@@ -0,0 +1,44 @@
+
+#ifndef MARLIN_NAMESPACE_NAME
+  #define MARLIN_NAMESPACE_NAME marlin_moe_wna16
+#endif
+
+#include "quantization/gptq_marlin/marlin.cuh"
+#include "quantization/gptq_marlin/marlin_dtypes.cuh"
+#include "core/scalar_type.hpp"
+
+#define MARLIN_KERNEL_PARAMS                                                \
+  const int4 *__restrict__ A, const int4 *__restrict__ B,                   \
+      int4 *__restrict__ C, int4 *__restrict__ C_tmp,                       \
+      const int4 *__restrict__ scales_ptr, const int4 *__restrict__ zp_ptr, \
+      const int *__restrict__ g_idx,                                        \
+      const int32_t *__restrict__ sorted_token_ids_ptr,                     \
+      const int32_t *__restrict__ expert_ids_ptr,                           \
+      const int32_t *__restrict__ num_tokens_past_padded_ptr,               \
+      const float *__restrict__ topk_weights_ptr, int top_k,                \
+      bool mul_topk_weights, bool is_ep, int num_groups, int prob_m,        \
+      int prob_n, int prob_k, int *locks, bool use_atomic_add,              \
+      bool use_fp32_reduce
+
+namespace MARLIN_NAMESPACE_NAME {
+template <typename scalar_t,  // compute dtype, half or nv_float16
+          const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
+          const int threads,          // number of threads in a threadblock
+          const int thread_m_blocks,  // number of 16x16 blocks in the m
+                                      // dimension (batchsize) of the
+                                      // threadblock
+          const int thread_n_blocks,  // same for n dimension (output)
+          const int thread_k_blocks,  // same for k dimension (reduction)
+          const bool m_block_size_8,  // whether m_block_size == 8
+                                      // only works when thread_m_blocks == 1
+          const int stages,  // number of stages for the async global->shared
+                             // fetch pipeline
+          const bool has_act_order,  // whether act_order is enabled
+          const bool has_zp,         // whether zero-points are enabled
+          const int group_blocks,    // number of consecutive 16x16 blocks
+                                     // with a separate quantization scale
+          const bool is_zp_float     // is zero point of float16 type?
+          >
+__global__ void Marlin(MARLIN_KERNEL_PARAMS);
+
+}

csrc/moe/marlin_moe_wna16/ops.cu

@@ -0,0 +1,927 @@
+/*
+ * Modified by Neural Magic
+ * Copyright (C) Marlin.2024 Elias Frantar
+ *
+ * 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.
+ */
+
+/*
+ * Adapted from https://github.com/IST-DASLab/marlin
+ */
+
+#ifndef MARLIN_NAMESPACE_NAME
+  #define MARLIN_NAMESPACE_NAME marlin_moe_wna16
+#endif
+
+#include "kernel.h"
+#include "core/registration.h"
+
+#define STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t)               \
+  static_assert(std::is_same<scalar_t, half>::value ||          \
+                    std::is_same<scalar_t, nv_bfloat16>::value, \
+                "only float16 and bfloat16 is supported");
+
+namespace MARLIN_NAMESPACE_NAME {
+
+__global__ void MarlinDefault(MARLIN_KERNEL_PARAMS){};
+
+using MarlinFuncPtr = void (*)(MARLIN_KERNEL_PARAMS);
+
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+
+template <int moe_block_size>
+__global__ void permute_cols_kernel(
+    int4 const* __restrict__ a_int4_ptr, int const* __restrict__ perm_int_ptr,
+    int4* __restrict__ out_int4_ptr,
+    const int32_t* __restrict__ sorted_token_ids_ptr,
+    const int32_t* __restrict__ expert_ids_ptr,
+    const int32_t* __restrict__ num_tokens_past_padded_ptr, int size_m,
+    int size_k, int top_k) {};
+
+}  // namespace marlin
+
+torch::Tensor moe_wna16_marlin_gemm(
+    torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
+    torch::Tensor& b_q_weight, torch::Tensor& b_scales,
+    std::optional<torch::Tensor> const& b_zeros_or_none,
+    std::optional<torch::Tensor> const& g_idx_or_none,
+    std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
+    torch::Tensor& sorted_token_ids, torch::Tensor& expert_ids,
+    torch::Tensor& num_tokens_past_padded, torch::Tensor& topk_weights,
+    int64_t moe_block_size, int64_t top_k, bool mul_topk_weights, bool is_ep,
+    vllm::ScalarTypeId const& b_q_type_id, int64_t size_m, int64_t size_n,
+    int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
+    bool is_zp_float) {
+  TORCH_CHECK_NOT_IMPLEMENTED(false,
+                              "marlin_gemm(..) requires CUDA_ARCH >= 8.0");
+  return torch::empty({1, 1});
+}
+
+#else
+
+// For a given "a" of size [M,K] performs a permutation of the K columns based
+// on the given "perm" indices.
+template <int moe_block_size>
+__global__ void permute_cols_kernel(
+    int4 const* __restrict__ a_int4_ptr, int const* __restrict__ perm_int_ptr,
+    int4* __restrict__ out_int4_ptr,
+    const int32_t* __restrict__ sorted_token_ids_ptr,
+    const int32_t* __restrict__ expert_ids_ptr,
+    const int32_t* __restrict__ num_tokens_past_padded_ptr, int size_m,
+    int size_k, int top_k) {
+  int num_tokens_past_padded = num_tokens_past_padded_ptr[0];
+  int num_moe_blocks = div_ceil(num_tokens_past_padded, moe_block_size);
+  int32_t block_sorted_ids[moe_block_size];
+  int block_num_valid_tokens = 0;
+  int64_t old_expert_id = 0;
+  int64_t expert_id = 0;
+  int row_stride = size_k * sizeof(half) / 16;
+
+  auto read_moe_block_data = [&](int block_id) {
+    block_num_valid_tokens = moe_block_size;
+    int4* tmp_block_sorted_ids = reinterpret_cast<int4*>(block_sorted_ids);
+    for (int i = 0; i < moe_block_size / 4; i++) {
+      tmp_block_sorted_ids[i] =
+          ((int4*)sorted_token_ids_ptr)[block_id * moe_block_size / 4 + i];
+    }
+    for (int i = 0; i < moe_block_size; i++) {
+      if (block_sorted_ids[i] >= size_m * top_k) {
+        block_num_valid_tokens = i;
+        break;
+      };
+    }
+  };
+
+  auto permute_row = [&](int row) {
+    int iters = size_k / default_threads;
+    int rest = size_k % default_threads;
+
+    int in_offset = (row / top_k) * row_stride;
+    int out_offset = row * row_stride;
+
+    half const* a_row_half =
+        reinterpret_cast<half const*>(a_int4_ptr + in_offset);
+    half* out_half = reinterpret_cast<half*>(out_int4_ptr + out_offset);
+
+    int base_k = 0;
+
+    for (int i = 0; i < iters; i++) {
+      int cur_k = base_k + threadIdx.x;
+      int src_pos = perm_int_ptr[cur_k];
+
+      out_half[cur_k] = a_row_half[src_pos];
+
+      base_k += default_threads;
+    }
+
+    if (rest) {
+      if (threadIdx.x < rest) {
+        int cur_k = base_k + threadIdx.x;
+        int src_pos = perm_int_ptr[cur_k];
+
+        out_half[cur_k] = a_row_half[src_pos];
+      }
+    }
+  };
+
+  for (int index = blockIdx.x; index < num_moe_blocks; index += gridDim.x) {
+    old_expert_id = expert_id;
+    int tmp_expert_id = expert_ids_ptr[index];
+    if (tmp_expert_id == -1) continue;
+    expert_id = tmp_expert_id;
+    perm_int_ptr += (expert_id - old_expert_id) * size_k;
+    read_moe_block_data(index);
+
+    for (int i = 0; i < block_num_valid_tokens; i++)
+      permute_row(block_sorted_ids[i]);
+  }
+}
+
+typedef struct {
+  int thread_k;
+  int thread_n;
+  int num_threads;
+} thread_config_t;
+
+thread_config_t small_batch_thread_configs[] = {
+    // Ordered by priority
+
+    // thread_k, thread_n, num_threads
+    {128, 128, 256},
+    {64, 128, 128}};
+
+thread_config_t large_batch_thread_configs[] = {
+    // Ordered by priority
+
+    // thread_k, thread_n, num_threads
+    {64, 256, 256},
+    {64, 128, 128}};
+
+typedef struct {
+  int blocks_per_sm;
+  thread_config_t tb_cfg;
+} exec_config_t;
+
+int get_scales_cache_size(thread_config_t const& th_config, int prob_m,
+                          int prob_n, int prob_k, int num_bits, int group_size,
+                          bool has_act_order, bool is_k_full) {
+  bool cache_scales_chunk = has_act_order && !is_k_full;
+
+  int tb_n = th_config.thread_n;
+  int tb_k = th_config.thread_k;
+
+  // Get max scale groups per thread-block
+  int tb_groups;
+  if (group_size == -1) {
+    tb_groups = 1;
+  } else if (group_size == 0) {
+    tb_groups = div_ceil(tb_k, 32);  // Worst case is 32 group size
+  } else {
+    tb_groups = div_ceil(tb_k, group_size);
+  }
+
+  if (cache_scales_chunk) {
+    int load_groups =
+        tb_groups * pipe_stages * 2;     // Chunk size is 2x pipeline over dim K
+    load_groups = max(load_groups, 32);  // We load at least 32 scale groups
+    return load_groups * tb_n * 2;
+
+  } else {
+    int tb_scales = tb_groups * tb_n * 2;
+
+    return tb_scales * pipe_stages;
+  }
+}
+
+int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
+                          int prob_m, int prob_n, int prob_k, int num_bits,
+                          int group_size, bool has_act_order, bool is_k_full,
+                          int has_zp, int is_zp_float) {
+  int pack_factor = 32 / num_bits;
+
+  // Get B size
+  int tb_k = th_config.thread_k;
+  int tb_n = th_config.thread_n;
+  int tb_m = thread_m_blocks * 16;
+
+  // shm size for block_sorted_ids/block_topk_weights
+  // both of them requires tb_m * 4 bytes (tb_m * int32 or tb_m * float32)
+  int sh_block_meta_size = tb_m * 4 * 2;
+  int sh_a_size = pipe_stages * (tb_m * tb_k) * 2;
+  int sh_b_size = pipe_stages * (tb_k * tb_n / pack_factor) * 4;
+  int sh_s_size =
+      get_scales_cache_size(th_config, prob_m, prob_n, prob_k, num_bits,
+                            group_size, has_act_order, is_k_full);
+  int sh_g_idx_size = has_act_order && !is_k_full ? pipe_stages * tb_k / 4 : 0;
+  int sh_zp_size = 0;
+  if (has_zp) {
+    if (is_zp_float)
+      sh_zp_size = sh_s_size;
+    else if (num_bits == 4)
+      sh_zp_size = sh_s_size / 4;
+    else if (num_bits == 8)
+      sh_zp_size = sh_s_size / 2;
+  }
+
+  int total_size = sh_a_size + sh_b_size + sh_s_size + sh_zp_size +
+                   sh_g_idx_size + sh_block_meta_size;
+
+  return total_size;
+}
+
+bool is_valid_config(thread_config_t const& th_config, int thread_m_blocks,
+                     int prob_m, int prob_n, int prob_k, int num_bits,
+                     int group_size, bool has_act_order, bool is_k_full,
+                     int has_zp, int is_zp_float, int max_shared_mem) {
+  // Sanity
+  if (th_config.thread_k == -1 || th_config.thread_n == -1 ||
+      th_config.num_threads == -1) {
+    return false;
+  }
+
+  // Verify K/N are divisible by thread K/N
+  if (prob_k % th_config.thread_k != 0 || prob_n % th_config.thread_n != 0) {
+    return false;
+  }
+
+  // Verify min for thread K/N
+  if (th_config.thread_n < min_thread_n || th_config.thread_k < min_thread_k) {
+    return false;
+  }
+
+  // num_threads must be at least 128 (= 4 warps)
+  if (th_config.num_threads < 128) {
+    return false;
+  }
+
+  // Check that pipeline fits into cache
+  int cache_size = get_kernel_cache_size(
+      th_config, thread_m_blocks, prob_m, prob_n, prob_k, num_bits, group_size,
+      has_act_order, is_k_full, has_zp, is_zp_float);
+  return cache_size <= max_shared_mem;
+}
+
+  #define __GET_IF(W_TYPE, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, \
+                   M_BLOCK_SIZE_8, HAS_ACT_ORDER, HAS_ZP, GROUP_BLOCKS,       \
+                   NUM_THREADS, IS_ZP_FLOAT)                                  \
+    else if (q_type == W_TYPE && thread_m_blocks == THREAD_M_BLOCKS &&        \
+             thread_n_blocks == THREAD_N_BLOCKS &&                            \
+             thread_k_blocks == THREAD_K_BLOCKS &&                            \
+             m_block_size_8 == M_BLOCK_SIZE_8 &&                              \
+             has_act_order == HAS_ACT_ORDER && has_zp == HAS_ZP &&            \
+             group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS &&    \
+             is_zp_float == IS_ZP_FLOAT) {                                    \
+      kernel = Marlin<scalar_t, W_TYPE.id(), NUM_THREADS, THREAD_M_BLOCKS,    \
+                      THREAD_N_BLOCKS, THREAD_K_BLOCKS, M_BLOCK_SIZE_8,       \
+                      pipe_stages, HAS_ACT_ORDER, HAS_ZP, GROUP_BLOCKS,       \
+                      IS_ZP_FLOAT>;                                           \
+    }
+
+  #define GPTQ_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)              \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, true, false, 0, NUM_THREADS, \
+             false)                                                            \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, true, false, 0,             \
+             NUM_THREADS, false)                                               \
+                                                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, -1,            \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, 2,             \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, 4,             \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, 8,             \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, -1,           \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, 2,            \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, 4,            \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, 8,            \
+             NUM_THREADS, false)
+
+  #define GPTQ_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)  \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, true, false, 0,   \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, true, false, 0,   \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, true, false, 0,   \
+             NUM_THREADS, false)                                     \
+                                                                     \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, -1, \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, 2,  \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, 4,  \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, 8,  \
+             NUM_THREADS, false)                                     \
+                                                                     \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, -1, \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, 2,  \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, 4,  \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, 8,  \
+             NUM_THREADS, false)                                     \
+                                                                     \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, -1, \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, 2,  \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, 4,  \
+             NUM_THREADS, false)                                     \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, 8,  \
+             NUM_THREADS, false)
+
+  #define AWQ_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, -1,             \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 2, NUM_THREADS, \
+             false)                                                            \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 4, NUM_THREADS, \
+             false)                                                            \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 8, NUM_THREADS, \
+             false)                                                            \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, -1,            \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 2,             \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
+             NUM_THREADS, false)                                               \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 8,             \
+             NUM_THREADS, false)
+
+  #define AWQ_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)  \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, -1, \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 2,  \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 4,  \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 8,  \
+             NUM_THREADS, false)                                    \
+                                                                    \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, -1, \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 2,  \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 4,  \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 8,  \
+             NUM_THREADS, false)                                    \
+                                                                    \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, -1, \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 2,  \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 4,  \
+             NUM_THREADS, false)                                    \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 8,  \
+             NUM_THREADS, false)
+
+  // We currently have 4-bit models only with group_blocks == 4
+  #define HQQ_GET_IF(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)                  \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 4, NUM_THREADS, \
+             true)                                                             \
+    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
+             NUM_THREADS, true)                                                \
+    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
+             NUM_THREADS, true)                                                \
+    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
+             NUM_THREADS, true)                                                \
+    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
+             NUM_THREADS, true)
+
+template <typename scalar_t>
+MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType q_type,
+                                int thread_m_blocks, int thread_n_blocks,
+                                int thread_k_blocks, bool m_block_size_8,
+                                bool has_act_order, bool has_zp,
+                                int group_blocks, int num_threads,
+                                bool is_zp_float) {
+  int num_bits = q_type.size_bits();
+  auto kernel = MarlinDefault;
+  if (false) {
+  }
+  GPTQ_GET_IF_M1(vllm::kU4B8, 8, 8, 256)
+  GPTQ_GET_IF_M1(vllm::kU4B8, 8, 4, 128)
+
+  GPTQ_GET_IF_M234(vllm::kU4B8, 16, 4, 256)
+  GPTQ_GET_IF_M234(vllm::kU4B8, 8, 4, 128)
+
+  GPTQ_GET_IF_M1(vllm::kU8B128, 8, 8, 256)
+  GPTQ_GET_IF_M1(vllm::kU8B128, 8, 4, 128)
+
+  GPTQ_GET_IF_M234(vllm::kU8B128, 16, 4, 256)
+  GPTQ_GET_IF_M234(vllm::kU8B128, 8, 4, 128)
+
+  AWQ_GET_IF_M1(vllm::kU4, 8, 8, 256)
+  AWQ_GET_IF_M1(vllm::kU4, 8, 4, 128)
+
+  AWQ_GET_IF_M234(vllm::kU4, 16, 4, 256)
+  AWQ_GET_IF_M234(vllm::kU4, 8, 4, 128)
+
+  return kernel;
+}
+
+template <typename scalar_t>
+exec_config_t determine_exec_config(const vllm::ScalarType& q_type, int prob_m,
+                                    int prob_n, int prob_k, int thread_m_blocks,
+                                    bool m_block_size_8, int num_bits,
+                                    int group_size, bool has_act_order,
+                                    bool is_k_full, bool has_zp,
+                                    bool is_zp_float, int max_shared_mem) {
+  exec_config_t exec_cfg = exec_config_t{1, thread_config_t{-1, -1, -1}};
+  thread_config_t* thread_configs = thread_m_blocks > 1
+                                        ? large_batch_thread_configs
+                                        : small_batch_thread_configs;
+  int thread_configs_size =
+      thread_m_blocks > 1
+          ? sizeof(large_batch_thread_configs) / sizeof(thread_config_t)
+          : sizeof(small_batch_thread_configs) / sizeof(thread_config_t);
+
+  int count = 0;
+  constexpr int device_max_reg_size = 255 * 1024;
+  for (int i = 0; i < thread_configs_size; i++) {
+    thread_config_t th_config = thread_configs[i];
+
+    if (!is_valid_config(th_config, thread_m_blocks, prob_m, prob_n, prob_k,
+                         num_bits, group_size, has_act_order, is_k_full, has_zp,
+                         is_zp_float, max_shared_mem)) {
+      continue;
+    }
+
+    int cache_size = get_kernel_cache_size(
+        th_config, thread_m_blocks, prob_m, prob_n, prob_k, num_bits,
+        group_size, has_act_order, is_k_full, has_zp, is_zp_float);
+
+    int group_blocks = 0;
+    if (!has_act_order) {
+      group_blocks = group_size == -1 ? -1 : group_size / 16;
+    }
+
+    auto kernel = get_marlin_kernel<scalar_t>(
+        q_type, thread_m_blocks, th_config.thread_n / 16,
+        th_config.thread_k / 16, m_block_size_8, has_act_order, has_zp,
+        group_blocks, th_config.num_threads, is_zp_float);
+
+    if (kernel == MarlinDefault) continue;
+
+    if (thread_m_blocks > 1) {
+      exec_cfg = {1, th_config};
+      break;
+    } else {
+      cudaFuncAttributes attr;
+      cudaFuncGetAttributes(&attr, kernel);
+      int reg_size = max(attr.numRegs, 1) * th_config.num_threads * 4;
+      int allow_count = min(device_max_reg_size / reg_size,
+                            max_shared_mem / (cache_size + 1024));
+      allow_count = max(min(allow_count, 4), 1);
+      if (allow_count > count) {
+        count = allow_count;
+        exec_cfg = {count, th_config};
+      };
+    }
+  }
+
+  return exec_cfg;
+}
+
+template <typename scalar_t>
+void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
+               void* zp, void* g_idx, void* perm, void* a_tmp,
+               void* sorted_token_ids, void* expert_ids,
+               void* num_tokens_past_padded, void* topk_weights,
+               int moe_block_size, int top_k, bool mul_topk_weights, bool is_ep,
+               int prob_m, int prob_n, int prob_k, void* workspace,
+               vllm::ScalarType const& q_type, bool has_act_order,
+               bool is_k_full, bool has_zp, int num_groups, int group_size,
+               int dev, cudaStream_t stream, int thread_k, int thread_n,
+               int sms, bool use_atomic_add, bool use_fp32_reduce,
+               bool is_zp_float) {
+  int thread_m_blocks = div_ceil(moe_block_size, 16);
+  bool m_block_size_8 = moe_block_size == 8;
+
+  if (has_zp) {
+    TORCH_CHECK(
+        q_type == vllm::kU4 || q_type == vllm::kU8,
+        "q_type must be u4 or u8 when has_zp = True. Got = ", q_type.str());
+  } else {
+    TORCH_CHECK(
+        q_type == vllm::kU4B8 || q_type == vllm::kU8B128,
+        "q_type must be uint4b8 or uint8b128 when has_zp = False. Got = ",
+        q_type.str());
+  }
+
+  TORCH_CHECK(prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [", prob_m,
+              ", ", prob_n, ", ", prob_k, "]");
+
+  int group_blocks = 0;
+  if (has_act_order) {
+    if (is_k_full) {
+      TORCH_CHECK(group_size != -1);
+      group_blocks = group_size / 16;
+      TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
+                  " is not divisible by group_blocks = ", group_blocks);
+    } else {
+      TORCH_CHECK(group_size == 0);
+      group_blocks = 0;
+    }
+  } else {
+    if (group_size == -1) {
+      group_blocks = -1;
+    } else {
+      group_blocks = group_size / 16;
+      TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
+                  " is not divisible by group_blocks = ", group_blocks);
+    }
+  }
+
+  int num_bits = q_type.size_bits();
+  const int4* A_ptr = (const int4*)A;
+  const int4* B_ptr = (const int4*)B;
+  int4* C_ptr = (int4*)C;
+  int4* C_tmp_ptr = (int4*)C_tmp;
+  const int4* s_ptr = (const int4*)s;
+  const int4* zp_ptr = (const int4*)zp;
+  const int* g_idx_ptr = (const int*)g_idx;
+  const int* perm_ptr = (const int*)perm;
+  int4* a_tmp_ptr = (int4*)a_tmp;
+  const int32_t* sorted_token_ids_ptr = (const int32_t*)sorted_token_ids;
+  const int32_t* expert_ids_ptr = (const int32_t*)expert_ids;
+  const int32_t* num_tokens_past_padded_ptr =
+      (const int32_t*)num_tokens_past_padded;
+  const float* topk_weights_ptr = (const float*)topk_weights;
+  int* locks = (int*)workspace;
+
+  if (has_act_order) {
+    // Permute A columns
+    auto kernel = permute_cols_kernel<8>;
+    if (moe_block_size == 8) {
+    } else if (moe_block_size == 16)
+      kernel = permute_cols_kernel<16>;
+    else if (moe_block_size == 32)
+      kernel = permute_cols_kernel<32>;
+    else if (moe_block_size == 48)
+      kernel = permute_cols_kernel<48>;
+    else if (moe_block_size == 64)
+      kernel = permute_cols_kernel<64>;
+    else
+      TORCH_CHECK(false, "unsupported moe_block_size ", moe_block_size);
+
+    // avoid ">>>" being formatted to "> > >"
+    // clang-format off
+    kernel<<<sms, default_threads, 0, stream>>>(
+        A_ptr, perm_ptr, a_tmp_ptr, sorted_token_ids_ptr, expert_ids_ptr,
+        num_tokens_past_padded_ptr, prob_m, prob_k, top_k);
+    // clang-format on
+    A_ptr = a_tmp_ptr;
+    prob_m = prob_m * top_k;
+    top_k = 1;
+
+    // If we have a full K, then we can run the non-act-order version of Marlin
+    // (since the weight rows are reordered by increasing group ids, and by
+    // having a full K, we have full original groups)
+    if (is_k_full) has_act_order = false;
+  }
+
+  int max_shared_mem = 0;
+  cudaDeviceGetAttribute(&max_shared_mem,
+                         cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
+  TORCH_CHECK(max_shared_mem > 0);
+
+  // Set thread config
+  exec_config_t exec_cfg;
+  thread_config_t thread_tfg;
+  if (thread_k != -1 && thread_n != -1) {
+    thread_tfg = thread_config_t{thread_k, thread_n, default_threads};
+    exec_cfg = exec_config_t{1, thread_tfg};
+    TORCH_CHECK(prob_n % thread_n == 0, "prob_n = ", prob_n,
+                " is not divisible by thread_n = ", thread_n);
+    TORCH_CHECK(prob_k % thread_k == 0, "prob_k = ", prob_k,
+                " is not divisible by thread_k = ", thread_k);
+  } else {
+    // Auto config
+    exec_cfg = determine_exec_config<scalar_t>(
+        q_type, prob_m, prob_n, prob_k, thread_m_blocks, m_block_size_8,
+        num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float,
+        max_shared_mem);
+    thread_tfg = exec_cfg.tb_cfg;
+  }
+
+  int num_threads = thread_tfg.num_threads;
+  thread_k = thread_tfg.thread_k;
+  thread_n = thread_tfg.thread_n;
+  int blocks = sms * exec_cfg.blocks_per_sm;
+  if (exec_cfg.blocks_per_sm > 1)
+    max_shared_mem = max_shared_mem / exec_cfg.blocks_per_sm - 1024;
+
+  int thread_k_blocks = thread_k / 16;
+  int thread_n_blocks = thread_n / 16;
+
+  TORCH_CHECK(is_valid_config(thread_tfg, thread_m_blocks, prob_m, prob_n,
+                              prob_k, num_bits, group_size, has_act_order,
+                              is_k_full, has_zp, is_zp_float, max_shared_mem),
+              "Invalid thread config: thread_m_blocks = ", thread_m_blocks,
+              ", thread_k = ", thread_tfg.thread_k,
+              ", thread_n = ", thread_tfg.thread_n,
+              ", num_threads = ", thread_tfg.num_threads, " for MKN = [",
+              prob_m, ", ", prob_k, ", ", prob_n, "] and num_bits = ", num_bits,
+              ", group_size = ", group_size,
+              ", has_act_order = ", has_act_order, ", is_k_full = ", is_k_full,
+              ", has_zp = ", has_zp, ", is_zp_float = ", is_zp_float,
+              ", max_shared_mem = ", max_shared_mem);
+
+  auto kernel = get_marlin_kernel<scalar_t>(
+      q_type, thread_m_blocks, thread_n_blocks, thread_k_blocks, m_block_size_8,
+      has_act_order, has_zp, group_blocks, num_threads, is_zp_float);
+
+  if (kernel == MarlinDefault) {
+    TORCH_CHECK(false, "Unsupported shapes: MNK = [", prob_m, ", ", prob_n,
+                ", ", prob_k, "]", ", has_act_order = ", has_act_order,
+                ", num_groups = ", num_groups, ", group_size = ", group_size,
+                ", thread_m_blocks = ", thread_m_blocks,
+                ", thread_n_blocks = ", thread_n_blocks,
+                ", thread_k_blocks = ", thread_k_blocks,
+                ", num_bits = ", num_bits);
+  }
+
+  cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize,
+                       max_shared_mem);
+  // avoid ">>>" being formatted to "> > >"
+  // clang-format off
+  kernel<<<blocks, num_threads, max_shared_mem, stream>>>(
+      A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, zp_ptr, g_idx_ptr,
+      sorted_token_ids_ptr, expert_ids_ptr, num_tokens_past_padded_ptr,
+      topk_weights_ptr, top_k, mul_topk_weights, is_ep, num_groups, prob_m,
+      prob_n, prob_k, locks, use_atomic_add, use_fp32_reduce);
+  // clang-format on
+}
+
+}  // namespace MARLIN_NAMESPACE_NAME
+
+torch::Tensor moe_wna16_marlin_gemm(
+    torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
+    torch::Tensor& b_q_weight, torch::Tensor& b_scales,
+    std::optional<torch::Tensor> const& b_zeros_or_none,
+    std::optional<torch::Tensor> const& g_idx_or_none,
+    std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
+    torch::Tensor& sorted_token_ids, torch::Tensor& expert_ids,
+    torch::Tensor& num_tokens_past_padded, torch::Tensor& topk_weights,
+    int64_t moe_block_size, int64_t top_k, bool mul_topk_weights, bool is_ep,
+    vllm::ScalarTypeId const& b_q_type_id, int64_t size_m, int64_t size_n,
+    int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
+    bool is_zp_float) {
+  vllm::ScalarType const b_q_type = vllm::ScalarType::from_id(b_q_type_id);
+  int pack_factor = 32 / b_q_type.size_bits();
+
+  if (moe_block_size != 8) {
+    TORCH_CHECK(moe_block_size % 16 == 0,
+                "unsupported moe_block_size=", moe_block_size);
+    TORCH_CHECK(moe_block_size >= 16 && moe_block_size <= 64,
+                "unsupported moe_block_size=", moe_block_size);
+  }
+
+  // Verify A
+  TORCH_CHECK(a.size(0) == size_m, "Shape mismatch: a.size(0) = ", a.size(0),
+              ", size_m = ", size_m);
+  TORCH_CHECK(a.size(1) == size_k, "Shape mismatch: a.size(1) = ", a.size(1),
+              ", size_k = ", size_k);
+
+  // Verify B
+  TORCH_CHECK(
+      size_k % MARLIN_NAMESPACE_NAME::tile_size == 0, "size_k = ", size_k,
+      " is not divisible by tile_size = ", MARLIN_NAMESPACE_NAME::tile_size);
+  TORCH_CHECK((size_k / MARLIN_NAMESPACE_NAME::tile_size) == b_q_weight.size(1),
+              "Shape mismatch: b_q_weight.size(1) = ", b_q_weight.size(1),
+              ", size_k = ", size_k,
+              ", tile_size = ", MARLIN_NAMESPACE_NAME::tile_size);
+  TORCH_CHECK(
+      b_q_weight.size(2) % MARLIN_NAMESPACE_NAME::tile_size == 0,
+      "b_q_weight.size(2) = ", b_q_weight.size(2),
+      " is not divisible by tile_size = ", MARLIN_NAMESPACE_NAME::tile_size);
+  int actual_size_n =
+      (b_q_weight.size(2) / MARLIN_NAMESPACE_NAME::tile_size) * pack_factor;
+  TORCH_CHECK(size_n == actual_size_n, "size_n = ", size_n,
+              ", actual_size_n = ", actual_size_n);
+
+  // Verify device and strides
+  TORCH_CHECK(a.device().is_cuda(), "A is not on GPU");
+  TORCH_CHECK(a.is_contiguous(), "A is not contiguous");
+
+  TORCH_CHECK(b_q_weight.device().is_cuda(), "b_q_weight is not on GPU");
+  TORCH_CHECK(b_q_weight.is_contiguous(), "b_q_weight is not contiguous");
+
+  TORCH_CHECK(b_scales.device().is_cuda(), "b_scales is not on GPU");
+  TORCH_CHECK(b_scales.is_contiguous(), "b_scales is not contiguous");
+
+  // thread_k: `k` size of a thread_tile in `weights` (can usually be left as
+  // auto -1)
+  int thread_k = -1;
+  // thread_n: `n` size of a thread_tile in `weights` (can usually be left as
+  // auto -1)
+  int thread_n = -1;
+  // sms: number of SMs to use for the kernel
+  int sms = -1;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, a.get_device());
+
+  // Alloc buffers
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(a));
+  auto options = torch::TensorOptions().dtype(a.dtype()).device(a.device());
+  torch::Tensor c;
+  if (c_or_none.has_value()) {
+    c = c_or_none.value();
+    TORCH_CHECK(c.device().is_cuda(), "c is not on GPU");
+    TORCH_CHECK(c.is_contiguous(), "c is not contiguous");
+    TORCH_CHECK(c.size(0) == size_m * top_k,
+                "Shape mismatch: c.size(0) = ", c.size(0),
+                ", size_m * topk = ", size_m * top_k);
+    TORCH_CHECK(c.size(1) == size_n, "Shape mismatch: c.size(1) = ", c.size(1),
+                ", size_n = ", size_n);
+  } else {
+    c = torch::empty({size_m * top_k, size_n}, options);
+  }
+
+  // Alloc C tmp buffer that is going to be used for the global reduce
+  torch::Tensor c_tmp;
+  auto options_fp32 =
+      torch::TensorOptions().dtype(at::kFloat).device(a.device());
+  if (use_fp32_reduce && !use_atomic_add) {
+    // max num of threadblocks is sms * 4
+    long max_c_tmp_size = min(
+        (long)size_n * sorted_token_ids.size(0),
+        (long)sms * 4 * moe_block_size * MARLIN_NAMESPACE_NAME::max_thread_n);
+    if (moe_block_size == 8) max_c_tmp_size *= 2;
+    c_tmp = torch::empty({max_c_tmp_size}, options_fp32);
+  } else {
+    c_tmp = torch::empty({0}, options_fp32);
+  }
+
+  // Detect groupsize and act_order
+  int num_groups = -1;
+  int group_size = -1;
+
+  int rank = b_scales.sizes().size();
+  TORCH_CHECK(rank == 3, "b_scales rank = ", rank, " is not 3");
+  TORCH_CHECK(b_scales.size(2) == size_n, "b_scales dim 2 = ", b_scales.size(2),
+              " is not size_n = ", size_n);
+  num_groups = b_scales.size(1);
+
+  torch::Tensor g_idx, perm, a_tmp;
+  ;
+  if (g_idx_or_none.has_value() && perm_or_none.has_value()) {
+    g_idx = g_idx_or_none.value();
+    perm = perm_or_none.value();
+
+    TORCH_CHECK(g_idx.device().is_cuda(), "g_idx is not on GPU");
+    TORCH_CHECK(g_idx.is_contiguous(), "g_idx is not contiguous");
+    TORCH_CHECK(perm.device().is_cuda(), "perm is not on GPU");
+    TORCH_CHECK(perm.is_contiguous(), "perm is not contiguous");
+
+    // Verify g_idx and perm
+    TORCH_CHECK((g_idx.size(-1) == 0 && perm.size(-1) == 0) ||
+                    (g_idx.size(-1) == size_k && perm.size(-1) == size_k),
+                "Unexpected g_idx.size(-1) = ", g_idx.size(-1),
+                " and perm.size(-1) = ", perm.size(-1),
+                ", where size_k = ", size_k);
+  } else {
+    g_idx = torch::empty({0}, options);
+    perm = torch::empty({0}, options);
+    a_tmp = torch::empty({0}, options);
+  }
+  bool has_act_order = g_idx.size(-1) > 0 && perm.size(-1) > 0;
+
+  if (has_act_order) {
+    a_tmp = torch::empty({size_m * top_k, size_k}, options);
+    if (is_k_full) {
+      TORCH_CHECK(num_groups > 1, "For act_order, num_groups must be > 1");
+      TORCH_CHECK(size_k % num_groups == 0, "size_k = ", size_k,
+                  ", is not divisible by num_groups = ", num_groups);
+      group_size = size_k / num_groups;
+    } else {
+      group_size = 0;
+    }
+
+  } else {
+    a_tmp = torch::empty({0}, options);
+    if (num_groups > 1) {
+      TORCH_CHECK(
+          size_k % num_groups == 0, "size_k = ", size_k,
+          ", is not divisible by b_scales.size(1) = ", b_scales.size(1));
+      group_size = size_k / num_groups;
+    } else {
+      group_size = -1;
+    }
+  }
+
+  torch::Tensor b_zeros;
+  if (b_zeros_or_none.has_value()) {
+    b_zeros = b_zeros_or_none.value();
+    TORCH_CHECK(b_zeros.device().is_cuda(), "b_zeros is not on GPU");
+    TORCH_CHECK(b_zeros.is_contiguous(), "b_zeros is not contiguous");
+  } else {
+    b_zeros = torch::empty({0}, options);
+  }
+  bool has_zp = b_zeros.size(-1) > 0;
+
+  if (has_zp) {
+    TORCH_CHECK(
+        b_q_type == vllm::kU4,
+        "b_q_type must be u4 when has_zp = True. Got = ", b_q_type.str());
+  } else {
+    TORCH_CHECK(
+        b_q_type == vllm::kU4B8 || b_q_type == vllm::kU8B128,
+        "b_q_type must be uint4b8 or uint8b128 when has_zp = False. Got = ",
+        b_q_type.str());
+  }
+
+  if (has_zp && is_zp_float) {
+    TORCH_CHECK(a.scalar_type() == at::ScalarType::Half,
+                "Computation type must be float16 (half) when using float zero "
+                "points.");
+  }
+
+  // Verify b_zeros
+  if (has_zp) {
+    int rank = b_zeros.sizes().size();
+    TORCH_CHECK(rank == 3, "b_zeros rank = ", rank, " is not 3");
+    if (is_zp_float) {
+      TORCH_CHECK(b_zeros.size(2) == size_n,
+                  "b_zeros dim 2 = ", b_zeros.size(2),
+                  " is not size_n = ", size_n);
+      TORCH_CHECK(num_groups == b_zeros.size(1),
+                  "b_zeros dim 1 = ", b_zeros.size(1),
+                  " is not num_groups = ", num_groups);
+      TORCH_CHECK(num_groups != -1, "num_groups must be != -1");
+    } else {
+      TORCH_CHECK(b_zeros.size(1) == num_groups,
+                  "b_zeros dim 1 = ", b_zeros.size(1),
+                  " is not num_groups = ", num_groups);
+      TORCH_CHECK(b_zeros.size(2) == size_n / pack_factor,
+                  "b_zeros dim 2 = ", b_zeros.size(2),
+                  " is not size_n / pack_factor = ", size_n / pack_factor);
+    }
+  }
+
+  // Verify workspace size
+  TORCH_CHECK(size_n % MARLIN_NAMESPACE_NAME::min_thread_n == 0,
+              "size_n = ", size_n, ", is not divisible by min_thread_n = ",
+              MARLIN_NAMESPACE_NAME::min_thread_n);
+
+  int max_n_tiles = size_n / MARLIN_NAMESPACE_NAME::min_thread_n;
+  int min_workspace_size = min(
+      max_n_tiles * (int)(sorted_token_ids.size(0) / moe_block_size), sms * 4);
+  TORCH_CHECK(workspace.numel() >= min_workspace_size,
+              "workspace.numel = ", workspace.numel(),
+              " is below min_workspace_size = ", min_workspace_size);
+
+  int dev = a.get_device();
+  if (a.scalar_type() == at::ScalarType::Half) {
+    MARLIN_NAMESPACE_NAME::marlin_mm<half>(
+        a.data_ptr<at::Half>(), b_q_weight.data_ptr(), c.data_ptr<at::Half>(),
+        c_tmp.data_ptr<float>(), b_scales.data_ptr<at::Half>(),
+        b_zeros.data_ptr(), g_idx.data_ptr(), perm.data_ptr(),
+        a_tmp.data_ptr<at::Half>(), sorted_token_ids.data_ptr(),
+        expert_ids.data_ptr(), num_tokens_past_padded.data_ptr(),
+        topk_weights.data_ptr(), moe_block_size, top_k, mul_topk_weights, is_ep,
+        size_m, size_n, size_k, workspace.data_ptr(), b_q_type, has_act_order,
+        is_k_full, has_zp, num_groups, group_size, dev,
+        at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
+        use_atomic_add, use_fp32_reduce, is_zp_float);
+  } else if (a.scalar_type() == at::ScalarType::BFloat16) {
+    MARLIN_NAMESPACE_NAME::marlin_mm<nv_bfloat16>(
+        a.data_ptr<at::BFloat16>(), b_q_weight.data_ptr(),
+        c.data_ptr<at::BFloat16>(), c_tmp.data_ptr<float>(),
+        b_scales.data_ptr<at::BFloat16>(), b_zeros.data_ptr(), g_idx.data_ptr(),
+        perm.data_ptr(), a_tmp.data_ptr<at::BFloat16>(),
+        sorted_token_ids.data_ptr(), expert_ids.data_ptr(),
+        num_tokens_past_padded.data_ptr(), topk_weights.data_ptr(),
+        moe_block_size, top_k, mul_topk_weights, is_ep, size_m, size_n, size_k,
+        workspace.data_ptr(), b_q_type, has_act_order, is_k_full, has_zp,
+        num_groups, group_size, dev, at::cuda::getCurrentCUDAStream(dev),
+        thread_k, thread_n, sms, use_atomic_add, use_fp32_reduce, is_zp_float);
+  } else {
+    TORCH_CHECK(false,
+                "moe_wna16_marlin_gemm only supports bfloat16 and float16");
+  }
+
+  return c;
+}
+
+#endif
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("moe_wna16_marlin_gemm", &moe_wna16_marlin_gemm);
+}

csrc/moe/moe_wna16.cu

@@ -13,7 +13,6 @@
 template <typename scalar_t, int bit, int GROUPS>
 __global__ void moe_wna16_gemm_kernel(
     const scalar_t* __restrict__ input, scalar_t* __restrict__ output,
-
     const uint32_t* __restrict__ qweight, const scalar_t* __restrict__ scales,
     const uint32_t* __restrict__ qzeros,
 
@@ -54,8 +53,6 @@ __global__ void moe_wna16_gemm_kernel(
       if (token_index / top_k >= size_m) break;
 
       num_valid_tokens = m + 1;
-      if (blockIdx.z == 0 && offset_n < size_n)
-        output[token_index * size_n + offset_n] = Dtype::int2num(0);
 
       if (expert_id != -1) {
         int k_per_thread = DIVIDE(BLOCK_SIZE_K, BLOCK_SIZE_N);
@@ -284,8 +281,7 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
                              int64_t BLOCK_SIZE_K, int64_t bit) {
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
-  auto options =
-      torch::TensorOptions().dtype(input.dtype()).device(input.device());
+  output.zero_();
 
   const int num_experts = b_qweight.size(0);
   const int size_m = input.size(0);
@@ -302,9 +298,9 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
   const uint32_t* b_qzeros_ptr;
   if (b_qzeros.has_value())
     b_qzeros_ptr = (const uint32_t*)b_qzeros.value().data_ptr<uint8_t>();
-  const float* topk_weights_ptr;
+  const float* topk_weights_ptr = nullptr;
   if (topk_weights.has_value())
-    topk_weights_ptr = (const float*)topk_weights.value().data_ptr();
+    topk_weights_ptr = (const float*)topk_weights.value().data_ptr<float>();
 
   int groups_per_block_row = BLOCK_SIZE_K / group_size;
   TORCH_CHECK(bit == 4 || bit == 8, "bit must be 4 or 8");

csrc/moe/torch_bindings.cpp

@@ -43,14 +43,17 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
   m.impl("moe_wna16_gemm", torch::kCUDA, &moe_wna16_gemm);
 
   m.def(
-      "marlin_gemm_moe(Tensor! a, Tensor! b_q_weights, Tensor! sorted_ids, "
-      "Tensor! topk_weights, Tensor! topk_ids, Tensor! b_scales, Tensor! "
-      "b_zeros, Tensor! g_idx, Tensor! perm, Tensor! workspace, "
-      "int b_q_type, SymInt size_m, "
-      "SymInt size_n, SymInt size_k, bool is_k_full, int num_experts, int "
-      "topk, "
-      "int moe_block_size, bool replicate_input, bool apply_weights)"
-      " -> Tensor");
+      "moe_wna16_marlin_gemm(Tensor! a, Tensor? c_or_none,"
+      "Tensor! b_q_weight, Tensor! b_scales, Tensor? b_zeros_or_none,"
+      "Tensor? g_idx_or_none, Tensor? perm_or_none, Tensor! workspace,"
+      "Tensor sorted_token_ids,"
+      "Tensor! expert_ids, Tensor! num_tokens_past_padded,"
+      "Tensor! topk_weights, int moe_block_size, int top_k, "
+      "bool mul_topk_weights, bool is_ep, int b_q_type_id,"
+      "int size_m, int size_n, int size_k,"
+      "bool is_full_k, bool use_atomic_add,"
+      "bool use_fp32_reduce, bool is_zp_float) -> Tensor");
+
   // conditionally compiled so impl registration is in source file
 
 #endif

csrc/ops.h

@@ -52,6 +52,15 @@ void paged_attention_v2(
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step);
 
+#ifndef USE_ROCM
+void merge_attn_states(torch::Tensor& output,
+                       std::optional<torch::Tensor> output_lse,
+                       const torch::Tensor& prefix_output,
+                       const torch::Tensor& prefix_lse,
+                       const torch::Tensor& suffix_output,
+                       const torch::Tensor& suffix_lse);
+#endif
+
 void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
               double epsilon);
 
@@ -119,6 +128,8 @@ void advance_step_flashinfer(
     torch::Tensor& paged_kv_indices, torch::Tensor& paged_kv_indptr,
     torch::Tensor& paged_kv_last_page_len, torch::Tensor& block_table_bounds);
 
+torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor);
+
 #ifndef USE_ROCM
 torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
                         const torch::Tensor& codebooks,
@@ -143,7 +154,8 @@ torch::Tensor permute_cols(torch::Tensor const& A, torch::Tensor const& perm);
 #endif
 
 torch::Tensor ggml_dequantize(torch::Tensor W, int64_t type, int64_t m,
-                              int64_t n);
+                              int64_t n,
+                              std::optional<at::ScalarType> const& dtype);
 
 torch::Tensor ggml_mul_mat_vec_a8(torch::Tensor W, torch::Tensor X,
                                   int64_t type, int64_t row);
@@ -164,6 +176,7 @@ int64_t ggml_moe_get_block_size(int64_t type);
 bool cutlass_scaled_mm_supports_fp4(int64_t cuda_device_capability);
 bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
 bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability);
+bool cutlass_group_gemm_supported(int64_t cuda_device_capability);
 
 void cutlass_scaled_fp4_mm(torch::Tensor& D, torch::Tensor const& A,
                            torch::Tensor const& B, torch::Tensor const& A_sf,
@@ -175,6 +188,19 @@ void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,
                        torch::Tensor const& b_scales,
                        std::optional<torch::Tensor> const& bias);
 
+void cutlass_moe_mm(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides);
+
+void get_cutlass_moe_mm_data(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k);
+
 void cutlass_scaled_mm_azp(torch::Tensor& out, torch::Tensor const& a,
                            torch::Tensor const& b,
                            torch::Tensor const& a_scales,
@@ -251,10 +277,10 @@ void causal_conv1d_fwd(const at::Tensor& x, const at::Tensor& weight,
                        const std::optional<at::Tensor>& has_initial_state,
                        bool silu_activation, int64_t pad_slot_id);
 
-#ifndef USE_ROCM
 using fptr_t = int64_t;
 fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
-                      torch::Tensor& rank_data, int64_t rank, bool full_nvlink);
+                      torch::Tensor& rank_data, int64_t rank,
+                      bool fully_connected);
 void all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
                 fptr_t reg_buffer, int64_t reg_buffer_sz_bytes);
 void dispose(fptr_t _fa);
@@ -265,4 +291,7 @@ get_graph_buffer_ipc_meta(fptr_t _fa);
 void register_graph_buffers(fptr_t _fa,
                             const std::vector<std::vector<int64_t>>& handles,
                             const std::vector<std::vector<int64_t>>& offsets);
-#endif
+std::tuple<int64_t, torch::Tensor> allocate_shared_buffer_and_handle(
+    int64_t size);
+int64_t open_mem_handle(torch::Tensor& mem_handle);
+void free_shared_buffer(int64_t buffer);

csrc/quantization/cutlass_w8a8/moe/get_group_starts.cuh

@@ -0,0 +1,80 @@
+#pragma once
+
+#include <cuda.h>
+#include <torch/all.h>
+#include <c10/cuda/CUDAStream.h>
+
+#include "core/scalar_type.hpp"
+#include "cutlass/bfloat16.h"
+#include "cutlass/float8.h"
+
+template <typename ElementAB, typename ElementC, typename ElementAccumulator>
+__global__ void get_group_gemm_starts(
+    int32_t* expert_offsets, ElementAB** a_offsets, ElementAB** b_offsets,
+    ElementC** out_offsets, ElementAccumulator** a_scales_offsets,
+    ElementAccumulator** b_scales_offsets, ElementAB* a_base_as_int,
+    ElementAB* b_base_as_int, ElementC* out_base_as_int,
+    ElementAccumulator* a_scales_base_as_int,
+    ElementAccumulator* b_scales_base_as_int, int64_t n, int64_t k,
+    bool per_act_token, bool per_out_ch) {
+  int expert_id = threadIdx.x;
+
+  int64_t expert_offset = expert_offsets[expert_id];
+
+  a_offsets[expert_id] = a_base_as_int + expert_offset * k;
+  b_offsets[expert_id] = b_base_as_int + expert_id * k * n;
+  out_offsets[expert_id] = out_base_as_int + expert_offset * n;
+  a_scales_offsets[expert_id] =
+      a_scales_base_as_int + (per_act_token ? expert_offset : 0);
+  b_scales_offsets[expert_id] =
+      b_scales_base_as_int + (per_out_ch ? n * expert_id : expert_id);
+}
+
+#define __CALL_GET_STARTS_KERNEL(TENSOR_C_TYPE, C_TYPE)                    \
+  else if (out_tensors.dtype() == TENSOR_C_TYPE) {                         \
+    get_group_gemm_starts<cutlass::float_e4m3_t, C_TYPE, float>            \
+        <<<1, num_experts, 0, stream>>>(                                   \
+            static_cast<int32_t*>(expert_offsets.data_ptr()),              \
+            static_cast<cutlass::float_e4m3_t**>(a_ptrs.data_ptr()),       \
+            static_cast<cutlass::float_e4m3_t**>(b_ptrs.data_ptr()),       \
+            static_cast<C_TYPE**>(out_ptrs.data_ptr()),                    \
+            static_cast<float**>(a_scales_ptrs.data_ptr()),                \
+            static_cast<float**>(b_scales_ptrs.data_ptr()),                \
+            static_cast<cutlass::float_e4m3_t*>(a_tensors.data_ptr()),     \
+            static_cast<cutlass::float_e4m3_t*>(b_tensors.data_ptr()),     \
+            static_cast<C_TYPE*>(out_tensors.data_ptr()),                  \
+            static_cast<float*>(a_scales.data_ptr()),                      \
+            static_cast<float*>(b_scales.data_ptr()), out_tensors.size(1), \
+            a_tensors.size(1), per_act_token, per_out_ch);                 \
+  }
+
+namespace {
+
+void run_get_group_gemm_starts(
+    torch::Tensor const& expert_offsets, torch::Tensor& a_ptrs,
+    torch::Tensor& b_ptrs, torch::Tensor& out_ptrs,
+    torch::Tensor& a_scales_ptrs, torch::Tensor& b_scales_ptrs,
+    torch::Tensor const& a_tensors, torch::Tensor const& b_tensors,
+    torch::Tensor& out_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales) {
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
+  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+
+  int num_experts = static_cast<int>(expert_offsets.size(0));
+  bool per_act_token = a_scales.numel() != 1;
+  bool per_out_ch = b_scales.numel() != num_experts;
+
+  auto stream = at::cuda::getCurrentCUDAStream(a_tensors.device().index());
+
+  if (false) {
+  }
+  __CALL_GET_STARTS_KERNEL(torch::kBFloat16, cutlass::bfloat16_t)
+  __CALL_GET_STARTS_KERNEL(torch::kFloat16, half)
+  else {
+    TORCH_CHECK(false, "Invalid output type (must be float16 or bfloat16)");
+  }
+}
+
+}  // namespace

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu

@@ -0,0 +1,160 @@
+#include <cudaTypedefs.h>
+
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/all.h>
+
+#include "cutlass/cutlass.h"
+#include "grouped_mm_c3x.cuh"
+
+using namespace cute;
+
+namespace {
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_default {
+  // M in (16, inf)
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_64, cute::_256, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_2, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_M16 {
+  // M in [1, 16]
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_64, cute::_64, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_4, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_K8192 {
+  // K in [8192, inf)
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_128, cute::_128, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_8, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_N8192 {
+  // N in [8192, inf)
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_64, cute::_128, cute::_256>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_8, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType>
+void run_cutlass_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  TORCH_CHECK(a_tensors.size(0) > 0, "No input A tensors provided.");
+  TORCH_CHECK(b_tensors.size(0) > 0, "No input B tensors provided.");
+  TORCH_CHECK(out_tensors.size(0) > 0, "No output tensors provided.");
+
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn,
+              "A tensors must be of type float8_e4m3fn.");
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn,
+              "B tensors must be of type float8_e4m3fn.");
+
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn);
+
+  using Cutlass3xGemmN8192 = typename sm90_fp8_config_N8192<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+  using Cutlass3xGemmK8192 = typename sm90_fp8_config_K8192<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+  using Cutlass3xGemmM16 = typename sm90_fp8_config_M16<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+  using Cutlass3xGemmDefault = typename sm90_fp8_config_default<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+
+  uint32_t const m = a_tensors.size(0);
+  uint32_t const n = out_tensors.size(1);
+  uint32_t const k = a_tensors.size(1);
+
+  if (n >= 8192) {
+    cutlass_group_gemm_caller<Cutlass3xGemmN8192>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else if (k >= 8192) {
+    cutlass_group_gemm_caller<Cutlass3xGemmK8192>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else if (m <= 16) {
+    cutlass_group_gemm_caller<Cutlass3xGemmM16>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else {
+    cutlass_group_gemm_caller<Cutlass3xGemmDefault>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  }
+}
+
+void dispatch_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  if (out_tensors.dtype() == torch::kBFloat16) {
+    run_cutlass_moe_mm_sm90<cutlass::float_e4m3_t, cutlass::bfloat16_t>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else {
+    run_cutlass_moe_mm_sm90<cutlass::float_e4m3_t, cutlass::half_t>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  }
+}
+
+}  // namespace
+
+void cutlass_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  dispatch_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
+                       expert_offsets, problem_sizes, a_strides, b_strides,
+                       c_strides);
+}

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cuh

@@ -0,0 +1,149 @@
+#pragma once
+
+#include "cutlass/cutlass.h"
+
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+
+#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
+#include "cutlass_extensions/common.hpp"
+#include "get_group_starts.cuh"
+
+using namespace cute;
+
+namespace {
+
+using ProblemShape =
+    cutlass::gemm::GroupProblemShape<cute::Shape<int, int, int>>;
+
+using ElementAccumulator = float;
+using ArchTag = cutlass::arch::Sm90;
+using OperatorClass = cutlass::arch::OpClassTensorOp;
+
+using LayoutA = cutlass::layout::RowMajor;
+using LayoutB = cutlass::layout::ColumnMajor;
+using LayoutC = cutlass::layout::RowMajor;
+
+template <typename ElementAB_, typename ElementC_,
+          template <typename, typename, typename> typename Epilogue_,
+          typename TileShape, typename ClusterShape, typename KernelSchedule,
+          typename EpilogueSchedule>
+struct cutlass_3x_group_gemm {
+  using ElementAB = ElementAB_;
+  using ElementC = void;
+  using ElementD = ElementC_;
+  using ElementAccumulator = float;
+
+  using Epilogue = Epilogue_<ElementAccumulator, ElementD, TileShape>;
+
+  using StrideC =
+      cute::remove_pointer_t<cute::Stride<int64_t, cute::Int<1>, cute::Int<0>>>;
+
+  static constexpr int AlignmentAB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+  static constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementD>::value;
+
+  using EVTCompute = typename Epilogue::EVTCompute;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, TileShape, ClusterShape,
+          cutlass::epilogue::collective::EpilogueTileAuto, ElementAccumulator,
+          ElementAccumulator, ElementC, LayoutC*, AlignmentC, ElementD,
+          LayoutC*, AlignmentC, EpilogueSchedule, EVTCompute>::CollectiveOp;
+
+  static constexpr size_t CEStorageSize =
+      sizeof(typename CollectiveEpilogue::SharedStorage);
+  using Stages = typename cutlass::gemm::collective::StageCountAutoCarveout<
+      static_cast<int>(CEStorageSize)>;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, ElementAB, LayoutA*, AlignmentAB, ElementAB,
+          LayoutB*, AlignmentAB, ElementAccumulator, TileShape, ClusterShape,
+          Stages, KernelSchedule>::CollectiveOp;
+
+  using KernelType = enable_sm90_only<cutlass::gemm::kernel::GemmUniversal<
+      ProblemShape, CollectiveMainloop, CollectiveEpilogue>>;
+
+  struct GemmKernel : public KernelType {};
+};
+
+template <typename Gemm>
+void cutlass_group_gemm_caller(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  using ElementAB = typename Gemm::ElementAB;
+  using ElementD = typename Gemm::ElementD;
+
+  int num_experts = static_cast<int>(expert_offsets.size(0));
+  int k_size = a_tensors.size(1);
+  int n_size = out_tensors.size(1);
+
+  bool per_act_token = a_scales.numel() != 1;
+  bool per_out_ch = b_scales.numel() != num_experts;
+
+  auto stream = at::cuda::getCurrentCUDAStream(a_tensors.device().index());
+
+  auto options_int =
+      torch::TensorOptions().dtype(torch::kInt64).device(a_tensors.device());
+
+  torch::Tensor a_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor b_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor out_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor a_scales_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor b_scales_ptrs = torch::empty(num_experts, options_int);
+
+  run_get_group_gemm_starts(expert_offsets, a_ptrs, b_ptrs, out_ptrs,
+                            a_scales_ptrs, b_scales_ptrs, a_tensors, b_tensors,
+                            out_tensors, a_scales, b_scales);
+
+  using GemmKernel = typename Gemm::GemmKernel;
+  using StrideA = Stride<int64_t, Int<1>, Int<0>>;
+  using StrideB = Stride<int64_t, Int<1>, Int<0>>;
+  using StrideC = typename GemmKernel::InternalStrideC;
+
+  ProblemShape::UnderlyingProblemShape* problem_sizes_as_shapes =
+      static_cast<ProblemShape::UnderlyingProblemShape*>(
+          problem_sizes.data_ptr());
+  ProblemShape prob_shape{num_experts, problem_sizes_as_shapes, nullptr};
+
+  typename GemmKernel::MainloopArguments mainloop_args{
+      static_cast<const ElementAB**>(a_ptrs.data_ptr()),
+      static_cast<StrideA*>(a_strides.data_ptr()),
+      static_cast<const ElementAB**>(b_ptrs.data_ptr()),
+      static_cast<StrideB*>(b_strides.data_ptr())};
+
+  // Currently, we are only able to do broadcast on either all or none a_scales
+  // and on either all or none b_scales
+  typename GemmKernel::EpilogueArguments epilogue_args{
+      Gemm::Epilogue::prepare_args(
+          static_cast<const ElementAccumulator**>(a_scales_ptrs.data_ptr()),
+          static_cast<const ElementAccumulator**>(b_scales_ptrs.data_ptr()),
+          per_act_token, per_out_ch),
+      nullptr, static_cast<StrideC*>(c_strides.data_ptr()),
+      static_cast<ElementD**>(out_ptrs.data_ptr()),
+      static_cast<StrideC*>(c_strides.data_ptr())};
+
+  typename GemmKernel::Arguments args{
+      cutlass::gemm::GemmUniversalMode::kGrouped, prob_shape, mainloop_args,
+      epilogue_args};
+
+  using GemmOp = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  GemmOp gemm_op;
+  CUTLASS_CHECK(gemm_op.can_implement(args));
+
+  size_t workspace_size = gemm_op.get_workspace_size(args);
+  auto const workspace_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(a_tensors.device());
+  auto workspace = torch::empty(workspace_size, workspace_options);
+
+  cutlass::Status status = gemm_op.run(args, workspace.data_ptr(), stream);
+  CUTLASS_CHECK(status);
+}
+
+}  // namespace

csrc/quantization/cutlass_w8a8/moe/moe_data.cu

@@ -0,0 +1,90 @@
+#include <cudaTypedefs.h>
+
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/all.h>
+
+#include <iostream>
+
+constexpr uint64_t THREADS_PER_EXPERT = 512;
+
+__global__ void compute_problem_sizes(const int* __restrict__ topk_ids,
+                                      int32_t* problem_sizes1,
+                                      int32_t* problem_sizes2,
+                                      int32_t* atomic_buffer,
+                                      const int topk_length, const int n,
+                                      const int k) {
+  int expert_id = blockIdx.x;
+
+  int occurrences = 0;
+  for (int i = threadIdx.x; i < topk_length; i += THREADS_PER_EXPERT) {
+    occurrences += (topk_ids[i] == expert_id);
+  }
+  atomicAdd(&atomic_buffer[expert_id], occurrences);
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+    int final_occurrences = atomic_buffer[expert_id];
+    problem_sizes1[expert_id * 3] = final_occurrences;
+    problem_sizes1[expert_id * 3 + 1] = 2 * n;
+    problem_sizes1[expert_id * 3 + 2] = k;
+    problem_sizes2[expert_id * 3] = final_occurrences;
+    problem_sizes2[expert_id * 3 + 1] = k;
+    problem_sizes2[expert_id * 3 + 2] = n;
+  }
+}
+
+__global__ void compute_expert_offsets(
+    const int32_t* __restrict__ problem_sizes1, int32_t* expert_offsets,
+    int32_t* atomic_buffer, const int num_experts) {
+  int32_t tot_offset = 0;
+  expert_offsets[0] = 0;
+  for (int i = 0; i < num_experts; ++i) {
+    atomic_buffer[i] = tot_offset;
+    tot_offset += problem_sizes1[i * 3];
+    expert_offsets[i + 1] = tot_offset;
+  }
+}
+
+__global__ void compute_arg_sorts(const int* __restrict__ topk_ids,
+                                  int32_t* input_permutation,
+                                  int32_t* output_permutation,
+                                  int32_t* atomic_buffer, const int topk_length,
+                                  const int topk) {
+  int expert_id = blockIdx.x;
+
+  for (int i = threadIdx.x; i < topk_length; i += THREADS_PER_EXPERT) {
+    if (topk_ids[i] == expert_id) {
+      int start = atomicAdd(&atomic_buffer[expert_id], 1);
+      input_permutation[start] = i / topk;
+      output_permutation[i] = start;
+    }
+  }
+}
+
+void get_cutlass_moe_mm_data_caller(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k) {
+  auto stream = at::cuda::getCurrentCUDAStream(topk_ids.device().index());
+  auto options_int32 =
+      torch::TensorOptions().dtype(torch::kInt32).device(topk_ids.device());
+  torch::Tensor atomic_buffer = torch::zeros(num_experts, options_int32);
+
+  int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
+  compute_problem_sizes<<<num_experts, num_threads, 0, stream>>>(
+      static_cast<const int32_t*>(topk_ids.data_ptr()),
+      static_cast<int32_t*>(problem_sizes1.data_ptr()),
+      static_cast<int32_t*>(problem_sizes2.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n, k);
+  compute_expert_offsets<<<1, 1, 0, stream>>>(
+      static_cast<const int32_t*>(problem_sizes1.data_ptr()),
+      static_cast<int32_t*>(expert_offsets.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
+  compute_arg_sorts<<<num_experts, num_threads, 0, stream>>>(
+      static_cast<const int32_t*>(topk_ids.data_ptr()),
+      static_cast<int32_t*>(input_permutation.data_ptr()),
+      static_cast<int32_t*>(output_permutation.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(),
+      topk_ids.size(1));
+}

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -29,6 +29,20 @@ void cutlass_scaled_mm_sm90(torch::Tensor& c, torch::Tensor const& a,
                             torch::Tensor const& a_scales,
                             torch::Tensor const& b_scales,
                             std::optional<torch::Tensor> const& bias);
+
+void cutlass_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides);
+
+void get_cutlass_moe_mm_data_caller(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k);
+
 #endif
 
 #if defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM100
@@ -102,6 +116,19 @@ bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) {
   return false;
 }
 
+bool cutlass_group_gemm_supported(int64_t cuda_device_capability) {
+  // CUTLASS groped FP8 kernels need at least CUDA 12.3
+  // and SM90 (Hopper)
+
+#if defined CUDA_VERSION
+  if (cuda_device_capability == 90) {
+    return CUDA_VERSION >= 12030;
+  }
+#endif
+
+  return false;
+}
+
 void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
                        torch::Tensor const& b, torch::Tensor const& a_scales,
                        torch::Tensor const& b_scales,
@@ -168,6 +195,46 @@ void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
       version_num);
 }
 
+void cutlass_moe_mm(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  int32_t version_num = get_sm_version_num();
+#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
+  cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
+                      expert_offsets, problem_sizes, a_strides, b_strides,
+                      c_strides);
+  return;
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      false,
+      "No compiled cutlass_scaled_mm for CUDA device capability: ", version_num,
+      ". Required capability: 90");
+}
+
+void get_cutlass_moe_mm_data(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k) {
+  // This function currently gets compiled only if we have a valid cutlass moe
+  // mm to run it for.
+  int32_t version_num = get_sm_version_num();
+#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
+  get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1,
+                                 problem_sizes2, input_permutation,
+                                 output_permutation, num_experts, n, k);
+  return;
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      false,
+      "No compiled get_cutlass_moe_mm_data: no cutlass_scaled_mm kernel for "
+      "CUDA device capability: ",
+      version_num, ". Required capability: 90");
+}
+
 void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a,
                            torch::Tensor const& b,
                            torch::Tensor const& a_scales,

csrc/quantization/fp8/common.cu

@@ -30,9 +30,6 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel(
     fp8_type* __restrict__ out, float* __restrict__ scale,
     scalar_t const* __restrict__ input, float const* __restrict__ scale_ub,
     const int hidden_size) {
-  float const min_scaling_factor =
-      1.0f / (fp8_e4m3_adjusted_max_v<fp8_type> * 512.f);
-
   int const tid = threadIdx.x;
   int const token_idx = blockIdx.x;
 
@@ -67,8 +64,8 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel(
       token_scale = block_absmax_val_maybe;
     }
     // token scale computation
-    token_scale = max(token_scale / fp8_e4m3_adjusted_max_v<fp8_type>,
-                      min_scaling_factor);
+    token_scale = max(token_scale / quant_type_max_v<fp8_type>,
+                      min_scaling_factor<fp8_type>::val());
     scale[token_idx] = token_scale;
   }
   __syncthreads();

csrc/quantization/fp8/common.cuh

@@ -1,20 +1,12 @@
 #pragma once
 
 #include "quantization/vectorization.cuh"
+#include "quantization/utils.cuh"
 
 #include <cmath>
-#include <c10/core/ScalarType.h>
 
-#ifndef USE_ROCM
-  #include <c10/util/Float8_e4m3fn.h>
-  #define MAYBE_HOST_DEVICE C10_HOST_DEVICE
-#else
-  #include <ATen/hip/HIPContext.h>
-  #include <c10/util/Float8_e4m3fn.h>
-  #include <c10/util/Float8_e4m3fnuz.h>
+#ifdef USE_ROCM
   #include "amd/quant_utils.cuh"
-  // ROCm doesn't seem to need C10_HOST_DEVICE for static constexpr
-  #define MAYBE_HOST_DEVICE
 #endif
 
 // Determines the preferred FP8 type for the current platform.
@@ -31,29 +23,6 @@ static bool is_fp8_ocp() {
 #endif
 }
 
-template <typename T>
-struct fp8_e4m3_adjusted_max;
-
-template <>
-struct fp8_e4m3_adjusted_max<c10::Float8_e4m3fn> {
-  static constexpr c10::Float8_e4m3fn val() {
-    return std::numeric_limits<c10::Float8_e4m3fn>::max();
-  }
-};
-
-// Using the default max value from pytorch (240.0 0x7F) will cause accuracy
-// issues when running dynamic quantization. Here use 224.0 0x7E for rocm.
-template <>
-struct fp8_e4m3_adjusted_max<c10::Float8_e4m3fnuz> {
-  static constexpr c10::Float8_e4m3fnuz val() {
-    return c10::Float8_e4m3fnuz(0x7E, c10::Float8_e4m3fnuz::from_bits());
-  }
-};
-
-template <typename T>
-MAYBE_HOST_DEVICE static constexpr T fp8_e4m3_adjusted_max_v =
-    fp8_e4m3_adjusted_max<T>::val();
-
 namespace vllm {
 
 __device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
@@ -76,8 +45,8 @@ __device__ __forceinline__ fp8_type scaled_fp8_conversion(float const val,
     x = val / scale;
   }
 
-  float r = fmax(-fp8_e4m3_adjusted_max_v<fp8_type>,
-                 fmin(x, fp8_e4m3_adjusted_max_v<fp8_type>));
+  float r =
+      fmax(-quant_type_max_v<fp8_type>, fmin(x, quant_type_max_v<fp8_type>));
 #ifndef USE_ROCM
   return static_cast<fp8_type>(r);
 #else
@@ -123,7 +92,7 @@ __global__ void segmented_max_reduction(float* __restrict__ scale,
   // Finally, since cache[0] contains the maximum for this thread block,
   // atomically write the max to the target location
   if (threadIdx.x == 0) {
-    atomicMaxFloat(scale, cache[0] / fp8_e4m3_adjusted_max_v<fp8_type>);
+    atomicMaxFloat(scale, cache[0] / quant_type_max_v<fp8_type>);
   }
 }
 

csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -14,8 +14,7 @@ __device__ void rms_norm_dynamic_per_token_quant_vec(
     float* __restrict__ scales,           // [num_tokens]
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
-    float const* scale_ub, float const var_epsilon,
-    float const min_scaling_factor, int32_t const hidden_size,
+    float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
     scalar_t* __restrict__ residual = nullptr) {
   float rms = 0.0f;
   float token_scale = 0.0f;
@@ -27,8 +26,8 @@ __device__ void rms_norm_dynamic_per_token_quant_vec(
   // Compute scale
   vllm::vectorized::compute_dynamic_per_token_scales<scalar_t, scalar_out_t,
                                                      has_residual>(
-      &token_scale, scales, input, weight, rms, scale_ub, min_scaling_factor,
-      hidden_size, residual);
+      &token_scale, scales, input, weight, rms, scale_ub, hidden_size,
+      residual);
 
   // RMS Norm + Quant
   if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
@@ -50,8 +49,7 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
     float* __restrict__ scales,           // [num_tokens]
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
-    float const* scale_ub, float const var_epsilon,
-    float const min_scaling_factor, int32_t const hidden_size,
+    float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
     scalar_t* __restrict__ residual = nullptr) {
   // For vectorization, token_input and token_output pointers need to be
   // aligned at 8-byte and 4-byte addresses respectively.
@@ -60,8 +58,8 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
   if (can_vectorize) {
     return rms_norm_dynamic_per_token_quant_vec<scalar_t, scalar_out_t,
                                                 has_residual>(
-        out, scales, input, weight, scale_ub, var_epsilon, min_scaling_factor,
-        hidden_size, residual);
+        out, scales, input, weight, scale_ub, var_epsilon, hidden_size,
+        residual);
   }
 
   float rms = 0.0f;
@@ -72,8 +70,8 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
                                             var_epsilon, residual);
   // Compute Scale
   vllm::compute_dynamic_per_token_scales<scalar_t, scalar_out_t, has_residual>(
-      &token_scale, scales, input, weight, rms, scale_ub, min_scaling_factor,
-      hidden_size, residual);
+      &token_scale, scales, input, weight, rms, scale_ub, hidden_size,
+      residual);
 
   // RMS Norm + Quant
   if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
@@ -105,11 +103,6 @@ void rms_norm_dynamic_per_token_quant_dispatch(
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  const float min_scaling_factor =
-      out.dtype() == torch::kInt8
-          ? std::numeric_limits<float>::epsilon()
-          : 1.0f / (std::numeric_limits<c10::Float8_e4m3fn>::max() * 512.f);
-
   if (residual.has_value()) {
     VLLM_DISPATCH_QUANT_TYPES(
         out.scalar_type(), "rms_norm_dynamic_per_token_quant_kernel", [&] {
@@ -119,8 +112,7 @@ void rms_norm_dynamic_per_token_quant_dispatch(
                   out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
                   input.data_ptr<scalar_in_t>(), weight.data_ptr<scalar_in_t>(),
                   scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
-                  var_epsilon, min_scaling_factor, hidden_size,
-                  residual->data_ptr<scalar_in_t>());
+                  var_epsilon, hidden_size, residual->data_ptr<scalar_in_t>());
         });
 
   } else {
@@ -132,7 +124,7 @@ void rms_norm_dynamic_per_token_quant_dispatch(
                   out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
                   input.data_ptr<scalar_in_t>(), weight.data_ptr<scalar_in_t>(),
                   scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
-                  var_epsilon, min_scaling_factor, hidden_size, nullptr);
+                  var_epsilon, hidden_size, nullptr);
         });
   }
 }

csrc/quantization/fused_kernels/layernorm_utils.cuh

@@ -5,6 +5,7 @@
  */
 
 #include "quantization/vectorization.cuh"
+#include "quantization/utils.cuh"
 #include "quant_conversions.cuh"
 
 #ifndef USE_ROCM
@@ -51,11 +52,11 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    float const min_scaling_factor, int32_t const hidden_size,
+    int32_t const hidden_size,
     scalar_t const* __restrict__ residual = nullptr) {
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   ;
-  constexpr scalar_out_t qmax{std::numeric_limits<scalar_out_t>::max()};
+  constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
 
   float block_absmax_val_maybe = 0.0f;
   for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
@@ -83,7 +84,7 @@ __device__ void compute_dynamic_per_token_scales(
       scale = block_absmax_val_maybe;
     }
     // token scale computation
-    scale = max(scale / qmax, min_scaling_factor);
+    scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
     s_token_scale = scale;                 // Shared memory store
     all_token_scales[blockIdx.x] = scale;  // Global output store
   }
@@ -184,7 +185,7 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    float const min_scaling_factor, int32_t const hidden_size,
+    int32_t const hidden_size,
     scalar_t const* __restrict__ residual = nullptr) {
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   ;
@@ -200,7 +201,7 @@ __device__ void compute_dynamic_per_token_scales(
         reinterpret_cast<vec4_t<scalar_t> const*>(&residual[token_offset]);
   }
 
-  constexpr scalar_out_t qmax{std::numeric_limits<scalar_out_t>::max()};
+  constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
 
   int32_t const num_vec_elems = hidden_size >> 2;
   float block_absmax_val_maybe = 0.0f;
@@ -248,7 +249,7 @@ __device__ void compute_dynamic_per_token_scales(
       scale = block_absmax_val_maybe;
     }
     // token scale computation
-    scale = max(scale / qmax, min_scaling_factor);
+    scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
     s_token_scale = scale;                 // shared memory store
     all_token_scales[blockIdx.x] = scale;  // global output store
   }

csrc/quantization/fused_kernels/quant_conversions.cuh

@@ -33,8 +33,8 @@ static __device__ __forceinline__ int8_t float_to_int8_rn(float const x) {
 
 template <typename fp8_type>
 static __device__ __forceinline__ fp8_type float_to_fp8(float const x) {
-  float const r = fmax(-fp8_e4m3_adjusted_max_v<fp8_type>,
-                       fmin(x, fp8_e4m3_adjusted_max_v<fp8_type>));
+  float const r =
+      fmax(-quant_type_max_v<fp8_type>, fmin(x, quant_type_max_v<fp8_type>));
   return static_cast<fp8_type>(r);
 }
 

csrc/quantization/gguf/dequantize.cuh

@@ -94,8 +94,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     dfloat2 v;
     dequantize_kernel(vx, ib, iqs, v);
 
-    y[iybs + iqs + 0]        = v.x;
-    y[iybs + iqs + y_offset] = v.y;
+    y[iybs + iqs + 0]        = convert_from_half<dst_t>(v.x);
+    y[iybs + iqs + y_offset] = convert_from_half<dst_t>(v.y);
 }
 
 template<typename dst_t>
@@ -114,10 +114,10 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
 
     half dall = __low2half(x[i].dm);
     half dmin = __high2half(x[i].dm);
-    y[l+ 0] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+0] >> 4)));
-    y[l+32] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+2] >> 4)));
-    y[l+64] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+4] >> 4)));
-    y[l+96] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+6] >> 4)));
+    y[l+ 0] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+0] >> 4))));
+    y[l+32] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+2] >> 4))));
+    y[l+64] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+4] >> 4))));
+    y[l+96] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+6] >> 4))));
 }
 
 template<typename dst_t>
@@ -148,7 +148,9 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const uint8_t * q = x[i].qs + 32*n;
     const uint8_t * hm = x[i].hmask;
 
-    for (int l = l0; l < l0+4; ++l) y[l] = __hmul(dl,  __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)));
+    for (int l = l0; l < l0+4; ++l) {
+        y[l] = convert_from_half<dst_t>(__hmul(dl,  __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4))));
+    }
 }
 
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
@@ -188,8 +190,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
     const half d2 = __hmul(dall, __int2half_rn(sc));
     const half m2 = __hmul(dmin, __int2half_rn(m));
     for (int l = 0; l < n; ++l) {
-        y[l + 0] = __hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1);
-        y[l +32] = __hsub(__hmul(d2,  __int2half_rn(q[l] >> 4)), m2);
+        y[l + 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1));
+        y[l +32] = convert_from_half<dst_t>(__hsub(__hmul(d2,  __int2half_rn(q[l] >> 4)), m2));
     }
 }
 
@@ -220,11 +222,11 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
     const half d2 = __hmul(dall, __int2half_rn(sc)); const half m2 = __hmul(dmin, __int2half_rn(m));
 
     uint8_t   hm  = 1 << (2*il);
-    y[ 0] = __hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1);
-    y[ 1] = __hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1);
+    y[ 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1));
+    y[ 1] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1));
     hm <<= 1;
-    y[32] = __hsub(__hmul(d2, __int2half_rn((ql[0] >>  4) + (qh[0] & hm ? 16 : 0))), m2);
-    y[33] = __hsub(__hmul(d2, __int2half_rn((ql[1] >>  4) + (qh[1] & hm ? 16 : 0))), m2);
+    y[32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[0] >>  4) + (qh[0] & hm ? 16 : 0))), m2));
+    y[33] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[1] >>  4) + (qh[1] & hm ? 16 : 0))), m2));
 }
 
 template<typename dst_t>
@@ -247,10 +249,10 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     const uint8_t   qh = x[i].qh[32*ip + il];
     const int8_t  * sc = x[i].scales + is;
 
-    y[ 0] = __hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32)));
-    y[32] = __hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32)));
-    y[64] = __hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32)));
-    y[96] = __hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32)));
+    y[ 0] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32))));
+    y[32] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32))));
+    y[64] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32))));
+    y[96] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32))));
 }
 
 template<typename dst_t>
@@ -269,7 +271,7 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const uint32_t aux32 = q2[2] | (q2[3] << 16);
     const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 }
 
 template<typename dst_t>
@@ -286,7 +288,7 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
     const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
-    for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 
 }
 
@@ -303,7 +305,7 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
     const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
-    for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 }
 
 template<typename dst_t>
@@ -324,8 +326,8 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
     const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.5f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 4; ++j) {
-        y[j+0] = __float2half(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
-        y[j+4] = __float2half(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
 }
 
@@ -345,8 +347,8 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
     const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f;
     const uint8_t signs = x[i].signs[4*ib + il];
     for (int j = 0; j < 4; ++j) {
-        y[j+0] = __float2half(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
-        y[j+4] = __float2half(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
 }
 
@@ -367,7 +369,7 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
     grid32[0] &= 0x0f0f0f0f;
     for (int j = 0; j < 8; ++j) {
-        y[j] = __float2half(d * (q[j] + delta));
+        y[j] = d * (q[j] + delta);
     }
 }
 
@@ -392,7 +394,7 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
     grid32[0] &= 0x0f0f0f0f;
     for (int j = 0; j < 8; ++j) {
-        y[j] = __float2half(d * (q[j] + delta));
+        y[j] = d * (q[j] + delta);
     }
 }
 
@@ -409,8 +411,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
     const uint8_t  * q4 = x[ib].qs + 4*il;
     const float d = __half2float(x[ib].d);
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = __float2half(d * kvalues_iq4nl[q4[j] & 0xf]);
-        y[j+16] = __float2half(d * kvalues_iq4nl[q4[j] >>  4]);
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 
 }
@@ -427,8 +429,8 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
     const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
     const float d = __half2float(x[i].d) * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = __float2half(d * kvalues_iq4nl[q4[j] & 0xf]);
-        y[j+16] = __float2half(d * kvalues_iq4nl[q4[j] >>  4]);
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 }
 
@@ -522,7 +524,8 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
-static to_fp16_cuda_t ggml_get_to_fp16_cuda(int64_t type) {
+template<typename dst_t>
+static to_cuda_ggml_t<dst_t> ggml_get_to_cuda(int64_t type) {
     switch (type) {
         case 2:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;

csrc/quantization/gguf/ggml-common.h

@@ -1063,7 +1063,8 @@ static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -
 typedef half dfloat; // dequantize float
 typedef half2 dfloat2;
 typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
-typedef void (*to_fp16_cuda_t)(const void * __restrict__ x, dfloat * __restrict__ y, int k, cudaStream_t stream);
+template<typename dst_t>
+using to_cuda_ggml_t = void (*)(const void * __restrict__ x, dst_t * __restrict__ y, int k, cudaStream_t stream);
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
 typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
 typedef void (*load_tiles_cuda_t)(
@@ -1075,6 +1076,25 @@ typedef float (*vec_dot_q_mul_mat_cuda_t)(
 
 // Utility function
 
+template<typename dst_t>
+static __device__ __forceinline__ dst_t convert_from_half(half val) {
+    return val;
+}
+
+template<>
+__device__ __forceinline__ c10::BFloat16 convert_from_half<c10::BFloat16>(half val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+    return __float2bfloat16(__half2float(val));
+#else
+    return __half2float(val);
+#endif  // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+}
+
+template<>
+__device__ __forceinline__ float convert_from_half<float>(half val) {
+    return __half2float(val);
+}
+
 #if defined(USE_ROCM)
 
 #ifndef __has_builtin

csrc/quantization/gguf/gguf_kernel.cu

@@ -71,14 +71,19 @@ static void quantize_row_q8_1_cuda(const scalar_t* x, void* vy, const int kx,
 }
 
 torch::Tensor ggml_dequantize(torch::Tensor W,  // quant weight
-                              int64_t type, int64_t m, int64_t n) {
+                              int64_t type, int64_t m, int64_t n,
+                              std::optional<at::ScalarType> const& dtype) {
   const at::cuda::OptionalCUDAGuard device_guard(device_of(W));
-  auto options =
-      torch::TensorOptions().dtype(torch::kFloat16).device(W.device());
+  auto dtype_ = dtype.value_or(torch::kFloat16);
+  auto options = torch::TensorOptions().dtype(dtype_).device(W.device());
   at::Tensor DW = torch::empty({m, n}, options);
   cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
-  const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(type);
-  to_fp16_cuda((void*)W.data_ptr(), (half*)DW.data_ptr(), m * n, stream);
+
+  VLLM_DISPATCH_FLOATING_TYPES(DW.scalar_type(), "ggml_dequantize", [&] {
+    auto to_cuda = ggml_get_to_cuda<scalar_t>(type);
+    to_cuda((void*)W.data_ptr(), (scalar_t*)DW.data_ptr(), m * n, stream);
+  });
+
   return DW;
 }
 
@@ -375,25 +380,25 @@ torch::Tensor ggml_moe_a8(torch::Tensor X,  // input
 int64_t ggml_moe_get_block_size(int64_t type) {
   switch (type) {
     case 2:
-      return MMQ_X_Q4_0;
+      return MOE_X_Q4_0;
     case 3:
-      return MMQ_X_Q4_1;
+      return MOE_X_Q4_1;
     case 6:
-      return MMQ_X_Q5_0;
+      return MOE_X_Q5_0;
     case 7:
-      return MMQ_X_Q5_1;
+      return MOE_X_Q5_1;
     case 8:
-      return MMQ_X_Q8_0;
+      return MOE_X_Q8_0;
     case 10:
-      return MMQ_X_Q2_K;
+      return MOE_X_Q2_K;
     case 11:
-      return MMQ_X_Q3_K;
+      return MOE_X_Q3_K;
     case 12:
-      return MMQ_X_Q4_K;
+      return MOE_X_Q4_K;
     case 13:
-      return MMQ_X_Q5_K;
+      return MOE_X_Q5_K;
     case 14:
-      return MMQ_X_Q6_K;
+      return MOE_X_Q6_K;
   }
   return 0;
 }

csrc/quantization/gguf/moe.cuh

@@ -129,12 +129,12 @@ static __device__ __forceinline__ void moe_q(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q4_0 64
-  #define MMQ_Y_Q4_0 128
+  #define MOE_X_Q4_0 8
+  #define MOE_Y_Q4_0 128
   #define NWARPS_Q4_0 8
 #else
-  #define MMQ_X_Q4_0 4
-  #define MMQ_Y_Q4_0 32
+  #define MOE_X_Q4_0 4
+  #define MOE_Y_Q4_0 32
   #define NWARPS_Q4_0 4
 #endif
 
@@ -149,8 +149,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q4_0, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q4_0;
-  const int mmq_y = MMQ_Y_Q4_0;
+  const int mmq_x = MOE_X_Q4_0;
+  const int mmq_y = MOE_Y_Q4_0;
   const int nwarps = NWARPS_Q4_0;
 
   moe_q<scalar_t, QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps,
@@ -167,8 +167,8 @@ static void ggml_moe_q4_0_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  int mmq_x = MMQ_X_Q4_0;
-  int mmq_y = MMQ_Y_Q4_0;
+  int mmq_x = MOE_X_Q4_0;
+  int mmq_y = MOE_Y_Q4_0;
   int nwarps = NWARPS_Q4_0;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -190,12 +190,12 @@ static void ggml_moe_q4_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q4_1 64
-  #define MMQ_Y_Q4_1 128
+  #define MOE_X_Q4_1 8
+  #define MOE_Y_Q4_1 128
   #define NWARPS_Q4_1 8
 #else
-  #define MMQ_X_Q4_1 4
-  #define MMQ_Y_Q4_1 32
+  #define MOE_X_Q4_1 4
+  #define MOE_Y_Q4_1 32
   #define NWARPS_Q4_1 4
 #endif
 
@@ -210,8 +210,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q4_1, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q4_1;
-  const int mmq_y = MMQ_Y_Q4_1;
+  const int mmq_x = MOE_X_Q4_1;
+  const int mmq_y = MOE_Y_Q4_1;
   const int nwarps = NWARPS_Q4_1;
 
   moe_q<scalar_t, QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps,
@@ -228,8 +228,8 @@ static void ggml_moe_q4_1_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  int mmq_x = MMQ_X_Q4_1;
-  int mmq_y = MMQ_Y_Q4_1;
+  int mmq_x = MOE_X_Q4_1;
+  int mmq_y = MOE_Y_Q4_1;
   int nwarps = NWARPS_Q4_1;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -251,12 +251,12 @@ static void ggml_moe_q4_1_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q5_0 64
-  #define MMQ_Y_Q5_0 128
+  #define MOE_X_Q5_0 8
+  #define MOE_Y_Q5_0 128
   #define NWARPS_Q5_0 8
 #else
-  #define MMQ_X_Q5_0 4
-  #define MMQ_Y_Q5_0 32
+  #define MOE_X_Q5_0 4
+  #define MOE_Y_Q5_0 32
   #define NWARPS_Q5_0 4
 #endif
 
@@ -271,8 +271,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q5_0, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q5_0;
-  const int mmq_y = MMQ_Y_Q5_0;
+  const int mmq_x = MOE_X_Q5_0;
+  const int mmq_y = MOE_Y_Q5_0;
   const int nwarps = NWARPS_Q5_0;
 
   moe_q<scalar_t, QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps,
@@ -289,8 +289,8 @@ static void ggml_moe_q5_0_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q5_0;
-  const int mmq_y = MMQ_Y_Q5_0;
+  const int mmq_x = MOE_X_Q5_0;
+  const int mmq_y = MOE_Y_Q5_0;
   const int nwarps = NWARPS_Q5_0;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -312,12 +312,12 @@ static void ggml_moe_q5_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q5_1 64
-  #define MMQ_Y_Q5_1 128
+  #define MOE_X_Q5_1 8
+  #define MOE_Y_Q5_1 128
   #define NWARPS_Q5_1 8
 #else
-  #define MMQ_X_Q5_1 4
-  #define MMQ_Y_Q5_1 32
+  #define MOE_X_Q5_1 4
+  #define MOE_Y_Q5_1 32
   #define NWARPS_Q5_1 4
 #endif
 
@@ -332,8 +332,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q5_1, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q5_1;
-  const int mmq_y = MMQ_Y_Q5_1;
+  const int mmq_x = MOE_X_Q5_1;
+  const int mmq_y = MOE_Y_Q5_1;
   const int nwarps = NWARPS_Q5_1;
 
   moe_q<scalar_t, QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps,
@@ -350,8 +350,8 @@ static void ggml_moe_q5_1_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q5_1;
-  const int mmq_y = MMQ_Y_Q5_1;
+  const int mmq_x = MOE_X_Q5_1;
+  const int mmq_y = MOE_Y_Q5_1;
   const int nwarps = NWARPS_Q5_1;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -373,12 +373,12 @@ static void ggml_moe_q5_1_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q8_0 64
-  #define MMQ_Y_Q8_0 128
+  #define MOE_X_Q8_0 8
+  #define MOE_Y_Q8_0 128
   #define NWARPS_Q8_0 8
 #else
-  #define MMQ_X_Q8_0 4
-  #define MMQ_Y_Q8_0 32
+  #define MOE_X_Q8_0 4
+  #define MOE_Y_Q8_0 32
   #define NWARPS_Q8_0 4
 #endif
 
@@ -393,8 +393,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q8_0, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q8_0;
-  const int mmq_y = MMQ_Y_Q8_0;
+  const int mmq_x = MOE_X_Q8_0;
+  const int mmq_y = MOE_Y_Q8_0;
   const int nwarps = NWARPS_Q8_0;
 
   moe_q<scalar_t, QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps,
@@ -411,8 +411,8 @@ static void ggml_moe_q8_0_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q8_0;
-  const int mmq_y = MMQ_Y_Q8_0;
+  const int mmq_x = MOE_X_Q8_0;
+  const int mmq_y = MOE_Y_Q8_0;
   const int nwarps = NWARPS_Q8_0;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -434,12 +434,12 @@ static void ggml_moe_q8_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q2_K 64
-  #define MMQ_Y_Q2_K 128
+  #define MOE_X_Q2_K 8
+  #define MOE_Y_Q2_K 128
   #define NWARPS_Q2_K 8
 #else
-  #define MMQ_X_Q2_K 4
-  #define MMQ_Y_Q2_K 32
+  #define MOE_X_Q2_K 4
+  #define MOE_Y_Q2_K 32
   #define NWARPS_Q2_K 4
 #endif
 
@@ -454,8 +454,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q2_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q2_K;
-  const int mmq_y = MMQ_Y_Q2_K;
+  const int mmq_x = MOE_X_Q2_K;
+  const int mmq_y = MOE_Y_Q2_K;
   const int nwarps = NWARPS_Q2_K;
 
   moe_q<scalar_t, QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps,
@@ -472,8 +472,8 @@ static void ggml_moe_q2_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q2_K;
-  const int mmq_y = MMQ_Y_Q2_K;
+  const int mmq_x = MOE_X_Q2_K;
+  const int mmq_y = MOE_Y_Q2_K;
   const int nwarps = NWARPS_Q2_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -495,12 +495,12 @@ static void ggml_moe_q2_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q3_K 64
-  #define MMQ_Y_Q3_K 128
+  #define MOE_X_Q3_K 8
+  #define MOE_Y_Q3_K 128
   #define NWARPS_Q3_K 8
 #else
-  #define MMQ_X_Q3_K 4
-  #define MMQ_Y_Q3_K 32
+  #define MOE_X_Q3_K 4
+  #define MOE_Y_Q3_K 32
   #define NWARPS_Q3_K 4
 #endif
 
@@ -516,8 +516,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q3_K, 2)
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
 
-  const int mmq_x = MMQ_X_Q3_K;
-  const int mmq_y = MMQ_Y_Q3_K;
+  const int mmq_x = MOE_X_Q3_K;
+  const int mmq_y = MOE_Y_Q3_K;
   const int nwarps = NWARPS_Q3_K;
 
   moe_q<scalar_t, QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps,
@@ -533,8 +533,8 @@ static void ggml_moe_q3_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q3_K;
-  const int mmq_y = MMQ_Y_Q3_K;
+  const int mmq_x = MOE_X_Q3_K;
+  const int mmq_y = MOE_Y_Q3_K;
   const int nwarps = NWARPS_Q3_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -556,12 +556,12 @@ static void ggml_moe_q3_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q4_K 64
-  #define MMQ_Y_Q4_K 128
+  #define MOE_X_Q4_K 8
+  #define MOE_Y_Q4_K 128
   #define NWARPS_Q4_K 8
 #else
-  #define MMQ_X_Q4_K 4
-  #define MMQ_Y_Q4_K 32
+  #define MOE_X_Q4_K 4
+  #define MOE_Y_Q4_K 32
   #define NWARPS_Q4_K 4
 #endif
 
@@ -576,8 +576,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q4_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q4_K;
-  const int mmq_y = MMQ_Y_Q4_K;
+  const int mmq_x = MOE_X_Q4_K;
+  const int mmq_y = MOE_Y_Q4_K;
   const int nwarps = NWARPS_Q4_K;
 
   moe_q<scalar_t, QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps,
@@ -594,8 +594,8 @@ static void ggml_moe_q4_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q4_K;
-  const int mmq_y = MMQ_Y_Q4_K;
+  const int mmq_x = MOE_X_Q4_K;
+  const int mmq_y = MOE_Y_Q4_K;
   const int nwarps = NWARPS_Q4_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -617,12 +617,12 @@ static void ggml_moe_q4_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q5_K 64
-  #define MMQ_Y_Q5_K 128
+  #define MOE_X_Q5_K 8
+  #define MOE_Y_Q5_K 128
   #define NWARPS_Q5_K 8
 #else
-  #define MMQ_X_Q5_K 4
-  #define MMQ_Y_Q5_K 32
+  #define MOE_X_Q5_K 4
+  #define MOE_Y_Q5_K 32
   #define NWARPS_Q5_K 4
 #endif
 
@@ -637,8 +637,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q5_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q5_K;
-  const int mmq_y = MMQ_Y_Q5_K;
+  const int mmq_x = MOE_X_Q5_K;
+  const int mmq_y = MOE_Y_Q5_K;
   const int nwarps = NWARPS_Q5_K;
 
   moe_q<scalar_t, QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps,
@@ -655,8 +655,8 @@ static void ggml_moe_q5_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q5_K;
-  const int mmq_y = MMQ_Y_Q5_K;
+  const int mmq_x = MOE_X_Q5_K;
+  const int mmq_y = MOE_Y_Q5_K;
   const int nwarps = NWARPS_Q5_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -678,12 +678,12 @@ static void ggml_moe_q5_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q6_K 64
-  #define MMQ_Y_Q6_K 128
+  #define MOE_X_Q6_K 8
+  #define MOE_Y_Q6_K 128
   #define NWARPS_Q6_K 8
 #else
-  #define MMQ_X_Q6_K 4
-  #define MMQ_Y_Q6_K 32
+  #define MOE_X_Q6_K 4
+  #define MOE_Y_Q6_K 32
   #define NWARPS_Q6_K 4
 #endif
 
@@ -698,8 +698,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q6_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q6_K;
-  const int mmq_y = MMQ_Y_Q6_K;
+  const int mmq_x = MOE_X_Q6_K;
+  const int mmq_y = MOE_Y_Q6_K;
   const int nwarps = NWARPS_Q6_K;
 
   moe_q<scalar_t, QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps,
@@ -716,8 +716,8 @@ static void ggml_moe_q6_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q6_K;
-  const int mmq_y = MMQ_Y_Q6_K;
+  const int mmq_x = MOE_X_Q6_K;
+  const int mmq_y = MOE_Y_Q6_K;
   const int nwarps = NWARPS_Q6_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;

csrc/quantization/gptq_marlin/awq_marlin_repack.cu

@@ -14,7 +14,7 @@ __global__ void awq_marlin_repack_kernel(
   int n_tiles = size_n / tile_n_size;
   int block_k_tiles = div_ceil(k_tiles, gridDim.x);
 
-  int start_k_tile = blockIdx.x * block_k_tiles;
+  auto start_k_tile = blockIdx.x * block_k_tiles;
   if (start_k_tile >= k_tiles) {
     return;
   }
@@ -51,8 +51,8 @@ __global__ void awq_marlin_repack_kernel(
     int4* sh_ptr = sh + stage_size * pipe;
 
     if (threadIdx.x < stage_size) {
-      int k_id = threadIdx.x / stage_n_threads;
-      int n_id = threadIdx.x % stage_n_threads;
+      auto k_id = threadIdx.x / stage_n_threads;
+      auto n_id = threadIdx.x % stage_n_threads;
 
       int first_k = k_tile_id * tile_k_size;
 
@@ -70,8 +70,8 @@ __global__ void awq_marlin_repack_kernel(
       return;
     }
 
-    int warp_id = threadIdx.x / 32;
-    int th_id = threadIdx.x % 32;
+    auto warp_id = threadIdx.x / 32;
+    auto th_id = threadIdx.x % 32;
 
     if (warp_id >= 4) {
       return;
@@ -265,4 +265,4 @@ TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
 
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, Meta, m) {
   m.impl("awq_marlin_repack", &awq_marlin_repack_meta);
-}
+}

csrc/quantization/gptq_marlin/gptq_marlin.cu

@@ -460,7 +460,7 @@ __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
                                     int const* __restrict__ perm_int_ptr,
                                     int4* __restrict__ out_int4_ptr, int size_m,
                                     int size_k, int lda, int block_rows) {
-  int start_row = block_rows * blockIdx.x;
+  auto start_row = block_rows * blockIdx.x;
   int finish_row = start_row + block_rows;
   if (finish_row > size_m) {
     finish_row = size_m;
@@ -484,7 +484,7 @@ __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
     int base_k = 0;
 
     for (int i = 0; i < iters; i++) {
-      int cur_k = base_k + threadIdx.x;
+      auto cur_k = base_k + threadIdx.x;
       int src_pos = perm_int_ptr[cur_k];
 
       out_half[cur_k] = a_row_half[src_pos];
@@ -494,7 +494,7 @@ __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
 
     if (rest) {
       if (threadIdx.x < rest) {
-        int cur_k = base_k + threadIdx.x;
+        auto cur_k = base_k + threadIdx.x;
         int src_pos = perm_int_ptr[cur_k];
 
         out_half[cur_k] = a_row_half[src_pos];
@@ -723,8 +723,8 @@ __global__ void Marlin(
                 (threadIdx.x % b_sh_stride_threads) * b_thread_vecs;
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x * b_thread_vecs;
-  int b_sh_rd = threadIdx.x * b_thread_vecs;
+  auto b_sh_wr = threadIdx.x * b_thread_vecs;
+  auto b_sh_rd = threadIdx.x * b_thread_vecs;
 
   // For act_order
   constexpr int k_iter_size = tb_k / b_sh_wr_iters;
@@ -743,7 +743,7 @@ __global__ void Marlin(
                 s_sh_stride * slice_col + threadIdx.x;
     }
   }
-  int s_sh_wr = threadIdx.x;
+  auto s_sh_wr = threadIdx.x;
   bool s_sh_wr_pred = threadIdx.x < s_sh_stride;
 
   // Zero-points
@@ -756,7 +756,7 @@ __global__ void Marlin(
                  zp_sh_stride * slice_col + threadIdx.x;
     }
   }
-  int zp_sh_wr = threadIdx.x;
+  auto zp_sh_wr = threadIdx.x;
   bool zp_sh_wr_pred = threadIdx.x < zp_sh_stride;
 
   // We use a different scale layout for grouped and column-wise quantization as
@@ -1047,7 +1047,7 @@ __global__ void Marlin(
           int4* sh_s_stage = sh_s + s_sh_stage * pipe;
           reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
         } else {
-          int warp_id = threadIdx.x / 32;
+          auto warp_id = threadIdx.x / 32;
           int n_warps = thread_n_blocks / 4;
 
           int warp_row = warp_id / n_warps;
@@ -1085,7 +1085,7 @@ __global__ void Marlin(
 
     // Determine "position" inside the thread-block (based on warp and
     // thread-id)
-    int warp_id = threadIdx.x / 32;
+    auto warp_id = threadIdx.x / 32;
     int n_warps =
         thread_n_blocks / 4;  // Each warp processes 4 16-size tiles over N
 
@@ -1094,7 +1094,7 @@ __global__ void Marlin(
 
     cur_k += warp_row * 16;
 
-    int th_id = threadIdx.x % 32;
+    auto th_id = threadIdx.x % 32;
     cur_k += (th_id % 4) * 2;  // Due to tensor-core layout for fp16 B matrix
 
     int s_col_shift =
@@ -1159,7 +1159,7 @@ __global__ void Marlin(
               (reinterpret_cast<int*>(sh_zp_stage))[zp_sh_rd + i];
         }
       } else {
-        int warp_id = threadIdx.x / 32;
+        auto warp_id = threadIdx.x / 32;
         int n_warps = thread_n_blocks / 4;
 
         int warp_row = warp_id / n_warps;
@@ -1197,7 +1197,7 @@ __global__ void Marlin(
                                      (pipe / (group_blocks / thread_k_blocks)));
           reinterpret_cast<int4*>(&frag_zpf[k % 2])[0] = sh_zp_stage[zp_sh_rd];
         } else {
-          int warp_id = threadIdx.x / 32;
+          auto warp_id = threadIdx.x / 32;
           int n_warps = thread_n_blocks / 4;
 
           int warp_row = warp_id / n_warps;
@@ -1323,7 +1323,7 @@ __global__ void Marlin(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride_threads / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride_threads;
+      auto red_idx = threadIdx.x / b_sh_stride_threads;
       constexpr int red_sh_stride = b_sh_stride_threads * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride_threads;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride_threads) +
@@ -1390,7 +1390,7 @@ __global__ void Marlin(
                     4 * (threadIdx.x / 32) + threadIdx.x % 4;
       c_gl_wr += (2 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads;
-      int c_sh_wr = threadIdx.x;
+      auto c_sh_wr = threadIdx.x;
 
       int row = (threadIdx.x % 32) / 4;
 
@@ -1785,7 +1785,7 @@ __global__ void Marlin(
             <<<blocks, NUM_THREADS, max_shared_mem, stream>>>(                 \
                 A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, zp_ptr, g_idx_ptr,      \
                 num_groups, prob_m, prob_n, prob_k, lda, locks,                \
-                use_atomic_add, use_fp32_reduce);                              \
+                part_use_atomic_add, use_fp32_reduce);                         \
       }                                                                        \
     }
 
@@ -2215,6 +2215,10 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
       thread_m_blocks = exec_cfg.max_m_blocks;
     }
 
+    // atomic add reduce have better performance only when m * n is small
+    bool part_use_atomic_add =
+        use_atomic_add && div_ceil(prob_m, 64) * prob_n <= 2048;
+
     if (false) {
     }
     GPTQ_CALL_IF(vllm::kU4B8, 16, 4, 256)

csrc/quantization/gptq_marlin/gptq_marlin_repack.cu

@@ -15,7 +15,7 @@ __global__ void gptq_marlin_repack_kernel(
   int n_tiles = size_n / tile_n_size;
   int block_k_tiles = div_ceil(k_tiles, gridDim.x);
 
-  int start_k_tile = blockIdx.x * block_k_tiles;
+  auto start_k_tile = blockIdx.x * block_k_tiles;
   if (start_k_tile >= k_tiles) {
     return;
   }
@@ -71,8 +71,8 @@ __global__ void gptq_marlin_repack_kernel(
 
     if constexpr (has_perm) {
       if (threadIdx.x < stage_size) {
-        int k_id = threadIdx.x / stage_n_threads;
-        int n_id = threadIdx.x % stage_n_threads;
+        auto k_id = threadIdx.x / stage_n_threads;
+        auto n_id = threadIdx.x % stage_n_threads;
 
         uint32_t const* sh_perm_int_ptr =
             reinterpret_cast<uint32_t const*>(sh_perm_ptr);
@@ -88,8 +88,8 @@ __global__ void gptq_marlin_repack_kernel(
 
     } else {
       if (threadIdx.x < stage_size) {
-        int k_id = threadIdx.x / stage_n_threads;
-        int n_id = threadIdx.x % stage_n_threads;
+        auto k_id = threadIdx.x / stage_n_threads;
+        auto n_id = threadIdx.x % stage_n_threads;
 
         int first_k = k_tile_id * tile_k_size;
         int first_k_packed = first_k / pack_factor;
@@ -109,8 +109,8 @@ __global__ void gptq_marlin_repack_kernel(
       return;
     }
 
-    int warp_id = threadIdx.x / 32;
-    int th_id = threadIdx.x % 32;
+    auto warp_id = threadIdx.x / 32;
+    auto th_id = threadIdx.x % 32;
 
     if (warp_id >= 4) {
       return;
@@ -339,4 +339,4 @@ TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
 
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, Meta, m) {
   m.impl("gptq_marlin_repack", &gptq_marlin_repack_meta);
-}
+}

csrc/quantization/gptq_marlin/marlin.cuh

@@ -9,7 +9,11 @@
 #include <cuda_runtime.h>
 #include <iostream>
 
-namespace marlin {
+#ifndef MARLIN_NAMESPACE_NAME
+  #define MARLIN_NAMESPACE_NAME marlin
+#endif
+
+namespace MARLIN_NAMESPACE_NAME {
 
 // Marlin params
 
@@ -23,6 +27,7 @@ static constexpr int pipe_stages =
 
 static constexpr int min_thread_n = 64;
 static constexpr int min_thread_k = 64;
+static constexpr int max_thread_n = 256;
 
 static constexpr int tile_size = 16;
 static constexpr int max_par = 16;
@@ -84,4 +89,4 @@ __device__ inline void cp_async_wait() {
 
 #endif
 
-}  // namespace marlin
+}  // namespace MARLIN_NAMESPACE_NAME

csrc/quantization/gptq_marlin/marlin_dtypes.cuh

@@ -5,7 +5,11 @@
 #include <cuda_fp16.h>
 #include <cuda_bf16.h>
 
-namespace marlin {
+#ifndef MARLIN_NAMESPACE_NAME
+  #define MARLIN_NAMESPACE_NAME marlin
+#endif
+
+namespace MARLIN_NAMESPACE_NAME {
 
 template <typename scalar_t>
 class ScalarType {};
@@ -54,7 +58,7 @@ class ScalarType<nv_bfloat16> {
   using FragS = Vec<nv_bfloat162, 1>;
   using FragZP = Vec<nv_bfloat162, 4>;
 
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
   static __device__ float inline num2float(const nv_bfloat16 x) {
     return __bfloat162float(x);
   }
@@ -74,6 +78,6 @@ class ScalarType<nv_bfloat16> {
 #endif
 };
 
-}  // namespace marlin
+}  // namespace MARLIN_NAMESPACE_NAME
 
 #endif

csrc/quantization/marlin/dense/marlin_cuda_kernel.cu

@@ -277,12 +277,12 @@ __global__ void Marlin(
       b_gl_stride * (threadIdx.x / b_sh_stride) + (threadIdx.x % b_sh_stride);
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x;
-  int b_sh_rd = threadIdx.x;
+  auto b_sh_wr = threadIdx.x;
+  auto b_sh_rd = threadIdx.x;
 
   int s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
                 s_sh_stride * slice_col + threadIdx.x;
-  int s_sh_wr = threadIdx.x;
+  auto s_sh_wr = threadIdx.x;
   int s_sh_rd;
   // We use a different scale layout for grouped and column-wise quantization as
   // we scale a `half2` tile in column-major layout in the former and in
@@ -455,7 +455,7 @@ __global__ void Marlin(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride;
+      auto red_idx = threadIdx.x / b_sh_stride;
       constexpr int red_sh_stride = b_sh_stride * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride) +
@@ -522,7 +522,7 @@ __global__ void Marlin(
                     4 * (threadIdx.x / 32) + threadIdx.x % 4;
       c_gl_wr += (2 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads;
-      int c_sh_wr = threadIdx.x;
+      auto c_sh_wr = threadIdx.x;
 
       int row = (threadIdx.x % 32) / 4;
 

csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu

@@ -353,10 +353,10 @@ __global__ void Marlin(
       b_gl_stride * (threadIdx.x / b_sh_stride) + (threadIdx.x % b_sh_stride);
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x;
-  int b_sh_rd = threadIdx.x;
+  auto b_sh_wr = threadIdx.x;
+  auto b_sh_rd = threadIdx.x;
 
-  int s_tok_gl_rd = threadIdx.x;
+  auto s_tok_gl_rd = threadIdx.x;
   // NOTE(HandH1998): activation scale s_tok need shuffle to [0, 8, 1, 9, 2, 10,
   // 3, 11, 4, 12, 5, 13, 6, 14, 7, 15] for example, 0, 8 row scales serve for
   // thread 0, 1, 2, 3. For more details, refer to mma operand A layout as
@@ -368,8 +368,8 @@ __global__ void Marlin(
   int s_tok_sh_rd = (threadIdx.x % 32) / 4;
   bool s_tok_sh_wr_pred = threadIdx.x < prob_m;
 
-  int s_ch_gl_rd = s_ch_sh_stride * slice_col + threadIdx.x;
-  int s_ch_sh_wr = threadIdx.x;
+  auto s_ch_gl_rd = s_ch_sh_stride * slice_col + threadIdx.x;
+  auto s_ch_sh_wr = threadIdx.x;
   int s_ch_sh_rd = 16 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
                    2 * ((threadIdx.x % 32) % 4);
   bool s_ch_sh_wr_pred = threadIdx.x < s_ch_sh_stride;
@@ -558,7 +558,7 @@ __global__ void Marlin(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride;
+      auto red_idx = threadIdx.x / b_sh_stride;
       constexpr int red_sh_stride = b_sh_stride * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride) +
@@ -628,7 +628,7 @@ __global__ void Marlin(
                     8 * (threadIdx.x / 32) + (threadIdx.x % 4) * 2;
       c_gl_wr += (4 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads * 2;
-      int c_sh_wr = 2 * threadIdx.x;
+      auto c_sh_wr = 2 * threadIdx.x;
 
       int row = (threadIdx.x % 32) / 4;
 

csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu

@@ -273,15 +273,15 @@ __global__ void Marlin_24(
                 (threadIdx.x % b_sh_stride_threads) * b_thread_vecs;
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x * b_thread_vecs;
-  int b_sh_rd = threadIdx.x * b_thread_vecs;
+  auto b_sh_wr = threadIdx.x * b_thread_vecs;
+  auto b_sh_rd = threadIdx.x * b_thread_vecs;
 
   int m_gl_rd = m_gl_stride * (threadIdx.x / (m_sh_stride)) +
                 (threadIdx.x % (m_sh_stride));
   m_gl_rd += (m_sh_stride)*slice_col;
   m_gl_rd += m_gl_rd_delta_o * slice_row;
-  int m_sh_wr = threadIdx.x;
-  int m_sh_rd = threadIdx.x % 16 + (threadIdx.x / 32) * 16;
+  auto m_sh_wr = threadIdx.x;
+  auto m_sh_rd = threadIdx.x % 16 + (threadIdx.x / 32) * 16;
 
   int s_gl_rd;
   if constexpr (group_blocks == -1) {
@@ -291,7 +291,7 @@ __global__ void Marlin_24(
               s_sh_stride * slice_col + threadIdx.x;
   }
 
-  int s_sh_wr = threadIdx.x;
+  auto s_sh_wr = threadIdx.x;
   int s_sh_rd;
   // We use a different scale layout for grouped and column-wise quantization as
   // we scale a `half2` tile in column-major layout in the former and in
@@ -516,7 +516,7 @@ __global__ void Marlin_24(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride_threads / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride_threads;
+      auto red_idx = threadIdx.x / b_sh_stride_threads;
       constexpr int red_sh_stride = b_sh_stride_threads * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride_threads;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride_threads) +
@@ -583,7 +583,7 @@ __global__ void Marlin_24(
                     8 * (threadIdx.x / 32) + (threadIdx.x % 32) / 4;
       c_gl_wr += (2 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads;
-      int c_sh_wr = threadIdx.x;
+      auto c_sh_wr = threadIdx.x;
 
       int col = 2 * ((threadIdx.x % 32) % 4);
 

csrc/quantization/utils.cuh

@@ -0,0 +1,59 @@
+#pragma once
+
+/**
+ * Quantization utilities including:
+ *   Adjusted maximum values for qtypes.
+ *   Minimum scaling factors for qtypes.
+ */
+
+#include <cmath>
+#include <torch/types.h>
+
+#ifndef USE_ROCM
+  #include <c10/util/Float8_e4m3fn.h>
+  #define MAYBE_HOST_DEVICE C10_HOST_DEVICE
+#else
+  #include <ATen/hip/HIPContext.h>
+  #include <c10/util/Float8_e4m3fn.h>
+  #include <c10/util/Float8_e4m3fnuz.h>
+  // ROCm doesn't seem to need C10_HOST_DEVICE for static constexpr
+  #define MAYBE_HOST_DEVICE
+#endif
+
+template <typename T,
+          typename = std::enable_if_t<std::is_same_v<T, c10::Float8_e4m3fn> ||
+                                      std::is_same_v<T, c10::Float8_e4m3fnuz> ||
+                                      std::is_same_v<T, int8_t>>>
+struct quant_type_max {
+  static constexpr T val() { return std::numeric_limits<T>::max(); }
+};
+
+// Using the default max value from pytorch (240.0 0x7F) will cause accuracy
+// issues when running dynamic quantization. Here use 224.0 0x7E for rocm.
+template <>
+struct quant_type_max<c10::Float8_e4m3fnuz> {
+  static constexpr c10::Float8_e4m3fnuz val() {
+    return c10::Float8_e4m3fnuz(0x7E, c10::Float8_e4m3fnuz::from_bits());
+  }
+};
+
+template <typename T>
+MAYBE_HOST_DEVICE static constexpr T quant_type_max_v =
+    quant_type_max<T>::val();
+
+template <typename T,
+          typename = std::enable_if_t<std::is_same_v<T, c10::Float8_e4m3fn> ||
+                                      std::is_same_v<T, c10::Float8_e4m3fnuz> ||
+                                      std::is_same_v<T, int8_t>>>
+struct min_scaling_factor {
+  C10_DEVICE C10_ALWAYS_INLINE static float val() {
+    return 1.0f / (quant_type_max_v<T> * 512.0f);
+  }
+};
+
+template <>
+struct min_scaling_factor<int8_t> {
+  C10_DEVICE C10_ALWAYS_INLINE static float val() {
+    return std::numeric_limits<float>::epsilon();
+  }
+};

csrc/rocm/attention.cu

@@ -272,6 +272,7 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
     const float scale,    
     const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,   // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,   // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes, // [num_heads]
     const int q_stride,
@@ -284,18 +285,25 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
     int max_ctx_blocks, const float* k_scale, const float* v_scale) {
   // clang-format on
   constexpr int NWARPS = NUM_THREADS / WARP_SIZE;
-  const int warpid = threadIdx.x / WARP_SIZE;
-  const int laneid = threadIdx.x % WARP_SIZE;
+  const auto warpid = threadIdx.x / WARP_SIZE;
+  const auto laneid = threadIdx.x % WARP_SIZE;
   const int lane4id = laneid % 4;
   const int lane16id = laneid % 16;
   const int rowid = laneid / 16;
 
-  const int seq_idx = blockIdx.x;
-  const int partition_idx = blockIdx.y;
+  const auto seq_idx = blockIdx.x;
+  // NOTE queries with sequence len > 1 are prefills and taken care by another
+  // kernel.
+  if (query_start_loc_ptr != nullptr &&
+      (query_start_loc_ptr[seq_idx + 1] - query_start_loc_ptr[seq_idx]) != 1) {
+    return;
+  }
+
+  const auto partition_idx = blockIdx.y;
 
   constexpr int T_PAR_SIZE = 256;  // token partition size set to 256
 
-  const int max_num_partitions = gridDim.y;
+  const auto max_num_partitions = gridDim.y;
 
   const int context_len = context_lens[seq_idx];
 
@@ -346,9 +354,9 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
   // can be interpreted as B8x16 for 8 bit types
   _B16x8 Klocal[TLOOP][QKHELOOP];
 
-  const int wg_start_head_idx = blockIdx.z * GQA_RATIO;
-  const int wg_start_kv_head_idx = blockIdx.z;
-  const int total_num_heads = gridDim.z * GQA_RATIO;
+  const auto wg_start_head_idx = blockIdx.z * GQA_RATIO;
+  const auto wg_start_kv_head_idx = blockIdx.z;
+  const auto total_num_heads = gridDim.z * GQA_RATIO;
 
   // for QK mfma, tokens in multiples of TOKENS_PER_WARP are spread across warps
   // each mfma takes QH16xT16x16HE across warp
@@ -377,9 +385,10 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
   // fetch Q in shared across warps and then write to registers
   const int local_qhead_idx = 4 * warpid + rowid;
   const int global_qhead_idx = wg_start_head_idx + local_qhead_idx;
-  const int64_t seq_idx64 = static_cast<int64_t>(seq_idx);
+  const int64_t query_start_off = static_cast<int64_t>(
+      query_start_loc_ptr ? query_start_loc_ptr[seq_idx] : seq_idx);
   const scalar_t* q_ptr =
-      q + seq_idx64 * q_stride + global_qhead_idx * HEAD_SIZE;
+      q + query_start_off * q_stride + global_qhead_idx * HEAD_SIZE;
 
   const int qhead_element = lane16id * CONTIGUOUS_SCALAR_ELEMS_16B;
   if ((local_qhead_idx < GQA_RATIO) && (qhead_element < HEAD_SIZE)) {
@@ -777,6 +786,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     const float scale,
     const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,   // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,   // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes, // [num_heads]
     const int q_stride,
@@ -789,14 +799,20 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     int max_ctx_blocks, const float* k_scale, const float* v_scale) {
   // clang-format on
   constexpr int NWARPS = NUM_THREADS / WARP_SIZE;
-  const int warpid = threadIdx.x / WARP_SIZE;
-  const int laneid = threadIdx.x % WARP_SIZE;
+  const auto warpid = threadIdx.x / WARP_SIZE;
+  const auto laneid = threadIdx.x % WARP_SIZE;
   const int lane4id = laneid % 4;
 
-  const int seq_idx = blockIdx.x;
-  const int partition_idx = blockIdx.y;
-  const int partition_size = blockDim.x;
-  const int max_num_partitions = gridDim.y;
+  const auto seq_idx = blockIdx.x;
+  // NOTE queries with sequence len > 1 are prefills and taken care by another
+  // kernel.
+  if (query_start_loc_ptr != nullptr &&
+      (query_start_loc_ptr[seq_idx + 1] - query_start_loc_ptr[seq_idx] != 1)) {
+    return;
+  }
+  const auto partition_idx = blockIdx.y;
+  const auto partition_size = blockDim.x;
+  const auto max_num_partitions = gridDim.y;
 
   const int context_len = context_lens[seq_idx];
   const int partition_start_token_idx = partition_idx * partition_size;
@@ -838,8 +854,8 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     qk_max[h] = -FLT_MAX;
   }
 
-  const int wg_start_head_idx = blockIdx.z * GQA_RATIO;
-  const int wg_start_kv_head_idx = blockIdx.z;
+  const auto wg_start_head_idx = blockIdx.z * GQA_RATIO;
+  const auto wg_start_kv_head_idx = blockIdx.z;
 
   const int warp_start_token_idx =
       partition_start_token_idx + warpid * WARP_SIZE;
@@ -857,7 +873,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
 
     const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
     // token id within partition
-    const int local_token_idx = threadIdx.x;
+    const auto local_token_idx = threadIdx.x;
     // token id within sequence
     const int global_token_idx = partition_start_token_idx + local_token_idx;
 
@@ -882,9 +898,11 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     }
 
     // fetch q elements
-    // every 4 lanes fetch 8 elems, so warp fetches 8*16 = 128 elems
+    // every 4 lanes fetch 8 elems, so warp fetches 8*16 = 128 elemsc
+    const int64_t query_start_off = static_cast<int64_t>(
+        query_start_loc_ptr ? query_start_loc_ptr[seq_idx] : seq_idx);
     const scalar_t* q_ptr =
-        q + seq_idx * q_stride + wg_start_head_idx * HEAD_SIZE;
+        q + query_start_off * q_stride + wg_start_head_idx * HEAD_SIZE;
     const _B16x8* q_ptrh8 = reinterpret_cast<const _B16x8*>(q_ptr);
     const int qhead_elemh8 = laneid / 4;
 
@@ -1126,7 +1144,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
 
   __syncthreads();
 
-  const int num_heads = gridDim.z * GQA_RATIO;
+  const auto num_heads = gridDim.z * GQA_RATIO;
   float* max_logits_ptr =
       max_logits + seq_idx * num_heads * max_num_partitions + partition_idx;
   float* exp_sums_ptr =
@@ -1267,15 +1285,24 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     const scalar_t* __restrict__ tmp_out,  // [num_seqs, num_heads,
                                            // max_num_partitions, head_size]
     const int* __restrict__ context_lens,  // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_partitions) {
-  const int num_heads = gridDim.x;
-  const int head_idx = blockIdx.x;
-  const int seq_idx = blockIdx.y;
+  const auto num_heads = gridDim.x;
+  const auto head_idx = blockIdx.x;
+  const auto seq_idx = blockIdx.y;
+
+  // NOTE queries with sequence len > 1 are prefills and taken care by another
+  // kernel.
+  if (query_start_loc_ptr != nullptr &&
+      (query_start_loc_ptr[seq_idx + 1] - query_start_loc_ptr[seq_idx] != 1)) {
+    return;
+  }
+
   const int context_len = context_lens[seq_idx];
   const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
   [[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-  const int warpid = threadIdx.x / WARP_SIZE;
-  [[maybe_unused]] const int laneid = threadIdx.x % WARP_SIZE;
+  const auto warpid = threadIdx.x / WARP_SIZE;
+  [[maybe_unused]] const auto laneid = threadIdx.x % WARP_SIZE;
 
   __shared__ float shared_global_exp_sum;
   // max num partitions supported is warp_size * NPAR_LOOPS
@@ -1294,7 +1321,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
 
   #pragma unroll
     for (int i = 0; i < NPAR_LOOPS; i++) {
-      const int partition_no = i * WARP_SIZE + threadIdx.x;
+      const auto partition_no = i * WARP_SIZE + threadIdx.x;
       valid_partition[i] =
           (partition_no < num_partitions) ? partition_no : last_valid_partition;
     }
@@ -1324,7 +1351,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     }
   #pragma unroll
     for (int i = 0; i < NPAR_LOOPS; i++) {
-      const int partition_no = i * WARP_SIZE + threadIdx.x;
+      const auto partition_no = i * WARP_SIZE + threadIdx.x;
       rescaled_exp_sum[i] *= (partition_no < num_partitions)
                                  ? expf(reg_max_logit[i] - max_logit)
                                  : 0.0f;
@@ -1336,7 +1363,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     }
   #pragma unroll
     for (int i = 0; i < NPAR_LOOPS; i++) {
-      const int partition_no = i * WARP_SIZE + threadIdx.x;
+      const auto partition_no = i * WARP_SIZE + threadIdx.x;
       shared_exp_sums[partition_no] = rescaled_exp_sum[i];
     }
 
@@ -1439,7 +1466,9 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
       __fdividef(1.0f, shared_global_exp_sum + 1e-6f);
   acc *= inv_global_exp_sum;
 
-  OUTT* out_ptr = out + static_cast<int64_t>(seq_idx) * num_heads * HEAD_SIZE +
+  const int64_t query_start_off = static_cast<int64_t>(
+      query_start_loc_ptr ? query_start_loc_ptr[seq_idx] : seq_idx);
+  OUTT* out_ptr = out + query_start_off * num_heads * HEAD_SIZE +
                   static_cast<int64_t>(head_idx) * HEAD_SIZE;
   if constexpr (std::is_same<OUTT, bit8_t>::value) {
     out_ptr[threadIdx.x] =
@@ -1466,6 +1495,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma16_kernel(
     const float scale,
     const int* __restrict__ block_tables,    // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,    // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride,
@@ -1492,6 +1522,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     const float scale,
     const int* __restrict__ block_tables,    // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,    // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride,
@@ -1515,6 +1546,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     const float* __restrict__ max_logits,  // [num_seqs, num_heads, max_num_partitions]
     const scalar_t* __restrict__ tmp_out,  // [num_seqs, num_heads, max_num_partitions, head_size]
     const int* __restrict__ context_lens,  // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_partitions) {
   UNREACHABLE_CODE
 }
@@ -1522,34 +1554,34 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
 
 #endif  // defined(__HIP__MI300_MI250__) TODO: Add NAVI support
 
-#define LAUNCH_CUSTOM_ATTENTION_MFMA16(GQA_RATIO)                             \
-  paged_attention_ll4mi_QKV_mfma16_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE, \
-                                          HEAD_SIZE, NTHR, ALIBI_ENABLED,     \
-                                          GQA_RATIO>                          \
-      <<<grid, block, 0, stream>>>(                                           \
-          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,     \
-          block_tables_ptr, context_lens_ptr, max_num_blocks_per_seq,         \
-          alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,        \
-          exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr, max_ctx_blocks, \
-          k_scale_ptr, v_scale_ptr);
+#define LAUNCH_CUSTOM_ATTENTION_MFMA16(GQA_RATIO)                              \
+  paged_attention_ll4mi_QKV_mfma16_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE,  \
+                                          HEAD_SIZE, NTHR, ALIBI_ENABLED,      \
+                                          GQA_RATIO>                           \
+      <<<grid, block, 0, stream>>>(                                            \
+          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,      \
+          block_tables_ptr, context_lens_ptr, query_start_loc_ptr,             \
+          max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, kv_block_stride, \
+          kv_head_stride, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr,  \
+          max_ctx_blocks, k_scale_ptr, v_scale_ptr);
 
-#define LAUNCH_CUSTOM_ATTENTION_MFMA4(GQA_RATIO)                              \
-  paged_attention_ll4mi_QKV_mfma4_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE,  \
-                                         HEAD_SIZE, NTHR, ALIBI_ENABLED,      \
-                                         GQA_RATIO>                           \
-      <<<grid, block, 0, stream>>>(                                           \
-          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,     \
-          block_tables_ptr, context_lens_ptr, max_num_blocks_per_seq,         \
-          alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,        \
-          exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr, max_ctx_blocks, \
-          k_scale_ptr, v_scale_ptr);
+#define LAUNCH_CUSTOM_ATTENTION_MFMA4(GQA_RATIO)                               \
+  paged_attention_ll4mi_QKV_mfma4_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE,   \
+                                         HEAD_SIZE, NTHR, ALIBI_ENABLED,       \
+                                         GQA_RATIO>                            \
+      <<<grid, block, 0, stream>>>(                                            \
+          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,      \
+          block_tables_ptr, context_lens_ptr, query_start_loc_ptr,             \
+          max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, kv_block_stride, \
+          kv_head_stride, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr,  \
+          max_ctx_blocks, k_scale_ptr, v_scale_ptr);
 
 #define LAUNCH_CUSTOM_REDUCTION(NPAR_LOOPS)                          \
   paged_attention_ll4mi_reduce_kernel<T, OUTT, HEAD_SIZE, HEAD_SIZE, \
                                       PARTITION_SIZE, NPAR_LOOPS>    \
       <<<reduce_grid, reduce_block, 0, stream>>>(                    \
           out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr,        \
-          context_lens_ptr, max_num_partitions);
+          context_lens_ptr, query_start_loc_ptr, max_num_partitions);
 
 template <typename T, typename KVT, vllm::Fp8KVCacheDataType KV_DTYPE,
           int BLOCK_SIZE, int HEAD_SIZE, typename OUTT, int PARTITION_SIZE_OLD,
@@ -1559,9 +1591,10 @@ void paged_attention_custom_launcher(
     torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, const int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& context_lens,
-    int max_context_len, const std::optional<torch::Tensor>& alibi_slopes,
-    torch::Tensor& k_scale, torch::Tensor& v_scale) {
-  int num_seqs = query.size(0);
+    const std::optional<torch::Tensor>& query_start_loc, int max_context_len,
+    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
+    torch::Tensor& v_scale) {
+  int num_seqs = block_tables.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
   int max_num_blocks_per_seq = block_tables.size(1);
@@ -1569,6 +1602,13 @@ void paged_attention_custom_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
+  // NOTE: query start location is optional for V0 decode should not be used.
+  // If batch contains mix of prefills and decode, prefills should be skipped.
+  const int* query_start_loc_ptr =
+      query_start_loc
+          ? reinterpret_cast<const int*>(query_start_loc.value().data_ptr())
+          : nullptr;
+
   // NOTE: alibi_slopes is optional.
   const float* alibi_slopes_ptr =
       alibi_slopes
@@ -1700,8 +1740,8 @@ void paged_attention_custom_launcher(
   paged_attention_custom_launcher<T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, T, \
                                   PSIZE, ALIBI_ENABLED>(                    \
       out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache,    \
-      num_kv_heads, scale, block_tables, context_lens, max_context_len,     \
-      alibi_slopes, k_scale, v_scale);
+      num_kv_heads, scale, block_tables, context_lens, query_start_loc,     \
+      max_context_len, alibi_slopes, k_scale, v_scale);
 
 #define CALL_CUSTOM_LAUNCHER_ALIBI(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE,      \
                                    PSIZE)                                      \
@@ -1750,6 +1790,7 @@ void paged_attention(
     double scale,
     torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& context_lens, // [num_seqs]
+    const std::optional<torch::Tensor>& query_start_loc, // [num_seqs]
     int64_t block_size, int64_t max_context_len,
     const std::optional<torch::Tensor>& alibi_slopes,
     const std::string& kv_cache_dtype, torch::Tensor& k_scale,

csrc/rocm/ops.h

@@ -7,8 +7,9 @@ void paged_attention(torch::Tensor& out, torch::Tensor& exp_sums,
                      torch::Tensor& query, torch::Tensor& key_cache,
                      torch::Tensor& value_cache, int64_t num_kv_heads,
                      double scale, torch::Tensor& block_tables,
-                     torch::Tensor& context_lens, int64_t block_size,
-                     int64_t max_context_len,
+                     torch::Tensor& context_lens,
+                     const std::optional<torch::Tensor>& query_start_loc,
+                     int64_t block_size, int64_t max_context_len,
                      const std::optional<torch::Tensor>& alibi_slopes,
                      const std::string& kv_cache_dtype, torch::Tensor& k_scale,
                      torch::Tensor& v_scale);

csrc/rocm/torch_bindings.cpp

@@ -23,7 +23,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
       "                Tensor query, Tensor key_cache,"
       "                Tensor value_cache, int num_kv_heads,"
       "                float scale, Tensor block_tables,"
-      "                Tensor context_lens, int block_size,"
+      "                Tensor context_lens,"
+      "                Tensor? query_start_loc,"
+      "                int block_size,"
       "                int max_context_len,"
       "                Tensor? alibi_slopes,"
       "                str kv_cache_dtype,"

csrc/torch_bindings.cpp

@@ -31,6 +31,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("weak_ref_tensor(Tensor input) -> Tensor");
   ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);
 
+  ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
+  ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
+           &get_cuda_view_from_cpu_tensor);
+
   // Attention ops
   // Compute the attention between an input query and the cached
   // keys/values using PagedAttention.
@@ -60,6 +64,21 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    int blocksparse_head_sliding_step) -> ()");
   ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);
 
+#ifndef USE_ROCM
+  // Merge attn states
+  // Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+  // can be used to combine partial attention results (in the split-KV case)
+  ops.def(
+      "merge_attn_states("
+      "    Tensor! output,"
+      "    Tensor!? output_lse,"
+      "    Tensor prefix_output,"
+      "    Tensor prefix_lse,"
+      "    Tensor suffix_output,"
+      "    Tensor suffix_lse) -> ()");
+  ops.impl("merge_attn_states", torch::kCUDA, &merge_attn_states);
+#endif
+
   // Activation ops
   // Activation function used in SwiGLU.
   ops.def("silu_and_mul(Tensor! out, Tensor input) -> ()");
@@ -291,7 +310,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif
 
   // Dequantization for GGML.
-  ops.def("ggml_dequantize(Tensor W, int type, SymInt m, SymInt n) -> Tensor");
+  ops.def(
+      "ggml_dequantize(Tensor W, int type, SymInt m, SymInt n, ScalarType? "
+      "dtype) -> Tensor");
   ops.impl("ggml_dequantize", torch::kCUDA, &ggml_dequantize);
 
   // mmvq kernel for GGML.
@@ -365,6 +386,35 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("cutlass_scaled_mm_supports_fp8(int cuda_device_capability) -> bool");
   ops.impl("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);
 
+  // Check if cutlass grouped gemm is supported for CUDA devices of the given
+  // capability
+  ops.def("cutlass_group_gemm_supported(int cuda_device_capability) -> bool");
+  ops.impl("cutlass_group_gemm_supported", &cutlass_group_gemm_supported);
+
+  // CUTLASS w8a8 grouped GEMM
+  ops.def(
+      "cutlass_moe_mm(Tensor! out_tensors, Tensor a_tensors, Tensor b_tensors, "
+      "               Tensor a_scales, Tensor b_scales, Tensor expert_offsets, "
+      "               Tensor problem_sizes, Tensor a_strides, "
+      "               Tensor b_strides, Tensor c_strides) -> ()",
+      {stride_tag});
+  ops.impl("cutlass_moe_mm", torch::kCUDA, &cutlass_moe_mm);
+
+  // A function that computes data required to run fused MoE with w8a8 grouped
+  // GEMM. It takes topk_ids as an input, and computes expert_offsets
+  // (token start indices of each expert). In addition to this, it computes
+  // problem sizes for each expert's multiplication used by the two mms called
+  // from fused MoE operation, and arrays with permutations required to shuffle
+  // and de-shuffle the input/output of the fused operation.
+  ops.def(
+      "get_cutlass_moe_mm_data(Tensor topk_ids, Tensor! expert_offsets, "
+      "                        Tensor! problem_sizes1, Tensor! problem_sizes2, "
+      "                        Tensor! input_permutation, "
+      "                        Tensor! output_permutation, int num_experts, "
+      "                        int n, int k) -> ()",
+      {stride_tag});
+  ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);
+
   // Check if cutlass scaled_mm supports block quantization (used by DeepSeekV3)
   ops.def(
       "cutlass_scaled_mm_supports_block_fp8(int cuda_device_capability) -> "
@@ -581,12 +631,11 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
                   &get_max_shared_memory_per_block_device_attribute);
 }
 
-#ifndef USE_ROCM
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
   // Custom all-reduce kernels
   custom_ar.def(
       "init_custom_ar(int[] ipc_tensors, Tensor rank_data, "
-      "int rank, bool full_nvlink) -> int");
+      "int rank, bool fully_connected) -> int");
   custom_ar.impl("init_custom_ar", torch::kCUDA, &init_custom_ar);
   custom_ar.def(
       "all_reduce(int fa, Tensor inp, Tensor! out, int reg_buffer, "
@@ -599,7 +648,13 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
   custom_ar.def("register_buffer", &register_buffer);
   custom_ar.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta);
   custom_ar.def("register_graph_buffers", &register_graph_buffers);
+
+  custom_ar.def("allocate_shared_buffer_and_handle",
+                &allocate_shared_buffer_and_handle);
+  custom_ar.def("open_mem_handle(Tensor mem_handle) -> int", &open_mem_handle);
+  custom_ar.impl("open_mem_handle", torch::kCPU, &open_mem_handle);
+
+  custom_ar.def("free_shared_buffer", &free_shared_buffer);
 }
-#endif
 
 REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

docker/Dockerfile

@@ -240,6 +240,8 @@ if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
     uv pip install --system https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.1.post2/flashinfer_python-0.2.1.post2+cu124torch2.6-cp38-abi3-linux_x86_64.whl ; \
 fi
 COPY examples examples
+COPY benchmarks benchmarks
+COPY ./vllm/collect_env.py .
 
 # Although we build Flashinfer with AOT mode, there's still
 # some issues w.r.t. JIT compilation. Therefore we need to

docker/Dockerfile.cpu

@@ -0,0 +1,142 @@
+# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
+#
+# Build targets:
+#   vllm-openai (default): used for serving deployment
+#   vllm-test: used for CI tests
+#   vllm-dev: used for development
+#
+# Build arguments:
+#   PYTHON_VERSION=3.12 (default)|3.11|3.10|3.9
+#   VLLM_CPU_DISABLE_AVX512=false (default)|true
+#
+
+######################### BASE IMAGE #########################
+FROM ubuntu:22.04 AS base
+
+WORKDIR /workspace/
+
+ARG PYTHON_VERSION=3.12
+ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
+
+ENV LD_PRELOAD=""
+
+# Install minimal dependencies and uv
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt,sharing=locked \
+    apt-get update -y \
+    && apt-get install -y --no-install-recommends ccache git curl wget ca-certificates \
+        gcc-12 g++-12 libtcmalloc-minimal4 libnuma-dev ffmpeg libsm6 libxext6 libgl1 \
+    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12 \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh
+
+ENV CCACHE_DIR=/root/.cache/ccache
+ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
+
+ENV PATH="/root/.local/bin:$PATH"
+ENV VIRTUAL_ENV="/opt/venv"
+ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
+RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+
+ENV UV_HTTP_TIMEOUT=500
+
+# Install Python dependencies 
+ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+ENV UV_LINK_MODE="copy"
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/common.txt,target=requirements/common.txt \
+    --mount=type=bind,src=requirements/cpu.txt,target=requirements/cpu.txt \
+    uv pip install --upgrade pip && \
+    uv pip install -r requirements/cpu.txt
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install intel-openmp==2024.2.1 intel_extension_for_pytorch==2.6.0
+
+ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/opt/venv/lib/libiomp5.so:$LD_PRELOAD"
+
+RUN echo 'ulimit -c 0' >> ~/.bashrc
+
+######################### BUILD IMAGE #########################
+FROM base AS vllm-build
+
+ARG GIT_REPO_CHECK=0
+# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
+ARG VLLM_CPU_DISABLE_AVX512
+ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
+
+WORKDIR /workspace/vllm
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
+    uv pip install -r requirements/build.txt
+
+COPY . .
+RUN --mount=type=bind,source=.git,target=.git \
+    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel 
+
+######################### DEV IMAGE #########################
+FROM vllm-build AS vllm-dev
+
+WORKDIR /workspace/vllm
+
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt,sharing=locked \
+    apt-get install -y --no-install-recommends vim numactl
+
+# install development dependencies (for testing)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -e tests/vllm_test_utils 
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py develop 
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -r requirements/dev.txt && \
+    pre-commit install --hook-type pre-commit --hook-type commit-msg
+
+ENTRYPOINT ["bash"]
+
+######################### TEST IMAGE #########################
+FROM base AS vllm-test
+
+WORKDIR /workspace/
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/test.txt,target=requirements/test.txt \
+    uv pip install -r requirements/test.txt
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
+    uv pip install dist/*.whl
+
+ADD ./tests/ ./tests/
+ADD ./examples/ ./examples/
+ADD ./benchmarks/ ./benchmarks/
+ADD ./vllm/collect_env.py .
+
+# install development dependencies (for testing)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -e tests/vllm_test_utils 
+
+ENTRYPOINT ["bash"]
+
+######################### RELEASE IMAGE #########################
+FROM base AS vllm-openai
+
+WORKDIR /workspace/
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
+    uv pip install dist/*.whl
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.hpu

@@ -1,4 +1,4 @@
-FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest
+FROM vault.habana.ai/gaudi-docker/1.20.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
 
 COPY ./ /workspace/vllm
 

docker/Dockerfile.neuron

@@ -1,6 +1,6 @@
 # default base image
 # https://gallery.ecr.aws/neuron/pytorch-inference-neuronx
-ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.21.0-ubuntu22.04"
+ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.22.0-ubuntu22.04"
 
 FROM $BASE_IMAGE
 
@@ -21,9 +21,9 @@ VOLUME [ ${APP_MOUNT} ]
 WORKDIR ${APP_MOUNT}/vllm
 
 RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
-RUN python3 -m pip install sentencepiece transformers==4.45.2 -U
-RUN python3 -m pip install neuronx-cc==2.16.345.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas tenacity
+RUN python3 -m pip install sentencepiece transformers==4.48.0 -U
+RUN python3 -m pip install neuronx-cc==2.17.194.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
 RUN python3 -m pip install pytest
 
 # uninstall transformers-neuronx package explicitly to avoid version conflict

docker/Dockerfile.ppc64le

@@ -0,0 +1,277 @@
+ARG BASE_UBI_IMAGE_TAG=9.5-1741850109
+
+###############################################################
+# base stage with basic dependencies
+###############################################################
+
+FROM registry.access.redhat.com/ubi9/ubi-minimal:${BASE_UBI_IMAGE_TAG} AS base-builder
+
+ARG PYTHON_VERSION=3.12
+ARG OPENBLAS_VERSION=0.3.29
+
+# Set Environment Variables for venv, cargo & openblas
+ENV VIRTUAL_ENV=/opt/vllm
+ENV PATH=${VIRTUAL_ENV}/bin:/root/.cargo/bin:$PATH
+ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig/
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib64:/usr/local/lib:/usr/lib64:/usr/lib
+ENV UV_LINK_MODE=copy
+
+# install gcc-13, python, rust, openblas
+# Note: A symlink for libatomic.so is created for gcc-13 (linker fails to find libatomic otherwise - reqd. for sentencepiece)
+# Note: A dummy file 'control' is created in /tmp/ to artificially create dependencies between stages when building stages in parallel
+#       when `--jobs=<N>` is passed with podman build command
+RUN microdnf install -y openssl-devel dnf \
+    && dnf install -y https://mirror.stream.centos.org/9-stream/BaseOS/`arch`/os/Packages/centos-gpg-keys-9.0-24.el9.noarch.rpm \
+        https://mirror.stream.centos.org/9-stream/BaseOS/`arch`/os/Packages/centos-stream-repos-9.0-24.el9.noarch.rpm \
+        https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm \
+    && dnf config-manager --add-repo https://mirror.stream.centos.org/9-stream/BaseOS/`arch`/os \
+    && dnf config-manager --add-repo https://mirror.stream.centos.org/9-stream/AppStream/`arch`/os \
+    && dnf config-manager --set-enabled crb \
+    && dnf install -y \
+       git tar gcc-toolset-13 automake libtool numactl-devel lapack-devel \
+       pkgconfig xsimd zeromq-devel kmod findutils protobuf* \
+       libtiff-devel libjpeg-devel openjpeg2-devel zlib-devel \
+       freetype-devel lcms2-devel libwebp-devel tcl-devel tk-devel \
+       harfbuzz-devel fribidi-devel libraqm-devel libimagequant-devel libxcb-devel \
+       python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip \
+    && dnf clean all \
+    && ln -sf /usr/lib64/libatomic.so.1 /usr/lib64/libatomic.so \
+    && python${PYTHON_VERSION} -m venv ${VIRTUAL_ENV} \
+    && python -m pip install -U pip uv \
+    && uv pip install wheel build "setuptools<70" setuptools_scm setuptools_rust meson-python 'cmake<4' ninja cython scikit_build_core scikit_build \
+    && curl -sL https://ftp2.osuosl.org/pub/ppc64el/openblas/latest/Openblas_${OPENBLAS_VERSION}_ppc64le.tar.gz | tar xvf - -C /usr/local \
+    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y \
+    && cd /tmp && touch control
+
+###############################################################
+# Stage to build torch family
+###############################################################
+
+FROM base-builder AS torch-builder
+
+ARG MAX_JOBS
+ARG TORCH_VERSION=2.6.0
+ARG _GLIBCXX_USE_CXX11_ABI=1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable &&  \
+    git clone --recursive https://github.com/pytorch/pytorch.git -b v${TORCH_VERSION} && \
+    cd pytorch && \
+    uv pip install -r requirements.txt && \
+    python setup.py develop && \
+    rm -f dist/torch*+git*whl && \
+    MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+    PYTORCH_BUILD_VERSION=${TORCH_VERSION} PYTORCH_BUILD_NUMBER=1 uv build --wheel --out-dir /torchwheels/
+
+ARG TORCHVISION_VERSION=0.21.0
+ARG TORCHVISION_USE_NVJPEG=0
+ARG TORCHVISION_USE_FFMPEG=0
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/pytorch/vision.git -b v${TORCHVISION_VERSION} && \
+    cd vision && \
+    MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+    BUILD_VERSION=${TORCHVISION_VERSION} \
+    uv build --wheel --out-dir /torchwheels/ --no-build-isolation
+
+ARG TORCHAUDIO_VERSION=2.6.0
+ARG BUILD_SOX=1
+ARG BUILD_KALDI=1
+ARG BUILD_RNNT=1
+ARG USE_FFMPEG=0
+ARG USE_ROCM=0
+ARG USE_CUDA=0
+ARG TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG=1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/pytorch/audio.git -b v${TORCHAUDIO_VERSION} && \
+    cd audio && \
+    MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+    BUILD_VERSION=${TORCHAUDIO_VERSION} \
+    uv build --wheel --out-dir /torchwheels/ --no-build-isolation
+
+###############################################################
+# Stage to build pyarrow
+###############################################################
+
+FROM base-builder AS arrow-builder
+
+ARG MAX_JOBS
+ARG PYARROW_PARALLEL
+ARG PYARROW_VERSION=19.0.1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/apache/arrow.git -b apache-arrow-${PYARROW_VERSION} && \
+    cd arrow/cpp && \
+    mkdir build && cd build && \
+    cmake -DCMAKE_BUILD_TYPE=release \
+        -DCMAKE_INSTALL_PREFIX=/usr/local \
+        -DARROW_PYTHON=ON \
+        -DARROW_BUILD_TESTS=OFF \
+        -DARROW_JEMALLOC=ON \
+        -DARROW_BUILD_STATIC="OFF" \
+        -DARROW_PARQUET=ON \
+        .. && \
+    make install -j ${MAX_JOBS:-$(nproc)} && \
+    cd ../../python/ && \
+    uv pip install -v -r requirements-wheel-build.txt && \
+    PYARROW_PARALLEL=${PYARROW_PARALLEL:-$(nproc)} \
+    python setup.py build_ext \
+    --build-type=release --bundle-arrow-cpp \
+    bdist_wheel --dist-dir /arrowwheels/
+
+###############################################################
+# Stage to build opencv
+###############################################################
+
+FROM base-builder AS cv-builder
+
+ARG MAX_JOBS
+ARG OPENCV_VERSION=86
+# patch for version 4.11.0.86
+ARG OPENCV_PATCH=97f3f39
+ARG ENABLE_HEADLESS=1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/opencv/opencv-python.git -b ${OPENCV_VERSION} && \
+    cd opencv-python && \
+    sed -i -E -e 's/"setuptools.+",/"setuptools",/g' pyproject.toml && \
+    cd opencv && git cherry-pick --no-commit $OPENCV_PATCH && cd .. && \
+    python -m build --wheel --installer=uv --outdir /opencvwheels/
+
+###############################################################
+# Stage to build vllm - this stage builds and installs
+# vllm, tensorizer and vllm-tgis-adapter and builds uv cache
+# for transitive dependencies - eg. grpcio
+###############################################################
+
+FROM base-builder AS vllmcache-builder
+
+COPY --from=torch-builder /tmp/control /dev/null
+COPY --from=arrow-builder /tmp/control /dev/null
+COPY --from=cv-builder /tmp/control /dev/null
+
+ARG VLLM_TARGET_DEVICE=cpu
+ARG GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=1
+
+# this step installs vllm and populates uv cache
+# with all the transitive dependencies
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone https://github.com/huggingface/xet-core.git && cd xet-core/hf_xet/ && \
+    uv pip install maturin && \
+    uv build --wheel --out-dir /hf_wheels/
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
+    --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
+    --mount=type=bind,from=cv-builder,source=/opencvwheels/,target=/opencvwheels/,ro \
+    --mount=type=bind,src=.,dst=/src/,rw \
+    source /opt/rh/gcc-toolset-13/enable && \
+    uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl && \
+    sed -i -e 's/.*torch.*//g' /src/pyproject.toml /src/requirements/*.txt && \
+    uv pip install pandas pythran pybind11 /hf_wheels/*.whl && \
+    # sentencepiece.pc is in some pkgconfig inside uv cache
+    export PKG_CONFIG_PATH=$(find / -type d -name "pkgconfig" 2>/dev/null | tr '\n' ':') && \
+    uv pip install -r /src/requirements/common.txt -r /src/requirements/cpu.txt -r /src/requirements/build.txt --no-build-isolation && \
+    cd /src/ && \
+    uv build --wheel --out-dir /vllmwheel/ --no-build-isolation && \
+    uv pip install /vllmwheel/*.whl
+
+
+###############################################################
+# Stage to build numactl
+###############################################################
+
+FROM base-builder AS numa-builder
+
+# Note: Building numactl with gcc-11. Compiling with gcc-13 in this builder stage will
+# trigger recompilation with gcc-11 (and require libtool) in the final stage where we do not have gcc-13
+ARG MAX_JOBS
+ARG NUMACTL_VERSION=2.0.19
+RUN git clone --recursive https://github.com/numactl/numactl.git -b v${NUMACTL_VERSION} \
+    && cd numactl \
+    && autoreconf -i && ./configure \
+    && make -j ${MAX_JOBS:-$(nproc)}
+
+###############################################################
+# Stage to build lapack
+###############################################################
+
+FROM base-builder AS lapack-builder
+
+ARG MAX_JOBS
+ARG LAPACK_VERSION=3.12.1
+RUN git clone --recursive https://github.com/Reference-LAPACK/lapack.git -b v${LAPACK_VERSION} \
+    && cd lapack && source /opt/rh/gcc-toolset-13/enable \
+    && cmake -B build -S . \
+    && cmake --build build -j ${MAX_JOBS:-$(nproc)}
+
+
+###############################################################
+#                   FINAL VLLM IMAGE STAGE                    #
+###############################################################
+
+FROM registry.access.redhat.com/ubi9/ubi-minimal:${BASE_UBI_IMAGE_TAG} AS vllm-openai
+
+ARG PYTHON_VERSION=3.12
+ARG OPENBLAS_VERSION=0.3.29
+
+# Set Environment Variables for venv & openblas
+ENV VIRTUAL_ENV=/opt/vllm
+ENV PATH=${VIRTUAL_ENV}/bin:$PATH
+ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig/
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib64:/usr/local/lib:/usr/lib64:/usr/lib
+ENV UV_LINK_MODE=copy
+
+# create artificial dependencies between stages for independent stages to build in parallel
+COPY --from=torch-builder /tmp/control /dev/null
+COPY --from=arrow-builder /tmp/control /dev/null
+COPY --from=cv-builder /tmp/control /dev/null
+COPY --from=vllmcache-builder /tmp/control /dev/null
+COPY --from=numa-builder /tmp/control /dev/null
+COPY --from=lapack-builder /tmp/control /dev/null
+
+# install gcc-11, python, openblas, numactl, lapack
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=numa-builder,source=/numactl/,target=/numactl/,rw \
+    --mount=type=bind,from=lapack-builder,source=/lapack/,target=/lapack/,rw \
+    rpm -ivh https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm && \
+    microdnf install --nodocs -y \
+    tar findutils openssl \
+    pkgconfig xsimd g++ gcc-fortran libsndfile \
+    libtiff libjpeg openjpeg2 zlib zeromq \
+    freetype lcms2 libwebp tcl tk utf8proc \
+    harfbuzz fribidi libraqm libimagequant libxcb \
+    python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip \
+    && microdnf clean all \
+    && python${PYTHON_VERSION} -m venv ${VIRTUAL_ENV} \
+    && python -m pip install -U pip uv --no-cache \
+    && curl -sL https://ftp2.osuosl.org/pub/ppc64el/openblas/latest/Openblas_${OPENBLAS_VERSION}_ppc64le.tar.gz | tar xvf - -C /usr/local \
+    && make -C /numactl install \
+    && uv pip install 'cmake<4' \
+    && cmake --install /lapack/build \
+    && uv pip uninstall cmake
+
+# consume previously built wheels (including vllm)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
+    --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
+    --mount=type=bind,from=cv-builder,source=/opencvwheels/,target=/opencvwheels/,ro \
+    --mount=type=bind,from=vllmcache-builder,source=/hf_wheels/,target=/hf_wheels/,ro \
+    --mount=type=bind,from=vllmcache-builder,source=/vllmwheel/,target=/vllmwheel/,ro \
+    HOME=/root uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl /hf_wheels/*.whl /vllmwheel/*.whl
+
+COPY ./ /workspace/vllm
+WORKDIR /workspace/vllm
+ARG GIT_REPO_CHECK=0
+RUN --mount=type=bind,source=.git,target=.git \
+    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
+
+# install development dependencies (for testing)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]