[VLM] Update Qwen3-VL max_num_video_tokens calculation for configurable video profiling (#25557)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Roger Wang <hey@rogerw.io>
Signed-off-by: simon-mo <simon.mo@hey.com>

.buildkite/release-pipeline.yaml

@@ -76,7 +76,7 @@ steps:
       queue: arm64_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=12.9.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "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=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
 
   # Add job to create multi-arch manifest

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

@@ -58,11 +58,8 @@ function cpu_tests() {
     # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
     # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
 
-    # Note: disable Bart until supports V1
-    pytest -x -v -s tests/models/language/generation -m cpu_model \
-                --ignore=tests/models/language/generation/test_bart.py
-    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model \
-                --ignore=tests/models/language/generation/test_bart.py
+    pytest -x -v -s tests/models/language/generation -m cpu_model
+    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model
 
     pytest -x -v -s tests/models/language/pooling -m cpu_model
     pytest -x -v -s tests/models/multimodal/generation \

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

@@ -35,7 +35,7 @@ docker run \
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    VLLM_ATTENTION_BACKEND=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
     cd tests
     pytest -v -s v1/core
     pytest -v -s v1/engine

.buildkite/scripts/run-prime-rl-test.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Setup script for Prime-RL integration tests
+# This script prepares the environment for running Prime-RL tests with nightly vLLM
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
+PRIME_RL_REPO="https://github.com/PrimeIntellect-ai/prime-rl.git"
+PRIME_RL_DIR="${REPO_ROOT}/prime-rl"
+
+echo "Setting up Prime-RL integration test environment..."
+
+# Clean up any existing Prime-RL directory
+if [ -d "${PRIME_RL_DIR}" ]; then
+    echo "Removing existing Prime-RL directory..."
+    rm -rf "${PRIME_RL_DIR}"
+fi
+
+# Install UV if not available
+if ! command -v uv &> /dev/null; then
+    echo "Installing UV package manager..."
+    curl -LsSf https://astral.sh/uv/install.sh | sh
+    source $HOME/.local/bin/env
+fi
+
+# Clone Prime-RL repository at specific branch for reproducible tests
+PRIME_RL_BRANCH="integ-vllm-main"
+echo "Cloning Prime-RL repository at branch: ${PRIME_RL_BRANCH}..."
+git clone --branch "${PRIME_RL_BRANCH}" --single-branch "${PRIME_RL_REPO}" "${PRIME_RL_DIR}"
+cd "${PRIME_RL_DIR}"
+
+echo "Setting up UV project environment..."
+export UV_PROJECT_ENVIRONMENT=/usr/local
+ln -s /usr/bin/python3 /usr/local/bin/python
+
+# Remove vllm pin from pyproject.toml
+echo "Removing vllm pin from pyproject.toml..."
+sed -i '/vllm==/d' pyproject.toml
+
+# Sync Prime-RL dependencies
+echo "Installing Prime-RL dependencies..."
+uv sync --inexact && uv sync --inexact --all-extras
+
+# Verify installation
+echo "Verifying installations..."
+uv run python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
+uv run python -c "import prime_rl; print('Prime-RL imported successfully')"
+
+echo "Prime-RL integration test environment setup complete!"
+
+echo "Running Prime-RL integration tests..."
+export WANDB_MODE=offline # this makes this test not require a WANDB_API_KEY
+uv run pytest -vs tests/integration/test_rl.py -m gpu
+
+echo "Prime-RL integration tests completed!"

.buildkite/test-pipeline.yaml

@@ -164,6 +164,7 @@ steps:
   - tests/v1/test_internal_lb_dp.py
   - tests/v1/test_hybrid_lb_dp.py
   - tests/v1/engine/test_engine_core_client.py
+  - tests/distributed/test_symm_mem_allreduce.py
   commands:
   # test with torchrun tp=2 and external_dp=2
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
@@ -188,6 +189,7 @@ steps:
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
+  - pytest -v -s distributed/test_symm_mem_allreduce.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
@@ -298,10 +300,12 @@ steps:
     - pytest -v -s v1/spec_decode
     - pytest -v -s v1/kv_connector/unit
     - pytest -v -s v1/metrics
+    - pytest -v -s v1/test_kv_sharing.py
+    - pytest -v -s v1/test_metrics_reader.py
+    - pytest -v -s v1/test_oracle.py
+    - pytest -v -s v1/test_request.py
     - pytest -v -s v1/test_serial_utils.py
     - pytest -v -s v1/test_utils.py
-    - pytest -v -s v1/test_oracle.py
-    - pytest -v -s v1/test_metrics_reader.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -329,6 +333,8 @@ steps:
     - python3 offline_inference/basic/classify.py
     - python3 offline_inference/basic/embed.py
     - python3 offline_inference/basic/score.py
+    - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
+    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
 
 - label: Platform Tests (CUDA) # 4min
   timeout_in_minutes: 15
@@ -766,8 +772,9 @@ steps:
     - pytest -v -s tests/models/multimodal/processing/
     - pytest -v -s tests/models/multimodal/test_mapping.py
     - python3 examples/offline_inference/basic/chat.py
-    - python3 examples/offline_inference/audio_language.py --model-type whisper
     - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+    # Whisper needs spawn method to avoid deadlock
+    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
 - label: Blackwell Test # 38 min
   timeout_in_minutes: 60
@@ -865,24 +872,28 @@ steps:
     - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
     - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
 
-- label: Distributed Tests (2 GPUs) # 110min
-  timeout_in_minutes: 150
+- label: Distributed Tests (2 GPUs) # 68min
+  timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
+  - vllm/compilation/
   - vllm/distributed/
   - vllm/engine/
   - vllm/executor/
-  - vllm/model_executor/models/
-  - tests/distributed/
-  - vllm/compilation
   - vllm/worker/worker_base.py
-  - entrypoints/llm/test_collective_rpc.py
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/compile/test_basic_correctness.py
+  - tests/compile/test_wrapper.py
+  - tests/distributed/
+  - tests/entrypoints/llm/test_collective_rpc.py
   - tests/v1/test_async_llm_dp.py
   - tests/v1/test_external_lb_dp.py
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
-  - vllm/v1/engine/
+  - tests/v1/shutdown
+  - tests/v1/worker/test_worker_memory_snapshot.py
   commands:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
@@ -891,19 +902,29 @@ steps:
   - 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'
+  - pytest -v -s distributed/test_sequence_parallel.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
+  - pytest -v -s v1/worker/test_worker_memory_snapshot.py
+
+- label: Distributed Model Tests (2 GPUs) # 37min
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  source_file_dependencies:
+  - vllm/model_executor/model_loader/sharded_state_loader.py
+  - vllm/model_executor/models/
+  - tests/basic_correctness/
+  - tests/model_executor/model_loader/test_sharded_state_loader.py
+  - tests/models/
+  commands:
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py
   # Avoid importing model tests that cause CUDA reinitialization error
   - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/language -v -s -m 'distributed(num_gpus=2)'
   - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
   - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
-  # test sequence parallel
-  - pytest -v -s distributed/test_sequence_parallel.py
-  # this test fails consistently.
-  # TODO: investigate and fix
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
-  - pytest -v -s models/multimodal/generation/test_maverick.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   timeout_in_minutes: 60
@@ -1037,3 +1058,16 @@ steps:
   num_gpus: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
+    - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
+
+##### RL Integration Tests #####
+- label: Prime-RL Integration Test # 15min
+  timeout_in_minutes: 30
+  optional: true
+  num_gpus: 2
+  working_dir: "/vllm-workspace"
+  source_file_dependencies:
+  - vllm/
+  - .buildkite/scripts/run-prime-rl-test.sh
+  commands:
+    - bash .buildkite/scripts/run-prime-rl-test.sh

.readthedocs.yaml

@@ -13,6 +13,7 @@ build:
 
 mkdocs:
   configuration: mkdocs.yaml
+  fail_on_warning: true
 
 # Optionally declare the Python requirements required to build your docs
 python:

benchmarks/auto_tune/auto_tune.sh

@@ -103,10 +103,15 @@ start_server() {
         VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
             vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
     fi
+    local server_pid=$!
 
     # wait for 10 minutes...
     server_started=0
     for i in {1..60}; do
+        # This line checks whether the server is still alive or not,
+        # since that we should always have permission to send signal to the server process.
+        kill -0 $server_pid 2> /dev/null || break
+
         RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
         STATUS_CODE=$(echo "$RESPONSE" | tail -n 1)
         if [[ "$STATUS_CODE" -eq 200 ]]; then
@@ -118,7 +123,7 @@ start_server() {
     done
 
     if (( ! server_started )); then
-        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
+        echo "server did not start within 10 minutes or crashed. Please check server log at $vllm_log".
         return 1
     else
         return 0

benchmarks/benchmark_ngram_proposer.py

@@ -1,17 +1,31 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import gc
+import time
+from unittest import mock
 
 import numpy as np
 from tabulate import tabulate
 
 from benchmark_utils import TimeCollector
-from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
+from vllm.config import (
+    CacheConfig,
+    DeviceConfig,
+    LoadConfig,
+    ModelConfig,
+    ParallelConfig,
+    SchedulerConfig,
+    SpeculativeConfig,
+    VllmConfig,
+)
+from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
+from vllm.v1.worker.gpu_input_batch import InputBatch
+from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 
 
-def main(args):
+def benchmark_propose(args):
     rows = []
     for max_ngram in args.max_ngram:
         collector = TimeCollector(TimeCollector.US)
@@ -69,10 +83,88 @@ def main(args):
     )
 
 
+def benchmark_batched_propose(args):
+    NUM_SPECULATIVE_TOKENS_NGRAM = 10
+    PROMPT_LOOKUP_MIN = 5
+    PROMPT_LOOKUP_MAX = 15
+    MAX_MODEL_LEN = int(1e7)
+    DEVICE = current_platform.device_type
+
+    model_config = ModelConfig(model="facebook/opt-125m", runner="generate")
+
+    speculative_config = SpeculativeConfig(
+        target_model_config=model_config,
+        target_parallel_config=ParallelConfig(),
+        method="ngram",
+        num_speculative_tokens=NUM_SPECULATIVE_TOKENS_NGRAM,
+        prompt_lookup_max=PROMPT_LOOKUP_MAX,
+        prompt_lookup_min=PROMPT_LOOKUP_MIN,
+    )
+
+    vllm_config = VllmConfig(
+        model_config=model_config,
+        cache_config=CacheConfig(),
+        speculative_config=speculative_config,
+        device_config=DeviceConfig(device=current_platform.device_type),
+        parallel_config=ParallelConfig(),
+        load_config=LoadConfig(),
+        scheduler_config=SchedulerConfig(),
+    )
+
+    # monkey patch vllm.v1.worker.gpu_model_runner.get_pp_group
+    mock_pp_group = mock.MagicMock()
+    mock_pp_group.world_size = 1
+    with mock.patch(
+        "vllm.v1.worker.gpu_model_runner.get_pp_group", return_value=mock_pp_group
+    ):
+        runner = GPUModelRunner(vllm_config, DEVICE)
+
+        # hack max model len
+        runner.max_model_len = MAX_MODEL_LEN
+        runner.drafter.max_model_len = MAX_MODEL_LEN
+
+        dummy_input_batch = InputBatch(
+            max_num_reqs=args.num_req,
+            max_model_len=MAX_MODEL_LEN,
+            max_num_batched_tokens=args.num_req * args.num_token,
+            device=DEVICE,
+            pin_memory=False,
+            vocab_size=256000,
+            block_sizes=[16],
+        )
+        dummy_input_batch._req_ids = list(str(id) for id in range(args.num_req))
+        dummy_input_batch.spec_decode_unsupported_reqs = ()
+        dummy_input_batch.num_tokens_no_spec = [args.num_token] * args.num_req
+        dummy_input_batch.token_ids_cpu = np.random.randint(
+            0, 20, (args.num_req, args.num_token)
+        )
+
+        runner.input_batch = dummy_input_batch
+
+        sampled_token_ids = [[0]] * args.num_req
+
+        print("Starting benchmark")
+        # first run is warmup so ignore it
+        for _ in range(args.num_iteration):
+            start = time.time()
+            runner.drafter.propose(
+                sampled_token_ids,
+                dummy_input_batch.req_ids,
+                dummy_input_batch.num_tokens_no_spec,
+                dummy_input_batch.token_ids_cpu,
+                dummy_input_batch.spec_decode_unsupported_reqs,
+            )
+            end = time.time()
+            print(f"Iteration time (s): {end - start}")
+
+
 def invoke_main() -> None:
     parser = FlexibleArgumentParser(
         description="Benchmark the performance of N-gram speculative decode drafting"
     )
+    parser.add_argument(
+        "--batched", action="store_true", help="consider time to prepare batch"
+    )  # noqa: E501
     parser.add_argument(
         "--num-iteration",
         type=int,
@@ -105,8 +197,17 @@ def invoke_main() -> None:
         help="Number of speculative tokens to generate",
     )
     args = parser.parse_args()
-    main(args)
+
+    if not args.batched:
+        benchmark_propose(args)
+    else:
+        benchmark_batched_propose(args)
 
 
+"""
+# Example command lines:
+# time python3 benchmarks/benchmark_ngram_proposer.py
+# time python3 benchmarks/benchmark_ngram_proposer.py --batched --num-iteration 4 --num-token 1000000 --num-req 128
+"""  # noqa: E501
 if __name__ == "__main__":
     invoke_main()  # pragma: no cover

benchmarks/benchmark_serving_structured_output.py

@@ -449,7 +449,8 @@ async def benchmark(
     def prepare_extra_body(request) -> dict:
         extra_body = {}
         # Add the schema to the extra_body
-        extra_body[request.structure_type] = request.schema
+        extra_body["structured_outputs"] = {}
+        extra_body["structured_outputs"][request.structure_type] = request.schema
         return extra_body
 
     print("Starting initial single prompt test run...")

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -3,6 +3,7 @@
 import argparse
 import copy
 import itertools
+import os
 
 import torch
 from weight_shapes import WEIGHT_SHAPES
@@ -23,21 +24,45 @@ PROVIDER_CFGS = {
     "torch-bf16": dict(enabled=True),
     "nvfp4": dict(no_a_quant=False, enabled=True),
     "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
+    "fbgemm-nvfp4": dict(fbgemm=True, no_a_quant=False, enabled=True),
+    "fbgemm-nvfp4-noquant": dict(fbgemm=True, no_a_quant=True, enabled=True),
 }
 
+_needs_fbgemm = any(
+    v.get("fbgemm", False) for v in PROVIDER_CFGS.values() if v.get("enabled", False)
+)
+if _needs_fbgemm:
+    try:
+        from fbgemm_gpu.experimental.gemm.triton_gemm.fp4_quantize import (
+            triton_scale_nvfp4_quant,
+        )
+    except ImportError:
+        print(
+            "WARNING: FBGEMM providers are enabled but fbgemm_gpu is not installed. "
+            "These providers will be skipped. Please install fbgemm_gpu with: "
+            "'pip install fbgemm-gpu-genai' to run them."
+        )
+        # Disable FBGEMM providers so the benchmark can run.
+        for cfg in PROVIDER_CFGS.values():
+            if cfg.get("fbgemm"):
+                cfg["enabled"] = False
+
 _enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
 
 
-def _quant_weight_nvfp4(b: torch.Tensor, device: str):
+def _quant_weight_nvfp4(b: torch.Tensor, device: str, cfg):
     # Compute global scale for weight
     b_amax = torch.abs(b).max().to(torch.float32)
     b_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / b_amax
-    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    if "fbgemm" in cfg and cfg["fbgemm"]:
+        b_fp4, scale_b_fp4 = triton_scale_nvfp4_quant(b, b_global_scale)
+    else:
+        b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
     return b_fp4, scale_b_fp4, b_global_scale
 
 
 def build_nvfp4_runner(cfg, a, b, dtype, device):
-    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device)
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device, cfg)
 
     # Compute global scale for activation
     # NOTE: This is generally provided ahead-of-time by the model checkpoint.
@@ -46,6 +71,35 @@ def build_nvfp4_runner(cfg, a, b, dtype, device):
 
     # Alpha for the GEMM operation
     alpha = 1.0 / (a_global_scale * b_global_scale)
+    if "fbgemm" in cfg and cfg["fbgemm"]:
+        if cfg["no_a_quant"]:
+            a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
+
+            def run():
+                return torch.ops.fbgemm.f4f4bf16(
+                    a_fp4,
+                    b_fp4,
+                    scale_a_fp4,
+                    scale_b_fp4,
+                    global_scale=alpha,
+                    use_mx=False,
+                )
+
+            return run
+        else:
+
+            def run():
+                a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
+                return torch.ops.fbgemm.f4f4bf16(
+                    a_fp4,
+                    b_fp4,
+                    scale_a_fp4,
+                    scale_b_fp4,
+                    global_scale=alpha,
+                    use_mx=False,
+                )
+
+            return run
 
     if cfg["no_a_quant"]:
         # Pre-quantize activation
@@ -130,10 +184,13 @@ if __name__ == "__main__":
 
     for K, N, model in prepare_shapes(args):
         print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
+        save_dir = f"bench_nvfp4_res_n{N}_k{K}"
+        os.makedirs(save_dir, exist_ok=True)
+
         benchmark.run(
             print_data=True,
             show_plots=True,
-            save_path=f"bench_nvfp4_res_n{N}_k{K}",
+            save_path=save_dir,
             N=N,
             K=K,
         )

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -51,7 +51,7 @@ def calculate_diff(
 ):
     """Calculate the difference between Inductor and CUDA implementations."""
     device = torch.device("cuda")
-    x = torch.rand((batch_size * hidden_size, 4096), dtype=dtype, device=device)
+    x = torch.randn((batch_size, hidden_size), dtype=dtype, device=device)
 
     quant_fp8 = QuantFP8(False, group_shape, column_major_scales=False)
 
@@ -59,23 +59,25 @@ def calculate_diff(
     torch_eager_out, torch_eager_scale = quant_fp8.forward_native(x)
     cuda_out, cuda_scale = quant_fp8.forward_cuda(x)
 
-    out_allclose = lambda o1, o2: torch.allclose(
-        o1.to(torch.float32),
-        o2.to(torch.float32),
-        rtol=1e-3,
-        atol=1e-5,
-    )
-    scale_allclose = lambda s1, s2: torch.allclose(s1, s2, rtol=1e-3, atol=1e-5)
-
-    if (
-        out_allclose(cuda_out, torch_out)
-        and scale_allclose(cuda_scale, torch_scale)
-        and out_allclose(cuda_out, torch_eager_out)
-        and scale_allclose(cuda_scale, torch_eager_scale)
-    ):
+    try:
+        torch.testing.assert_close(
+            cuda_out.to(torch.float32),
+            torch_out.to(torch.float32),
+            rtol=1e-3,
+            atol=1e-5,
+        )
+        torch.testing.assert_close(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5)
+        torch.testing.assert_close(
+            cuda_out.to(torch.float32),
+            torch_eager_out.to(torch.float32),
+            rtol=1e-3,
+            atol=1e-5,
+        )
+        torch.testing.assert_close(cuda_scale, torch_eager_scale, rtol=1e-3, atol=1e-5)
         print("✅ All implementations match")
-    else:
+    except AssertionError as e:
         print("❌ Implementations differ")
+        print(e)
 
 
 configs = []
@@ -91,7 +93,7 @@ def benchmark_quantization(
 ):
     device = torch.device("cuda")
 
-    x = torch.randn(batch_size * hidden_size, 4096, device=device, dtype=dtype)
+    x = torch.randn(batch_size, hidden_size, device=device, dtype=dtype)
 
     quantiles = [0.5, 0.2, 0.8]
     quant_fp8 = QuantFP8(False, group_shape, column_major_scales=col_major)
@@ -157,21 +159,21 @@ if __name__ == "__main__":
     )
     parser.add_argument("-c", "--check", action="store_true")
     parser.add_argument(
-        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
     )
     parser.add_argument(
         "--hidden-sizes",
         type=int,
         nargs="+",
-        default=None,
-        help="Hidden sizes to benchmark (default: 1,16,64,128,256,512,1024,2048,4096)",
+        default=[896, 1024, 2048, 4096, 7168],
+        help="Hidden sizes to benchmark",
     )
     parser.add_argument(
         "--batch-sizes",
         type=int,
         nargs="+",
-        default=None,
-        help="Batch sizes to benchmark (default: 1,16,32,64,128)",
+        default=[1, 16, 128, 512, 1024],
+        help="Batch sizes to benchmark",
     )
     parser.add_argument(
         "--group-sizes",
@@ -192,8 +194,8 @@ if __name__ == "__main__":
 
     dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
 
-    hidden_sizes = args.hidden_sizes or [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
-    batch_sizes = args.batch_sizes or [1, 16, 32, 64, 128]
+    hidden_sizes = args.hidden_sizes
+    batch_sizes = args.batch_sizes
 
     if args.group_sizes is not None:
         group_shapes = []

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -0,0 +1,406 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark the performance of the cutlass_moe_fp8 kernel vs the triton_moe
+kernel. Both kernels take in fp8 quantized weights and 16-bit activations,
+but use different quantization strategies and backends.
+"""
+
+import nvtx
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
+from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
+from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+# Weight shapes for different models: [num_experts, topk, hidden_size,
+# intermediate_size]
+WEIGHT_SHAPES_MOE = {
+    "mixtral-8x7b": [
+        [8, 2, 4096, 14336],
+    ],
+    "deepseek-v2": [
+        [160, 6, 5120, 12288],
+    ],
+    "custom-small": [
+        [8, 2, 2048, 7168],
+    ],
+    "glm45-fp8": [
+        [128, 8, 4096, 1408],
+    ],
+    "Llama-4-Maverick-17B-128E-Instruct-FP8": [
+        [128, 1, 5120, 8192],
+    ],
+}
+
+DEFAULT_MODELS = [
+    "mixtral-8x7b",
+]
+
+DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]
+DEFAULT_TP_SIZES = [1]
+
+PER_ACT_TOKEN_OPTS = [False, True]
+PER_OUT_CH_OPTS = [False, True]
+
+FP8_DTYPE = current_platform.fp8_dtype()
+
+
+def bench_run(
+    results: list,
+    model: str,
+    num_experts: int,
+    topk: int,
+    per_act_token: bool,
+    per_out_ch: bool,
+    mkn: tuple[int, int, int],
+):
+    (m, k, n) = mkn
+
+    dtype = torch.half
+    device = "cuda"
+
+    # Create input activations
+    a = torch.randn((m, k), device=device, dtype=dtype) / 10
+
+    # Create weights
+    w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10
+    w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10
+
+    # Create FP8 quantized weights and scales for both kernels
+    w1_fp8q = torch.empty((num_experts, 2 * n, k), device=device, dtype=FP8_DTYPE)
+    w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=FP8_DTYPE)
+
+    # Create scales based on quantization strategy
+    if per_out_ch:
+        # Per-channel quantization
+        w1_scale = torch.empty(
+            (num_experts, 2 * n, 1), device=device, dtype=torch.float32
+        )
+        w2_scale = torch.empty((num_experts, k, 1), device=device, dtype=torch.float32)
+    else:
+        # Per-tensor quantization
+        w1_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
+        w2_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
+
+    # Quantize weights
+    for expert in range(num_experts):
+        if per_out_ch:
+            # Per-channel quantization - not yet implemented properly
+            # For now, fall back to per-tensor quantization
+            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
+            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
+            # Expand scalar scales to the expected per-channel shape
+            w1_scale[expert] = w1_scale_temp.expand(2 * n, 1)
+            w2_scale[expert] = w2_scale_temp.expand(k, 1)
+        else:
+            # Per-tensor quantization
+            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
+            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
+            # Store scalar scales in [1, 1] tensors
+            w1_scale[expert, 0, 0] = w1_scale_temp
+            w2_scale[expert, 0, 0] = w2_scale_temp
+
+    # Prepare weights for CUTLASS (no transpose needed)
+    w1_fp8q_cutlass = w1_fp8q  # Keep original [E, 2N, K]
+    w2_fp8q_cutlass = w2_fp8q  # Keep original [E, K, N]
+
+    # Create router scores and get topk
+    score = torch.randn((m, num_experts), device=device, dtype=dtype)
+    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
+
+    # WORKAROUND: CUTLASS MoE FP8 has issues with per-token quantization
+    # Force per-tensor quantization for all cases to match working e2e setup
+    a1_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
+    a2_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
+
+    # Force per-tensor quantization for all cases
+    per_act_token = False
+
+    # Create stride tensors for CUTLASS
+    ab_strides1 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
+    ab_strides2 = torch.full((num_experts,), n, dtype=torch.int64, device=device)
+    c_strides1 = torch.full((num_experts,), 2 * n, dtype=torch.int64, device=device)
+    c_strides2 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
+
+    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,
+        a1_scale: torch.Tensor,
+        a2_scale: torch.Tensor,
+        num_repeats: int,
+    ):
+        quant_config = fp8_w8a8_moe_quant_config(
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+            a1_scale=a1_scale,
+            a2_scale=a2_scale,
+            per_act_token_quant=per_act_token,
+            per_out_ch_quant=per_out_ch,
+        )
+
+        for _ in range(num_repeats):
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                quant_config=quant_config,
+            )
+
+    def run_cutlass_moe_fp8(
+        a: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
+        w1_scale: torch.Tensor,
+        w2_scale: torch.Tensor,
+        a1_scale: torch.Tensor,
+        a2_scale: torch.Tensor,
+        num_repeats: int,
+    ):
+        quant_config = fp8_w8a8_moe_quant_config(
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+            a1_scale=a1_scale,
+            a2_scale=a2_scale,
+            per_act_token_quant=per_act_token,
+            per_out_ch_quant=per_out_ch,
+        )
+
+        for _ in range(num_repeats):
+            with nvtx.annotate("cutlass_moe_fp8", color="blue"):
+                cutlass_moe_fp8(
+                    a=a,
+                    w1_q=w1,
+                    w2_q=w2,
+                    topk_weights=topk_weights,
+                    topk_ids=topk_ids,
+                    ab_strides1=ab_strides1,
+                    ab_strides2=ab_strides2,
+                    c_strides1=c_strides1,
+                    c_strides2=c_strides2,
+                    quant_config=quant_config,
+                    activation="silu",
+                    global_num_experts=num_experts,
+                )
+
+    # Pre-create quantization config to avoid creating it inside CUDA graph
+    quant_config = fp8_w8a8_moe_quant_config(
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        a1_scale=a1_scale,
+        a2_scale=a2_scale,
+        per_act_token_quant=per_act_token,
+        per_out_ch_quant=per_out_ch,
+    )
+
+    # Create CUDA graphs for CUTLASS (match benchmark_moe.py pattern exactly)
+    cutlass_stream = torch.cuda.Stream()
+    cutlass_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
+        # Capture 10 invocations like benchmark_moe.py
+        for _ in range(10):
+            cutlass_moe_fp8(
+                a=a,
+                w1_q=w1_fp8q_cutlass,
+                w2_q=w2_fp8q_cutlass,
+                topk_weights=topk_weights,
+                topk_ids=topk_ids,
+                ab_strides1=ab_strides1,
+                ab_strides2=ab_strides2,
+                c_strides1=c_strides1,
+                c_strides2=c_strides2,
+                quant_config=quant_config,
+                activation="silu",
+                global_num_experts=num_experts,
+            )
+    torch.cuda.synchronize()
+
+    # Create CUDA graphs for Triton (match benchmark_moe.py pattern exactly)
+    triton_stream = torch.cuda.Stream()
+    triton_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(triton_graph, stream=triton_stream):
+        # Capture 10 invocations like benchmark_moe.py
+        for _ in range(10):
+            fused_experts(
+                a,
+                w1_fp8q,
+                w2_fp8q,
+                topk_weights,
+                topk_ids,
+                quant_config=quant_config,
+            )
+    torch.cuda.synchronize()
+
+    def bench_cuda_graph(graph, num_warmup=5, num_iters=100):
+        """Benchmark CUDA graph using events like benchmark_moe.py"""
+        # Warmup
+        for _ in range(num_warmup):
+            graph.replay()
+        torch.cuda.synchronize()
+
+        # Timing
+        start_event = torch.cuda.Event(enable_timing=True)
+        end_event = torch.cuda.Event(enable_timing=True)
+
+        latencies = []
+        for _ in range(num_iters):
+            torch.cuda.synchronize()
+            start_event.record()
+            graph.replay()
+            end_event.record()
+            end_event.synchronize()
+            latencies.append(start_event.elapsed_time(end_event))
+
+        # Divide by 10 since graph contains 10 calls
+        return sum(latencies) / (num_iters * 10)
+
+    # Benchmark parameters
+    num_warmup = 5
+    num_iters = 100
+
+    # Benchmark only CUDA graphs (more reliable and faster)
+    # Benchmark Triton MoE with CUDA graphs
+    triton_graph_time = bench_cuda_graph(
+        triton_graph, num_warmup=num_warmup, num_iters=num_iters
+    )
+
+    # Benchmark CUTLASS MoE with CUDA graphs
+    cutlass_graph_time = bench_cuda_graph(
+        cutlass_graph, num_warmup=num_warmup, num_iters=num_iters
+    )
+
+    # Convert ms to us and return results
+    triton_time_us = triton_graph_time * 1000
+    cutlass_time_us = cutlass_graph_time * 1000
+
+    return {
+        "batch_size": m,
+        "triton_time_us": triton_time_us,
+        "cutlass_time_us": cutlass_time_us,
+    }
+
+
+def main(args):
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    all_results = []
+
+    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 args.per_act_token_opts:
+                    for per_out_ch in args.per_out_ch_opts:
+                        print(
+                            f"\n=== {model}, experts={num_experts}, topk={topk},"
+                            f"per_act={per_act_token}, per_out_ch={per_out_ch} ==="
+                        )
+
+                        config_results = []
+                        for size_m in args.batch_sizes:
+                            mkn = (size_m, size_k, size_n)
+                            result = bench_run(
+                                [],  # Not used anymore
+                                model,
+                                num_experts,
+                                topk,
+                                per_act_token,
+                                per_out_ch,
+                                mkn,
+                            )
+                            if result:
+                                config_results.append(result)
+
+                        # Print results table for this configuration
+                        if config_results:
+                            print(
+                                f"\n{'Batch Size':<12}"
+                                f"{'Triton (us)':<15}"
+                                f"{'CUTLASS (us)':<15}"
+                            )
+                            print("-" * 45)
+                            for result in config_results:
+                                print(
+                                    f"{result['batch_size']:<12}"
+                                    f"{result['triton_time_us']:<15.2f}"
+                                    f"{result['cutlass_time_us']:<15.2f}"
+                                )
+
+                            all_results.extend(config_results)
+
+    print(f"\nTotal benchmarks completed: {len(all_results)}")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="""Benchmark CUTLASS FP8 MOE vs Triton FP8 FUSED MOE
+         across specified models/shapes/batches
+
+        Example usage:
+        python benchmark_cutlass_moe_fp8.py  \
+            --model "Llama-4-Maverick-17B-128E-Instruct-FP8"  \
+            --tp-sizes 8 \
+            --batch-size 2 4 8  \
+            --per-act-token-opts false \
+            --per-out-ch-opts false
+
+        """
+    )
+    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(
+        "--per-act-token-opts",
+        nargs="+",
+        type=lambda x: x.lower() == "true",
+        default=[False, True],
+        help="Per-activation token quantization options (true/false)",
+    )
+    parser.add_argument(
+        "--per-out-ch-opts",
+        nargs="+",
+        type=lambda x: x.lower() == "true",
+        default=[False, True],
+        help="Per-output channel quantization options (true/false)",
+    )
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_device_communicators.py

@@ -7,6 +7,10 @@ Benchmark script for device communicators:
 CustomAllreduce (oneshot, twoshot), PyNcclCommunicator,
 and SymmMemCommunicator (multimem, two-shot).
 
+for NCCL symmetric memory you need to set the environment variables
+NCCL_NVLS_ENABLE=1 NCCL_CUMEM_ENABLE=1 VLLM_USE_NCCL_SYMM_MEM=1, otherwise NCCL does
+not use fast NVLS implementation for all reduce.
+
 Usage:
     torchrun --nproc_per_node=<N> benchmark_device_communicators.py [options]
 
@@ -26,7 +30,13 @@ import torch.distributed as dist
 from torch.distributed import ProcessGroup
 
 from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
-from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
+from vllm.distributed.device_communicators.pynccl import (
+    PyNcclCommunicator,
+    register_nccl_symmetric_ops,
+)
+from vllm.distributed.device_communicators.pynccl_allocator import (
+    set_graph_pool_id,
+)
 from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
 from vllm.logger import init_logger
 from vllm.utils import FlexibleArgumentParser
@@ -98,6 +108,7 @@ class CommunicatorBenchmark:
             )
             if not self.pynccl_comm.disabled:
                 logger.info("Rank %s: PyNcclCommunicator initialized", self.rank)
+                register_nccl_symmetric_ops(self.pynccl_comm)
             else:
                 logger.info("Rank %s: PyNcclCommunicator disabled", self.rank)
                 self.pynccl_comm = None
@@ -194,6 +205,15 @@ class CommunicatorBenchmark:
                     None,  # no env variable needed
                 )
             )
+            communicators.append(
+                (
+                    "pynccl-symm",
+                    lambda t: torch.ops.vllm.all_reduce_symmetric_with_copy(t),
+                    lambda t: True,  # Always available if initialized
+                    nullcontext(),
+                    None,  # no env variable needed
+                )
+            )
 
         if self.symm_mem_comm_multimem is not None:
             comm = self.symm_mem_comm_multimem
@@ -271,7 +291,9 @@ class CommunicatorBenchmark:
                 # Capture the graph using context manager
                 with context:
                     graph = torch.cuda.CUDAGraph()
-                    with torch.cuda.graph(graph):
+                    graph_pool = torch.cuda.graph_pool_handle()
+                    set_graph_pool_id(graph_pool)
+                    with torch.cuda.graph(graph, pool=graph_pool):
                         for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
                             allreduce_fn(graph_input)
 

benchmarks/kernels/benchmark_lora.py

@@ -79,9 +79,9 @@ def make_rand_lora_weight_tensor(
 
 
 def make_rand_tensors(
-    a_shape: tuple[int],
-    b_shape: tuple[int],
-    c_shape: tuple[int],
+    a_shape: tuple[int, ...],
+    b_shape: tuple[int, ...],
+    c_shape: tuple[int, ...],
     a_dtype: torch.dtype,
     b_dtype: torch.dtype,
     c_dtype: torch.dtype,
@@ -243,7 +243,7 @@ class OpType(Enum):
         lora_rank: int,
         num_loras: int,
         num_slices: int,
-    ) -> tuple[tuple[int], tuple[int], tuple[int]]:
+    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
         in A x B = C, for the op_type

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -9,6 +9,9 @@ import torch
 from tabulate import tabulate
 
 from vllm import _custom_ops as ops
+from vllm.attention.ops.triton_reshape_and_cache_flash import (
+    triton_reshape_and_cache_flash,
+)
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (
@@ -31,6 +34,8 @@ def run_benchmark(
     kv_cache_dtype: str,
     kv_cache_layout: str,
     num_iters: int,
+    implementation: str,
+    benchmark_mode: str,
     device: str = "cuda",
 ) -> float:
     """Return latency (seconds) for given num_tokens."""
@@ -38,6 +43,14 @@ def run_benchmark(
     if kv_cache_dtype == "fp8" and head_size % 16:
         raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
 
+    if implementation not in ("cuda", "triton"):
+        raise ValueError(
+            f"Unsupported implementation: {implementation}. "
+            "Only 'cuda' and 'triton' are supported."
+        )
+    if implementation == "triton" and kv_cache_layout == "HND":
+        return float("nan")  # Triton does not support HND layout yet.
+
     current_platform.seed_everything(42)
     torch.set_default_device(device)
 
@@ -65,27 +78,49 @@ def run_benchmark(
         cache_layout=kv_cache_layout,
     )
     key_cache, value_cache = key_caches[0], value_caches[0]
+    # to free unused memory
+    del key_caches, value_caches
 
     # compute per-kernel scaling factors for fp8 conversion (if used).
     k_scale = (key.amax() / 64.0).to(torch.float32)
     v_scale = (value.amax() / 64.0).to(torch.float32)
 
+    if implementation == "cuda":
+        function_under_test = lambda: ops.reshape_and_cache_flash(
+            key,  # noqa: F821
+            value,  # noqa: F821
+            key_cache,  # noqa: F821
+            value_cache,  # noqa: F821
+            slot_mapping,  # noqa: F821
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )
+    else:
+        function_under_test = lambda: triton_reshape_and_cache_flash(
+            key,  # noqa: F821
+            value,  # noqa: F821
+            key_cache,  # noqa: F821
+            value_cache,  # noqa: F821
+            slot_mapping,  # noqa: F821
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )
+    if benchmark_mode == "cudagraph":
+        g = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(g):
+            function_under_test()
+        torch.cuda.synchronize()
+        function_under_test = lambda: g.replay()
+
     def run_cuda_benchmark(n_iters: int) -> float:
         nonlocal key, value, key_cache, value_cache, slot_mapping
         torch.cuda.synchronize()
         start = time.perf_counter()
         for _ in range(n_iters):
-            ops.reshape_and_cache_flash(
-                key,
-                value,
-                key_cache,
-                value_cache,
-                slot_mapping,
-                kv_cache_dtype,
-                k_scale,
-                v_scale,
-            )
-        torch.cuda.synchronize()
+            function_under_test()
+            torch.cuda.synchronize()
         end = time.perf_counter()
         return (end - start) / n_iters
 
@@ -116,10 +151,16 @@ def main(args):
                 kv_cache_dtype=args.kv_cache_dtype,
                 kv_cache_layout=layout,
                 num_iters=args.iters,
+                implementation=args.implementation,
+                benchmark_mode=args.mode,
                 device="cuda",
             )
             rows.append([n_tok, layout, f"{lat * 1e6:.3f}"])
 
+    print(
+        f"Benchmark results for implementation {args.implementation}"
+        f" (measuring with {args.mode}):"
+    )
     print(tabulate(rows, headers=["num_tokens", "layout", "latency (µs)"]))
 
 
@@ -151,6 +192,21 @@ if __name__ == "__main__":
     )
 
     parser.add_argument("--iters", type=int, default=100)
+
+    parser.add_argument(
+        "--implementation",
+        type=str,
+        choices=["cuda", "triton"],
+        default="cuda",
+    )
+
+    parser.add_argument(
+        "--mode",
+        type=str,
+        choices=["cudagraph", "no_graph"],
+        default="cudagraph",
+    )
+
     args = parser.parse_args()
 
     main(args)

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -8,12 +8,16 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    get_col_major_tma_aligned_tensor,
     per_token_group_quant_fp8,
     w8a8_block_fp8_matmul,
 )
 from vllm.triton_utils import triton
-from vllm.utils.deep_gemm import calc_diff, fp8_gemm_nt, per_block_cast_to_fp8
+from vllm.utils.deep_gemm import (
+    calc_diff,
+    fp8_gemm_nt,
+    get_col_major_tma_aligned_tensor,
+    per_block_cast_to_fp8,
+)
 
 
 def benchmark_shape(m: int,

cmake/cpu_extension.cmake

@@ -101,6 +101,7 @@ else()
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
     find_isa(${CPUINFO} "S390" S390_FOUND)
+    find_isa(${CPUINFO} "v" RVV_FOUND) # Check for RISC-V RVV support
 endif()
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
@@ -177,8 +178,14 @@ elseif (S390_FOUND)
         "-mzvector"
         "-march=native"
         "-mtune=native")
+elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
+    if(RVV_FOUND)
+	    message(FAIL_ERROR "Can't support rvv now.")
+    else()
+        list(APPEND CXX_COMPILE_FLAGS "-march=rv64gc")
+    endif()
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA or ARMv8 support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
 endif()
 
 #
@@ -258,7 +265,8 @@ set(VLLM_EXT_SRC
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/cpu/torch_bindings.cpp")
+    "csrc/cpu/torch_bindings.cpp"
+    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC

csrc/attention/mla/cutlass_sm100_mla/device/sm100_mla.hpp

@@ -135,10 +135,10 @@ public:
     max_splits = min(16, max_splits);
 
     // TODO: This avoids a hang when the batch size larger than 1 and 
-    // there is more than 4 kv_splits. 
+    // there is more than 1 kv_splits. 
     // Discuss with NVIDIA how this can be fixed.
     if (B > 1) {
-      max_splits = min(2, max_splits);
+      max_splits = min(1, max_splits);
     }
     
     // printf("    max_splits = %d\n", max_splits);

csrc/cpu/cpu_types.hpp

@@ -14,7 +14,8 @@
   // arm implementation
   #include "cpu_types_arm.hpp"
 #else
-  #warning "unsupported vLLM cpu implementation"
+  #warning "unsupported vLLM cpu implementation, vLLM will compile with scalar"
+  #include "cpu_types_scalar.hpp"
 #endif
 
 #ifdef _OPENMP

csrc/cpu/cpu_types_scalar.hpp

@@ -0,0 +1,513 @@
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <torch/all.h>
+#include "float_convert.hpp"
+
+namespace vec_op {
+
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)            \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#ifndef CPU_OP_GUARD
+  #define CPU_KERNEL_GUARD_IN(NAME)
+  #define CPU_KERNEL_GUARD_OUT(NAME)
+#else
+  #define CPU_KERNEL_GUARD_IN(NAME) \
+    std::cout << #NAME << " invoked." << std::endl;
+  #define CPU_KERNEL_GUARD_OUT(NAME) \
+    std::cout << #NAME << " exit." << std::endl;
+#endif
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+#define __max(a, b) ((a) > (b) ? (a) : (b))
+#define __min(a, b) ((a) < (b) ? (a) : (b))
+#define __abs(a) ((a) < (0) ? (0 - a) : (a))
+
+typedef struct f16x8_t {
+  uint16_t val[8];
+} f16x8_t;
+
+typedef struct f16x16_t {
+  uint16_t val[16];
+} f16x16_t;
+
+typedef struct f16x32_t {
+  uint16_t val[32];
+} f16x32_t;
+
+typedef struct f32x4_t {
+  float val[4];
+} f32x4_t;
+
+typedef struct f32x8_t {
+  float val[8];
+} f32x8_t;
+
+typedef struct f32x16_t {
+  float val[16];
+} f32x16_t;
+
+namespace {
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+};
+};  // namespace
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T> > >
+constexpr void unroll_loop(F&& f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T>
+struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
+};
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  f16x8_t reg;
+
+  explicit FP16Vec8(const void* ptr)
+      : reg(*reinterpret_cast<const f16x8_t*>(ptr)) {};
+
+  explicit FP16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const { *reinterpret_cast<f16x8_t*>(ptr) = reg; }
+};
+
+struct FP16Vec16 : public Vec<FP16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  f16x16_t reg;
+
+  explicit FP16Vec16(const void* ptr)
+      : reg(*reinterpret_cast<const f16x16_t*>(ptr)) {};
+
+  explicit FP16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
+
+  void save(void* ptr, const int elem_num) const {
+    int num = __min(elem_num, VEC_ELEM_NUM);
+    std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
+  }
+};
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  f16x8_t reg;
+
+  explicit BF16Vec8(const void* ptr)
+      : reg(*reinterpret_cast<const f16x8_t*>(ptr)) {};
+
+  explicit BF16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const { *reinterpret_cast<f16x8_t*>(ptr) = reg; }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  f16x16_t reg;
+
+  explicit BF16Vec16(const void* ptr)
+      : reg(*reinterpret_cast<const f16x16_t*>(ptr)) {};
+
+  explicit BF16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
+
+  void save(void* ptr, const int elem_num) const {
+    int num = __min(elem_num, VEC_ELEM_NUM);
+    std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
+  }
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+  f16x32_t reg;
+
+  explicit BF16Vec32(const void* ptr)
+      : reg(*reinterpret_cast<const f16x32_t*>(ptr)) {};
+
+  explicit BF16Vec32(f16x32_t data) : reg(data) {};
+
+  explicit BF16Vec32(BF16Vec8& vec8_data) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = vec8_data.reg.val[i % BF16Vec8::VEC_ELEM_NUM];
+    }
+  }
+
+  void save(void* ptr) const { *reinterpret_cast<f16x32_t*>(ptr) = reg; }
+};
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+
+  f32x4_t reg;
+
+  explicit FP32Vec4(float v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = v;
+    }
+  }
+
+  explicit FP32Vec4() {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = 0.0f;
+    }
+  }
+
+  explicit FP32Vec4(const float* ptr)
+      : reg(*reinterpret_cast<const f32x4_t*>(ptr)) {};
+
+  explicit FP32Vec4(f32x4_t data) : reg(data) {};
+
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  f32x8_t reg;
+
+  explicit FP32Vec8(float v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = v;
+    }
+  }
+
+  explicit FP32Vec8() {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = 0.0f;
+    }
+  }
+
+  explicit FP32Vec8(const float* ptr)
+      : reg(*reinterpret_cast<const f32x8_t*>(ptr)) {};
+
+  explicit FP32Vec8(f32x8_t data) : reg(data) {};
+
+  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
+
+  explicit FP32Vec8(const FP16Vec8& v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = fp16_to_float(v.reg.val[i]);
+    }
+  }
+
+  FP32Vec8(const BF16Vec8& v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = bf16_to_float(v.reg.val[i]);
+    }
+  }
+
+  float reduce_sum() const {
+    float result = 0;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result += reg.val[i];
+    }
+    return result;
+  }
+
+  FP32Vec8 exp() const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = expf(reg.val[i]);
+    }
+    return FP32Vec8(ret);
+  }
+
+  FP32Vec8 tanh() const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = tanhf(reg.val[i]);
+    }
+    return FP32Vec8(ret);
+  }
+
+  FP32Vec8 er() const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = erf(reg.val[i]);
+    }
+    return FP32Vec8(ret);
+  }
+
+  FP32Vec8 operator*(const FP32Vec8& b) const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] * b.reg.val[i];
+    }
+    return FP32Vec8(ret);
+  }
+
+  FP32Vec8 operator+(const FP32Vec8& b) const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] + b.reg.val[i];
+    }
+    return FP32Vec8(ret);
+  }
+
+  FP32Vec8 operator-(const FP32Vec8& b) const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] - b.reg.val[i];
+    }
+    return FP32Vec8(ret);
+  }
+
+  FP32Vec8 operator/(const FP32Vec8& b) const {
+    f32x8_t ret;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] / b.reg.val[i];
+    }
+    return FP32Vec8(ret);
+  }
+
+  void save(void* ptr) const { *reinterpret_cast<f32x8_t*>(ptr) = reg; }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  f32x16_t reg;
+
+  explicit FP32Vec16(float v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = v;
+    }
+  }
+
+  explicit FP32Vec16() {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = 0.0f;
+    }
+  }
+
+  explicit FP32Vec16(const float* ptr)
+      : reg(*reinterpret_cast<const f32x16_t*>(ptr)) {};
+
+  explicit FP32Vec16(f32x16_t data) : reg(data) {};
+
+  FP32Vec16(const FP32Vec4& data) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = data.reg.val[i % FP32Vec4::VEC_ELEM_NUM];
+    }
+  }
+
+  FP32Vec16(const FP32Vec8& data) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = data.reg.val[i % FP32Vec8::VEC_ELEM_NUM];
+    }
+  }
+
+  FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
+
+  explicit FP32Vec16(const FP16Vec16& v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = fp16_to_float(v.reg.val[i]);
+    }
+  }
+
+  explicit FP32Vec16(const BF16Vec16& v) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = bf16_to_float(v.reg.val[i]);
+    }
+  }
+
+  explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
+
+  FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
+
+  FP32Vec16 operator*(const FP32Vec16& b) const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] * b.reg.val[i];
+    }
+    return result;
+  }
+
+  FP32Vec16 operator+(const FP32Vec16& b) const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] + b.reg.val[i];
+    }
+    return result;
+  }
+
+  FP32Vec16 operator-(const FP32Vec16& b) const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] - b.reg.val[i];
+    }
+    return result;
+  }
+
+  FP32Vec16 operator/(const FP32Vec16& b) const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] / b.reg.val[i];
+    }
+    return result;
+  }
+
+  FP32Vec16 max(const FP32Vec16& b) const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = __max(reg.val[i], b.reg.val[i]);
+    }
+    return result;
+  }
+
+  FP32Vec16 min(const FP32Vec16& b) const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = __min(reg.val[i], b.reg.val[i]);
+    }
+    return result;
+  }
+
+  FP32Vec16 abs() const {
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = __abs(reg.val[i]);
+    }
+    return result;
+  }
+
+  float reduce_sum() const {
+    float result = 0.0f;
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result += reg.val[i];
+    }
+    return result;
+  }
+
+  float reduce_max() const {
+    float result = reg.val[0];
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result = __max(reg.val[i], result);
+    }
+    return result;
+  }
+
+  float reduce_min() const {
+    float result = reg.val[0];
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result = __min(reg.val[i], result);
+    }
+    return result;
+  }
+
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+    float sum = 0.0;
+    int start = idx * group_size;
+    int end = (idx + 1) * group_size;
+
+    for (; (start < VEC_ELEM_NUM) && (start < end); ++start) {
+      sum += reg.val[start];
+    }
+
+    return sum;
+  }
+
+  void save(void* ptr) const { *reinterpret_cast<f32x16_t*>(ptr) = reg; }
+};
+
+template <typename T>
+struct VecType {
+  using vec_type = void;
+};
+
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
+
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+};
+
+template <>
+struct VecType<c10::Half> {
+  using vec_type = FP16Vec8;
+};
+
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+};
+
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
+}
+
+/*
+template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
+  c10::Half __attribute__((__may_alias__)) *v_ptr =
+      reinterpret_cast<c10::Half *>(&v);
+  *ptr = *(v_ptr + 1);
+}
+*/
+
+template <>
+inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
+  uint16_t fp16 = float_to_fp16(v);
+  *reinterpret_cast<uint16_t*>(ptr) = fp16;
+}
+
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  c10::BFloat16 __attribute__((__may_alias__))* v_ptr =
+      reinterpret_cast<c10::BFloat16*>(&v);
+  *ptr = *(v_ptr + 1);
+}
+
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
+  int i = 0;
+  for (i = 0; i < FP16Vec16::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_fp16(v.reg.val[i]);
+  }
+}
+
+inline FP16Vec8 ::FP16Vec8(const FP32Vec8& v) {
+  int i = 0;
+  for (i = 0; i < FP16Vec8::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_fp16(v.reg.val[i]);
+  }
+}
+
+inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
+  acc = acc + a * b;
+}
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
+  int i = 0;
+  for (i = 0; i < BF16Vec8::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_bf16(v.reg.val[i]);
+  }
+}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
+  int i = 0;
+  for (i = 0; i < BF16Vec16::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_bf16(v.reg.val[i]);
+  }
+}
+
+inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 3); }
+
+};  // namespace vec_op

csrc/cpu/float_convert.hpp

@@ -0,0 +1,106 @@
+
+static float bf16_to_float(uint16_t bf16) {
+  uint32_t bits = static_cast<uint32_t>(bf16) << 16;
+  float fp32;
+  std::memcpy(&fp32, &bits, sizeof(fp32));
+  return fp32;
+}
+
+static uint16_t float_to_bf16(float fp32) {
+  uint32_t bits;
+  std::memcpy(&bits, &fp32, sizeof(fp32));
+  return static_cast<uint16_t>(bits >> 16);
+}
+
+/************************************************
+ * Copyright (c) 2015 Princeton Vision Group
+ * Licensed under the MIT license.
+ * Codes below copied from
+ * https://github.com/PrincetonVision/marvin/tree/master/tools/tensorIO_matlab
+ *************************************************/
+static uint16_t float_to_fp16(float fp32) {
+  uint16_t fp16;
+
+  unsigned x;
+  unsigned u, remainder, shift, lsb, lsb_s1, lsb_m1;
+  unsigned sign, exponent, mantissa;
+
+  std::memcpy(&x, &fp32, sizeof(fp32));
+  u = (x & 0x7fffffff);
+
+  // Get rid of +NaN/-NaN case first.
+  if (u > 0x7f800000) {
+    fp16 = 0x7fffU;
+    return fp16;
+  }
+
+  sign = ((x >> 16) & 0x8000);
+
+  // Get rid of +Inf/-Inf, +0/-0.
+  if (u > 0x477fefff) {
+    fp16 = sign | 0x7c00U;
+    return fp16;
+  }
+  if (u < 0x33000001) {
+    fp16 = (sign | 0x0000);
+    return fp16;
+  }
+
+  exponent = ((u >> 23) & 0xff);
+  mantissa = (u & 0x7fffff);
+
+  if (exponent > 0x70) {
+    shift = 13;
+    exponent -= 0x70;
+  } else {
+    shift = 0x7e - exponent;
+    exponent = 0;
+    mantissa |= 0x800000;
+  }
+  lsb = (1 << shift);
+  lsb_s1 = (lsb >> 1);
+  lsb_m1 = (lsb - 1);
+
+  // Round to nearest even.
+  remainder = (mantissa & lsb_m1);
+  mantissa >>= shift;
+  if (remainder > lsb_s1 || (remainder == lsb_s1 && (mantissa & 0x1))) {
+    ++mantissa;
+    if (!(mantissa & 0x3ff)) {
+      ++exponent;
+      mantissa = 0;
+    }
+  }
+
+  fp16 = (sign | (exponent << 10) | mantissa);
+
+  return fp16;
+}
+
+static float fp16_to_float(uint16_t fp16) {
+  unsigned sign = ((fp16 >> 15) & 1);
+  unsigned exponent = ((fp16 >> 10) & 0x1f);
+  unsigned mantissa = ((fp16 & 0x3ff) << 13);
+  int temp;
+  float fp32;
+  if (exponent == 0x1f) { /* NaN or Inf */
+    mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);
+    exponent = 0xff;
+  } else if (!exponent) { /* Denorm or Zero */
+    if (mantissa) {
+      unsigned int msb;
+      exponent = 0x71;
+      do {
+        msb = (mantissa & 0x400000);
+        mantissa <<= 1; /* normalize */
+        --exponent;
+      } while (!msb);
+      mantissa &= 0x7fffff; /* 1.mantissa is implicit */
+    }
+  } else {
+    exponent += 0x70;
+  }
+  temp = ((sign << 31) | (exponent << 23) | mantissa);
+  std::memcpy(&fp32, &temp, sizeof(temp));
+  return fp32;
+}

csrc/cpu/torch_bindings.cpp

@@ -88,8 +88,18 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    int tp_rank, int blocksparse_local_blocks,"
       "    int blocksparse_vert_stride, int blocksparse_block_size,"
       "    int blocksparse_head_sliding_step) -> ()");
+
   ops.impl("paged_attention_v1", torch::kCPU, &paged_attention_v1);
 
+  ops.def(
+      "dynamic_4bit_int_moe("
+      "Tensor x, Tensor topk_ids, Tensor topk_weights,"
+      "Tensor w13_packed, Tensor w2_packed, int H, int I, int I2,"
+      "int group_size, bool apply_router_weight_on_input, int activation_kind"
+      ") -> Tensor");
+
+  ops.impl("dynamic_4bit_int_moe", torch::kCPU, &dynamic_4bit_int_moe_cpu);
+
   // PagedAttention V2.
   ops.def(
       "paged_attention_v2("

csrc/moe/dynamic_4bit_int_moe_cpu.cpp

@@ -0,0 +1,156 @@
+#include <ATen/ATen.h>
+#include <ATen/Parallel.h>
+#include <torch/all.h>
+
+// _dyn_quant_matmul_4bit is only available on AArch64.
+#if defined(__aarch64__)
+  #include <ATen/ops/_dyn_quant_matmul_4bit.h>
+#endif
+
+inline torch::Tensor mm(const torch::Tensor& a, const torch::Tensor& packed_w,
+                        int64_t group_size_eff, int64_t in_features,
+                        int64_t out_features) {
+#if defined(__aarch64__)
+  return at::_ops::_dyn_quant_matmul_4bit::call(a, packed_w, group_size_eff,
+                                                in_features, out_features);
+#else
+  TORCH_CHECK(false,
+              "dynamic 4-bit int MoE path requires AArch64 (ARM64); "
+              "_dyn_quant_matmul_4bit is unavailable on this architecture");
+  return {};
+#endif
+}
+
+enum ActivationKind : int64_t {
+  SwiGLU_Gu = 0,  // act = SiLU(g) * u
+  SwiGLUOAI = 1,  // act = SiLU(u) * g
+  SiLU = 2        // SiLU
+};
+
+torch::Tensor dynamic_4bit_int_moe_cpu(
+    torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
+    torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,
+    int64_t I2, int64_t group_size, bool apply_router_weight_on_input,
+    int64_t activation_kind) {
+  TORCH_CHECK(x.dim() == 2, "x must be 2D");
+  TORCH_CHECK(topk_ids.dim() == 2 && topk_weights.dim() == 2,
+              "topk tensors must be [T, K]");
+  TORCH_CHECK(
+      w13_packed.size(0) == w2_packed.size(0),
+      "w13_packed and w2_packed must have same number of experts in dim 0");
+  TORCH_CHECK(I2 == 2 * I, "I2 must equal 2*I");
+
+  const int64_t T = x.size(0);
+  const int64_t K = topk_ids.size(1);
+  const int64_t E = w13_packed.size(0);
+  const int64_t N = T * K;
+
+  auto x_c = x.contiguous();
+  auto ids_c = topk_ids.contiguous();
+  auto gates_c = topk_weights.to(at::kFloat).contiguous();
+
+  // bucketing tokens -> experts
+  c10::SmallVector<int64_t, 64> counts(
+      E, 0);  // Small vector uses stack allocation
+  {
+    const auto* ids_ptr = ids_c.data_ptr<int64_t>();
+    for (int64_t i = 0; i < N; ++i) {
+      const int64_t e_id = ids_ptr[i];
+      TORCH_CHECK(0 <= e_id && e_id < E, "expert id out of range");
+      counts[e_id]++;
+    }
+  }
+  c10::SmallVector<int64_t, 65> offsets(E + 1, 0);  // ( E +1 )
+  for (int64_t e = 0; e < E; ++e) offsets[e + 1] = offsets[e] + counts[e];
+
+  auto expert_tokens = at::empty({offsets[E]}, ids_c.options());
+  auto expert_gates = at::empty({offsets[E]}, gates_c.options());
+  {
+    c10::SmallVector<int64_t, 64> cursor(E, 0);
+    const auto* ids_ptr = ids_c.data_ptr<int64_t>();
+    const auto* gts_ptr = gates_c.data_ptr<float>();
+    auto* tok_ptr = expert_tokens.data_ptr<int64_t>();
+    auto* gate_ptr = expert_gates.data_ptr<float>();
+
+    for (int64_t t = 0; t < T; ++t) {
+      const int64_t base = t * K;
+      for (int64_t k = 0; k < K; ++k) {
+        const int64_t idx = base + k;
+        const int64_t e = ids_ptr[idx];
+        const int64_t p = offsets[e] + (cursor[e]++);
+        tok_ptr[p] = t;
+        gate_ptr[p] = gts_ptr[idx];
+      }
+    }
+  }
+
+  const int64_t g_eff_13 = (group_size != -1) ? group_size : H;
+  const int64_t g_eff_2 = (group_size != -1) ? group_size : I;
+
+  // Per-expert outputs filled in parallel
+  std::vector<torch::Tensor> y_list(E);
+  y_list.resize(E);
+
+  at::parallel_for(0, E, 1, [&](int64_t e_begin, int64_t e_end) {
+    for (int64_t e = e_begin; e < e_end; ++e) {
+      const int64_t te = counts[e];
+      if (te == 0) {
+        y_list[e] = at::empty({0, H}, x_c.options());
+        continue;
+      }
+
+      const int64_t start = offsets[e];
+
+      auto sel_tokens =
+          expert_tokens.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
+      auto gates_e =
+          expert_gates.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
+
+      auto x_e = x_c.index_select(/*dim=*/0, sel_tokens);
+
+      if (apply_router_weight_on_input) {
+        x_e = x_e.mul(gates_e.unsqueeze(1));
+      }
+
+      auto w13_e = w13_packed.select(/*dim=*/0, e);
+      auto w2_e = w2_packed.select(/*dim=*/0, e);
+
+      // W13
+      auto y13 =
+          mm(x_e, w13_e, g_eff_13, /*in_features=*/H, /*out_features=*/I2);
+
+      auto g_part = y13.narrow(/*dim=*/1, /*start=*/0, /*length=*/I);
+      auto u_part = y13.narrow(/*dim=*/1, /*start=*/I, /*length=*/I);
+
+      torch::Tensor act;
+      if (activation_kind == ActivationKind::SwiGLUOAI) {  // SwiGLUOAI
+        constexpr double kAlpha = 1.702;                   // GPT-OSS default
+        constexpr double kLimit = 7.0;                     // GPT-OSS default
+        auto gate_c = at::clamp_max(g_part, kLimit);
+        auto up_c = at::clamp(u_part, -kLimit, kLimit);
+        auto glu = gate_c.mul(at::sigmoid(gate_c.mul(kAlpha)));
+        act = up_c.add(1.0).mul(glu);
+      } else {  // SiLU , SwiGLU_GU, vLLM maps silu to SiluAndMul()
+        act = at::silu(g_part).mul(u_part);
+      }
+
+      // W2
+      auto y = mm(act, w2_e, g_eff_2, /*in_features=*/I, /*out_features=*/H);
+
+      if (!apply_router_weight_on_input) {
+        y = y.mul(gates_e.unsqueeze(1));
+      }
+
+      // Store per-expert result
+      y_list[e] = y;
+    }
+  });
+
+  // Concatenate all expert outputs to match expert_tokens order
+  auto Y_all = at::cat(y_list, /*dim=*/0);
+  auto out = at::zeros({T, H}, x.options());
+  out =
+      at::index_add(out, /*dim=*/0, /*index=*/expert_tokens, /*source=*/Y_all);
+
+  return out;
+}

csrc/moe/grouped_topk_kernels.cu

@@ -418,6 +418,15 @@ __device__ inline T neg_inf() {
   return cuda_cast<T, float>(-cuda::std::numeric_limits<float>::infinity());
 }
 
+template <typename T>
+__device__ inline bool is_finite(const T val) {
+#if (__CUDACC_VER_MAJOR__ * 10000 + __CUDACC_VER_MINOR__ * 100 >= 120800)
+  return cuda::std::isfinite(val);
+#else
+  return isfinite(cuda_cast<float, T>(val));
+#endif
+}
+
 template <typename T>
 __device__ void topk_with_k2(T* output, T const* input,
                              cg::thread_block_tile<32> const& tile,
@@ -533,7 +542,7 @@ __global__ void group_idx_and_topk_idx_kernel(
     // calculate group_idx
     int32_t target_num_min = WARP_SIZE - n_group + topk_group;
     // The check is necessary to avoid abnormal input
-    if (lane_id < n_group && cuda::std::isfinite(group_scores[lane_id])) {
+    if (lane_id < n_group && is_finite(group_scores[lane_id])) {
       value = group_scores[lane_id];
     }
 
@@ -568,11 +577,10 @@ __global__ void group_idx_and_topk_idx_kernel(
         int32_t offset = i_group * num_experts_per_group;
         for (int32_t i = lane_id; i < align_num_experts_per_group;
              i += WARP_SIZE) {
-          T candidates =
-              (i < num_experts_per_group) &&
-                      cuda::std::isfinite(scores_with_bias[offset + i])
-                  ? scores_with_bias[offset + i]
-                  : neg_inf<T>();
+          T candidates = (i < num_experts_per_group) &&
+                                 is_finite(scores_with_bias[offset + i])
+                             ? scores_with_bias[offset + i]
+                             : neg_inf<T>();
           queue.add(candidates, offset + i);
         }
         if (group_scores[i_group] == topk_group_value) {

csrc/moe/moe_align_sum_kernels.cu

@@ -44,6 +44,9 @@ __global__ void moe_align_block_size_kernel(
 
   for (size_t i = tid; i < numel; i += stride) {
     int expert_id = topk_ids[i];
+    if (expert_id >= num_experts) {
+      continue;
+    }
     int warp_idx = expert_id / experts_per_warp;
     int expert_offset = expert_id % experts_per_warp;
     atomicAdd(&shared_counts[warp_idx * experts_per_warp + expert_offset], 1);
@@ -95,12 +98,15 @@ template <typename scalar_t>
 __global__ void count_and_sort_expert_tokens_kernel(
     const scalar_t* __restrict__ topk_ids,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ cumsum_buffer,
-    size_t numel) {
+    size_t numel, int32_t num_experts) {
   const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
   const size_t stride = blockDim.x * gridDim.x;
 
   for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
+    if (expert_id >= num_experts) {
+      continue;
+    }
     int32_t rank_post_pad = atomicAdd(&cumsum_buffer[expert_id], 1);
     sorted_token_ids[rank_post_pad] = i;
   }
@@ -269,7 +275,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
           sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
               topk_ids.data_ptr<scalar_t>(),
               sorted_token_ids.data_ptr<int32_t>(),
-              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
+              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel(), num_experts);
         }
       });
 }

csrc/ops.h

@@ -328,6 +328,12 @@ void selective_scan_fwd(const torch::Tensor& u, const torch::Tensor& delta,
                         const std::optional<torch::Tensor>& has_initial_state,
                         const torch::Tensor& ssm_states, int64_t pad_slot_id);
 
+torch::Tensor dynamic_4bit_int_moe_cpu(
+    torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
+    torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,
+    int64_t I2, int64_t group_size, bool apply_router_weight_on_input,
+    int64_t activation_kind);
+
 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,

csrc/quantization/activation_kernels.cu

@@ -23,9 +23,14 @@
 typedef __hip_bfloat162 __nv_bfloat162;
 typedef __hip_bfloat16 __nv_bfloat16;
 typedef __hip_bfloat16_raw __nv_bfloat16_raw;
-
+  #if defined(HIP_FP8_TYPE_OCP)
 typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
 typedef __hip_fp8x4_e4m3 __nv_fp8x4_e4m3;
+  #else
+// ROCm 6.2 fallback: only *_fnuz types exist
+typedef __hip_fp8_e4m3_fnuz __nv_fp8_e4m3;
+typedef __hip_fp8x4_e4m3_fnuz __nv_fp8x4_e4m3;
+  #endif
 #endif
 
 #include "core/registration.h"

csrc/quantization/fp8/per_token_group_quant.cu

@@ -12,8 +12,8 @@
 #include "../vectorization_utils.cuh"
 #include "../../dispatch_utils.h"
 
-__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
-  unsigned mask = 0xffff;
+__device__ __forceinline__ float GroupReduceMax(float val) {
+  unsigned mask = threadIdx.x % 32 >= 16 ? 0xffff0000 : 0x0000ffff;
 
   val = fmaxf(val, __shfl_xor_sync(mask, val, 8));
   val = fmaxf(val, __shfl_xor_sync(mask, val, 4));
@@ -86,7 +86,7 @@ __global__ void per_token_group_quant_8bit_kernel(
       threads_per_group,  // stride in group
       scalar_op_cache);   // scalar handler
 
-  local_absmax = GroupReduceMax(local_absmax, lane_id);
+  local_absmax = GroupReduceMax(local_absmax);
 
   float y_s = local_absmax / max_8bit;
   if constexpr (SCALE_UE8M0) {

csrc/rocm/attention.cu

@@ -25,6 +25,12 @@
 #include "../attention/dtype_fp8.cuh"
 #include "../quantization/fp8/amd/quant_utils.cuh"
 
+// ROCm 6.2 compatibility: map OCP fp8 types to FNUZ variants if OCP is absent
+#if !defined(HIP_FP8_TYPE_OCP)
+using __hip_fp8_e4m3 = __hip_fp8_e4m3_fnuz;
+using __hip_fp8_e5m2 = __hip_fp8_e5m2_fnuz;
+#endif
+
 #if defined(__HIPCC__) && \
     (defined(__gfx90a__) || defined(__gfx942__) || defined(__gfx950__))
   #define __HIP__GFX9__

csrc/rocm/ops.h

@@ -5,11 +5,14 @@
 torch::Tensor LLMM1(at::Tensor& in_a, at::Tensor& in_b,
                     const int64_t rows_per_block);
 
-torch::Tensor wvSplitK(at::Tensor& in_a, at::Tensor& in_b,
+torch::Tensor wvSplitK(const at::Tensor& in_a, const at::Tensor& in_b,
+                       const std::optional<at::Tensor>& in_bias,
                        const int64_t CuCount);
 
-void wvSplitKQ(at::Tensor& in_a, at::Tensor& in_b, at::Tensor& out_c,
-               at::Tensor& scale_a, at::Tensor& scale_b, const int64_t CuCount);
+void wvSplitKQ(const at::Tensor& in_a, const at::Tensor& in_b,
+               const std::optional<at::Tensor>& in_bias, at::Tensor& out_c,
+               const at::Tensor& scale_a, const at::Tensor& scale_b,
+               const int64_t CuCount);
 
 void paged_attention(
     torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,

csrc/rocm/skinny_gemms.cu

@@ -292,8 +292,9 @@ torch::Tensor LLMM1(at::Tensor& in_a, at::Tensor& in_b,
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitK_hf_sml_(const int K, const int M, const scalar_t* B,
-                     const scalar_t* __restrict__ A, scalar_t* C,
+    wvSplitK_hf_sml_(const int K, const int M, const int Bx, const int By,
+                     const scalar_t* B, const scalar_t* __restrict__ A,
+                     const scalar_t* __restrict__ BIAS, scalar_t* C,
                      const int _WvPrGrp, const int CuCount) {
   constexpr int max_lds_len = LDS_SIZE / 2;
   #if defined(__HIP__MI3XX__)
@@ -484,7 +485,14 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (threadIdx.x == 63) {
         for (int n = 0; n < N; n++) {
           for (int i = 0; i < YTILE; i++) {
-            // if (commitColumn[i]) C[m + i + n * M] = __float2half(sum[n][i]);
+            if constexpr (std::is_same_v<scalar_t, half>) {
+              if (BIAS)
+                sum[n][i] += __half2float(BIAS[(m + i) % Bx + (n % By) * M]);
+            } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
+              if (BIAS)
+                sum[n][i] +=
+                    __bfloat162float(BIAS[(m + i) % Bx + (n % By) * M]);
+            }
             C[m + i + n * M] = __float2s<scalar_t>(sum[n][i]);
           }
         }
@@ -529,7 +537,9 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (threadIdx.x == 63) {
         for (int n = 0; n < N; n++) {
           for (int i = 0; i < YTILE; i++) {
-            // if (commitColumn[i]) C[n + i + m * N] = __float2half(sum[n][i]);
+            if (BIAS)
+              sum4[n][i][0] +=
+                  __bfloat162float(BIAS[(m + i) % Bx + (n % By) * M]);
             C[m + i + n * M] = __float2bfloat16(sum4[n][i][0]);
           }
         }
@@ -541,8 +551,10 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 #else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
-__global__ void wvSplitK_hf_sml_(const int K, const int M, const scalar_t* B,
-                                 const scalar_t* __restrict__ A, scalar_t* C,
+__global__ void wvSplitK_hf_sml_(const int K, const int M, const int Bx,
+                                 const int By, const scalar_t* B,
+                                 const scalar_t* __restrict__ A,
+                                 const scalar_t* __restrict__ BIAS, scalar_t* C,
                                  const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
@@ -553,8 +565,9 @@ __global__ void wvSplitK_hf_sml_(const int K, const int M, const scalar_t* B,
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitK_hf_(const int K, const int M, const scalar_t* B,
-                 const scalar_t* __restrict__ A, scalar_t* C,
+    wvSplitK_hf_(const int K, const int M, const int Bx, const int By,
+                 const scalar_t* B, const scalar_t* __restrict__ A,
+                 const scalar_t* __restrict__ BIAS, scalar_t* C,
                  const int _WvPrGrp, const int CuCount) {
   constexpr int max_lds_len = LDS_SIZE / 2;
   #if defined(__HIP__MI3XX__)
@@ -772,8 +785,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (threadIdx.x == 63) {
         for (int n = 0; n < N; n++) {
           for (int i = 0; i < YTILE; i++) {
-            if (commitColumn[i])
+            if (commitColumn[i]) {
+              if constexpr (std::is_same_v<scalar_t, half>) {
+                if (BIAS)
+                  sum[n][i] += __half2float(BIAS[(m + i) % Bx + (n % By) * M]);
+              } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
+                if (BIAS)
+                  sum[n][i] +=
+                      __bfloat162float(BIAS[(m + i) % Bx + (n % By) * M]);
+              }
               C[m + i + n * M] = __float2s<scalar_t>(sum[n][i]);
+            }
           }
         }
       }
@@ -818,8 +840,12 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (threadIdx.x == 63) {
         for (int n = 0; n < N; n++) {
           for (int i = 0; i < YTILE; i++) {
-            // if (commitColumn[i]) C[n + i + m * N] = __float2half(sum[n][i]);
-            C[m + i + n * M] = __float2bfloat16(sum4[n][i][0]);
+            if (commitColumn[i]) {
+              if (BIAS)
+                sum4[n][i][0] +=
+                    __bfloat162float(BIAS[(m + i) % Bx + (n % By) * M]);
+              C[m + i + n * M] = __float2bfloat16(sum4[n][i][0]);
+            }
           }
         }
       }
@@ -842,8 +868,10 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 #else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
-__global__ void wvSplitK_hf_(const int K, const int M, const scalar_t* B,
-                             const scalar_t* __restrict__ A, scalar_t* C,
+__global__ void wvSplitK_hf_(const int K, const int M, const int Bx,
+                             const int By, const scalar_t* B,
+                             const scalar_t* __restrict__ A,
+                             const scalar_t* __restrict__ BIAS, scalar_t* C,
                              const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
@@ -854,8 +882,9 @@ __global__ void wvSplitK_hf_(const int K, const int M, const scalar_t* B,
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitK_hf_big_(const int K, const int M, const scalar_t* B,
-                     const scalar_t* __restrict__ A, scalar_t* C,
+    wvSplitK_hf_big_(const int K, const int M, const int Bx, const int By,
+                     const scalar_t* B, const scalar_t* __restrict__ A,
+                     const scalar_t* __restrict__ BIAS, scalar_t* C,
                      const int _WvPrGrp, const int CuCount) {
   constexpr int max_lds_len = LDS_SIZE / 2;
   #if defined(__HIP__MI3XX__)
@@ -1124,8 +1153,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (threadIdx.x == 63) {
         for (int n = 0; n < N; n++) {
           for (int i = 0; i < YTILE; i++) {
-            if (commitColumn[i])
+            if (commitColumn[i]) {
+              if constexpr (std::is_same_v<scalar_t, half>) {
+                if (BIAS)
+                  sum[n][i] += __half2float(BIAS[(m + i) % Bx + (n % By) * M]);
+              } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
+                if (BIAS)
+                  sum[n][i] +=
+                      __bfloat162float(BIAS[(m + i) % Bx + (n % By) * M]);
+              }
               C[m + i + n * M] = __float2s<scalar_t>(sum[n][i]);
+            }
           }
         }
       }
@@ -1166,8 +1204,12 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (threadIdx.x == 63) {
         for (int n = 0; n < N; n++) {
           for (int i = 0; i < YTILE; i++) {
-            // if (commitColumn[i]) C[n + i + m * N] = __float2half(sum[n][i]);
-            C[m + i + n * M] = __float2bfloat16(sum4[n][i][0]);
+            if (commitColumn[i]) {
+              if (BIAS)
+                sum4[n][i][0] +=
+                    __bfloat162float(BIAS[(m + i) % Bx + (n % By) * M]);
+              C[m + i + n * M] = __float2bfloat16(sum4[n][i][0]);
+            }
           }
         }
       }
@@ -1190,8 +1232,10 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 #else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
-__global__ void wvSplitK_hf_big_(const int K, const int M, const scalar_t* B,
-                                 const scalar_t* __restrict__ A, scalar_t* C,
+__global__ void wvSplitK_hf_big_(const int K, const int M, const int Bx,
+                                 const int By, const scalar_t* B,
+                                 const scalar_t* __restrict__ A,
+                                 const scalar_t* __restrict__ BIAS, scalar_t* C,
                                  const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
@@ -1226,11 +1270,20 @@ int mindiv(int N, int div1, int div2) {
   return rtn;
 }
 
-torch::Tensor wvSplitK(at::Tensor& in_a, at::Tensor& in_b,
+torch::Tensor wvSplitK(const at::Tensor& in_a, const at::Tensor& in_b,
+                       const std::optional<at::Tensor>& in_bias,
                        const int64_t CuCount) {
   auto M_in = in_a.size(0);
   auto K_in = in_a.size(1);
   auto N_in = in_b.size(0);
+  auto Bx_in =
+      (in_bias.has_value() && in_bias->numel() > 0)
+          ? (in_bias->sizes().size() == 2) ? in_bias->size(1) : in_bias->size(0)
+          : 1;
+  auto By_in = (in_bias.has_value() && in_bias->numel() > 0 &&
+                in_bias->sizes().size() == 2)
+                   ? in_bias->size(0)
+                   : 1;
 
   TORCH_CHECK(in_a.dtype() == in_b.dtype());
   TORCH_CHECK(K_in % 8 == 0, "k % 8 == 0");
@@ -1254,18 +1307,18 @@ torch::Tensor wvSplitK(at::Tensor& in_a, at::Tensor& in_b,
     if ((K_in * N_in <= max_lds_len) && (M_in % _YTILEs == 0)) {              \
       int __wvPrGrp = mindiv(M_in, CuCount * _YTILEs, _WvPrGrp);              \
       wvSplitK_hf_sml_<fptype, 64, _YTILEs, _WvPrGrp, 8, _UNRLs, _N>          \
-          <<<grid, block, 0, stream>>>(K_in, M_in, af4, bf4, c, __wvPrGrp,    \
-                                       CuCount);                              \
+          <<<grid, block, 0, stream>>>(K_in, M_in, Bx_in, By_in, af4, bf4,    \
+                                       biasf4, c, __wvPrGrp, CuCount);        \
     } else if (K_in * N_in <= max_lds_len * 1.2) {                            \
       int __wvPrGrp = mindiv(M_in, CuCount * _YTILEm, _WvPrGrp);              \
       wvSplitK_hf_<fptype, 64, _YTILEm, _WvPrGrp, 8, _UNRLm, _N>              \
-          <<<grid, block, 0, stream>>>(K_in, M_in, af4, bf4, c, __wvPrGrp,    \
-                                       CuCount);                              \
+          <<<grid, block, 0, stream>>>(K_in, M_in, Bx_in, By_in, af4, bf4,    \
+                                       biasf4, c, __wvPrGrp, CuCount);        \
     } else {                                                                  \
       int __wvPrGrp = mindiv(M_in, CuCount * _YTILEb, _WvPrGrp);              \
       wvSplitK_hf_big_<fptype, 64, _YTILEb, _WvPrGrp, 8, _UNRLb, _N>          \
-          <<<grid, block, 0, stream>>>(K_in, M_in, af4, bf4, c, __wvPrGrp,    \
-                                       CuCount);                              \
+          <<<grid, block, 0, stream>>>(K_in, M_in, Bx_in, By_in, af4, bf4,    \
+                                       biasf4, c, __wvPrGrp, CuCount);        \
     }                                                                         \
   }
 
@@ -1273,6 +1326,10 @@ torch::Tensor wvSplitK(at::Tensor& in_a, at::Tensor& in_b,
     using fptype = typename scalar<scalar_t>::type;
     fptype* af4 = reinterpret_cast<fptype*>(in_a.data_ptr());
     const fptype* bf4 = reinterpret_cast<const fptype*>(in_b.data_ptr());
+    const fptype* biasf4 =
+        (in_bias.has_value() && in_bias->numel() > 0)
+            ? reinterpret_cast<const fptype*>(in_bias->data_ptr())
+            : nullptr;
     fptype* c = reinterpret_cast<fptype*>(out_c.data_ptr());
     switch (N_in) {
       case 1:
@@ -1300,8 +1357,9 @@ torch::Tensor wvSplitK(at::Tensor& in_a, at::Tensor& in_b,
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitKQ_hf_sml_(const int K, const int Kp, const int M, const fp8_t* B,
-                      const fp8_t* __restrict__ A, scalar_t* C,
+    wvSplitKQ_hf_sml_(const int K, const int Kp, const int M, const int Bx,
+                      const int By, const fp8_t* B, const fp8_t* __restrict__ A,
+                      const scalar_t* __restrict__ BIAS, scalar_t* C,
                       const float* __restrict__ s_A,
                       const float* __restrict__ s_B, const int _WvPrGrp,
                       const int CuCount) {
@@ -1453,7 +1511,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     if (threadIdx.x == 0) {
       for (int n = 0; n < N; n++) {
         for (int y = 0; y < YTILE; y++) {
-          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0] * sA * sB);
+          if (y + m >= M) break;  // To avoid mem access fault.
+          sum[n][y][0] *= sA * sB;
+          if constexpr (std::is_same_v<scalar_t, half>) {
+            if (BIAS)
+              sum[n][y][0] += __half2float(BIAS[(m + y) % Bx + (n % By) * M]);
+          } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
+            if (BIAS)
+              sum[n][y][0] +=
+                  __bfloat162float(BIAS[(m + y) % Bx + (n % By) * M]);
+          }
+          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);  // * sA * sB);
         }
       }
     }
@@ -1465,7 +1533,9 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void wvSplitKQ_hf_sml_(const int K, const int Kp, const int M,
-                                  const fp8_t* B, const fp8_t* __restrict__ A,
+                                  const int Bx, const int By, const fp8_t* B,
+                                  const fp8_t* __restrict__ A,
+                                  const scalar_t* __restrict__ BIAS,
                                   scalar_t* C, const float* __restrict__ s_A,
                                   const float* __restrict__ s_B,
                                   const int _WvPrGrp, const int CuCount) {
@@ -1477,8 +1547,9 @@ __global__ void wvSplitKQ_hf_sml_(const int K, const int Kp, const int M,
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitKQ_hf_(const int K, const int Kp, const int M, const fp8_t* B,
-                  const fp8_t* __restrict__ A, scalar_t* C,
+    wvSplitKQ_hf_(const int K, const int Kp, const int M, const int Bx,
+                  const int By, const fp8_t* B, const fp8_t* __restrict__ A,
+                  const scalar_t* __restrict__ BIAS, scalar_t* C,
                   const float* __restrict__ s_A, const float* __restrict__ s_B,
                   const int _WvPrGrp, const int CuCount) {
   constexpr int max_lds_len = LDS_SIZE;
@@ -1626,7 +1697,16 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       for (int n = 0; n < N; n++) {
         for (int y = 0; y < YTILE; y++) {
           if (y + m >= M) break;  // To avoid mem access fault.
-          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0] * sA * sB);
+          sum[n][y][0] *= sA * sB;
+          if constexpr (std::is_same_v<scalar_t, half>) {
+            if (BIAS)
+              sum[n][y][0] += __half2float(BIAS[(m + y) % Bx + (n % By) * M]);
+          } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
+            if (BIAS)
+              sum[n][y][0] +=
+                  __bfloat162float(BIAS[(m + y) % Bx + (n % By) * M]);
+          }
+          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);
         }
       }
     }
@@ -1638,16 +1718,19 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void wvSplitKQ_hf_(const int K, const int Kp, const int M,
-                              const fp8_t* B, const fp8_t* __restrict__ A,
-                              scalar_t* C, const float* __restrict__ s_A,
+                              const int Bx, const int By, const fp8_t* B,
+                              const fp8_t* __restrict__ A,
+                              const scalar_t* __restrict__ BIAS, scalar_t* C,
+                              const float* __restrict__ s_A,
                               const float* __restrict__ s_B, const int _WvPrGrp,
                               const int CuCount) {
   UNREACHABLE_CODE
 }
 #endif  // defined(__HIP__MI3XX__) TODO: Add NAVI support
 
-void wvSplitKQ(at::Tensor& in_a, at::Tensor& in_b, at::Tensor& out_c,
-               at::Tensor& scale_a, at::Tensor& scale_b,
+void wvSplitKQ(const at::Tensor& in_a, const at::Tensor& in_b,
+               const std::optional<at::Tensor>& in_bias, at::Tensor& out_c,
+               const at::Tensor& scale_a, const at::Tensor& scale_b,
                const int64_t CuCount) {
   static c10::ScalarType kFp8Type = is_fp8_ocp()
                                         ? c10::ScalarType::Float8_e4m3fn
@@ -1656,6 +1739,15 @@ void wvSplitKQ(at::Tensor& in_a, at::Tensor& in_b, at::Tensor& out_c,
   auto K_in = in_a.size(1);
   auto N_in = in_b.size(0);
   auto Kp_in = in_a.stride(0);
+  auto Bx_in =
+      (in_bias.has_value() && in_bias->numel() > 0)
+          ? (in_bias->sizes().size() == 2) ? in_bias->size(1) : in_bias->size(0)
+          : 1;
+  auto By_in = (in_bias.has_value() && in_bias->numel() > 0 &&
+                in_bias->sizes().size() == 2)
+                   ? in_bias->size(0)
+                   : 1;
+
   TORCH_CHECK(K_in % 16 == 0, "k % 16 == 0");
   TORCH_CHECK(in_a.dtype() == in_b.dtype() && in_a.dtype() == kFp8Type);
   TORCH_CHECK(out_c.dtype() == torch::kFloat16 ||
@@ -1673,13 +1765,15 @@ void wvSplitKQ(at::Tensor& in_a, at::Tensor& in_b, at::Tensor& out_c,
     if ((K_in * N_in <= max_lds_len) && (M_in % _YTILEs == 0)) {               \
       int __wvPrGrp = mindiv(M_in, CuCount * _YTILEs, _WvPrGrp);               \
       wvSplitKQ_hf_sml_<fptype, fp8_t, 64, _YTILEs, _WvPrGrp, 16, _UNRLs, _N>  \
-          <<<grid, block, 0, stream>>>(K_in, Kp_in, M_in, a_ptr, b_ptr, c_ptr, \
-                                       s_a, s_b, __wvPrGrp, CuCount);          \
+          <<<grid, block, 0, stream>>>(K_in, Kp_in, M_in, Bx_in, By_in, a_ptr, \
+                                       b_ptr, bias_ptr, c_ptr, s_a, s_b,       \
+                                       __wvPrGrp, CuCount);                    \
     } else {                                                                   \
       int __wvPrGrp = mindiv(M_in, CuCount * _YTILEm, _WvPrGrp);               \
       wvSplitKQ_hf_<fptype, fp8_t, 64, _YTILEm, _WvPrGrp, 16, _UNRLm, _N>      \
-          <<<grid, block, 0, stream>>>(K_in, Kp_in, M_in, a_ptr, b_ptr, c_ptr, \
-                                       s_a, s_b, __wvPrGrp, CuCount);          \
+          <<<grid, block, 0, stream>>>(K_in, Kp_in, M_in, Bx_in, By_in, a_ptr, \
+                                       b_ptr, bias_ptr, c_ptr, s_a, s_b,       \
+                                       __wvPrGrp, CuCount);                    \
     }                                                                          \
   }
 
@@ -1691,6 +1785,9 @@ void wvSplitKQ(at::Tensor& in_a, at::Tensor& in_b, at::Tensor& out_c,
     VLLM_DISPATCH_FP8_TYPES(in_a.scalar_type(), "wvSplitKQ", [&] {
       auto a_ptr = in_a.data_ptr<fp8_t>();
       auto b_ptr = in_b.data_ptr<fp8_t>();
+      auto bias_ptr = (in_bias.has_value() && in_bias->numel() > 0)
+                          ? reinterpret_cast<fptype*>(in_bias->data_ptr())
+                          : nullptr;
       switch (N_in) {
         case 1:
           WVSPLITKQ(16, 2, 2, 2, 2, 2, 2, 1)

csrc/rocm/torch_bindings.cpp

@@ -22,13 +22,14 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
 
   // Custom gemm op for skinny matrix-matrix multiplication
   rocm_ops.def(
-      "wvSplitK(Tensor in_a, Tensor in_b, int CuCount) -> "
+      "wvSplitK(Tensor in_a, Tensor in_b, Tensor? in_bias, int CuCount) -> "
       "Tensor");
   rocm_ops.impl("wvSplitK", torch::kCUDA, &wvSplitK);
 
   // wvSplitK for fp8
   rocm_ops.def(
-      "wvSplitKQ(Tensor in_a, Tensor in_b, Tensor! out_c, Tensor scale_a, "
+      "wvSplitKQ(Tensor in_a, Tensor in_b, Tensor? in_bias, Tensor! out_c, "
+      "Tensor scale_a, "
       "          Tensor scale_b, int CuCount) -> ()");
   rocm_ops.impl("wvSplitKQ", torch::kCUDA, &wvSplitKQ);
 

docker/Dockerfile

@@ -404,6 +404,9 @@ RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
                 FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0a 12.0"
             fi
             echo "🏗️  Installing FlashInfer with AOT compilation for arches: ${FI_TORCH_CUDA_ARCH_LIST}"
+            export FLASHINFER_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}"
+            # HACK: We need these to run flashinfer.aot before installing flashinfer, get from the package in the future
+            uv pip install --system cuda-python==$(echo $CUDA_VERSION | cut -d. -f1,2) pynvml==$(echo $CUDA_VERSION | cut -d. -f1) nvidia-nvshmem-cu$(echo $CUDA_VERSION | cut -d. -f1)
             # Build AOT kernels
             TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
                 python3 -m flashinfer.aot

docker/Dockerfile.cpu

@@ -114,9 +114,6 @@ WORKDIR /workspace/vllm
 RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
     cp requirements/test.in requirements/cpu-test.in && \
     sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
-    sed -i 's/^torch==.*/torch==2.6.0/g' requirements/cpu-test.in && \
-    sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \
-    sed -i 's/torchvision.*/torchvision/g' requirements/cpu-test.in && \
     uv pip compile requirements/cpu-test.in -o requirements/cpu-test.txt --index-strategy unsafe-best-match --torch-backend cpu
 
 RUN --mount=type=cache,target=/root/.cache/uv \

docker/Dockerfile.rocm_base

@@ -65,8 +65,6 @@ ARG PYTORCH_BRANCH
 ARG PYTORCH_VISION_BRANCH
 ARG PYTORCH_REPO
 ARG PYTORCH_VISION_REPO
-ARG FA_BRANCH
-ARG FA_REPO
 RUN git clone ${PYTORCH_REPO} pytorch
 RUN cd pytorch && git checkout ${PYTORCH_BRANCH} && \
     pip install -r requirements.txt && git submodule update --init --recursive \
@@ -77,14 +75,20 @@ RUN git clone ${PYTORCH_VISION_REPO} vision
 RUN cd vision && git checkout ${PYTORCH_VISION_BRANCH} \
     && python3 setup.py bdist_wheel --dist-dir=dist \
     && pip install dist/*.whl
+RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \
+    && cp /app/vision/dist/*.whl /app/install
+
+FROM base AS build_fa
+ARG FA_BRANCH
+ARG FA_REPO
+RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
+    pip install /install/*.whl
 RUN git clone ${FA_REPO}
 RUN cd flash-attention \
     && git checkout ${FA_BRANCH} \
     && git submodule update --init \
     && GPU_ARCHS=$(echo ${PYTORCH_ROCM_ARCH} | sed -e 's/;gfx1[0-9]\{3\}//g') python3 setup.py bdist_wheel --dist-dir=dist
-RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \
-    && cp /app/vision/dist/*.whl /app/install \
-    && cp /app/flash-attention/dist/*.whl /app/install
+RUN mkdir -p /app/install && cp /app/flash-attention/dist/*.whl /app/install
 
 FROM base AS build_aiter
 ARG AITER_BRANCH
@@ -103,6 +107,8 @@ FROM base AS debs
 RUN mkdir /app/debs
 RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
     cp /install/*.whl /app/debs
+RUN --mount=type=bind,from=build_fa,src=/app/install/,target=/install \
+    cp /install/*.whl /app/debs
 RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
     cp /install/*.whl /app/debs
 RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
@@ -111,13 +117,7 @@ RUN --mount=type=bind,from=build_aiter,src=/app/install/,target=/install \
     cp /install/*.whl /app/debs
 
 FROM base AS final
-RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
-    pip install /install/*.whl
-RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
-    pip install /install/*.whl
-RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
-    pip install /install/*.whl
-RUN --mount=type=bind,from=build_aiter,src=/app/install/,target=/install \
+RUN --mount=type=bind,from=debs,src=/app/debs,target=/install \
     pip install /install/*.whl
 
 ARG BASE_IMAGE

docs/configuration/optimization.md

@@ -139,9 +139,9 @@ there is relatively little gain from TP. On the other hand, TP incurs significan
 overhead because of all-reduce being performed after every layer.
 
 Given this, it may be advantageous to instead shard the batched input data using TP, essentially
-performing batch-level DP. This has been shown to improve the throughput by around 10% for
+performing batch-level DP. This has been shown to improve the throughput and TTFT by around 10% for
 `tensor_parallel_size=8`. For vision encoders that use hardware-unoptimized Conv3D operations,
-batch-level DP can provide another 40% increase to throughput compared to regular TP.
+batch-level DP can provide another 40% improvement compared to regular TP.
 
 Nevertheless, since the weights of the multi-modal encoder are replicated across each TP rank,
 there will be a minor increase in memory consumption and may cause OOM if you can barely fit the model already.
@@ -172,14 +172,15 @@ Batch-level DP needs to be implemented on a per-model basis,
 and enabled by setting `supports_encoder_tp_data = True` in the model class.
 Regardless, you need to set `mm_encoder_tp_mode="data"` in engine arguments to use this feature.
 
-Known supported models:
+Known supported models (with corresponding benchmarks):
 
-- GLM-4.5V GLM-4.1V (<gh-pr:23168>)
+- dots_ocr (<gh-pr:25466>)
+- GLM-4.1V or above (<gh-pr:23168>)
 - InternVL (<gh-pr:23909>)
 - Kimi-VL (<gh-pr:23817>)
 - Llama4 (<gh-pr:18368>)
 - MiniCPM-V-2.5 or above (<gh-pr:23327>, <gh-pr:23948>)
-- Qwen2.5-VL (<gh-pr:22742>)
+- Qwen2-VL or above (<gh-pr:22742>, <gh-pr:24955>, <gh-pr:25445>)
 - Step3 (<gh-pr:22697>)
 
 ## Input Processing

docs/contributing/benchmarks.md

@@ -680,7 +680,7 @@ vllm bench serve \
   --save-result \
   --result-dir ~/vllm_benchmark_results \
   --save-detailed \
-  --endpoint /v1/chat/completion
+  --endpoint /v1/chat/completions
 ```
 
 ##### Videos (ShareGPT4Video)
@@ -707,7 +707,7 @@ vllm bench serve \
   --save-result \
   --result-dir ~/vllm_benchmark_results \
   --save-detailed \
-  --endpoint /v1/chat/completion
+  --endpoint /v1/chat/completions
 ```
 
 ##### Synthetic Random Images (random-mm)

docs/features/disagg_prefill.md

@@ -23,7 +23,7 @@ Now supports 5 types of connectors:
 
 - **SharedStorageConnector**: refer to <gh-file:examples/offline_inference/disaggregated-prefill-v1/run.sh> for the example usage of SharedStorageConnector disaggregated prefilling.
 - **LMCacheConnectorV1**: refer to <gh-file:examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh> for the example usage of LMCacheConnectorV1 disaggregated prefilling which uses NIXL as the underlying KV transmission.
-- **NixlConnector**: refer to <gh-file:tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh> for the example usage of NixlConnector disaggregated prefilling which support fully async send/recv.
+- **NixlConnector**: refer to <gh-file:tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh> for the example usage of NixlConnector disaggregated prefilling which support fully async send/recv. For detailed usage guide, see [NixlConnector Usage Guide](nixl_connector_usage.md).
 - **P2pNcclConnector**: refer to <gh-file:examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh> for the example usage of P2pNcclConnector disaggregated prefilling.
 - **MultiConnector**: take advantage of the kv_connector_extra_config: dict[str, Any] already present in KVTransferConfig to stash all the connectors we want in an ordered list of kwargs.such as:
 
@@ -31,6 +31,12 @@ Now supports 5 types of connectors:
   --kv-transfer-config '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both"},{"kv_connector":"SharedStorageConnector","kv_role":"kv_both","kv_connector_extra_config":{"shared_storage_path":"local_storage"}}]}}'
   ```
 
+For NixlConnector, you may also specify one or multiple NIXL_Backend. Such as:
+
+  ```bash
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both", "kv_buffer_device":"cuda", "kv_connector_extra_config":{"backends":["UCX", "GDS"]}}'
+  ```
+
 - **OffloadingConnector**: enable offloading of KV data to CPU memory, customizing the CPU block size (in tokens) and number of blocks to allocate (per worker):
 
   ```bash

docs/features/multimodal_inputs.md

@@ -6,6 +6,10 @@ This page teaches you how to pass multi-modal inputs to [multi-modal models][sup
     We are actively iterating on multi-modal support. See [this RFC](gh-issue:4194) for upcoming changes,
     and [open an issue on GitHub](https://github.com/vllm-project/vllm/issues/new/choose) if you have any feedback or feature requests.
 
+!!! tip
+    When serving multi-modal models, consider setting `--allowed-media-domains` to restrict domain that vLLM can access to prevent it from accessing arbitrary endpoints that can potentially be vulnerable to Server-Side Request Forgery (SSRF) attacks. You can provide a list of domains for this arg. For example: `--allowed-media-domains upload.wikimedia.org github.com www.bogotobogo.com`
+    This restriction is especially important if you run vLLM in a containerized environment where the vLLM pods may have unrestricted access to internal networks.
+
 ## Offline Inference
 
 To input multi-modal data, follow this schema in [vllm.inputs.PromptType][]:

docs/features/nixl_connector_usage.md

@@ -0,0 +1,159 @@
+# NixlConnector Usage Guide
+
+NixlConnector is a high-performance KV cache transfer connector for vLLM's disaggregated prefilling feature. It provides fully asynchronous send/receive operations using the NIXL library for efficient cross-process KV cache transfer.
+
+## Prerequisites
+
+### Installation
+
+Install the NIXL library: `uv pip install nixl`, as a quick start.
+
+- Refer to [NIXL official repository](https://github.com/ai-dynamo/nixl) for more installation instructions
+- The specified required NIXL version can be found in [requirements/kv_connectors.txt](gh-file:requirements/kv_connectors.txt) and other relevant config files
+
+### Transport Configuration
+
+NixlConnector uses NIXL library for underlying communication, which supports multiple transport backends. UCX (Unified Communication X) is the primary default transport library used by NIXL. Configure transport environment variables:
+
+```bash
+# Example UCX configuration, adjust according to your enviroment
+export UCX_TLS=all  # or specify specific transports like "rc,ud,sm,^cuda_ipc" ..etc
+export UCX_NET_DEVICES=all  # or specify network devices like "mlx5_0:1,mlx5_1:1"
+```
+
+!!! tip
+    When using UCX as the transport backend, NCCL environment variables (like `NCCL_IB_HCA`, `NCCL_SOCKET_IFNAME`) are not applicable to NixlConnector, so configure UCX-specific environment variables instead of NCCL variables.
+
+## Basic Usage (on the same host)
+
+### Producer (Prefiller) Configuration
+
+Start a prefiller instance that produces KV caches
+
+```bash
+# 1st GPU as prefiller
+CUDA_VISIBLE_DEVICES=0 \
+UCX_NET_DEVICES=all \
+VLLM_NIXL_SIDE_CHANNEL_PORT=5600 \
+vllm serve Qwen/Qwen3-0.6B \
+  --port 8100 \
+  --enforce-eager \
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+```
+
+### Consumer (Decoder) Configuration
+
+Start a decoder instance that consumes KV caches:
+
+```bash
+# 2nd GPU as decoder
+CUDA_VISIBLE_DEVICES=1 \
+UCX_NET_DEVICES=all \
+VLLM_NIXL_SIDE_CHANNEL_PORT=5601 \
+vllm serve Qwen/Qwen3-0.6B \
+  --port 8200 \
+  --enforce-eager \
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+```
+
+### Proxy Server
+
+Use a proxy server to route requests between prefiller and decoder:
+
+```bash
+python tests/v1/kv_connector/nixl_integration/toy_proxy_server.py \
+  --port 8192 \
+  --prefiller-hosts localhost \
+  --prefiller-ports 8100 \
+  --decoder-hosts localhost \
+  --decoder-ports 8200
+```
+
+## Environment Variables
+
+- `VLLM_NIXL_SIDE_CHANNEL_PORT`: Port for NIXL handshake communication
+    - Default: 5600
+    - **Required for both prefiller and decoder instances**
+    - Each vLLM worker needs a unique port on its host; using the same port number across different hosts is fine
+    - For TP/DP deployments, each worker's port on a node is computed as: base_port + dp_rank * tp_size + tp_rank (e.g., with `--tensor-parallel-size=4` and base_port=5600, tp_rank 0..3 use ports 5600, 5601, 5602, 5603 on that node).
+    - Used for the initial NIXL handshake between the prefiller and the decoder
+
+- `VLLM_NIXL_SIDE_CHANNEL_HOST`: Host for side channel communication
+    - Default: "localhost"
+    - Set when prefiller and decoder are on different machines
+    - Connection info is passed via KVTransferParams from prefiller to decoder for handshake
+
+- `VLLM_NIXL_ABORT_REQUEST_TIMEOUT`: Timeout (in seconds) for automatically releasing the prefiller’s KV cache for a particular request. (Optional)
+    - Default: 120
+    - If a request is aborted and the decoder has not yet read the KV-cache blocks through the nixl channel, the prefill instance will release its KV-cache blocks after this timeout to avoid holding them indefinitely.
+
+## Multi-Instance Setup
+
+### Multiple Prefiller Instances on Different Machines
+
+```bash
+# Prefiller 1 on Machine A (example IP: ${IP1})
+VLLM_NIXL_SIDE_CHANNEL_HOST=${IP1} \
+VLLM_NIXL_SIDE_CHANNEL_PORT=5600 \
+UCX_NET_DEVICES=all \
+vllm serve Qwen/Qwen3-0.6B --port 8000 \
+  --tensor-parallel-size 8 \
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_producer"}'
+
+# Prefiller 2 on Machine B (example IP: ${IP2})
+VLLM_NIXL_SIDE_CHANNEL_HOST=${IP2} \
+VLLM_NIXL_SIDE_CHANNEL_PORT=5600 \
+UCX_NET_DEVICES=all \
+vllm serve Qwen/Qwen3-0.6B --port 8000 \
+  --tensor-parallel-size 8 \
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_producer"}'
+```
+
+### Multiple Decoder Instances on Different Machines
+
+```bash
+# Decoder 1 on Machine C (example IP: ${IP3})
+VLLM_NIXL_SIDE_CHANNEL_HOST=${IP3} \
+VLLM_NIXL_SIDE_CHANNEL_PORT=5600 \
+UCX_NET_DEVICES=all \
+vllm serve Qwen/Qwen3-0.6B --port 8000 \
+  --tensor-parallel-size 8 \
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_consumer"}'
+
+# Decoder 2 on Machine D (example IP: ${IP4})
+VLLM_NIXL_SIDE_CHANNEL_HOST=${IP4} \
+VLLM_NIXL_SIDE_CHANNEL_PORT=5600 \
+UCX_NET_DEVICES=all \
+vllm serve Qwen/Qwen3-0.6B --port 8000 \
+  --tensor-parallel-size 8 \
+  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_consumer"}'
+```
+
+### Proxy for Multiple Instances
+
+```bash
+python tests/v1/kv_connector/nixl_integration/toy_proxy_server.py \
+  --port 8192 \
+  --prefiller-hosts ${IP1} ${IP2} \
+  --prefiller-ports 8000 8000 \
+  --decoder-hosts ${IP3} ${IP4} \
+  --decoder-ports 8000 8000
+```
+
+### KV Role Options
+
+- **kv_producer**: For prefiller instances that generate KV caches
+- **kv_consumer**: For decoder instances that consume KV caches from prefiller
+- **kv_both**: Enables symmetric functionality where the connector can act as both producer and consumer. This provides flexibility for experimental setups and scenarios where the role distinction is not predetermined.
+
+!!! tip
+    NixlConnector currently does not distinguish `kv_role`; the actual prefiller/decoder roles are determined by the upper-level proxy (e.g., `toy_proxy_server.py` using `--prefiller-hosts` and `--decoder-hosts`).
+    Therefore, `kv_role` in `--kv-transfer-config` is effectively a placeholder and does not affect NixlConnector's behavior.
+
+## Example Scripts/Code
+
+Refer to these example scripts in the vLLM repository:
+
+- [run_accuracy_test.sh](gh-file:tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh)
+- [toy_proxy_server.py](gh-file:tests/v1/kv_connector/nixl_integration/toy_proxy_server.py)
+- [test_accuracy.py](gh-file:tests/v1/kv_connector/nixl_integration/test_accuracy.py)

docs/features/structured_outputs.md

@@ -6,6 +6,17 @@ vLLM supports the generation of structured outputs using
 This document shows you some examples of the different options that are
 available to generate structured outputs.
 
+!!! warning
+    If you are still using the following deprecated API fields, please update your code to use `structured_outputs` as demonstrated in the rest of this document:
+
+    - `guided_json` -> `{"structured_outputs": {"json": ...}}` or `StructuredOutputsParams(json=...)`
+    - `guided_regex` -> `{"structured_outputs": {"regex": ...}}` or `StructuredOutputsParams(regex=...)`
+    - `guided_choice` -> `{"structured_outputs": {"choice": ...}}` or `StructuredOutputsParams(choice=...)`
+    - `guided_grammar` -> `{"structured_outputs": {"grammar": ...}}` or `StructuredOutputsParams(grammar=...)`
+    - `guided_whitespace_pattern` -> `{"structured_outputs": {"whitespace_pattern": ...}}` or `StructuredOutputsParams(whitespace_pattern=...)`
+    - `structural_tag` -> `{"structured_outputs": {"structural_tag": ...}}` or `StructuredOutputsParams(structural_tag=...)`
+    - `guided_decoding_backend` -> Remove this field from your request
+
 ## Online Serving (OpenAI API)
 
 You can generate structured outputs using the OpenAI's [Completions](https://platform.openai.com/docs/api-reference/completions) and [Chat](https://platform.openai.com/docs/api-reference/chat) API.

docs/features/tool_calling.md

@@ -310,6 +310,15 @@ Flags:
 * For non-reasoning: `--tool-call-parser hunyuan_a13b`
 * For reasoning: `--tool-call-parser hunyuan_a13b --reasoning-parser hunyuan_a13b --enable_reasoning`
 
+### LongCat-Flash-Chat Models (`longcat`)
+
+Supported models:
+
+* `meituan-longcat/LongCat-Flash-Chat`
+* `meituan-longcat/LongCat-Flash-Chat-FP8`
+
+Flags: `--tool-call-parser longcat`
+
 ### GLM-4.5 Models (`glm45`)
 
 Supported models:
@@ -319,6 +328,15 @@ Supported models:
 
 Flags: `--tool-call-parser glm45`
 
+### Qwen3-Coder Models (`qwen3_xml`)
+
+Supported models:
+
+* `Qwen/Qwen3-480B-A35B-Instruct`
+* `Qwen/Qwen3-Coder-30B-A3B-Instruct`
+
+Flags: `--tool-call-parser qwen3_xml`
+
 ### Models with Pythonic Tool Calls (`pythonic`)
 
 A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models.

docs/getting_started/installation/cpu/x86.inc.md

@@ -20,7 +20,80 @@ vLLM supports basic model inferencing and serving on x86 CPU platform, with data
 # --8<-- [end:pre-built-wheels]
 # --8<-- [start:build-wheel-from-source]
 
---8<-- "docs/getting_started/installation/cpu/build.inc.md"
+Install recommended compiler. We recommend to use `gcc/g++ >= 12.3.0` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run:
+
+```bash
+sudo apt-get update -y
+sudo apt-get install -y gcc-12 g++-12 libnuma-dev python3-dev
+sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
+```
+
+Clone the vLLM project:
+
+```bash
+git clone https://github.com/vllm-project/vllm.git vllm_source
+cd vllm_source
+```
+
+Install the required dependencies:
+
+```bash
+uv pip install -r requirements/cpu-build.txt --torch-backend cpu
+uv pip install -r requirements/cpu.txt --torch-backend cpu
+```
+
+??? console "pip"
+    ```bash
+    pip install --upgrade pip
+    pip install -v -r requirements/cpu-build.txt --extra-index-url https://download.pytorch.org/whl/cpu
+    pip install -v -r requirements/cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
+    ```
+
+Build and install vLLM:
+
+```bash
+VLLM_TARGET_DEVICE=cpu uv pip install . --no-build-isolation
+```
+
+If you want to develop vLLM, install it in editable mode instead.
+
+```bash
+VLLM_TARGET_DEVICE=cpu uv pip install -e . --no-build-isolation
+```
+
+Optionally, build a portable wheel which you can then install elsewhere:
+
+```bash
+VLLM_TARGET_DEVICE=cpu uv build --wheel
+```
+
+```bash
+uv pip install dist/*.whl
+```
+
+??? console "pip"
+    ```bash
+    VLLM_TARGET_DEVICE=cpu python -m build --wheel --no-isolation
+    ```
+
+    ```bash
+    pip install dist/*.whl
+    ```
+
+!!! example "Troubleshooting"
+    - **NumPy ≥2.0 error**: Downgrade using `pip install "numpy<2.0"`.
+    - **CMake picks up CUDA**: Add `CMAKE_DISABLE_FIND_PACKAGE_CUDA=ON` to prevent CUDA detection during CPU builds, even if CUDA is installed.
+    - `AMD` requies at least 4th gen processors (Zen 4/Genoa) or higher to support [AVX512](https://www.phoronix.com/review/amd-zen4-avx512) to run vLLM on CPU.
+    - If you receive an error such as: `Could not find a version that satisfies the requirement torch==X.Y.Z+cpu+cpu`, consider updating [pyproject.toml](https://github.com/vllm-project/vllm/blob/main/pyproject.toml) to help pip resolve the dependency.
+    ```toml title="pyproject.toml"
+    [build-system]
+    requires = [
+      "cmake>=3.26.1",
+      ...
+      "torch==X.Y.Z+cpu"   # <-------
+    ]
+    ```
+    - If you are building vLLM from source and not using the pre-built images, remember to set `LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD"` on x86 machines before running vLLM.
 
 # --8<-- [end:build-wheel-from-source]
 # --8<-- [start:pre-built-images]
@@ -57,4 +130,4 @@ docker run --rm \
 
 # --8<-- [end:build-image-from-source]
 # --8<-- [start:extra-information]
-# --8<-- [end:extra-information]
+# --8<-- [end:extra-information]

docs/mkdocs/hooks/generate_argparse.py

@@ -32,8 +32,9 @@ def auto_mock(module, attr, max_mocks=50):
     for _ in range(max_mocks):
         try:
             # First treat attr as an attr, then as a submodule
-            return getattr(importlib.import_module(module), attr,
-                           importlib.import_module(f"{module}.{attr}"))
+            with patch("importlib.metadata.version", return_value="0.0.0"):
+                return getattr(importlib.import_module(module), attr,
+                               importlib.import_module(f"{module}.{attr}"))
         except importlib.metadata.PackageNotFoundError as e:
             raise e
         except ModuleNotFoundError as e:
@@ -167,5 +168,5 @@ def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
         doc_path = ARGPARSE_DOC_DIR / f"{stem}.md"
         # Specify encoding for building on Windows
         with open(doc_path, "w", encoding="utf-8") as f:
-            f.write(parser.format_help())
+            f.write(super(type(parser), parser).format_help())
         logger.info("Argparse generated: %s", doc_path.relative_to(ROOT_DIR))

docs/models/generative_models.md

@@ -4,7 +4,7 @@ vLLM provides first-class support for generative models, which covers most of LL
 
 In vLLM, generative models implement the[VllmModelForTextGeneration][vllm.model_executor.models.VllmModelForTextGeneration] interface.
 Based on the final hidden states of the input, these models output log probabilities of the tokens to generate,
-which are then passed through [Sampler][vllm.model_executor.layers.sampler.Sampler] to obtain the final text.
+which are then passed through [Sampler][vllm.v1.sample.sampler.Sampler] to obtain the final text.
 
 ## Configuration
 

docs/models/supported_models.md

@@ -29,7 +29,7 @@ _*Vision-language models currently accept only image inputs. Support for video i
 
 If the Transformers model implementation follows all the steps in [writing a custom model](#writing-custom-models) then, when used with the Transformers backend, it will be compatible with the following features of vLLM:
 
-- All the features listed in the [compatibility matrix](../features/compatibility_matrix.md#feature-x-feature)
+- All the features listed in the [compatibility matrix](../features/README.md#feature-x-feature)
 - Any combination of the following vLLM parallelisation schemes:
     - Pipeline parallel
     - Tensor parallel
@@ -428,6 +428,7 @@ th {
 | `MiniMaxM1ForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-M1-40k`, `MiniMaxAI/MiniMax-M1-80k`, etc. | | | ✅︎ |
 | `MiniMaxText01ForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-Text-01`, etc. | | | ✅︎ |
 | `Zamba2ForCausalLM` | Zamba2 | `Zyphra/Zamba2-7B-instruct`, `Zyphra/Zamba2-2.7B-instruct`, `Zyphra/Zamba2-1.2B-instruct`, etc. | | | ✅︎ |
+| `LongcatFlashForCausalLM` | LongCat-Flash | `meituan-longcat/LongCat-Flash-Chat`, `meituan-longcat/LongCat-Flash-Chat-FP8` | ✅︎ |✅︎ | ✅︎ |
 
 Some models are supported only via the [Transformers backend](#transformers). The purpose of the table below is to acknowledge models which we officially support in this way. The logs will say that the Transformers backend is being used, and you will see no warning that this is fallback behaviour. This means that, if you have issues with any of the models listed below, please [make an issue](https://github.com/vllm-project/vllm/issues/new/choose) and we'll do our best to fix it!
 

docs/serving/expert_parallel_deployment.md

@@ -193,7 +193,7 @@ For production deployments requiring strict SLA guarantees for time-to-first-tok
 
 1. **Install gdrcopy/ucx/nixl**: For maximum performance, run the [install_gdrcopy.sh](gh-file:tools/install_gdrcopy.sh) script to install `gdrcopy` (e.g., `install_gdrcopy.sh "${GDRCOPY_OS_VERSION}" "12.8" "x64"`). You can find available OS versions [here](https://developer.download.nvidia.com/compute/redist/gdrcopy/CUDA%2012.8/). If `gdrcopy` is not installed, things will still work with a plain `pip install nixl`, just with lower performance. `nixl` and `ucx` are installed as dependencies via pip.
 
-2. **Configure Both Instances**: Add this flag to both prefill and decode instances `--kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}`
+2. **Configure Both Instances**: Add this flag to both prefill and decode instances `--kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}`. Noted, you may also specify one or multiple NIXL_Backend. Such as: `--kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both", "kv_connector_extra_config":{"backends":["UCX", "GDS"]}}'`
 
 3. **Client Orchestration**: Use the client-side script below to coordinate prefill/decode operations. We are actively working on routing solutions.
 

docs/usage/README.md

@@ -1,6 +1,6 @@
 # Using vLLM
 
-First, vLLM must be [installed](../getting_started/installation) for your chosen device in either a Python or Docker environment.
+First, vLLM must be [installed](../getting_started/installation/) for your chosen device in either a Python or Docker environment.
 
 Then, vLLM supports the following usage patterns:
 

docs/usage/security.md

@@ -60,6 +60,12 @@ Key points from the PyTorch security guide:
 - Implement proper authentication and authorization for management interfaces
 - Follow the principle of least privilege for all system components
 
+### 4. **Restrict Domains Access for Media URLs:**
+
+Restrict domains that vLLM can access for media URLs by setting
+`--allowed-media-domains` to prevent Server-Side Request Forgery (SSRF) attacks.
+(e.g. `--allowed-media-domains upload.wikimedia.org github.com www.bogotobogo.com`)
+
 ## Security and Firewalls: Protecting Exposed vLLM Systems
 
 While vLLM is designed to allow unsafe network services to be isolated to

examples/offline_inference/basic/chat.py

@@ -87,6 +87,7 @@ def main(args: dict):
             use_tqdm=False,
             chat_template=chat_template,
         )
+        print_outputs(outputs)
 
 
 if __name__ == "__main__":

examples/offline_inference/data_parallel.py

@@ -101,6 +101,13 @@ def parse_args():
         "--quantization",
         type=str,
     )
+    parser.add_argument(
+        "--disable-expert-parallel",
+        dest="enable_expert_parallel",
+        action="store_false",
+        help="Disable expert parallel (default: enabled).",
+    )
+    parser.set_defaults(enable_expert_parallel=True)
     return parser.parse_args()
 
 
@@ -113,6 +120,7 @@ def main(
     dp_master_port,
     GPUs_per_dp_rank,
     enforce_eager,
+    enable_expert_parallel,
     trust_remote_code,
     max_num_seqs,
     max_model_len,
@@ -168,7 +176,7 @@ def main(
         model=model,
         tensor_parallel_size=GPUs_per_dp_rank,
         enforce_eager=enforce_eager,
-        enable_expert_parallel=True,
+        enable_expert_parallel=enable_expert_parallel,
         trust_remote_code=trust_remote_code,
         max_num_seqs=max_num_seqs,
         max_model_len=max_model_len,
@@ -229,6 +237,7 @@ if __name__ == "__main__":
                 dp_master_port,
                 tp_size,
                 args.enforce_eager,
+                args.enable_expert_parallel,
                 args.trust_remote_code,
                 args.max_num_seqs,
                 args.max_model_len,

examples/offline_inference/spec_decode.py

@@ -49,6 +49,7 @@ def get_custom_mm_prompts(num_prompts):
 def parse_args():
     parser = FlexibleArgumentParser()
     add_dataset_parser(parser)
+    parser.add_argument("--test", action="store_true")
     parser.add_argument(
         "--method",
         type=str,
@@ -60,6 +61,7 @@ def parse_args():
     parser.add_argument("--tp", type=int, default=1)
     parser.add_argument("--enforce-eager", action="store_true")
     parser.add_argument("--enable-chunked-prefill", action="store_true")
+    parser.add_argument("--max-model-len", type=int, default=16384)
     parser.add_argument("--temp", type=float, default=0)
     parser.add_argument("--top-p", type=float, default=1.0)
     parser.add_argument("--top-k", type=int, default=-1)
@@ -71,8 +73,7 @@ def parse_args():
     return parser.parse_args()
 
 
-def main():
-    args = parse_args()
+def main(args):
     args.endpoint_type = "openai-chat"
 
     model_dir = args.model_dir
@@ -134,7 +135,7 @@ def main():
         gpu_memory_utilization=0.8,
         speculative_config=speculative_config,
         disable_log_stats=False,
-        max_model_len=16384,
+        max_model_len=args.max_model_len,
         limit_mm_per_prompt={"image": 5},
         disable_chunked_mm_input=True,
     )
@@ -198,6 +199,39 @@ def main():
         acceptance_rate = acceptance_counts[i] / num_drafts if num_drafts > 0 else 0
         print(f"acceptance at token {i}: {acceptance_rate:.2f}")
 
+    return acceptance_length
+
 
 if __name__ == "__main__":
-    main()
+    args = parse_args()
+    acceptance_length = main(args)
+
+    if args.test:
+        # takes ~30s to run on 1xH100
+        assert args.method in ["eagle", "eagle3"]
+        assert args.tp == 1
+        assert args.num_spec_tokens == 3
+        assert args.dataset_name == "hf"
+        assert args.dataset_path == "philschmid/mt-bench"
+        assert args.num_prompts == 80
+        assert args.temp == 0
+        assert args.top_p == 1.0
+        assert args.top_k == -1
+        assert args.enable_chunked_prefill
+
+        # check acceptance length is within 2% of expected value
+        rtol = 0.02
+        expected_acceptance_length = 2.296 if args.method == "eagle" else 2.811
+
+        assert (
+            acceptance_length <= (1 + rtol) * expected_acceptance_length
+            and acceptance_length >= (1 - rtol) * expected_acceptance_length
+        ), (
+            f"acceptance_length {acceptance_length} is not "
+            f"within {rtol * 100}% of {expected_acceptance_length}"
+        )
+
+        print(
+            f"Test passed! Expected AL: "
+            f"{expected_acceptance_length}, got {acceptance_length}"
+        )

examples/online_serving/dashboards/grafana/README.md

@@ -11,9 +11,9 @@ vLLM performance and metrics.
 
 ## Dashboard Descriptions
 
-- **[performance_statistics.json](./performance_statistics.json)**: Tracks performance metrics including latency and
+- **performance_statistics.json**: Tracks performance metrics including latency and
   throughput for your vLLM service.
-- **[query_statistics.json](./query_statistics.json)**: Tracks query performance, request volume, and key
+- **query_statistics.json**: Tracks query performance, request volume, and key
   performance indicators for your vLLM service.
 
 ## Deployment Options

examples/online_serving/dashboards/perses/README.md

@@ -21,9 +21,9 @@ deployment methods:
 
 ## Dashboard Descriptions
 
-- **[performance_statistics.yaml](./performance_statistics.yaml)**: Performance metrics with aggregated latency
+- **performance_statistics.yaml**: Performance metrics with aggregated latency
   statistics
-- **[query_statistics.yaml](./query_statistics.yaml)**: Query performance and deployment metrics
+- **query_statistics.yaml**: Query performance and deployment metrics
 
 ## Deployment Options
 

requirements/common.txt

@@ -24,7 +24,7 @@ outlines_core == 0.2.11
 # required for outlines backend disk cache
 diskcache == 5.6.3
 lark == 1.2.2
-xgrammar == 0.1.24; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
+xgrammar == 0.1.25; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs

requirements/cpu-build.txt

@@ -1,12 +1,11 @@
-# Temporarily used for x86 CPU backend to avoid performance regression of torch>2.6.0+cpu,
-# see https://github.com/pytorch/pytorch/pull/151218
 cmake>=3.26.1
 ninja
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch==2.6.0+cpu
+torch==2.8.0+cpu; platform_machine == "x86_64"
+torch==2.8.0; platform_machine == "ppc64le" or platform_machine == "aarch64" or platform_system == "Darwin"
 wheel
 jinja2>=3.1.6
 regex

requirements/cpu.txt

@@ -8,7 +8,7 @@ numba == 0.61.2; python_version > '3.9' and platform_machine != "s390x"
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch==2.6.0+cpu; platform_machine == "x86_64" # torch>2.6.0+cpu has performance regression on x86 platform, see https://github.com/pytorch/pytorch/pull/151218
+torch==2.8.0+cpu; platform_machine == "x86_64"
 torch==2.8.0; platform_system == "Darwin"
 torch==2.8.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
 
@@ -23,7 +23,7 @@ datasets # for benchmark scripts
 
 # Intel Extension for PyTorch, only for x86_64 CPUs
 intel-openmp==2024.2.1; platform_machine == "x86_64"
-intel_extension_for_pytorch==2.6.0; platform_machine == "x86_64" # torch>2.6.0+cpu has performance regression on x86 platform, see https://github.com/pytorch/pytorch/pull/151218
+intel_extension_for_pytorch==2.8.0; platform_machine == "x86_64"
 triton==3.2.0; platform_machine == "x86_64" # Triton is required for torch 2.6+cpu, as it is imported in torch.compile.
 
 # Use this to gather CPU info and optimize based on ARM Neoverse cores

requirements/tpu.txt

@@ -14,14 +14,4 @@ nixl==0.3.0
 tpu_info==0.4.0
 
 # Install torch_xla
---pre
---extra-index-url https://download.pytorch.org/whl/nightly/cpu
---find-links https://storage.googleapis.com/libtpu-wheels/index.html
---find-links https://storage.googleapis.com/libtpu-releases/index.html
---find-links https://storage.googleapis.com/jax-releases/jax_nightly_releases.html
---find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
-torch==2.9.0.dev20250730
-torchvision==0.24.0.dev20250730
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.9.0.dev20250730-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.9.0.dev20250730-cp312-cp312-linux_x86_64.whl ; python_version == "3.12"
-
+torch_xla[tpu, pallas]==2.8.0

tests/compile/piecewise/test_full_cudagraph.py

@@ -103,7 +103,7 @@ backend_configs = {
     # Triton Attention
     "TritonAttn":
     BackendConfig(name="TritonAttn",
-                  env_vars={"VLLM_ATTENTION_BACKEND": "TRITON_ATTN_VLLM_V1"},
+                  env_vars={"VLLM_ATTENTION_BACKEND": "TRITON_ATTN"},
                   comp_config={
                       "cudagraph_mode": "FULL",
                   }),

tests/compile/test_fusion_attn.py

@@ -338,7 +338,7 @@ else:
 @pytest.mark.parametrize("model_name, model_class", MODELS)
 @pytest.mark.parametrize("backend",
                          [_Backend.FLASHINFER] if current_platform.is_cuda()
-                         else [_Backend.TRITON_ATTN_VLLM_V1])
+                         else [_Backend.TRITON_ATTN])
 @pytest.mark.parametrize(
     "split_attention",
     [False, True] if current_platform.is_rocm() else [False])

tests/conftest.py

@@ -1079,7 +1079,7 @@ def dummy_llava_path():
                           local_dir=_dummy_llava_path,
                           ignore_patterns=[
                               "*.bin", "*.bin.index.json", "*.pt", "*.h5",
-                              "*.msgpack"
+                              "*.msgpack", "*.safetensors"
                           ])
         assert os.path.exists(json_path)
         with open(json_path) as f:
@@ -1098,7 +1098,7 @@ def dummy_gemma2_embedding_path():
                           local_dir=_dummy_gemma2_embedding_path,
                           ignore_patterns=[
                               "*.bin", "*.bin.index.json", "*.pt", "*.h5",
-                              "*.msgpack"
+                              "*.msgpack", "*.safetensors"
                           ])
         assert os.path.exists(json_path)
         with open(json_path) as f:

tests/distributed/test_nccl_symm_mem_allreduce.py

@@ -0,0 +1,94 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import random
+import typing
+
+import pytest
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+
+import vllm.envs as envs
+from vllm.distributed import cleanup_dist_env_and_memory
+from vllm.distributed.device_communicators.cuda_communicator import (
+    CudaCommunicator)
+from vllm.distributed.device_communicators.pynccl import (
+    register_nccl_symmetric_ops)
+from vllm.distributed.device_communicators.pynccl_allocator import (
+    get_nccl_mem_pool, is_symmetric_memory_enabled)
+from vllm.distributed.parallel_state import (get_tp_group,
+                                             init_distributed_environment,
+                                             initialize_model_parallel)
+from vllm.platforms import current_platform
+from vllm.utils import update_environment_variables
+
+torch.manual_seed(42)
+random.seed(44)
+
+test_size_elements = 4 * 1024 * 1024
+
+
+def nccl_symm_mem_allreduce_worker(local_rank: int, world_size: int):
+    monkeypatch = pytest.MonkeyPatch()
+    with monkeypatch.context() as m:
+        m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
+        dtype = torch.bfloat16
+        device = torch.device(f"cuda:{local_rank}")
+        torch.cuda.set_device(device)
+        torch.set_default_device(device)
+        torch.set_default_dtype(dtype)
+        update_environment_variables({
+            "RANK": str(local_rank),
+            "LOCAL_RANK": str(local_rank),
+            "WORLD_SIZE": str(world_size),
+            "MASTER_ADDR": "localhost",
+            "MASTER_PORT": "12345",
+        })
+
+        init_distributed_environment()
+        initialize_model_parallel(tensor_model_parallel_size=world_size)
+
+        cuda_communicator = typing.cast(CudaCommunicator,
+                                        get_tp_group().device_communicator)
+        pynccl_comm = cuda_communicator.pynccl_comm
+        if get_nccl_mem_pool() is None:
+            pytest.skip("NCCL allocator compilation failed "
+                        "(probably missing NCCL headers).")
+        if not is_symmetric_memory_enabled():
+            pytest.skip("NCCL symmetric memory allreduce is disabled.")
+
+        register_nccl_symmetric_ops(pynccl_comm)
+        input = torch.randint(1,
+                              23, (test_size_elements, ),
+                              dtype=dtype,
+                              device=device)
+        input_clone = input.clone()
+        output = torch.ops.vllm.all_reduce_symmetric_with_copy(input)
+        assert output is not None
+
+        group = get_tp_group().device_group
+        dist.all_reduce(input_clone, group=group)
+        torch.testing.assert_close(output, input_clone, atol=2.5, rtol=0.1)
+
+
+@pytest.mark.skipif(
+    not current_platform.is_cuda(),
+    reason="NCCLSymmMemAllreduce is only available for CUDA platforms.",
+)
+@pytest.mark.parametrize("world_size", [2])
+@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"],
+                    reason="Only test on CUDA")
+def test_nccl_symm_mem_allreduce(monkeypatch: pytest.MonkeyPatch, world_size):
+    if world_size > torch.cuda.device_count():
+        pytest.skip("Not enough GPUs to run the test.")
+
+    # Enable SymmMemCommunicator
+    monkeypatch.setenv("VLLM_USE_NCCL_SYMM_MEM", "1")
+    monkeypatch.setenv("NCCL_NVLS_ENABLE", "1")
+    monkeypatch.setenv("NCCL_CUMEM_ENABLE", "1")
+
+    mp.spawn(nccl_symm_mem_allreduce_worker,
+             args=(world_size, ),
+             nprocs=world_size)
+    cleanup_dist_env_and_memory()

tests/distributed/test_pipeline_parallel.py

@@ -382,7 +382,6 @@ def test_tp_language_generation(
     test_options: PPTestOptions,
     num_gpus_available,
 ):
-    pytest.skip("Skipping the test until V1 passes it.")
     _compare_tp(model_id,
                 parallel_setup,
                 distributed_backend,
@@ -410,7 +409,6 @@ def test_tp_language_embedding(
     test_options: PPTestOptions,
     num_gpus_available,
 ):
-    pytest.skip("Skipping the test until V1 passes it.")
     _compare_tp(model_id,
                 parallel_setup,
                 distributed_backend,
@@ -438,7 +436,6 @@ def test_tp_multimodal_generation(
     test_options: PPTestOptions,
     num_gpus_available,
 ):
-    pytest.skip("Skipping the test until V1 passes it.")
     _compare_tp(model_id,
                 parallel_setup,
                 distributed_backend,

tests/distributed/test_symm_mem_allreduce.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import queue
 import random
 import typing
 
@@ -10,26 +11,31 @@ import torch.distributed as dist
 import torch.multiprocessing as mp
 
 import vllm.envs as envs
+from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
 from vllm.distributed import cleanup_dist_env_and_memory
 from vllm.distributed.communication_op import tensor_model_parallel_all_reduce
 from vllm.distributed.device_communicators.cuda_communicator import (
     CudaCommunicator)
-from vllm.distributed.parallel_state import (get_tensor_model_parallel_group,
-                                             get_tp_group,
+from vllm.distributed.parallel_state import (get_tp_group,
                                              init_distributed_environment,
                                              initialize_model_parallel)
+from vllm.engine.arg_utils import EngineArgs
+from vllm.engine.llm_engine import LLMEngine
 from vllm.platforms import current_platform
 from vllm.utils import update_environment_variables
 
 torch.manual_seed(42)
 random.seed(44)
 
-test_size_elements = 4 * 1024 * 1024
+test_size_elements = 1024 * 1024
 
 
-def symm_mem_allreduce_worker(local_rank: int, world_size: int):
+def symm_mem_allreduce_worker(local_rank: int, world_size: int, q: mp.Queue):
     monkeypatch = pytest.MonkeyPatch()
-    with monkeypatch.context() as m:
+    config = VllmConfig(parallel_config=ParallelConfig(
+        tensor_parallel_size=world_size))
+
+    with monkeypatch.context() as m, set_current_vllm_config(config):
         m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
         dtype = torch.bfloat16
         device = torch.device(f"cuda:{local_rank}")
@@ -51,22 +57,26 @@ def symm_mem_allreduce_worker(local_rank: int, world_size: int):
                                         get_tp_group().device_communicator)
         symm_mem_comm = cuda_communicator.symm_mem_comm
         if symm_mem_comm is None or symm_mem_comm.disabled:
-            pytest.skip("SymmMemCommunicator is not available or disabled.")
+            # can't use skip under multiprocessing
+            q.put("SymmMemCommunicator is not available or disabled.")
+            return
 
         inp_direct_symm_mem = torch.randint(1,
                                             23, (test_size_elements, ),
                                             dtype=dtype,
                                             device=device)
         if not symm_mem_comm.should_use_symm_mem(inp_direct_symm_mem):
-            pytest.skip(
+            # can't use skip under multiprocessing
+            q.put(
                 "SymmMemCommunicator isn't used for this world and input size."
             )
+            return
 
         original_inp_direct_symm_mem = inp_direct_symm_mem.clone()
         out_direct_symm_mem = symm_mem_comm.all_reduce(inp_direct_symm_mem)
         assert out_direct_symm_mem is not None
 
-        group = get_tensor_model_parallel_group().device_group
+        group = get_tp_group().device_group
         dist.all_reduce(original_inp_direct_symm_mem, group=group)
         torch.testing.assert_close(out_direct_symm_mem,
                                    original_inp_direct_symm_mem,
@@ -100,9 +110,34 @@ def test_symm_mem_allreduce(monkeypatch: pytest.MonkeyPatch, tp_size,
     world_size = tp_size * pipeline_parallel_size
     if world_size > torch.cuda.device_count():
         pytest.skip("Not enough GPUs to run the test.")
+    q = mp.get_context('spawn').Queue()
+    mp.spawn(symm_mem_allreduce_worker,
+             args=(world_size, q),
+             nprocs=world_size)
+    try:
+        val = q.get(timeout=1)
+    except queue.Empty:
+        val = None
+    finally:
+        cleanup_dist_env_and_memory()
+        if val is not None:
+            pytest.skip(val)
 
-    # Enable SymmMemCommunicator
-    monkeypatch.setenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1")
 
-    mp.spawn(symm_mem_allreduce_worker, args=(world_size, ), nprocs=world_size)
-    cleanup_dist_env_and_memory()
+@pytest.mark.skipif(
+    not current_platform.is_cuda(),
+    reason="SymmMemAllreduce is only available for CUDA platforms.")
+@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"],
+                    reason="Only test on CUDA")
+def test_dp_with_symm_mem_allreduce(monkeypatch: pytest.MonkeyPatch):
+    world_size = 4
+    if world_size > torch.cuda.device_count():
+        pytest.skip("Not enough GPUs to run the test.")
+    # Verify that the DataParallel runs without error
+    engine_args = EngineArgs(model="distilbert/distilgpt2",
+                             enforce_eager=True,
+                             enable_prefix_caching=True,
+                             data_parallel_size=2,
+                             tensor_parallel_size=2,
+                             data_parallel_backend="mp")
+    LLMEngine.from_engine_args(engine_args)

tests/engine/test_arg_utils.py

@@ -50,8 +50,11 @@ def test_is_type(type_hint, type, expected):
 
 @pytest.mark.parametrize(("type_hints", "type", "expected"), [
     ({float, int}, int, True),
+    ({int, tuple}, int, True),
     ({int, tuple[int]}, int, True),
+    ({int, tuple[int, ...]}, int, True),
     ({int, tuple[int]}, float, False),
+    ({int, tuple[int, ...]}, float, False),
     ({str, Literal["x", "y"]}, Literal, True),
 ])
 def test_contains_type(type_hints, type, expected):

tests/entrypoints/openai/test_completion_with_prompt_embeds.py

@@ -15,7 +15,7 @@ from transformers import AutoConfig
 from ...utils import RemoteOpenAIServer
 
 # any model with a chat template should work here
-MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
+MODEL_NAME = "facebook/opt-125m"
 
 CONFIG = AutoConfig.from_pretrained(MODEL_NAME)
 
@@ -27,7 +27,7 @@ def default_server_args() -> list[str]:
         "--dtype",
         "bfloat16",
         "--max-model-len",
-        "8192",
+        "2048",
         "--max-num-seqs",
         "128",
         "--enforce-eager",
@@ -36,6 +36,27 @@ def default_server_args() -> list[str]:
     ]
 
 
+EXAMPLE_PROMPTS = [
+    "Hello, my name is",
+    "What is an LLM?",
+]
+
+
+def _encode_embeds(embeds: torch.Tensor):
+    buffer = io.BytesIO()
+    torch.save(embeds, buffer)
+    return base64.b64encode(buffer.getvalue()).decode('utf-8')
+
+
+@pytest.fixture(scope="module")
+def example_prompt_embeds(hf_runner):
+    """Create example embeddings and return them as base64 encoded string."""
+    with hf_runner(MODEL_NAME) as hf_model:
+        example_embeddings = hf_model.get_prompt_embeddings(EXAMPLE_PROMPTS)
+
+    return [_encode_embeds(item) for item in example_embeddings]
+
+
 @pytest.fixture(scope="module",
                 params=["", "--disable-frontend-multiprocessing"])
 def server_with_prompt_embeds(default_server_args, request):
@@ -52,21 +73,16 @@ async def client_with_prompt_embeds(server_with_prompt_embeds):
         yield async_client
 
 
-def create_dummy_embeds(num_tokens: int = 5) -> str:
-    """Create dummy embeddings and return them as base64 encoded string."""
-    dummy_embeds = torch.randn(num_tokens, CONFIG.hidden_size)
-    buffer = io.BytesIO()
-    torch.save(dummy_embeds, buffer)
-    return base64.b64encode(buffer.getvalue()).decode('utf-8')
-
-
-@pytest.mark.skip("This test is skipped because it is flaky.")
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 async def test_completions_with_prompt_embeds(
-        client_with_prompt_embeds: openai.AsyncOpenAI, model_name: str):
+    example_prompt_embeds,
+    client_with_prompt_embeds: openai.AsyncOpenAI,
+    model_name: str,
+):
+    encoded_embeds, encoded_embeds2 = example_prompt_embeds
+
     # Test case: Single prompt embeds input
-    encoded_embeds = create_dummy_embeds()
     completion = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="",  # Add empty prompt as required parameter
@@ -77,7 +93,6 @@ async def test_completions_with_prompt_embeds(
     assert completion.choices[0].prompt_logprobs is None
 
     # Test case: batch completion with prompt_embeds
-    encoded_embeds2 = create_dummy_embeds()
     completion = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="",  # Add empty prompt as required parameter
@@ -89,7 +104,6 @@ async def test_completions_with_prompt_embeds(
     assert len(completion.choices[1].text) >= 1
 
     # Test case: streaming with prompt_embeds
-    encoded_embeds = create_dummy_embeds()
     single_completion = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="",  # Add empty prompt as required parameter
@@ -117,7 +131,6 @@ async def test_completions_with_prompt_embeds(
     assert "".join(chunks) == single_output
 
     # Test case: batch streaming with prompt_embeds
-    encoded_embeds2 = create_dummy_embeds()
     stream = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="",  # Add empty prompt as required parameter
@@ -139,7 +152,6 @@ async def test_completions_with_prompt_embeds(
     assert len(chunks_stream_embeds[1]) > 0
 
     # Test case: mixed text and prompt_embeds
-    encoded_embeds = create_dummy_embeds()
     completion_mixed = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="This is a prompt",
@@ -184,10 +196,14 @@ async def test_completions_errors_with_prompt_embeds(
 @pytest.mark.parametrize("logprobs_arg", [1, 0])
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 async def test_completions_with_logprobs_and_prompt_embeds(
-        client_with_prompt_embeds: openai.AsyncOpenAI, logprobs_arg: int,
-        model_name: str):
+    example_prompt_embeds,
+    client_with_prompt_embeds: openai.AsyncOpenAI,
+    logprobs_arg: int,
+    model_name: str,
+):
+    encoded_embeds, encoded_embeds2 = example_prompt_embeds
+
     # Test case: Logprobs using prompt_embeds
-    encoded_embeds = create_dummy_embeds()
     completion = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="",  # Add empty prompt as required parameter
@@ -207,7 +223,6 @@ async def test_completions_with_logprobs_and_prompt_embeds(
     assert len(logprobs.tokens) == 5
 
     # Test case: Log probs with batch completion and prompt_embeds
-    encoded_embeds2 = create_dummy_embeds()
     completion = await client_with_prompt_embeds.completions.create(
         model=model_name,
         prompt="",  # Add empty prompt as required parameter
@@ -232,9 +247,12 @@ async def test_completions_with_logprobs_and_prompt_embeds(
 
 @pytest.mark.asyncio
 async def test_prompt_logprobs_raises_error(
-        client_with_prompt_embeds: openai.AsyncOpenAI):
+    example_prompt_embeds,
+    client_with_prompt_embeds: openai.AsyncOpenAI,
+):
+    encoded_embeds, _ = example_prompt_embeds
+
     with pytest.raises(BadRequestError, match="not compatible"):
-        encoded_embeds = create_dummy_embeds()
         await client_with_prompt_embeds.completions.create(
             model=MODEL_NAME,
             prompt="",

tests/entrypoints/openai/test_lora_resolvers.py

@@ -45,6 +45,7 @@ class MockModelConfig:
     logits_processor_pattern: Optional[str] = None
     diff_sampling_param: Optional[dict] = None
     allowed_local_media_path: str = ""
+    allowed_media_domains: Optional[list[str]] = None
     encoder_config = None
     generation_config: str = "auto"
     skip_tokenizer_init: bool = False

tests/entrypoints/openai/test_response_api_mcp_tools.py

@@ -0,0 +1,106 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import pytest_asyncio
+from openai import OpenAI
+
+from ...utils import RemoteOpenAIServer
+
+MODEL_NAME = "openai/gpt-oss-20b"
+
+
+@pytest.fixture(scope="module")
+def monkeypatch_module():
+    from _pytest.monkeypatch import MonkeyPatch
+    mpatch = MonkeyPatch()
+    yield mpatch
+    mpatch.undo()
+
+
+@pytest.fixture(scope="module")
+def mcp_disabled_server(monkeypatch_module: pytest.MonkeyPatch):
+    args = ["--enforce-eager", "--tool-server", "demo"]
+
+    with monkeypatch_module.context() as m:
+        m.setenv("VLLM_ENABLE_RESPONSES_API_STORE", "1")
+        m.setenv("PYTHON_EXECUTION_BACKEND", "dangerously_use_uv")
+        with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+            yield remote_server
+
+
+@pytest.fixture(scope="function")
+def mcp_enabled_server(monkeypatch_module: pytest.MonkeyPatch):
+    args = ["--enforce-eager", "--tool-server", "demo"]
+
+    with monkeypatch_module.context() as m:
+        m.setenv("VLLM_ENABLE_RESPONSES_API_STORE", "1")
+        m.setenv("PYTHON_EXECUTION_BACKEND", "dangerously_use_uv")
+        m.setenv("GPT_OSS_SYSTEM_TOOL_MCP_LABELS",
+                 "code_interpreter,container")
+        with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+            yield remote_server
+
+
+@pytest_asyncio.fixture
+async def mcp_disabled_client(mcp_disabled_server):
+    async with mcp_disabled_server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest_asyncio.fixture
+async def mcp_enabled_client(mcp_enabled_server):
+    async with mcp_enabled_server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.skip(reason="Code interpreter tool is not available in CI yet.")
+async def test_mcp_tool_env_flag_enabled(mcp_enabled_client: OpenAI,
+                                         model_name: str):
+    response = await mcp_enabled_client.responses.create(
+        model=model_name,
+        # TODO: Ideally should be able to set max tool calls
+        # to prevent multi-turn, but it is not currently supported
+        # would speed up the test
+        input=("What's the first 4 digits after the decimal point of "
+               "cube root of `19910212 * 20250910`? "
+               "Show only the digits. The python interpreter is not stateful "
+               "and you must print to see the output."),
+        tools=[{
+            "type": "mcp",
+            "server_label": "code_interpreter",
+            # URL unused for DemoToolServer
+            "server_url": "http://localhost:8888"
+        }],
+    )
+    assert response is not None
+    assert response.status == "completed"
+    assert response.usage.output_tokens_details.tool_output_tokens > 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.skip(reason="Code interpreter tool is not available in CI yet.")
+async def test_mcp_tool_env_flag_disabled(mcp_disabled_client: OpenAI,
+                                          model_name: str):
+    response = await mcp_disabled_client.responses.create(
+        model=model_name,
+        # TODO: Ideally should be able to set max tool calls
+        # to prevent multi-turn, but it is not currently supported
+        # would speed up the test
+        input=("What's the first 4 digits after the decimal point of "
+               "cube root of `19910212 * 20250910`? "
+               "Show only the digits. The python interpreter is not stateful "
+               "and you must print to see the output."),
+        tools=[{
+            "type": "mcp",
+            "server_label": "code_interpreter",
+            # URL unused for DemoToolServer
+            "server_url": "http://localhost:8888"
+        }],
+    )
+    assert response is not None
+    assert response.status == "completed"
+    assert response.usage.output_tokens_details.tool_output_tokens == 0

tests/entrypoints/openai/test_response_api_with_harmony.py

@@ -454,7 +454,13 @@ async def test_web_search(client: OpenAI, model_name: str):
 async def test_code_interpreter(client: OpenAI, model_name: str):
     response = await client.responses.create(
         model=model_name,
-        input="Multiply 64548*15151 using builtin python interpreter.",
+        # TODO: Ideally should be able to set max tool calls
+        # to prevent multi-turn, but it is not currently supported
+        # would speed up the test
+        input=("What's the first 4 digits after the decimal point of "
+               "cube root of `19910212 * 20250910`? "
+               "Show only the digits. The python interpreter is not stateful "
+               "and you must print to see the output."),
         tools=[{
             "type": "code_interpreter",
             "container": {
@@ -464,6 +470,7 @@ async def test_code_interpreter(client: OpenAI, model_name: str):
     )
     assert response is not None
     assert response.status == "completed"
+    assert response.usage.output_tokens_details.tool_output_tokens > 0
 
 
 def get_weather(latitude, longitude):
@@ -516,6 +523,7 @@ async def test_function_calling(client: OpenAI, model_name: str):
         input="What's the weather like in Paris today?",
         tools=tools,
         temperature=0.0,
+        extra_body={"request_id": "test_function_calling_non_resp"},
     )
     assert response is not None
     assert response.status == "completed"

tests/entrypoints/openai/test_serving_chat.py

@@ -68,7 +68,7 @@ def default_server_args(with_tool_parser: bool):
 def gptoss_server(monkeypatch_module: pytest.MonkeyPatch,
                   default_server_args: list[str]):
     with monkeypatch_module.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN_VLLM_V1")
+        m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN")
         with RemoteOpenAIServer(GPT_OSS_MODEL_NAME,
                                 default_server_args) as remote_server:
             yield remote_server
@@ -194,6 +194,7 @@ async def test_gpt_oss_multi_turn_chat(gptoss_client: OpenAI,
     assert tc.function is not None and tc.function.name == "get_current_weather"
     args1 = tc.function.arguments
     assert args1 is not None and len(args1) > 0
+    assert not first_msg.content
 
     messages.append({"role": "assistant", "content": args1})
     messages.append({
@@ -239,6 +240,7 @@ class MockModelConfig:
     logits_processor_pattern = None
     diff_sampling_param: Optional[dict] = None
     allowed_local_media_path: str = ""
+    allowed_media_domains: Optional[list[str]] = None
     encoder_config = None
     generation_config: str = "auto"
     media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)

tests/entrypoints/openai/tool_parsers/test_hermes_tool_parser.py

@@ -5,6 +5,11 @@ import json
 
 import pytest
 
+from vllm.entrypoints.openai.protocol import ChatCompletionRequest
+from vllm.entrypoints.openai.tool_parsers.hermes_tool_parser import (
+    Hermes2ProToolParser)
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+
 from ....utils import RemoteOpenAIServer
 
 MODEL_NAME = "meta-llama/Llama-3.2-1B-Instruct"
@@ -37,7 +42,7 @@ TOOLS = [{
                 },
                 "unit": {
                     "type": "string",
-                    "enum": ["celsius", "fahrenheit"]
+                    "enum": ["celsius", "fahrenheit"],
                 },
             },
             "required": ["location"],
@@ -75,7 +80,7 @@ PRODUCT_MESSAGES = [{
     "user",
     "content":
     "Hi! Do you have any detailed information about the product id "
-    "7355608 and inserted true?"
+    "7355608 and inserted true?",
 }]
 
 
@@ -144,8 +149,8 @@ async def test_streaming_tool_call():
                 if tool_chunk.function.name:
                     tool_call_chunks[index]["name"] += tool_chunk.function.name
                 if tool_chunk.function.arguments:
-                    tool_call_chunks[index][
-                        "arguments"] += tool_chunk.function.arguments
+                    tool_call_chunks[index]["arguments"] += (
+                        tool_chunk.function.arguments)
 
         assert len(tool_call_chunks) == 1
         reconstructed_tool_call = tool_call_chunks[0]
@@ -234,8 +239,8 @@ async def test_streaming_product_tool_call():
                 if tool_chunk.function.name:
                     tool_call_chunks[index]["name"] += tool_chunk.function.name
                 if tool_chunk.function.arguments:
-                    tool_call_chunks[index][
-                        "arguments"] += tool_chunk.function.arguments
+                    tool_call_chunks[index]["arguments"] += (
+                        tool_chunk.function.arguments)
 
         assert len(tool_call_chunks) == 1
         reconstructed_tool_call = tool_call_chunks[0]
@@ -258,3 +263,195 @@ async def test_streaming_product_tool_call():
         print("\n[Streaming Product Test Passed]")
         print(f"Reconstructed Tool Call: {reconstructed_tool_call['name']}")
         print(f"Reconstructed Arguments: {arguments}")
+
+
+@pytest.fixture
+def qwen_tokenizer() -> AnyTokenizer:
+    from vllm.transformers_utils.tokenizer import get_tokenizer
+
+    return get_tokenizer("Qwen/Qwen3-32B")
+
+
+@pytest.fixture
+def hermes_parser(qwen_tokenizer: AnyTokenizer) -> Hermes2ProToolParser:
+    return Hermes2ProToolParser(qwen_tokenizer)
+
+
+@pytest.fixture
+def any_chat_request() -> ChatCompletionRequest:
+    return ChatCompletionRequest(
+        seed=42,
+        model="Qwen/Qwen3-32B",
+        messages=[],
+    )
+
+
+def test_hermes_parser_streaming_just_forward_text(
+    qwen_tokenizer: AnyTokenizer,
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    text = (
+        """This is some prior text that has nothing to do with tool calling."""
+    )
+    tokens = qwen_tokenizer.encode(text)
+    previous_text = ""
+    delta_messages = []
+    for token in tokens:
+        delta_text = qwen_tokenizer.decode([token])
+        current_text = previous_text + delta_text
+        delta = hermes_parser.extract_tool_calls_streaming(
+            previous_text=previous_text,
+            current_text=current_text,
+            delta_text=delta_text,
+            previous_token_ids=[],
+            current_token_ids=[],
+            delta_token_ids=[],
+            request=any_chat_request,
+        )
+        previous_text = current_text
+        delta_messages.append(delta)
+
+    for delta in delta_messages:
+        assert delta is not None
+        assert not delta.tool_calls
+
+    print(delta_messages)
+    assert "".join([delta.content for delta in delta_messages]) == text
+
+
+def test_hermes_parser_streaming_failure_case_bug_19056(
+    qwen_tokenizer: AnyTokenizer,
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    text = """<tool_call>
+{"name": "final_answer", "arguments": {"trigger": true}}
+</tool_call>"""
+    tokens = qwen_tokenizer.encode(text)
+    previous_text = ""
+    delta_messages = []
+    for token in tokens:
+        text = qwen_tokenizer.decode([token])
+        current_text = previous_text + text
+        delta = hermes_parser.extract_tool_calls_streaming(
+            previous_text=previous_text,
+            current_text=current_text,
+            delta_text=text,
+            previous_token_ids=[],
+            current_token_ids=[],
+            delta_token_ids=[],
+            request=any_chat_request,
+        )
+        previous_text = current_text
+        if delta is not None:
+            delta_messages.append(delta)
+
+    assert delta_messages[0].tool_calls[0].function.name == "final_answer"
+    tool_call_args = "".join(delta.tool_calls[0].function.arguments or ""
+                             for delta in delta_messages)
+    assert tool_call_args == '{"trigger": true}'
+
+
+def test_hermes_parser_streaming(
+    qwen_tokenizer: AnyTokenizer,
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    text = '<tool_call>\
+{"name": "get_current_temperature",\
+"arguments": {"location":\
+"San Francisco, California, United States", "unit": "celsius"}}\
+</tool_call>'
+
+    tokens = qwen_tokenizer.encode(text)
+    previous_text = ""
+    delta_messages = []
+    for token in tokens:
+        text = qwen_tokenizer.decode([token])
+        current_text = previous_text + text
+        delta = hermes_parser.extract_tool_calls_streaming(
+            previous_text=previous_text,
+            current_text=current_text,
+            delta_text=text,
+            previous_token_ids=[],
+            current_token_ids=[],
+            delta_token_ids=[],
+            request=any_chat_request,
+        )
+        previous_text = current_text
+        if delta is not None:
+            delta_messages.append(delta)
+    print(delta_messages)
+    assert (delta_messages[0].tool_calls[0].function.name ==
+            "get_current_temperature")
+    tool_call_args = "".join(delta.tool_calls[0].function.arguments or ""
+                             for delta in delta_messages)
+    assert tool_call_args == (
+        '{"location":"San Francisco, California, United States", '
+        '"unit": "celsius"}')
+
+
+def test_hermes_parser_non_streaming_no_tool_call(
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    text = """This is not a tool call."""
+    tool_call = hermes_parser.extract_tool_calls(
+        model_output=text,
+        request=any_chat_request,
+    )
+
+    assert tool_call is not None
+    assert not tool_call.tools_called
+
+
+def test_hermes_parser_non_streaming_tool_call_between_tags(
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    text = """<tool_call>
+{"name": "final_answer", "arguments": {"trigger": true}}
+</tool_call>"""
+    tool_call = hermes_parser.extract_tool_calls(
+        model_output=text,
+        request=any_chat_request,
+    )
+
+    assert tool_call is not None
+    assert tool_call.tools_called
+    assert tool_call.tool_calls[0].function.name == "final_answer"
+    assert tool_call.tool_calls[0].function.arguments == '{"trigger": true}'
+
+
+def test_hermes_parser_non_streaming_tool_call_until_eos(
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    text = """<tool_call>
+{"name": "final_answer", "arguments": {"trigger": true}}"""
+    tool_call = hermes_parser.extract_tool_calls(
+        model_output=text,
+        request=any_chat_request,
+    )
+
+    assert tool_call is not None
+    assert tool_call.tools_called
+    assert tool_call.tool_calls[0].function.name == "final_answer"
+    assert tool_call.tool_calls[0].function.arguments == '{"trigger": true}'
+
+
+def test_hermes_parser_non_streaming_tool_call_invalid_json(
+    hermes_parser: Hermes2ProToolParser,
+    any_chat_request: ChatCompletionRequest,
+) -> None:
+    # Missing closing brace to trigger exception
+    text = """<tool_call>
+{"name": "final_answer", "arguments": {"trigger": true}"""
+    tool_call = hermes_parser.extract_tool_calls(
+        model_output=text,
+        request=any_chat_request,
+    )
+
+    assert tool_call is not None
+    assert not tool_call.tools_called

tests/entrypoints/test_chat_utils.py

@@ -19,6 +19,7 @@ from vllm.entrypoints.chat_utils import (_try_extract_ast, load_chat_template,
                                          parse_chat_messages,
                                          parse_chat_messages_futures,
                                          resolve_chat_template_content_format,
+                                         resolve_chat_template_kwargs,
                                          resolve_hf_chat_template)
 from vllm.multimodal import MultiModalDataDict, MultiModalUUIDDict
 from vllm.multimodal.utils import (encode_audio_base64, encode_image_base64,
@@ -37,6 +38,7 @@ QWEN2AUDIO_MODEL_ID = "Qwen/Qwen2-Audio-7B-Instruct"
 QWEN2VL_MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
 QWEN25VL_MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
 QWEN25OMNI_MODEL_ID = "Qwen/Qwen2.5-Omni-7B"
+QWEN3_MODEL_ID = "Qwen/Qwen3-8B"
 LLAMA_GUARD_MODEL_ID = "meta-llama/Llama-Guard-3-1B"
 HERMES_MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
 MISTRAL_MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
@@ -2255,6 +2257,89 @@ def test_resolve_hf_chat_template(sample_json_schema, model, use_tools):
     assert isinstance(chat_template, str)
 
 
+@pytest.mark.parametrize(
+    "model, expected_kwargs",
+    [
+        (
+            QWEN2VL_MODEL_ID,
+            {
+                "add_vision_id", "add_generation_prompt",
+                "continue_final_message", "tools"
+            },
+        ),
+        (
+            QWEN3_MODEL_ID,
+            {
+                "enable_thinking", "add_generation_prompt",
+                "continue_final_message", "tools"
+            },
+        ),
+    ],
+)
+def test_resolve_hf_chat_template_kwargs(sample_json_schema, model,
+                                         expected_kwargs):
+    """checks that chat_template is a dict type for HF models."""
+    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
+    model_info.check_available_online(on_fail="skip")
+
+    tools = ([{
+        "type": "function",
+        "function": {
+            "name": "dummy_function_name",
+            "description": "This is a dummy function",
+            "parameters": sample_json_schema,
+        },
+    }])
+
+    chat_template_kwargs = {
+        # both unused
+        "unsed_kwargs_1": 123,
+        "unsed_kwargs_2": "abc",
+        # should not appear
+        "chat_template": "{% Hello world! %}",
+        # used by tokenizer
+        "continue_final_message": True,
+        "tools": tools,
+        # both used by Qwen2-VL and Qwen3
+        "add_generation_prompt": True,
+        # only used by Qwen2-VL
+        "add_vision_id": True,
+        # only used by Qwen3
+        "enable_thinking": True,
+    }
+
+    model_config = ModelConfig(
+        model,
+        tokenizer=model_info.tokenizer or model,
+        tokenizer_mode=model_info.tokenizer_mode,
+        revision=model_info.revision,
+        trust_remote_code=model_info.trust_remote_code,
+        hf_overrides=model_info.hf_overrides,
+        skip_tokenizer_init=model_info.skip_tokenizer_init,
+        enforce_eager=model_info.enforce_eager,
+        dtype=model_info.dtype)
+
+    # Build the tokenizer
+    tokenizer = get_tokenizer(
+        model,
+        trust_remote_code=model_config.trust_remote_code,
+    )
+
+    # Test detecting the tokenizer's chat_template
+    chat_template = resolve_hf_chat_template(
+        tokenizer,
+        chat_template=None,
+        tools=tools,
+        model_config=model_config,
+    )
+    resolved_chat_template_kwargs = resolve_chat_template_kwargs(
+        tokenizer,
+        chat_template=chat_template,
+        chat_template_kwargs=chat_template_kwargs,
+    )
+    assert set(resolved_chat_template_kwargs.keys()) == expected_kwargs
+
+
 # NOTE: Qwen2-Audio default chat template is specially defined inside
 # processor class instead of using `tokenizer_config.json`
 # yapf: disable

tests/kernels/attention/test_attention_selector.py

@@ -31,7 +31,7 @@ DEVICE_MLA_BACKENDS = {
 }
 
 DEVICE_REGULAR_ATTN_BACKENDS = {
-    "cuda": ["XFORMERS", "FLASHINFER"],
+    "cuda": ["XFORMERS", "FLASHINFER", "FLASH_ATTN"],
     "hip": ["ROCM_FLASH"],
     "cpu": ["TORCH_SDPA"],
 }
@@ -67,7 +67,6 @@ def generate_params():
     return params
 
 
-@pytest.mark.skip(reason="Skipped for now. Should be revisited.")
 @pytest.mark.parametrize("device, name, use_mla, block_size",
                          generate_params())
 def test_env(
@@ -86,9 +85,8 @@ def test_env(
         if device == "cpu":
             with patch("vllm.attention.selector.current_platform",
                        CpuPlatform()):
-                backend = get_attn_backend(16, torch.float16, None, block_size,
-                                           False)
-            assert backend.get_name() == "TORCH_SDPA_VLLM_V1"
+                backend = get_attn_backend(16, torch.float16, None, block_size)
+            assert backend.get_name() == "TORCH_SDPA"
 
         elif device == "hip":
             with patch("vllm.attention.selector.current_platform",
@@ -107,7 +105,6 @@ def test_env(
                                              torch.float16,
                                              None,
                                              block_size,
-                                             False,
                                              use_mla=use_mla)
                         assert f"The selected backend, {name}" in str(
                             exc_info.value)
@@ -118,7 +115,6 @@ def test_env(
                                              torch.float16,
                                              None,
                                              block_size,
-                                             False,
                                              use_mla=use_mla)
                         assert f"The selected backend, {name}" in str(
                             exc_info.value)
@@ -128,18 +124,16 @@ def test_env(
                                                    torch.float16,
                                                    None,
                                                    block_size,
-                                                   False,
                                                    use_mla=use_mla)
-                        expected = f"{name}_VLLM_V1"
+                        expected = name
                         assert backend.get_name() == expected
                 else:
                     backend = get_attn_backend(16,
                                                torch.float16,
                                                None,
                                                block_size,
-                                               False,
                                                use_mla=use_mla)
-                    expected = "TRITON_ATTN_VLLM_V1"
+                    expected = "TRITON_ATTN"
                     assert backend.get_name() == expected
 
         elif device == "cuda":
@@ -165,9 +159,8 @@ def test_env(
                                                        torch.float16,
                                                        None,
                                                        block_size,
-                                                       False,
                                                        use_mla=use_mla)
-                            expected = "CUTLASS_MLA_VLLM_V1"
+                            expected = "CUTLASS_MLA"
                             assert backend.get_name() == expected
                     elif name == "FLASHINFER_MLA":
                         if block_size not in [32, 64]:
@@ -180,7 +173,6 @@ def test_env(
                                                        torch.float16,
                                                        None,
                                                        block_size,
-                                                       False,
                                                        use_mla=use_mla)
                             expected = "FLASHINFER_MLA"
                             assert backend.get_name() == expected
@@ -189,7 +181,7 @@ def test_env(
                             # FlashMLA only supports block_size == 64
                             pytest.skip("FlashMLA only supports block_size 64")
                         else:
-                            from vllm.attention.backends.flashmla import (
+                            from vllm.v1.attention.backends.mla.flashmla import (  # noqa: E501
                                 is_flashmla_supported)
                             is_supported, _ = is_flashmla_supported()
                             if not is_supported:
@@ -200,16 +192,14 @@ def test_env(
                                                            torch.float16,
                                                            None,
                                                            block_size,
-                                                           False,
                                                            use_mla=use_mla)
-                                expected = f"{name}_VLLM_V1"
+                                expected = name
                                 assert backend.get_name() == expected
                     elif name == "FLASH_ATTN_MLA":
                         backend = get_attn_backend(16,
                                                    torch.float16,
                                                    None,
                                                    block_size,
-                                                   False,
                                                    use_mla=use_mla)
                         expected = "FLASH_ATTN_MLA"
                         assert backend.get_name() == expected
@@ -219,38 +209,33 @@ def test_env(
                                                    torch.float16,
                                                    None,
                                                    block_size,
-                                                   False,
                                                    use_mla=use_mla)
-                        expected = "TRITON_MLA_VLLM_V1"
+                        expected = "TRITON_MLA"
                         assert backend.get_name() == expected
                 elif name == "FLASHINFER":
                     backend = get_attn_backend(16,
                                                torch.float16,
                                                None,
                                                block_size,
-                                               False,
                                                use_mla=use_mla)
-                    expected = "FLASHINFER_VLLM_V1"
+                    expected = "FLASHINFER"
                     assert backend.get_name() == expected
-                else:
+                elif name == "XFORMERS":
                     backend = get_attn_backend(32,
                                                torch.float16,
                                                None,
                                                block_size,
-                                               False,
                                                use_mla=use_mla)
-                    expected = "FLASH_ATTN_VLLM_V1"
+                    expected = "XFORMERS"
                     assert backend.get_name() == expected
-
-                    backend = get_attn_backend(16,
+                elif name == "FLASH_ATTN":
+                    backend = get_attn_backend(32,
                                                torch.float16,
                                                None,
                                                block_size,
-                                               False,
                                                use_mla=use_mla)
-                    assert backend.get_name() == "FLEX_ATTENTION", (
-                        "Should fallback to FlexAttention if head size is "
-                        "not supported by FlashAttention")
+                    expected = "FLASH_ATTN"
+                    assert backend.get_name() == expected
 
 
 @pytest.mark.parametrize("device", ["cpu", "cuda"])
@@ -265,13 +250,13 @@ def test_fp32_fallback(
         if device == "cpu":
             with patch("vllm.attention.selector.current_platform",
                        CpuPlatform()):
-                backend = get_attn_backend(16, torch.float32, None, 16, False)
-            assert backend.get_name() == "TORCH_SDPA_VLLM_V1"
+                backend = get_attn_backend(16, torch.float32, None, 16)
+            assert backend.get_name() == "TORCH_SDPA"
 
         elif device == "cuda":
             with patch("vllm.attention.selector.current_platform",
                        CudaPlatform()):
-                backend = get_attn_backend(16, torch.float32, None, 16, False)
+                backend = get_attn_backend(16, torch.float32, None, 16)
             assert backend.get_name() == "FLEX_ATTENTION"
 
 
@@ -280,6 +265,9 @@ def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
     # TODO: When testing for v1, pipe in `use_v1` as an argument to
     # get_attn_backend
 
+    pytest.skip("Skipping as current backend selector does not " \
+                "handle fallbacks when a backend is set via env var.")
+
     with monkeypatch.context() as m:
         m.setenv(STR_BACKEND_ENV_VAR, STR_FLASH_ATTN_VAL)
 
@@ -287,29 +275,29 @@ def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
         monkeypatch.setattr(torch.cuda,
                             "get_device_capability",
                             lambda _=None: (7, 5))
-        backend = get_attn_backend(16, torch.float16, None, 16, False)
+        backend = get_attn_backend(16, torch.float16, None, 16)
         assert backend.get_name() != STR_FLASH_ATTN_VAL
 
         # Reset the monkeypatch for subsequent tests
         monkeypatch.undo()
 
         # Unsupported data type
-        backend = get_attn_backend(16, torch.float8_e4m3fn, None, 16, False)
+        backend = get_attn_backend(16, torch.float8_e4m3fn, None, 16)
         assert backend.get_name() != STR_FLASH_ATTN_VAL
 
         # Unsupported kv cache data type
-        backend = get_attn_backend(16, torch.float16, "fp8", 16, False)
+        backend = get_attn_backend(16, torch.float16, "fp8", 16)
         assert backend.get_name() != STR_FLASH_ATTN_VAL
 
         # Unsupported block size
-        backend = get_attn_backend(16, torch.float16, None, 8, False)
+        backend = get_attn_backend(16, torch.float16, None, 8)
         assert backend.get_name() != STR_FLASH_ATTN_VAL
 
         # flash-attn is not installed
         import sys
         original_module = sys.modules.get('vllm_flash_attn')
         monkeypatch.setitem(sys.modules, 'vllm_flash_attn', None)
-        backend = get_attn_backend(16, torch.float16, None, 16, False)
+        backend = get_attn_backend(16, torch.float16, None, 16)
         assert backend.get_name() != STR_FLASH_ATTN_VAL
 
         # Restore the original module if it existed
@@ -320,11 +308,7 @@ def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
             monkeypatch.delitem(sys.modules, 'vllm_flash_attn', raising=False)
 
         # Unsupported head size
-        backend = get_attn_backend(17, torch.float16, None, 16, False)
-        assert backend.get_name() != STR_FLASH_ATTN_VAL
-
-        # Attention-free models should bypass env and use PlaceholderAttention
-        backend = get_attn_backend(16, torch.float16, None, 16, True)
+        backend = get_attn_backend(17, torch.float16, None, 16)
         assert backend.get_name() != STR_FLASH_ATTN_VAL
 
 
@@ -337,5 +321,5 @@ def test_invalid_env(monkeypatch: pytest.MonkeyPatch):
 
         # Should raise ValueError for invalid backend
         with pytest.raises(ValueError) as exc_info:
-            get_attn_backend(32, torch.float16, None, 16, False)
+            get_attn_backend(32, torch.float16, None, 16)
         assert "Invalid value 'INVALID'" in str(exc_info.value)

tests/kernels/attention/test_cache.py

@@ -39,6 +39,8 @@ CUDA_DEVICES = [
 # We assume fp8 is always enabled for testing.
 KV_CACHE_DTYPE = ["auto", "fp8"]
 
+RESHAPE_FLASH_IMPLEMENTATIONS = ["cuda", "triton"]
+
 
 @pytest.mark.parametrize("num_mappings", NUM_MAPPINGS)
 @pytest.mark.parametrize("num_layers", NUM_LAYERS)
@@ -223,6 +225,7 @@ def test_reshape_and_cache(
 @pytest.mark.parametrize("device", CUDA_DEVICES)
 @pytest.mark.parametrize("kv_cache_dtype", KV_CACHE_DTYPE)
 @pytest.mark.parametrize("kv_cache_layout", CACHE_LAYOUTS)
+@pytest.mark.parametrize("implementation", RESHAPE_FLASH_IMPLEMENTATIONS)
 @torch.inference_mode()
 def test_reshape_and_cache_flash(
     kv_cache_factory_flashinfer,
@@ -236,9 +239,13 @@ def test_reshape_and_cache_flash(
     device: str,
     kv_cache_dtype: str,
     kv_cache_layout: str,
+    implementation: str,
 ) -> None:
     current_platform.seed_everything(seed)
     torch.set_default_device(device)
+    assert implementation in ["cuda", "triton"]
+    if implementation == "triton" and kv_cache_layout == "HND":
+        pytest.skip("Triton implementation only supports NHD layout.")
 
     # fp8 conversion requires continugous memory buffer. Reduce the number of
     # blocks and tokens to consume less memory.
@@ -298,12 +305,20 @@ def test_reshape_and_cache_flash(
         cloned_key_cache = key_cache_compact.clone()
         cloned_value_cache = value_cache_compact.clone()
     # Call the reshape_and_cache kernel.
-    opcheck(torch.ops._C_cache_ops.reshape_and_cache_flash,
-            (key, value, key_cache, value_cache, slot_mapping, kv_cache_dtype,
-             k_scale, v_scale),
-            cond=(head_size == HEAD_SIZES[0]))
-    ops.reshape_and_cache_flash(key, value, key_cache, value_cache,
-                                slot_mapping, kv_cache_dtype, k_scale, v_scale)
+    if implementation == "cuda":
+        opcheck(torch.ops._C_cache_ops.reshape_and_cache_flash,
+                (key, value, key_cache, value_cache, slot_mapping,
+                 kv_cache_dtype, k_scale, v_scale),
+                cond=(head_size == HEAD_SIZES[0]))
+        ops.reshape_and_cache_flash(key, value, key_cache, value_cache,
+                                    slot_mapping, kv_cache_dtype, k_scale,
+                                    v_scale)
+    elif implementation == "triton":
+        from vllm.attention.ops.triton_reshape_and_cache_flash import (
+            triton_reshape_and_cache_flash)
+        triton_reshape_and_cache_flash(key, value, key_cache, value_cache,
+                                       slot_mapping, kv_cache_dtype, k_scale,
+                                       v_scale)
     key_cache_compact = permute_and_compact(key_cache)
     value_cache_compact = permute_and_compact(value_cache)
 

tests/kernels/attention/test_rocm_attention_selector.py

@@ -28,7 +28,7 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
         # Test standard ROCm attention
         backend = get_attn_backend(16, torch.float16, torch.float16, 16, False)
         assert (backend.get_name() == "ROCM_FLASH"
-                or backend.get_name() == "TRITON_ATTN_VLLM_V1")
+                or backend.get_name() == "TRITON_ATTN")
 
         # MLA test for deepseek related
 
@@ -40,8 +40,7 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
                                    16,
                                    False,
                                    use_mla=True)
-        assert (backend.get_name() == "TRITON_MLA"
-                or backend.get_name() == "TRITON_MLA_VLLM_V1")
+        assert backend.get_name() == "TRITON_MLA"
 
         # If attention backend is None
         # If use_mla is true
@@ -53,8 +52,7 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
                                    16,
                                    False,
                                    use_mla=True)
-        assert (backend.get_name() == "TRITON_MLA"
-                or backend.get_name() == "TRITON_MLA_VLLM_V1")
+        assert backend.get_name() == "TRITON_MLA"
 
         # change the attention backend to AITER MLA
         m.setenv(STR_BACKEND_ENV_VAR, "ROCM_AITER_MLA")
@@ -64,8 +62,7 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
                                    1,
                                    False,
                                    use_mla=True)
-        assert (backend.get_name() == "ROCM_AITER_MLA"
-                or backend.get_name() == "ROCM_AITER_MLA_VLLM_V1")
+        assert backend.get_name() == "ROCM_AITER_MLA"
 
         # If attention backend is None
         # If use_mla is true
@@ -79,5 +76,4 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
                                    1,
                                    False,
                                    use_mla=True)
-        assert (backend.get_name() == "ROCM_AITER_MLA"
-                or backend.get_name() == "ROCM_AITER_MLA_VLLM_V1")
+        assert backend.get_name() == "ROCM_AITER_MLA"

tests/kernels/attention/test_triton_decode_attention.py

@@ -46,6 +46,8 @@ def test_decode_attention(B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE):
     # o will have the same shape as q
     o = torch.zeros(B, H_Q, D_V, dtype=dtype, device="cuda")
 
+    lse = torch.zeros(B, H_Q, dtype=dtype, device="cuda")
+
     b_seq_len = torch.full((B, ), seq_len, device="cuda")
 
     attn_logits = torch.empty(
@@ -60,6 +62,7 @@ def test_decode_attention(B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE):
         k_buffer,
         v_buffer,
         o,
+        lse,
         req_to_token,
         b_seq_len,
         attn_logits,
@@ -72,12 +75,14 @@ def test_decode_attention(B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE):
     v_buffer = v_buffer.view(CACHE_SIZE // PAGE_SIZE, PAGE_SIZE, H_KV, D_V)
 
     o1 = torch.zeros_like(o)
+    lse1 = torch.zeros_like(lse)
 
     decode_attention_fwd(
         q,
         k_buffer,
         v_buffer,
         o1,
+        lse1,
         req_to_page,
         b_seq_len,
         attn_logits,

tests/kernels/core/test_pos_encoding.py

@@ -60,7 +60,7 @@ TENSORS_SHAPES_FN = [
 @torch.inference_mode()
 def test_rotary_embedding(
     is_neox_style: bool,
-    tensor_shape_fn: Callable[[int, int, int, int], tuple[int]],
+    tensor_shape_fn: Callable[[int, int, int, int], tuple[int, ...]],
     batch_size: int,
     seq_len: int,
     num_heads: int,

tests/kernels/mamba/test_mamba_ssm_ssd.py

@@ -7,7 +7,7 @@ import torch.nn.functional as F
 from einops import rearrange, repeat
 
 from vllm.model_executor.layers.mamba.ops.ssd_combined import (
-    mamba_chunk_scan_combined)
+    mamba_chunk_scan_combined_varlen)
 from vllm.platforms import current_platform
 from vllm.v1.attention.backends.mamba2_attn import (
     _query_start_loc_to_chunk_indices_offsets)
@@ -185,9 +185,14 @@ def generate_continuous_batched_examples(example_lens_by_batch,
             IND_S = [x % full_length for x in IND_E]
         IND_E = [end_boundary(x + y) for x, y in zip(IND_S, spec)]
 
+        # varlen has implicit batch=1
+        dt2 = dt2.squeeze(0)
+        X2 = X2.squeeze(0)
+        B2 = B2.squeeze(0)
+        C2 = C2.squeeze(0)
         yield ([Y_min[s, IND_S[s]:IND_E[s]]
                 for s in range(num_examples)] if return_naive_ref else None,
-               cu_seqlens, seq_idx.unsqueeze(0), (A, dt2, X2, B2, C2))
+               cu_seqlens, seq_idx, (A, dt2, X2, B2, C2))
 
 
 @pytest.mark.parametrize("itype",
@@ -198,7 +203,7 @@ def generate_continuous_batched_examples(example_lens_by_batch,
 def test_mamba_chunk_scan_single_example(d_head, n_heads, seq_len_chunk_size,
                                          itype):
 
-    # this tests the kernels on a single example (no batching)
+    # this tests the kernels on a single example (bs=1)
 
     # TODO: the bfloat16 case requires higher thresholds. To be investigated
 
@@ -219,23 +224,40 @@ def test_mamba_chunk_scan_single_example(d_head, n_heads, seq_len_chunk_size,
 
     Y_min, final_state_min = ssd_minimal_discrete(X * dt.unsqueeze(-1), A * dt,
                                                   B, C, chunk_size)
+
+    cu_seqlens = torch.tensor((0, seqlen), device='cuda').cumsum(dim=0)
+    seq_idx = torch.zeros(seqlen, dtype=torch.int32, device=cu_seqlens.device)
+
+    chunk_indices, chunk_offsets = \
+            _query_start_loc_to_chunk_indices_offsets(
+                cu_seqlens, chunk_size, cu_seqlens[-1])
+
+    # varlen has implicit batch=1
+    X = X.squeeze(0)
+    dt = dt.squeeze(0)
+    A = A.squeeze(0)
+    B = B.squeeze(0)
+    C = C.squeeze(0)
     Y = torch.empty_like(X)
-    final_state = mamba_chunk_scan_combined(X,
-                                            dt,
-                                            A,
-                                            B,
-                                            C,
-                                            chunk_size,
-                                            D=None,
-                                            return_final_states=True,
-                                            out=Y)
+    final_state = mamba_chunk_scan_combined_varlen(X,
+                                                   dt,
+                                                   A,
+                                                   B,
+                                                   C,
+                                                   chunk_size,
+                                                   D=None,
+                                                   cu_seqlens=cu_seqlens,
+                                                   seq_idx=seq_idx,
+                                                   chunk_indices=chunk_indices,
+                                                   chunk_offsets=chunk_offsets,
+                                                   out=Y)
 
     # just test the last in sequence
-    torch.testing.assert_close(Y[:, -1], Y_min[:, -1], atol=atol, rtol=rtol)
+    torch.testing.assert_close(Y[-1], Y_min[0, -1], atol=atol, rtol=rtol)
 
     # just test the last head
     # NOTE, in the kernel we always cast states to fp32
-    torch.testing.assert_close(final_state[:, -1],
+    torch.testing.assert_close(final_state[:, -1].to(torch.float32),
                                final_state_min[:, -1].to(torch.float32),
                                atol=atol,
                                rtol=rtol)
@@ -300,7 +322,7 @@ def test_mamba_chunk_scan_cont_batch(d_head, n_heads, seq_len_chunk_size_cases,
                 cu_seqlens, chunk_size, cu_seqlens[-1])
 
         Y = torch.empty_like(X)
-        new_states = mamba_chunk_scan_combined(
+        new_states = mamba_chunk_scan_combined_varlen(
             X,
             dt,
             A,
@@ -312,7 +334,6 @@ def test_mamba_chunk_scan_cont_batch(d_head, n_heads, seq_len_chunk_size_cases,
             seq_idx=seq_idx,
             chunk_indices=chunk_indices,
             chunk_offsets=chunk_offsets,
-            return_varlen_states=True,
             initial_states=states,
             out=Y,
         )
@@ -321,7 +342,7 @@ def test_mamba_chunk_scan_cont_batch(d_head, n_heads, seq_len_chunk_size_cases,
         for i in range(num_examples):
 
             # just test one dim and dstate
-            Y_eg = Y[0, cu_seqlens[i]:cu_seqlens[i + 1], 0, 0]
+            Y_eg = Y[cu_seqlens[i]:cu_seqlens[i + 1], 0, 0]
             Y_min_eg = Y_min[i][:, 0, 0]
             torch.testing.assert_close(Y_eg, Y_min_eg, atol=atol, rtol=rtol)
 
@@ -386,7 +407,7 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
             _query_start_loc_to_chunk_indices_offsets(
                 cu_seqlens, chunk_size, cu_seqlens[-1])
     Y_ref = torch.empty_like(X)
-    state_ref = mamba_chunk_scan_combined(
+    state_ref = mamba_chunk_scan_combined_varlen(
         X,
         dt,
         A,
@@ -398,7 +419,6 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
         seq_idx=seq_idx,
         chunk_indices=chunk_indices,
         chunk_offsets=chunk_offsets,
-        return_varlen_states=True,
         initial_states=None,
         out=Y_ref,
     )
@@ -414,27 +434,27 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
     chunked_seq_idx = torch.repeat_interleave(
         torch.arange(len(chunked_seqlens), device=device),
         chunked_seqlens,
-        output_size=chunked_cu_seqlens[-1]).unsqueeze(0).to(torch.int32)
+        output_size=chunked_cu_seqlens[-1]).to(torch.int32)
     chunked_input_seq_len = chunked_cu_seqlens[-1]
-    X_chunked = torch.zeros_like(X)[:, :chunked_input_seq_len, ...]
-    dt_chunked = torch.zeros_like(dt)[:, :chunked_input_seq_len, ...]
-    B_chunked = torch.zeros_like(B)[:, :chunked_input_seq_len, ...]
-    C_chunked = torch.zeros_like(C)[:, :chunked_input_seq_len, ...]
+    X_chunked = torch.zeros_like(X)[:chunked_input_seq_len, ...]
+    dt_chunked = torch.zeros_like(dt)[:chunked_input_seq_len, ...]
+    B_chunked = torch.zeros_like(B)[:chunked_input_seq_len, ...]
+    C_chunked = torch.zeros_like(C)[:chunked_input_seq_len, ...]
     for i in range(num_sequences):
         # fmt: off
-        chunk_f = lambda x, i: x[:, cu_seqlens[i]:cu_seqlens[i] + chunked_seqlens[i], ...]  # noqa: E501
+        chunk_f = lambda x, i: x[cu_seqlens[i]:cu_seqlens[i] + chunked_seqlens[i], ...]  # noqa: E501
 
-        X_chunked[:, chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(X, i)  # noqa: E501
-        dt_chunked[:, chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(dt, i)  # noqa: E501
-        B_chunked[:, chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(B, i)  # noqa: E501
-        C_chunked[:, chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(C, i)  # noqa: E501
+        X_chunked[chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(X, i)  # noqa: E501
+        dt_chunked[chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(dt, i)  # noqa: E501
+        B_chunked[chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(B, i)  # noqa: E501
+        C_chunked[chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1], ...] = chunk_f(C, i)  # noqa: E501
         # fmt: on
 
     chunk_indices, chunk_offsets = \
             _query_start_loc_to_chunk_indices_offsets(
                 chunked_cu_seqlens, chunk_size, chunked_cu_seqlens[-1])
     Y_partial = torch.empty_like(X_chunked)
-    partial_state = mamba_chunk_scan_combined(
+    partial_state = mamba_chunk_scan_combined_varlen(
         X_chunked,
         dt_chunked,
         A,
@@ -446,7 +466,6 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
         seq_idx=chunked_seq_idx,
         chunk_indices=chunk_indices,
         chunk_offsets=chunk_offsets,
-        return_varlen_states=True,
         initial_states=None,
         out=Y_partial,
     )
@@ -461,29 +480,28 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
     remaining_chunked_seq_idx = torch.repeat_interleave(
         torch.arange(len(remaining_chunked_seqlens), device=device),
         remaining_chunked_seqlens,
-        output_size=remaining_chunked_cu_seqlens[-1]).unsqueeze(0).to(
-            torch.int32)
+        output_size=remaining_chunked_cu_seqlens[-1]).to(torch.int32)
     remaining_chunked_input_seq_len = remaining_chunked_cu_seqlens[-1]
     # fmt: off
-    remaining_X_chunked = torch.zeros_like(X)[:, :remaining_chunked_input_seq_len, ...]  # noqa: E501
-    remaining_dt_chunked = torch.zeros_like(dt)[:, :remaining_chunked_input_seq_len, ...]  # noqa: E501
-    remaining_B_chunked = torch.zeros_like(B)[:, :remaining_chunked_input_seq_len, ...]  # noqa: E501
-    remaining_C_chunked = torch.zeros_like(C)[:, :remaining_chunked_input_seq_len, ...]  # noqa: E501
+    remaining_X_chunked = torch.zeros_like(X)[:remaining_chunked_input_seq_len, ...]  # noqa: E501
+    remaining_dt_chunked = torch.zeros_like(dt)[:remaining_chunked_input_seq_len, ...]  # noqa: E501
+    remaining_B_chunked = torch.zeros_like(B)[:remaining_chunked_input_seq_len, ...]  # noqa: E501
+    remaining_C_chunked = torch.zeros_like(C)[:remaining_chunked_input_seq_len, ...]  # noqa: E501
     for i in range(num_sequences):
-        remaining_chunk_f = lambda x, i: x[:, cu_seqlens[i] + chunked_seqlens[i]:cu_seqlens[i+1], ...]  # noqa: E501
+        remaining_chunk_f = lambda x, i: x[cu_seqlens[i] + chunked_seqlens[i]:cu_seqlens[i+1], ...]  # noqa: E501
 
-        remaining_X_chunked[:, remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(X, i)  # noqa: E501
-        remaining_dt_chunked[:, remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(dt, i)  # noqa: E501
-        remaining_B_chunked[:, remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(B, i)  # noqa: E501
-        remaining_C_chunked[:, remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(C, i)  # noqa: E501
+        remaining_X_chunked[remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(X, i)  # noqa: E501
+        remaining_dt_chunked[remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(dt, i)  # noqa: E501
+        remaining_B_chunked[remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(B, i)  # noqa: E501
+        remaining_C_chunked[remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1], ...] = remaining_chunk_f(C, i)  # noqa: E501
 
     # assert input chunking is correct
     concat_chunk_f = lambda pt1, pt2, i: torch.cat([
-        pt1[:,chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1],...],
-        pt2[:,remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1],...],
+        pt1[chunked_cu_seqlens[i]:chunked_cu_seqlens[i+1],...],
+        pt2[remaining_chunked_cu_seqlens[i]:remaining_chunked_cu_seqlens[i+1],...],
         ],
-        dim=1)
-    concat_batch_f = lambda pt1, pt2: torch.cat([concat_chunk_f(pt1, pt2, i) for i in range(num_sequences)], dim=1)  # noqa: E501
+        dim=0)
+    concat_batch_f = lambda pt1, pt2: torch.cat([concat_chunk_f(pt1, pt2, i) for i in range(num_sequences)], dim=0)  # noqa: E501
     # fmt: on
 
     assert concat_batch_f(X_chunked, remaining_X_chunked).equal(X)
@@ -498,7 +516,7 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
                 remaining_chunked_cu_seqlens[-1])
 
     Y_chunked = torch.empty_like(remaining_X_chunked)
-    state_chunked = mamba_chunk_scan_combined(
+    state_chunked = mamba_chunk_scan_combined_varlen(
         remaining_X_chunked,
         remaining_dt_chunked,
         A,
@@ -510,7 +528,6 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
         seq_idx=remaining_chunked_seq_idx,
         chunk_indices=chunk_indices,
         chunk_offsets=chunk_offsets,
-        return_varlen_states=True,
         initial_states=partial_state,
         out=Y_chunked,
     )
@@ -518,17 +535,17 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
 
     # kernel chunked is same as kernel overall
     for i in range(num_sequences):
-        Y_seq = Y[:, cu_seqlens[i]:cu_seqlens[i + 1], ...]
-        Y_ref_seq = Y_ref[:, cu_seqlens[i]:cu_seqlens[i + 1], ...]
+        Y_seq = Y[cu_seqlens[i]:cu_seqlens[i + 1], ...]
+        Y_ref_seq = Y_ref[cu_seqlens[i]:cu_seqlens[i + 1], ...]
         torch.testing.assert_close(
-            Y_seq[:, :chunked_seqlens[i], ...],
-            Y_ref_seq[:, :chunked_seqlens[i], ...],
+            Y_seq[:chunked_seqlens[i], ...],
+            Y_ref_seq[:chunked_seqlens[i], ...],
             atol=atol,
             rtol=rtol,
             msg=lambda x: f"seq{i} output part1 " + x)  # noqa: B023
         torch.testing.assert_close(
-            Y_seq[:, chunked_seqlens[i]:, ...],
-            Y_ref_seq[:, chunked_seqlens[i]:, ...],
+            Y_seq[chunked_seqlens[i]:, ...],
+            Y_ref_seq[chunked_seqlens[i]:, ...],
             atol=atol,
             rtol=rtol,
             msg=lambda x: f"seq{i} output part2 " + x)  # noqa: B023

tests/kernels/moe/modular_kernel_tools/mk_objects.py

@@ -222,7 +222,8 @@ if (has_flashinfer_cutlass_fused_moe()
     from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (  # noqa: E501
         FlashInferExperts)
     from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize import (  # noqa: E501
-        FlashInferCutlassMoEPrepareAndFinalize)
+        FlashInferCutlassMoEPrepareAndFinalize,
+        create_flashinfer_prepare_finalize)
 
     register_prepare_and_finalize(
         FlashInferCutlassMoEPrepareAndFinalize,
@@ -373,7 +374,7 @@ def make_prepare_finalize(
         assert prepare_finalize is not None
         return prepare_finalize
     elif prepare_finalize_type == FlashInferCutlassMoEPrepareAndFinalize:
-        return FlashInferCutlassMoEPrepareAndFinalize(
+        return create_flashinfer_prepare_finalize(
             use_dp=moe.moe_parallel_config.dp_size > 1)
     else:
         return MoEPrepareAndFinalizeNoEP()

tests/kernels/moe/test_flashinfer.py

@@ -138,7 +138,7 @@ def test_flashinfer_per_tensor_moe_fp8_no_graph(
         td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=True)
 
         score = torch.randn((m, e), device="cuda", dtype=torch.bfloat16)
-        topk_weights, topk_ids = FusedMoE.select_experts(
+        topk_weights, topk_ids, _ = FusedMoE.select_experts(
             hidden_states=td.hidden_states,
             router_logits=score,
             use_grouped_topk=False,
@@ -206,7 +206,7 @@ def test_flashinfer_cutlass_moe_fp8_no_graph(
         td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=False)
 
         score = torch.randn((m, e), device="cuda", dtype=torch.bfloat16)
-        topk_weights, topk_ids = FusedMoE.select_experts(
+        topk_weights, topk_ids, _ = FusedMoE.select_experts(
             hidden_states=td.hidden_states,
             router_logits=score,
             use_grouped_topk=False,

tests/kernels/quantization/test_block_fp8.py

@@ -11,11 +11,12 @@ from tests.kernels.quant_utils import (native_per_token_group_quant_fp8,
                                        native_w8a8_block_matmul)
 from vllm.config import VllmConfig
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    cutlass_scaled_mm, get_col_major_tma_aligned_tensor,
-    per_token_group_quant_fp8, w8a8_block_fp8_matmul)
+    cutlass_scaled_mm, per_token_group_quant_fp8, w8a8_block_fp8_matmul)
 from vllm.platforms import current_platform
 from vllm.utils import has_deep_gemm
-from vllm.utils.deep_gemm import fp8_gemm_nt, per_block_cast_to_fp8
+from vllm.utils.deep_gemm import (fp8_gemm_nt,
+                                  get_col_major_tma_aligned_tensor,
+                                  per_block_cast_to_fp8)
 
 if current_platform.get_device_capability() < (9, 0):
     pytest.skip("FP8 Triton requires CUDA 9.0 or higher",

tests/kernels/quantization/test_rocm_skinny_gemms.py

@@ -1,12 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import math
+
 import pytest
 import torch
 
 import vllm._custom_ops as ops
 from tests.kernels.quant_utils import ref_dynamic_per_tensor_fp8_quant
-from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
-    rocm_per_tensor_w8a8_scaled_mm_impl)
 from vllm.platforms import current_platform
 
 DTYPES = [torch.bfloat16, torch.float16]
@@ -49,6 +49,7 @@ NKM_FACTORS_WVSPLITK_FP8 = [
     (2, 512, 512),
     (3, 2048, 2048),
     (4, 4096, 4096),
+    (4, 16400, 2048),
     # Extended FP8 dimensions not covered by WVSPLITK
     (1, 14336, 1024),
     (2, 24576, 2048),
@@ -67,6 +68,9 @@ SEEDS = [0]
 @torch.inference_mode()
 def test_rocm_llmm1_kernel(n, k, m, dtype, rows_per_block, seed):
     torch.manual_seed(seed)
+    #TODO: Zero-centering the inputs causes errors for LLMM1!
+    #      Without that the numbers quickly saturate, and may
+    #      be giving false matches.
     A = torch.rand(n, k, dtype=dtype, device="cuda")
     B = torch.rand(m, k, dtype=dtype, device="cuda")
 
@@ -85,11 +89,51 @@ def test_rocm_wvsplitk_kernel(n, k, m, dtype, seed):
     torch.manual_seed(seed)
     cu_count = current_platform.get_cu_count()
 
-    A = torch.rand(n, k, dtype=dtype, device="cuda")
-    B = torch.rand(m, k, dtype=dtype, device="cuda")
+    A = torch.rand(n, k, dtype=dtype, device="cuda") - .5
+    B = torch.rand(m, k, dtype=dtype, device="cuda") - .5
 
-    ref_out = torch.matmul(A, B.t())
-    out = ops.wvSplitK(B, A, cu_count)
+    ref_out = torch.nn.functional.linear(A, B)
+    out = ops.wvSplitK(B, A.view(-1, A.size(-1)), cu_count)
+
+    assert torch.allclose(out, ref_out, rtol=0.01)
+
+
+@pytest.mark.parametrize("n,k,m", NKM_FACTORS_WVSPLITK)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@pytest.mark.skipif(not current_platform.is_rocm(),
+                    reason="only test for rocm")
+def test_rocm_wvsplitk_bias1D_kernel(n, k, m, dtype, seed):
+    torch.manual_seed(seed)
+    cu_count = current_platform.get_cu_count()
+
+    xavier = math.sqrt(2 / k)  # normalize to avoid large output-bias deltas
+    A = (torch.rand(n, k, dtype=dtype, device="cuda") - .5) * xavier
+    B = (torch.rand(m, k, dtype=dtype, device="cuda") - .5) * xavier
+    BIAS = torch.rand(m, dtype=dtype, device="cuda") - .5
+
+    ref_out = torch.nn.functional.linear(A, B, BIAS)
+    out = ops.wvSplitK(B, A.view(-1, A.size(-1)), cu_count, BIAS)
+
+    assert torch.allclose(out, ref_out, rtol=0.01)
+
+
+@pytest.mark.parametrize("n,k,m", NKM_FACTORS_WVSPLITK)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@pytest.mark.skipif(not current_platform.is_rocm(),
+                    reason="only test for rocm")
+def test_rocm_wvsplitk_bias2D_kernel(n, k, m, dtype, seed):
+    torch.manual_seed(seed)
+    cu_count = current_platform.get_cu_count()
+
+    xavier = math.sqrt(2 / k)  # normalize to avoid large output-bias deltas
+    A = (torch.rand(n, k, dtype=dtype, device="cuda") - .5) * xavier
+    B = (torch.rand(m, k, dtype=dtype, device="cuda") - .5) * xavier
+    BIAS = torch.rand(n, m, dtype=dtype, device="cuda") - .5
+
+    ref_out = torch.nn.functional.linear(A, B, BIAS)
+    out = ops.wvSplitK(B, A.view(-1, A.size(-1)), cu_count, BIAS)
 
     assert torch.allclose(out, ref_out, rtol=0.01)
 
@@ -103,8 +147,8 @@ def test_rocm_wvsplitk_kernel(n, k, m, dtype, seed):
 def test_rocm_wvsplitk_fp8_kernel(n, k, m, dtype, seed):
     torch.manual_seed(seed)
 
-    A = torch.rand(n, k, device="cuda")
-    B = torch.rand(m, k, device="cuda")
+    A = torch.rand(n, k, device="cuda") - 0.5
+    B = torch.rand(m, k, device="cuda") - 0.5
 
     A, scale_a = ref_dynamic_per_tensor_fp8_quant(A)
     B, scale_b = ref_dynamic_per_tensor_fp8_quant(B)
@@ -123,27 +167,27 @@ def test_rocm_wvsplitk_fp8_kernel(n, k, m, dtype, seed):
 @pytest.mark.parametrize("n,k,m", NKM_FACTORS_WVSPLITK_FP8)
 @pytest.mark.parametrize("dtype", DTYPES)
 @pytest.mark.parametrize("seed", SEEDS)
-@pytest.mark.parametrize("use_bias", [True, False])
 @pytest.mark.skipif(
     not (current_platform.is_rocm() and current_platform.supports_fp8()),
     reason="only test for rocm fp8")
-def test_rocm_per_tensor_w8a8_scaled_mm_impl(n, k, m, dtype, seed, use_bias):
+def test_rocm_wvsplitk_fp8_bias1D_kernel(n, k, m, dtype, seed):
     torch.manual_seed(seed)
 
-    A = torch.rand(n, k, device="cuda")
-    B = torch.rand(m, k, device="cuda")
+    xavier = math.sqrt(2 / k)  # normalize to avoid large output-bias deltas
+    A = (torch.rand(n, k, device="cuda") - .5) * xavier
+    B = (torch.rand(m, k, device="cuda") - .5) * xavier
+    BIAS = torch.rand(m, dtype=dtype, device="cuda") - .5
 
     A, scale_a = ref_dynamic_per_tensor_fp8_quant(A)
     B, scale_b = ref_dynamic_per_tensor_fp8_quant(B)
 
-    bias = torch.rand(1, m, dtype=dtype, device="cuda") if use_bias else None
-
-    output = rocm_per_tensor_w8a8_scaled_mm_impl(A, B.t(), dtype, scale_a,
-                                                 scale_b, bias)
     ref_out = torch._scaled_mm(A,
                                B.t(),
                                out_dtype=dtype,
                                scale_a=scale_a,
                                scale_b=scale_b,
-                               bias=bias)
-    assert torch.allclose(output, ref_out, rtol=0.01)
+                               bias=BIAS)
+    out = ops.wvSplitKQ(B, A, dtype, scale_a, scale_b,
+                        current_platform.get_cu_count(), BIAS)
+
+    assert torch.allclose(out, ref_out, rtol=0.01)

tests/kernels/test_onednn.py

@@ -165,7 +165,7 @@ def onednn_gemm_test_helper(primitive_cache_size: int,
 def test_onednn_int8_scaled_gemm(
     n: int,
     k: int,
-    m_list: tuple[int],
+    m_list: tuple[int, ...],
     per_tensor_a_scale: bool,
     per_tensor_b_scale: bool,
     use_bias: bool,
@@ -196,7 +196,7 @@ def test_onednn_int8_scaled_gemm(
 def test_onednn_gemm(
     n: int,
     k: int,
-    m_list: tuple[int],
+    m_list: tuple[int, ...],
     use_bias: bool,
     use_stride: bool,
     dtype: torch.dtype,

tests/kernels/utils.py

@@ -524,14 +524,14 @@ def make_backend(backend_name: str) -> AttentionBackend:
 
     * Backend instance
     '''
-    if backend_name in (STR_XFORMERS_ATTN_VAL, "XFORMERS_VLLM_V1"):
+    if backend_name == STR_XFORMERS_ATTN_VAL:
         from vllm.v1.attention.backends.xformers import (
             XFormersAttentionBackend)
         return XFormersAttentionBackend()
-    if backend_name in (STR_FLASH_ATTN_VAL, "FLASH_ATTN_VLLM_V1"):
+    if backend_name == STR_FLASH_ATTN_VAL:
         from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
         return FlashAttentionBackend()
-    if backend_name == "TRITON_ATTN_VLLM_V1":
+    if backend_name == "TRITON_ATTN":
         from vllm.v1.attention.backends.triton_attn import (
             TritonAttentionBackend)
         return TritonAttentionBackend()
@@ -539,7 +539,7 @@ def make_backend(backend_name: str) -> AttentionBackend:
         from vllm.v1.attention.backends.flex_attention import (
             FlexAttentionBackend)
         return FlexAttentionBackend()
-    if backend_name in ("TORCH_SDPA", "TORCH_SDPA_VLLM_V1"):
+    if backend_name == "TORCH_SDPA":
         from vllm.v1.attention.backends.cpu_attn import TorchSDPABackend
         return TorchSDPABackend()
     if backend_name == "FLASHINFER":

tests/lora/test_layers.py

@@ -164,8 +164,8 @@ def populate_loras(
                         weight=layer_weights,
                         generate_embeddings_tensor=generate_embeddings_tensor,
                     )
-                sublora.lora_b = sublora.lora_b[:, (sublora_len *
-                                                    i):(sublora_len * (i + 1))]
+                sublora.lora_b = sublora.lora_b[(sublora_len *
+                                                 i):(sublora_len * (i + 1)), :]
                 sublora.optimize()
                 subloras.append(sublora)
 
@@ -304,9 +304,9 @@ def test_embeddings(dist_init, num_loras, device, vocab_size, stage) -> None:
             result = embedding(input_)
             after_a = F.embedding(
                 input_,
-                lora.lora_a,
+                lora.lora_a.T,
             )
-            result += (after_a @ lora.lora_b)
+            result += (after_a @ lora.lora_b.T)
             expected_results.append(result)
         expected_result = torch.cat(expected_results)
 
@@ -445,9 +445,9 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
             result = expanded_embedding(input_)
             after_a = F.embedding(
                 original_input_,
-                lora.lora_a,
+                lora.lora_a.T,
             )
-            result += (after_a @ lora.lora_b)
+            result += (after_a @ lora.lora_b.T)
             expected_results.append(result)
         expected_result = torch.cat(expected_results)
 
@@ -575,7 +575,7 @@ def test_lm_head_logits_processor(dist_init, num_loras, device, vocab_size,
                                                   lm_head=linear,
                                                   embedding_bias=None)
             result[:, vocab_size + embeddings_tensor_len:] = float("-inf")
-            result += input_ @ lora.lora_a @ lora.lora_b * lora.scaling
+            result += input_ @ lora.lora_a.T @ lora.lora_b.T * lora.scaling
             expected_results.append(result)
         expected_result = torch.cat(expected_results)
         logits_processor.org_vocab_size = vocab_size
@@ -692,9 +692,10 @@ def test_linear_replicated(
 
         expected_results: list[torch.Tensor] = []
         for input_, lora_id in zip(inputs, prompt_mapping):
+
             lora = lora_dict[lora_id]
             result = linear(input_)[0]
-            result += input_ @ lora.lora_a @ lora.lora_b * lora.scaling
+            result += input_ @ lora.lora_a.T @ lora.lora_b.T * lora.scaling
             expected_results.append(result)
         expected_result = torch.cat(expected_results)
 
@@ -817,7 +818,7 @@ def test_linear_parallel(dist_init, num_loras, orientation, fully_shard,
         for input_, lora_id in zip(inputs, prompt_mapping):
             lora = lora_dict[lora_id]
             result = linear(input_)[0]
-            result += input_ @ lora.lora_a @ lora.lora_b * lora.scaling
+            result += input_ @ lora.lora_a.T @ lora.lora_b.T * lora.scaling
             expected_results.append(result)
         expected_result = torch.cat(expected_results)
 
@@ -965,9 +966,10 @@ def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,
             result = linear(input_)[0]
             subloras = sublora_dict[lora_id]
             for i, sublora in enumerate(subloras):
-                result[:, sublora.lora_b.shape[1] * i:sublora.lora_b.shape[1] *
-                       (i + 1)] += (input_ @ sublora.lora_a @ sublora.lora_b *
-                                    sublora.scaling)
+                result[:, sublora.lora_b.shape[0] * i:sublora.lora_b.shape[0] *
+                       (i + 1)] += (
+                           input_ @ sublora.lora_a.T @ sublora.lora_b.T *
+                           sublora.scaling)
             expected_results.append(result)
         expected_result = torch.cat(expected_results)
 

tests/lora/test_lora_manager.py

@@ -63,9 +63,9 @@ def test_from_lora_tensors(sql_lora_files, device):
         assert lora.lora_b is not None
         assert lora.lora_a.device == torch.device(device)
         assert lora.lora_b.device == torch.device(device)
-        assert (lora.lora_a.shape[1] == lora.lora_b.shape[0]
+        assert (lora.lora_a.shape[0] == lora.lora_b.shape[1]
                 ), f"{lora.lora_a.shape=}, {lora.lora_b.shape=}"
-        assert lora.lora_a.shape[1] == 8
+        assert lora.lora_a.shape[0] == 8
         embeddings_module = next(
             (k for k in EMBEDDING_MODULES if k in module_name), None)
         if embeddings_module:
@@ -86,8 +86,8 @@ def create_lora(lora_id: int, model: nn.Module, sub_modules: list[str],
             name,
             8,
             16,
-            torch.rand([w.shape[1], 8], device=device),
-            torch.rand([8, w.shape[0]], device=device),
+            torch.rand([8, w.shape[1]], device=device),
+            torch.rand([w.shape[0], 8], device=device),
         )
     return LoRAModel(lora_id, 8, loras)
 
@@ -109,8 +109,8 @@ def create_packed_lora(
             replaced_module_name,
             8,
             16,
-            torch.rand([w.shape[1], 8], device=device),
-            torch.rand([8, w.shape[0] // len(replaced_module_names)],
+            torch.rand([8, w.shape[1]], device=device),
+            torch.rand([w.shape[0] // len(replaced_module_names), 8],
                        device=device),
         )
     return LoRAModel(lora_id, 8, loras)

tests/lora/utils.py

@@ -36,10 +36,10 @@ class DummyLoRAManager:
             module_name,
             rank=rank,
             lora_alpha=1,
-            lora_a=torch.rand([weight.shape[1], rank],
+            lora_a=torch.rand([rank, weight.shape[1]],
                               dtype=weight.dtype,
                               device=self._device),
-            lora_b=torch.rand([rank, weight.shape[0]],
+            lora_b=torch.rand([weight.shape[0], rank],
                               dtype=weight.dtype,
                               device=self._device),
         )
@@ -67,8 +67,8 @@ class DummyLoRAManager:
             module_name,
             rank=rank,
             lora_alpha=1,
-            lora_a=torch.rand([input_dim, rank], device="cuda"),
-            lora_b=torch.rand([rank, output_dim], device="cuda"),
+            lora_a=torch.rand([rank, input_dim], device="cuda"),
+            lora_b=torch.rand([output_dim, input_dim], device="cuda"),
             embeddings_tensor=embeddings_tensor,
         )
         self.set_module_lora(module_name, lora)

tests/model_executor/model_loader/test_sharded_state_loader.py

@@ -91,8 +91,7 @@ def _run_generate(input_dir, queue: mp.Queue, **kwargs):
 @pytest.mark.parametrize("enable_lora", [False, True])
 @pytest.mark.parametrize("tp_size", [1, 2])
 def test_sharded_state_loader(enable_lora, tp_size, num_gpus_available,
-                              llama_3p2_1b_files,
-                              monkeypatch: pytest.MonkeyPatch):
+                              llama_3p2_1b_files):
     if num_gpus_available < tp_size:
         pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
 

tests/model_executor/test_enabled_custom_ops.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Optional
 
 import pytest
 import torch
@@ -34,15 +35,15 @@ class Relu3(ReLUSquaredActivation):
     [
         # Default values based on compile level
         # - All by default (no Inductor compilation)
-        ("", 0, False, [True] * 4, True),
-        ("", 1, True, [True] * 4, True),
-        ("", 2, False, [True] * 4, True),
+        (None, 0, False, [True] * 4, True),
+        (None, 1, True, [True] * 4, True),
+        (None, 2, False, [True] * 4, True),
         # - None by default (with Inductor)
-        ("", 3, True, [False] * 4, False),
-        ("", 4, True, [False] * 4, False),
+        (None, 3, True, [False] * 4, False),
+        (None, 4, True, [False] * 4, False),
         # - All by default (without Inductor)
-        ("", 3, False, [True] * 4, True),
-        ("", 4, False, [True] * 4, True),
+        (None, 3, False, [True] * 4, True),
+        (None, 4, False, [True] * 4, True),
         # Explicitly enabling/disabling
         #
         # Default: all
@@ -54,7 +55,7 @@ class Relu3(ReLUSquaredActivation):
         # All but SiluAndMul
         ("all,-silu_and_mul", 2, True, [1, 0, 1, 1], True),
         # All but ReLU3 (even if ReLU2 is on)
-        ("-relu3,relu2", 3, False, [1, 1, 1, 0], True),
+        ("-relu3,+relu2", 3, False, [1, 1, 1, 0], True),
         # RMSNorm and SiluAndMul
         ("none,-relu3,+rms_norm,+silu_and_mul", 4, False, [1, 1, 0, 0], False),
         # All but RMSNorm
@@ -67,12 +68,13 @@ class Relu3(ReLUSquaredActivation):
         # All but RMSNorm
         ("all,-rms_norm", 4, True, [0, 1, 1, 1], True),
     ])
-def test_enabled_ops(env: str, torch_level: int, use_inductor: bool,
+def test_enabled_ops(env: Optional[str], torch_level: int, use_inductor: bool,
                      ops_enabled: list[int], default_on: bool):
+    custom_ops = env.split(',') if env else []
     vllm_config = VllmConfig(
         compilation_config=CompilationConfig(use_inductor=bool(use_inductor),
                                              level=torch_level,
-                                             custom_ops=env.split(",")))
+                                             custom_ops=custom_ops))
     with set_current_vllm_config(vllm_config):
         assert CustomOp.default_on() == default_on
 

tests/models/language/generation/test_hybrid.py

@@ -20,7 +20,9 @@ pytestmark = pytest.mark.hybrid_model
 SSM_MODELS = [
     "state-spaces/mamba-130m-hf",
     "tiiuae/falcon-mamba-tiny-dev",
-    "yujiepan/mamba2-codestral-v0.1-tiny-random",
+    # mamba2-codestral in transformers is broken pending:
+    # https://github.com/huggingface/transformers/pull/40861
+    #"yujiepan/mamba2-codestral-v0.1-tiny-random",
 ]
 
 HYBRID_MODELS = [
@@ -31,18 +33,7 @@ HYBRID_MODELS = [
     "ibm-granite/granite-4.0-tiny-preview",
     "tiiuae/Falcon-H1-0.5B-Base",
     "LiquidAI/LFM2-1.2B",
-]
-
-V1_SUPPORTED_MODELS = [
-    "state-spaces/mamba-130m-hf",
-    "ai21labs/Jamba-tiny-dev",
-    "pfnet/plamo-2-1b",
-    "yujiepan/mamba2-codestral-v0.1-tiny-random",
-    "Zyphra/Zamba2-1.2B-instruct",
-    "hmellor/tiny-random-BambaForCausalLM",
-    "ibm-granite/granite-4.0-tiny-preview",
-    "tiiuae/Falcon-H1-0.5B-Base",
-    "LiquidAI/LFM2-1.2B",
+    "tiny-random/qwen3-next-moe",
 ]
 
 FULL_CUDA_GRAPH_MODELS = [
@@ -51,10 +42,6 @@ FULL_CUDA_GRAPH_MODELS = [
     "Zyphra/Zamba2-1.2B-instruct",
 ]
 
-V0_UNSUPPORTED_MODELS = [
-    "LiquidAI/LFM2-1.2B",
-]
-
 FP32_STATE_MODELS = [
     "state-spaces/mamba-130m-hf",
     "Zyphra/Zamba2-1.2B-instruct",
@@ -88,20 +75,16 @@ def test_models(
         hf_outputs = hf_model.generate_greedy_logprobs_limit(
             example_prompts, max_tokens, num_logprobs)
 
-    if model in V1_SUPPORTED_MODELS:
-        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
-            vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
-                example_prompts, max_tokens, num_logprobs)
-    else:
-        vllm_v1_outputs = None
+    with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
 
-    if model in V1_SUPPORTED_MODELS:
-        check_logprobs_close(
-            outputs_0_lst=hf_outputs,
-            outputs_1_lst=vllm_v1_outputs,
-            name_0="hf",
-            name_1="vllm-v1",
-        )
+    check_logprobs_close(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
 
 
 @pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
@@ -299,14 +282,14 @@ def test_full_cuda_graph(
             example_prompts, max_tokens, num_logprobs)
 
     with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
-        vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
             example_prompts, max_tokens, num_logprobs)
 
     check_logprobs_close(
         outputs_0_lst=hf_outputs,
-        outputs_1_lst=vllm_v1_outputs,
+        outputs_1_lst=vllm_outputs,
         name_0="hf",
-        name_1="vllm-v1",
+        name_1="vllm",
     )
 
 
@@ -340,12 +323,12 @@ def test_fp32_cache_state(
     with vllm_runner(model,
                      max_num_seqs=MAX_NUM_SEQS,
                      **{cache_dtype_param: "float32"}) as vllm_model:
-        vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
             example_prompts, max_tokens, num_logprobs)
 
     check_logprobs_close(
         outputs_0_lst=hf_outputs,
-        outputs_1_lst=vllm_v1_outputs,
+        outputs_1_lst=vllm_outputs,
         name_0="hf",
-        name_1="vllm-v1",
+        name_1="vllm",
     )

tests/models/multimodal/generation/test_qwen2_5_vl.py

@@ -0,0 +1,132 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+
+from vllm.multimodal.video import sample_frames_from_video
+
+from ....conftest import VIDEO_ASSETS
+
+models = ["Qwen/Qwen2.5-VL-3B-Instruct"]
+target_dtype = "bfloat16"
+
+VIDEO_PLACEHOLDER = "<|vision_start|><|video_pad|><|vision_end|>"
+
+
+def qwen2_5_vl_chat_template(*query):
+    return f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{''.join(query)}<|im_end|><|im_start|>assistant\n"  # noqa: E501
+
+
+VIDEO_PROMPTS = VIDEO_ASSETS.prompts({
+    "baby_reading":
+    qwen2_5_vl_chat_template(
+        VIDEO_PLACEHOLDER,
+        "Describe this video with a short sentence ",
+        "(no more than 20 words)",
+    ),
+})
+
+
+@pytest.mark.core_model
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize("video_pruning_rate", [0.0, 0.75])
+@pytest.mark.parametrize("num_frames", [16])
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_qwen2_5_vl_evs_functionality(vllm_runner, video_assets, model,
+                                      video_pruning_rate: float,
+                                      num_frames: int, dtype: str,
+                                      max_tokens: int) -> None:
+    """Test EVS (Efficient Video Sampling) functionality with different
+    pruning rates.
+    """
+
+    # Sample frames from video assets
+    sampled_vids = [
+        sample_frames_from_video(asset.np_ndarrays, num_frames)
+        for asset in video_assets
+    ]
+
+    prompts = [VIDEO_PROMPTS[0]]
+    videos = [sampled_vids[0]]
+
+    # Initialize model with EVS configuration
+    with vllm_runner(model,
+                     runner="generate",
+                     max_model_len=4000,
+                     max_num_seqs=1,
+                     dtype=dtype,
+                     limit_mm_per_prompt={"video": 1},
+                     tensor_parallel_size=1,
+                     video_pruning_rate=video_pruning_rate) as vllm_model:
+
+        # Generate output - this should not crash
+        outputs = vllm_model.generate_greedy(prompts,
+                                             max_tokens,
+                                             videos=videos)
+
+        # Basic validation that we got a response
+        assert len(outputs) == 1
+        output_ids, output_text = outputs[0]
+
+        # Ensure we got some output
+        assert len(output_ids) > 0
+        assert len(output_text) > 0
+
+        # Ensure the output is a string
+        assert isinstance(output_text, str)
+
+
+@pytest.mark.core_model
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize("video_pruning_rate", [0.0, 0.75])
+@pytest.mark.parametrize("num_frames", [16])
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_qwen2_5_vl_evs_batched_videos(vllm_runner, video_assets, model,
+                                       video_pruning_rate: float,
+                                       num_frames: int, dtype: str,
+                                       max_tokens: int) -> None:
+    """Test EVS functionality with batched videos.
+
+    This test validates that:
+    1. The model handles batched video inputs correctly with EVS
+    2. Both pruning configurations work with multiple videos
+    3. The model doesn't crash when processing multiple videos simultaneously
+    """
+    # Sample frames from video assets
+    sampled_vids = [
+        sample_frames_from_video(asset.np_ndarrays, num_frames)
+        for asset in video_assets
+    ]
+
+    # Test batched videos
+    prompts = [VIDEO_PROMPTS[0], VIDEO_PROMPTS[0]]
+    videos = [sampled_vids[0],
+              sampled_vids[0]]  # Use same video twice for testing
+
+    # Initialize model with EVS configuration
+    with vllm_runner(model,
+                     runner="generate",
+                     max_model_len=4000,
+                     max_num_seqs=2,
+                     dtype=dtype,
+                     limit_mm_per_prompt={"video": 2},
+                     tensor_parallel_size=1,
+                     video_pruning_rate=video_pruning_rate) as vllm_model:
+
+        # Generate output - this should not crash
+        outputs = vllm_model.generate_greedy(prompts,
+                                             max_tokens,
+                                             videos=videos)
+
+        # Basic validation that we got responses for both videos
+        assert len(outputs) == 2
+
+        for output_ids, output_text in outputs:
+            # Ensure we got some output for each video
+            assert len(output_ids) > 0
+            assert len(output_text) > 0
+
+            # Ensure the output is a string
+            assert isinstance(output_text, str)

tests/models/multimodal/generation/vlm_utils/types.py

@@ -101,7 +101,7 @@ class VLMTestInfo(NamedTuple):
     # Function for converting ImageAssets to image embeddings;
     # We need to define this explicitly for embedding tests
     convert_assets_to_embeddings: Optional[Callable[[ImageTestAssets],
-                                                    torch.Tensor]] = None
+                                                    list[torch.Tensor]]] = None
 
     # Exposed options for vLLM runner; we change these in a several tests,
     # but the defaults are derived from VllmRunner & the engine defaults
@@ -137,12 +137,12 @@ class VLMTestInfo(NamedTuple):
     # Default expandable params per test; these defaults can be overridden in
     # instances of this object; the complete set of test cases for the model
     # is all combinations of .models + all fields below
-    max_tokens: Union[int, tuple[int]] = 128
-    num_logprobs: Union[int, tuple[int]] = 5
-    dtype: Union[str, Union[list[str], tuple[str, ...]]] = "auto"
-    distributed_executor_backend: Optional[Union[str, Iterable[str]]] = None
+    max_tokens: int = 128
+    num_logprobs: int = 5
+    dtype: str = "auto"
+    distributed_executor_backend: Optional[str] = None
     # Only expanded in video tests
-    num_video_frames: Union[int, tuple[int]] = 16
+    num_video_frames: int = 16
 
     # Fixed image sizes / image size factors; most tests use image_size_factors
     # The values provided for these two fields will be stacked and expanded

tests/models/multimodal/processing/test_common.py

@@ -12,11 +12,11 @@ from mistral_common.protocol.instruct.request import ChatCompletionRequest
 from PIL import Image
 
 from vllm.config import ModelConfig
-from vllm.inputs import InputProcessingContext
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
 from vllm.multimodal.cache import MultiModalProcessorOnlyCache
 from vllm.multimodal.inputs import MultiModalInputs
-from vllm.multimodal.processing import BaseMultiModalProcessor
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        InputProcessingContext)
 from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer,
                                                cached_tokenizer_from_config,
                                                encode_tokens)
@@ -213,6 +213,7 @@ _IGNORE_MM_KEYS = {
 MM_DATA_PATCHES = {
     # GLM4.1V and Qwen3-VL requires video metadata to be included in the input
     "glm4v": glm4_1v_patch_mm_data,
+    "glm4v_moe": glm4_1v_patch_mm_data,
     "qwen3_vl": qwen3_vl_patch_mm_data,
     "qwen3_vl_moe": qwen3_vl_patch_mm_data,
 }

tests/models/multimodal/processing/test_tensor_schema.py

@@ -18,10 +18,12 @@ from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
 from vllm.distributed import (cleanup_dist_env_and_memory,
                               init_distributed_environment,
                               initialize_model_parallel)
-from vllm.inputs import InputProcessingContext
 from vllm.model_executor.model_loader.utils import set_default_torch_dtype
+from vllm.model_executor.models.interfaces import (SupportsMultiModal,
+                                                   supports_multimodal)
 from vllm.multimodal import MULTIMODAL_REGISTRY, BatchedTensorInputs
-from vllm.multimodal.processing import BaseMultiModalProcessor
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        InputProcessingContext)
 from vllm.multimodal.utils import group_mm_kwargs_by_modality
 from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
 from vllm.utils import is_list_of
@@ -88,6 +90,7 @@ def resize_mm_data(
 
 
 def create_batched_mm_kwargs(
+    model_cls: type[SupportsMultiModal],
     model_config: ModelConfig,
     processor: BaseMultiModalProcessor,
     size_factors: tuple[float, ...] = (1.0, 0.5, 0.25),
@@ -127,16 +130,22 @@ def create_batched_mm_kwargs(
         mm_data=resized_mm_data,
         hf_processor_mm_kwargs=processor_inputs.hf_processor_mm_kwargs,
         tokenization_kwargs=processor_inputs.tokenization_kwargs,
-    )["mm_kwargs"]
+    )["mm_kwargs"].require_data()
     items = [
         item for modality in supported_mm_limits
         for item in mm_kwargs[modality]
     ]
-    return group_mm_kwargs_by_modality(items)
+    return group_mm_kwargs_by_modality(
+        items,
+        merge_by_field_config=model_cls.merge_by_field_config,
+    )
 
 
 @contextmanager
-def initialize_dummy_model(model_cls: nn.Module, model_config: ModelConfig):
+def initialize_dummy_model(
+    model_cls: type[nn.Module],
+    model_config: ModelConfig,
+):
     temp_file = tempfile.mkstemp()[1]
     init_distributed_environment(
         world_size=1,
@@ -198,8 +207,12 @@ def test_model_tensor_schema(model_arch: str, model_id: str):
         hf_overrides=hf_overrides_fn,
         skip_tokenizer_init=model_info.skip_tokenizer_init,
         enforce_eager=model_info.enforce_eager,
-        dtype=model_info.dtype)
+        dtype=model_info.dtype,
+    )
+
     model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
+    assert supports_multimodal(model_cls)
+
     factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
 
     inputs_parse_methods = []
@@ -228,7 +241,7 @@ def test_model_tensor_schema(model_arch: str, model_id: str):
 
     with initialize_dummy_model(model_cls, model_config) as model:
         for modality, _, mm_kwargs in create_batched_mm_kwargs(
-                model_config, processor):
+                model_cls, model_config, processor):
             for method_name in inputs_parse_methods:
                 print(f"Testing `{method_name}` with modality={modality} "
                       f"and mm_kwargs{list(mm_kwargs.keys())}")

tests/models/registry.py

@@ -196,6 +196,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
                                          trust_remote_code=True),
     "Cohere2ForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r7b-12-2024", # noqa: E501
                                          trust_remote_code=True),
+    "CwmForCausalLM": _HfExamplesInfo("facebook/cwm", # noqa: E501
+                                      trust_remote_code=True,
+                                      is_available_online=False),
     "DbrxForCausalLM": _HfExamplesInfo("databricks/dbrx-instruct"),
     "DeciLMForCausalLM": _HfExamplesInfo("nvidia/Llama-3_3-Nemotron-Super-49B-v1", # noqa: E501
                                          trust_remote_code=True),
@@ -273,6 +276,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
                                         is_available_online=False),
     "Llama4ForCausalLM": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
                                          is_available_online=False),
+    "LongcatFlashForCausalLM": _HfExamplesInfo
+                ("meituan-longcat/LongCat-Flash-Chat", trust_remote_code=True),
     "MambaForCausalLM": _HfExamplesInfo("state-spaces/mamba-130m-hf"),
     "Mamba2ForCausalLM": _HfExamplesInfo("mistralai/Mamba-Codestral-7B-v0.1",
                                          min_transformers_version="4.55.3",
@@ -312,14 +317,12 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
     "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
     "Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
-    "Phi4FlashForCausalLM": _HfExamplesInfo("microsoft/Phi-4-mini-flash-reasoning", # noqa: E501
-                                        trust_remote_code=True,
-                                        v0_only=True,
-                                        max_model_len=10240),
     "PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
                                          trust_remote_code=True),
     "Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",
-                                        trust_remote_code=True),
+                                         max_transformers_version="4.55.4",
+                                         transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
+                                         trust_remote_code=True),
     "QWenLMHeadModel": _HfExamplesInfo("Qwen/Qwen-7B-Chat",
                                        max_transformers_version="4.53",
                                        transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
@@ -330,7 +333,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
     "Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
     "Qwen3NextForCausalLM": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
-                                            min_transformers_version="4.56.2"),
+                                            extras={"tiny-random": "tiny-random/qwen3-next-moe"}, # noqa: E501
+                                            min_transformers_version="4.56.3"),
     "RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
     "SeedOssForCausalLM": _HfExamplesInfo("ByteDance-Seed/Seed-OSS-36B-Instruct", # noqa: E501
                                           trust_remote_code=True,
@@ -527,7 +531,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                               trust_remote_code=True),
     "Llama_Nemotron_Nano_VL" : _HfExamplesInfo("nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", # noqa: E501
                                                      trust_remote_code=True),
-    "NemotronH_Nano_VL": _HfExamplesInfo("nano_vl_dummy",
+    "NemotronH_Nano_VL_V2": _HfExamplesInfo("nano_vl_dummy",
                                           is_available_online=False,
                                           trust_remote_code=True),
     "Ovis": _HfExamplesInfo("AIDC-AI/Ovis2-1B", trust_remote_code=True,
@@ -562,10 +566,12 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-Omni-7B-AWQ"),  # noqa: E501
     "Qwen3VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-4B-Instruct", # noqa: E501
                                                         max_model_len=4096,
-                                                        min_transformers_version="4.57"),  # noqa: E501
+                                                        min_transformers_version="4.57",
+                                                        is_available_online=False),
     "Qwen3VLMoeForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-30B-A3B-Instruct", # noqa: E501
-                                                        max_model_len=4096,
-                                                        min_transformers_version="4.57"),
+                                                          max_model_len=4096,
+                                                          min_transformers_version="4.57",
+                                                          is_available_online=False),
     "RForConditionalGeneration": _HfExamplesInfo("YannQi/R-4B",
                                                  trust_remote_code=True),
     "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B",
@@ -638,11 +644,15 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
                                         speculative_model="zai-org/GLM-4.5",
                                         min_transformers_version="4.54",
                                         is_available_online=False),
+    "LongCatFlashMTPModel": _HfExamplesInfo(
+        "meituan-longcat/LongCat-Flash-Chat",
+        trust_remote_code=True,
+        speculative_model="meituan-longcat/LongCat-Flash-Chat"),
     "MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                     trust_remote_code=True,
                                     speculative_model="XiaomiMiMo/MiMo-7B-RL"),
     "Qwen3NextMTP": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
-                                     min_transformers_version="4.56.2"),
+                                     min_transformers_version="4.56.3"),
 }
 
 _TRANSFORMERS_BACKEND_MODELS = {

tests/models/test_initialization.py

@@ -84,7 +84,7 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
             # FIXME: A hack to bypass FA3 assertion because our CI's L4 GPU
             # has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
             # L4 supports FA3.
-            m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN_VLLM_V1")
+            m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN")
         if model_arch == "WhisperForConditionalGeneration":
             m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
         LLM(

tests/models/test_oot_registration.py

@@ -42,7 +42,6 @@ def test_oot_registration_text_generation(
             assert rest == ""
 
 
-@pytest.mark.skip(reason="This test is skipped because it failed on V1.")
 @create_new_process_for_each_test()
 def test_oot_registration_embedding(
     monkeypatch: pytest.MonkeyPatch,
@@ -63,7 +62,6 @@ def test_oot_registration_embedding(
 image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
 
 
-@pytest.mark.skip(reason="This test is skipped because it failed on V1.")
 @create_new_process_for_each_test()
 def test_oot_registration_multimodal(
     monkeypatch: pytest.MonkeyPatch,

tests/models/test_vision.py

@@ -1,10 +1,20 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import math
 
 import pytest
 import torch
+import torch.multiprocessing as mp
 
-from vllm.model_executor.models.vision import resolve_visual_encoder_outputs
+from tests.utils import multi_gpu_test
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.distributed.parallel_state import (init_distributed_environment,
+                                             initialize_model_parallel)
+from vllm.model_executor.models.vision import (
+    get_load_balance_assignment, resolve_visual_encoder_outputs,
+    run_dp_sharded_mrope_vision_model, run_dp_sharded_vision_model)
+from vllm.platforms import current_platform
+from vllm.utils import get_open_port, update_environment_variables
 
 
 @pytest.mark.parametrize(
@@ -33,3 +43,415 @@ def test_resolve_visual_encoder_outputs(feature_sample_layers,
         post_layer_norm=None,
         max_possible_layers=max_possible_layers)
     assert torch.equal(torch.tensor(expected_features), output_tensor)
+
+
+class SimpleLinearModel(torch.nn.Module):
+    """A simple linear vision model for testing."""
+
+    def __init__(self, input_dim: int = 3 * 224 * 224, output_dim: int = 32):
+        super().__init__()
+        self.flatten = torch.nn.Flatten()
+        self.linear = torch.nn.Linear(input_dim, output_dim)
+
+    def forward(self, x: torch.Tensor):
+        # Flatten the input and apply linear transformation
+        x = self.flatten(x)
+        return self.linear(x)
+
+
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize(
+    "batch_size",
+    [
+        1,  # Single image
+        4,  # Small batch
+        5,  # Odd batch size (for testing padding)
+    ],
+)
+def test_run_dp_sharded_vision_model(batch_size: int):
+    world_size = 2
+    # Launch processes
+    mp.spawn(
+        run_dp_sharded_vision_model_vs_direct,
+        args=(
+            world_size,
+            batch_size,
+            get_open_port(),
+        ),
+        nprocs=world_size,
+    )
+
+
+def run_dp_sharded_vision_model_vs_direct(local_rank: int, world_size: int,
+                                          batch_size: int, master_port: int):
+    """
+    Test that run_dp_sharded_vision_model produces the same results as 
+    calling the model directly.
+    """
+
+    # Set random seed for reproducibility
+    current_platform.seed_everything(0)
+
+    device = f"{current_platform.device_name}:{local_rank}"
+    current_platform.set_device(device)
+    torch.set_default_device(device)
+
+    update_environment_variables({
+        'RANK': str(local_rank),
+        'LOCAL_RANK': str(local_rank),
+        'WORLD_SIZE': str(world_size),
+        'MASTER_ADDR': 'localhost',
+        'MASTER_PORT': str(master_port),
+    })
+
+    # initialize distributed
+    init_distributed_environment()
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+
+    # Create a test input tensor
+    image_input = torch.randn(batch_size, 3, 224, 224)
+
+    # Create a simple linear model
+    vision_model = SimpleLinearModel()
+
+    # Run the model directly on the full input
+    with torch.inference_mode():
+        direct_output = vision_model(image_input)
+
+    # Run the model through the sharded function
+    with torch.inference_mode():
+        sharded_output = run_dp_sharded_vision_model(image_input, vision_model)
+
+    # Check that the world size is set up correctly
+    assert get_tensor_model_parallel_world_size() == world_size
+
+    # Check that the outputs have the same shape
+    assert direct_output.shape == sharded_output.shape
+
+    # Check that the outputs are close (they should be identical)
+    assert torch.allclose(direct_output, sharded_output, rtol=1e-5, atol=1e-5)
+
+
+@pytest.mark.parametrize(
+    "sizes,num_gpus,expected_shuffle_indices,expected_gpu_sample_counts,"
+    "expected_grouped_sizes_per_gpu,test_description",
+    [
+        # Empty input
+        ([], 2, [], [0, 0], [0, 0], "empty input"),
+
+        # Fewer samples than GPUs
+        ([100, 200], 4, [1, 0], [1, 1, 0, 0], [200, 100, 0, 0
+                                               ], "fewer samples than GPUs"),
+
+        # Single GPU
+        ([100, 200, 300], 1, [2, 1, 0], [3], [600], "single GPU"),
+
+        # Balanced assignment
+        ([100, 100, 100, 100
+          ], 2, [0, 2, 1, 3], [2, 2], [200, 200], "balanced assignment"),
+
+        # Unbalanced sizes - this one is trickier since the algorithm is greedy
+        ([1000, 100, 200, 50], 2, [0, 2, 1, 3
+                                   ], [1, 3], [1000, 350], "unbalanced sizes"),
+    ],
+)
+def test_get_load_balance_assignment_cases(sizes, num_gpus,
+                                           expected_shuffle_indices,
+                                           expected_gpu_sample_counts,
+                                           expected_grouped_sizes_per_gpu,
+                                           test_description):
+    """Test get_load_balance_assignment with various input cases."""
+    result = get_load_balance_assignment(sizes, num_gpus=num_gpus)
+    (shuffle_indices, gpu_sample_counts, grouped_sizes_per_gpu) = result
+
+    # Common assertions for all cases
+    assert len(shuffle_indices) == len(sizes)
+    assert len(gpu_sample_counts) == num_gpus
+    assert len(grouped_sizes_per_gpu) == num_gpus
+    assert sum(gpu_sample_counts) == len(sizes)
+
+    assert shuffle_indices == expected_shuffle_indices
+
+    assert gpu_sample_counts == expected_gpu_sample_counts
+    assert grouped_sizes_per_gpu == expected_grouped_sizes_per_gpu
+
+
+class SimpleMRopeVisionModel(torch.nn.Module):
+    """A simple vision model for testing mrope functionality."""
+
+    def __init__(self, spatial_merge_size: int = 2, out_hidden_size: int = 64):
+        super().__init__()
+        self.spatial_merge_size = spatial_merge_size
+        self.out_hidden_size = out_hidden_size
+        self.linear = torch.nn.Linear(768, out_hidden_size)
+
+    def forward(self, pixel_values: torch.Tensor,
+                grid_thw_list: list[list[int]]):
+        """Simple forward pass that simulates spatial merging."""
+        # Apply linear transformation
+        embeddings = self.linear(pixel_values)
+
+        # Simulate spatial merging by reducing the number of patches
+        merge_factor = self.spatial_merge_size * self.spatial_merge_size
+
+        # Group patches and merge spatially
+        merged_embeddings = []
+        start_idx = 0
+
+        for grid_thw in grid_thw_list:
+            num_patches = math.prod(grid_thw)
+            end_idx = start_idx + num_patches
+
+            # Get patches for this image
+            image_patches = embeddings[start_idx:end_idx]
+
+            # Simulate spatial merging by averaging groups of patches
+            merged_patches = num_patches // merge_factor
+            if merged_patches > 0:
+                # Reshape and average to simulate merging
+                reshaped = image_patches[:merged_patches * merge_factor].view(
+                    merged_patches, merge_factor, -1)
+                merged = reshaped.mean(dim=1)
+                merged_embeddings.append(merged)
+
+            start_idx = end_idx
+
+        if merged_embeddings:
+            return torch.cat(merged_embeddings, dim=0)
+        else:
+            return torch.empty((0, self.out_hidden_size),
+                               device=pixel_values.device,
+                               dtype=pixel_values.dtype)
+
+
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize(
+    "batch_size",
+    [
+        1,  # Single image
+        3,  # Small batch
+        5,  # Odd batch size (for testing padding)
+    ],
+)
+def test_run_dp_sharded_mrope_vision_model(batch_size: int):
+    world_size = 2
+    # Launch processes
+    mp.spawn(
+        run_dp_sharded_mrope_vision_model_vs_direct,
+        args=(
+            world_size,
+            batch_size,
+            get_open_port(),
+        ),
+        nprocs=world_size,
+    )
+
+
+def run_dp_sharded_mrope_vision_model_vs_direct(local_rank: int,
+                                                world_size: int,
+                                                batch_size: int,
+                                                master_port: int):
+    """
+    Test that run_dp_sharded_mrope_vision_model produces the same results as 
+    calling the model directly.
+    """
+    # Set random seed for reproducibility
+    current_platform.seed_everything(0)
+    device = f"{current_platform.device_name}:{local_rank}"
+    current_platform.set_device(device)
+    torch.set_default_device(device)
+
+    update_environment_variables({
+        'RANK': str(local_rank),
+        'LOCAL_RANK': str(local_rank),
+        'WORLD_SIZE': str(world_size),
+        'MASTER_ADDR': 'localhost',
+        'MASTER_PORT': str(master_port),
+    })
+
+    # initialize distributed
+    init_distributed_environment()
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+
+    # Create test data
+    grid_thw_list = []
+    pixel_values_list = []
+
+    for i in range(batch_size):
+        # Varying image sizes for better testing
+        t, h, w = 1, 4 + i, 4 + i
+        grid_thw_list.append([t, h, w])
+
+        num_patches = t * h * w
+        # Create random pixel values for this image
+        image_pixels = torch.randn(num_patches, 768)
+        pixel_values_list.append(image_pixels)
+
+    # Concatenate all pixel values
+    pixel_values = torch.cat(pixel_values_list, dim=0)
+
+    # Create a simple mrope vision model
+    vision_model = SimpleMRopeVisionModel()
+
+    # Run the model directly on the full input (only on rank 0)
+    if local_rank == 0:
+        with torch.inference_mode():
+            direct_output = vision_model(pixel_values, grid_thw_list)
+
+    # Run the model through the sharded function
+    with torch.inference_mode():
+        sharded_output = run_dp_sharded_mrope_vision_model(vision_model,
+                                                           pixel_values,
+                                                           grid_thw_list,
+                                                           rope_type="rope_3d")
+        sharded_output = torch.cat(sharded_output, dim=0)
+
+    # Check that the world size is set up correctly
+    assert get_tensor_model_parallel_world_size() == world_size
+
+    # Compare outputs (only on rank 0)
+    if local_rank == 0:
+        # Check that the outputs have the same shape
+        assert direct_output.shape == sharded_output.shape
+        # Check that the outputs are close (they should be identical)
+        assert torch.allclose(direct_output,
+                              sharded_output,
+                              rtol=1e-5,
+                              atol=1e-5)
+
+
+@multi_gpu_test(num_gpus=2)
+def test_run_dp_sharded_mrope_vision_model_empty_input():
+    world_size = 2
+    mp.spawn(
+        run_dp_sharded_mrope_vision_model_empty_input_worker,
+        args=(world_size, get_open_port()),
+        nprocs=world_size,
+    )
+
+
+def run_dp_sharded_mrope_vision_model_empty_input_worker(
+        local_rank: int, world_size: int, master_port: int):
+    """Test run_dp_sharded_mrope_vision_model with empty input."""
+    # Set up distributed environment
+    device = f"{current_platform.device_name}:{local_rank}"
+    current_platform.set_device(device)
+    torch.set_default_device(device)
+
+    update_environment_variables({
+        'RANK': str(local_rank),
+        'LOCAL_RANK': str(local_rank),
+        'WORLD_SIZE': str(world_size),
+        'MASTER_ADDR': 'localhost',
+        'MASTER_PORT': str(master_port),
+    })
+
+    init_distributed_environment()
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+
+    # Create empty inputs
+    pixel_values = torch.empty((0, 768))
+    grid_thw_list: list[list[int]] = []
+
+    vision_model = SimpleMRopeVisionModel()
+
+    # Should handle empty input gracefully
+    with torch.inference_mode():
+        output = run_dp_sharded_mrope_vision_model(vision_model,
+                                                   pixel_values,
+                                                   grid_thw_list,
+                                                   rope_type="rope_3d")
+
+    assert len(output) == 0
+
+
+@multi_gpu_test(num_gpus=4)
+def test_run_dp_sharded_mrope_vision_model_uneven_load():
+    world_size = 4
+    mp.spawn(
+        run_dp_sharded_mrope_vision_model_uneven_load_worker,
+        args=(world_size, get_open_port()),
+        nprocs=world_size,
+    )
+
+
+def run_dp_sharded_mrope_vision_model_uneven_load_worker(
+        local_rank: int, world_size: int, master_port: int):
+    """Test run_dp_sharded_mrope_vision_model with uneven load distribution."""
+    # Set up distributed environment
+    current_platform.seed_everything(123)
+    device = f"{current_platform.device_name}:{local_rank}"
+    current_platform.set_device(device)
+    torch.set_default_device(device)
+
+    update_environment_variables({
+        'RANK': str(local_rank),
+        'LOCAL_RANK': str(local_rank),
+        'WORLD_SIZE': str(world_size),
+        'MASTER_ADDR': 'localhost',
+        'MASTER_PORT': str(master_port),
+    })
+
+    init_distributed_environment()
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+
+    # Create images with very different sizes
+    grid_thw_list = [
+        [1, 2, 2],  # Small: 4 patches
+        [1, 8, 8],  # Large: 64 patches  
+        [1, 3, 3],  # Medium: 9 patches
+    ]
+
+    pixel_values_list = []
+    for grid_thw in grid_thw_list:
+        num_patches = math.prod(grid_thw)
+        image_pixels = torch.randn(num_patches, 768)
+        pixel_values_list.append(image_pixels)
+
+    pixel_values = torch.cat(pixel_values_list, dim=0)
+    vision_model = SimpleMRopeVisionModel()
+
+    # Should handle uneven distribution without errors
+    with torch.inference_mode():
+        output_tuple = run_dp_sharded_mrope_vision_model(vision_model,
+                                                         pixel_values,
+                                                         grid_thw_list,
+                                                         rope_type="rope_3d")
+
+    # Verify output shape is reasonable
+    merge_factor = vision_model.spatial_merge_size**2
+    expected_output_patches = list(
+        math.prod(grid_thw) // merge_factor for grid_thw in grid_thw_list)
+
+    for i, output in enumerate(output_tuple):
+        assert output.shape[0] == expected_output_patches[i]
+        assert output.shape[1] == vision_model.out_hidden_size
+
+
+@pytest.mark.parametrize("spatial_merge_size", [2, 4])
+def test_simple_mrope_vision_model_spatial_merge(spatial_merge_size: int):
+    """Test SimpleMRopeVisionModel with different spatial merge sizes."""
+    device = current_platform.device_type
+
+    grid_thw_list = [[1, 4, 4], [1, 6, 6]]  # Two images
+    pixel_values_list = []
+
+    for grid_thw in grid_thw_list:
+        num_patches = math.prod(grid_thw)
+        image_pixels = torch.randn(num_patches, 768, device=device)
+        pixel_values_list.append(image_pixels)
+
+    pixel_values = torch.cat(pixel_values_list, dim=0)
+    vision_model = SimpleMRopeVisionModel(
+        spatial_merge_size=spatial_merge_size).to(device)
+
+    with torch.inference_mode():
+        output = vision_model(pixel_values, grid_thw_list)
+
+    # Verify output dimensions based on spatial merging
+    total_patches = sum(math.prod(grid_thw) for grid_thw in grid_thw_list)
+    merge_factor = spatial_merge_size**2
+    expected_output_patches = total_patches // merge_factor
+
+    assert output.shape[0] == expected_output_patches
+    assert output.shape[1] == vision_model.out_hidden_size

tests/models/utils.py

@@ -11,8 +11,9 @@ import torch.nn.functional as F
 from transformers import PretrainedConfig
 
 from vllm.config import ModelConfig, ModelDType, RunnerOption
-from vllm.inputs import InputContext
 from vllm.logprobs import Logprob, PromptLogprobs, SampleLogprobs
+from vllm.multimodal.processing import InputProcessingContext
+from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
 
 from .registry import HF_EXAMPLE_MODELS
 
@@ -264,7 +265,7 @@ def build_model_context(
     limit_mm_per_prompt: Optional[dict[str, int]] = None,
     mm_processor_cache_gb: int = 0,
 ):
-    """Creates an InputContext for a given model.
+    """Creates an InputProcessingContext for a given model.
 
     Args:
         model_id: ID of the model being considered.
@@ -273,7 +274,7 @@ def build_model_context(
         limit_mm_per_prompt: Multimodal limits.
 
     Returns:
-        InputContext for the model being considered.
+        InputProcessingContext for the model being considered.
     """
     model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
     model_info.check_available_online(on_fail="skip")
@@ -298,7 +299,11 @@ def build_model_context(
         enforce_eager=model_info.enforce_eager,
         **model_config_kwargs,
     )
-    return InputContext(model_config)
+
+    return InputProcessingContext(
+        model_config,
+        tokenizer=cached_tokenizer_from_config(model_config),
+    )
 
 
 def check_embeddings_close(
@@ -423,9 +428,8 @@ def dummy_hf_overrides(
         num_hidden_layers = (3 if model_arch
                              == "Gemma3nForConditionalGeneration" else 1)
 
-    text_config.update({
+    update_dict = {
         "num_layers": num_layers,
-        "num_hidden_layers": num_hidden_layers,
         "num_experts": num_experts,
         "num_experts_per_tok": 2,
         "num_local_experts": num_experts,
@@ -435,7 +439,14 @@ def dummy_hf_overrides(
         "n_routed_experts": num_experts,
         # For Gemma-3n
         "num_kv_shared_layers": 1,
-    })
+    }
+
+    # Update num_hidden_layers for non-Longcat architectures
+    if model_arch != "LongcatFlashForCausalLM" \
+            and model_arch != "LongCatFlashMTPModel":
+        update_dict["num_hidden_layers"] = num_hidden_layers
+
+    text_config.update(update_dict)
 
     if hasattr(hf_config, "vision_config"):
         hf_config.vision_config.update({

tests/multimodal/test_processing.py

@@ -8,11 +8,11 @@ import numpy as np
 import pytest
 
 from vllm.config import ModelConfig
-from vllm.inputs import InputProcessingContext
 from vllm.multimodal import MULTIMODAL_REGISTRY
 # yapf conflicts with isort for this block
 # yapf: disable
-from vllm.multimodal.processing import (PlaceholderFeaturesInfo,
+from vllm.multimodal.processing import (InputProcessingContext,
+                                        PlaceholderFeaturesInfo,
                                         PromptIndexTargets, PromptInsertion,
                                         PromptReplacement, apply_text_matches,
                                         apply_token_matches,