Revert "[V0 deprecation] Remove V0 CPU/XPU/TPU backends (#20412)"

This reverts commit e202dd2736bc575b11250b15311512d19d3225d5.

.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh

@@ -46,6 +46,6 @@ while getopts "m:b:l:f:t:" OPT; do
 done
 
 lm_eval --model vllm \
-  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,trust_remote_code=true,max_model_len=4096" \
+  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,distributed_executor_backend=ray,trust_remote_code=true,max_model_len=4096" \
   --tasks gsm8k --num_fewshot "$FEWSHOT" --limit "$LIMIT" \
   --batch_size "$BATCH_SIZE"

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -18,14 +18,12 @@ RTOL = 0.08
 
 def launch_lm_eval(eval_config, tp_size):
     trust_remote_code = eval_config.get("trust_remote_code", False)
-    max_model_len = eval_config.get("max_model_len", 4096)
     model_args = (
         f"pretrained={eval_config['model_name']},"
         f"tensor_parallel_size={tp_size},"
         f"enforce_eager=true,"
         f"add_bos_token=true,"
-        f"trust_remote_code={trust_remote_code},"
-        f"max_model_len={max_model_len}"
+        f"trust_remote_code={trust_remote_code}"
     )
     results = lm_eval.simple_evaluate(
         model="vllm",

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

@@ -108,6 +108,7 @@ fi
 if [[ $commands == *" kernels/attention"* ]]; then
   commands="${commands} \
   --ignore=kernels/attention/test_attention_selector.py \
+  --ignore=kernels/attention/test_blocksparse_attention.py \
   --ignore=kernels/attention/test_encoder_decoder_attn.py \
   --ignore=kernels/attention/test_flash_attn.py \
   --ignore=kernels/attention/test_flashinfer.py \

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

@@ -6,7 +6,6 @@ set -ex
 
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
-# used for TP/PP E2E test
 OMP_CORE_RANGE=${OMP_CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
@@ -25,8 +24,8 @@ numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE
 numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
 
 function cpu_tests() {
   set -e
@@ -49,16 +48,10 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    # Note: disable until supports V1
-    # pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
-    # pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
-
-    # Note: disable Bart until supports V1
-    pytest -v -s tests/models/language/generation -m cpu_model \
-                --ignore=tests/models/language/generation/test_bart.py
-    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model \
-                --ignore=tests/models/language/generation/test_bart.py
-
+    pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
+    pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
+    pytest -v -s tests/models/language/generation -m cpu_model
+    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model
     pytest -v -s tests/models/language/pooling -m cpu_model
     pytest -v -s tests/models/multimodal/generation \
                 --ignore=tests/models/multimodal/generation/test_mllama.py \
@@ -69,26 +62,33 @@ function cpu_tests() {
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs[False-10-32-neuralmagic/Llama-3.2-1B-quantized.w8a8]" 
+    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
+    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
 
-  # Note: disable it until supports V1
   # Run AWQ test
-  # docker exec cpu-test-"$NUMA_NODE" bash -c "
-  #   set -e
-  #   VLLM_USE_V1=0 pytest -s -v \
-  #   tests/quantization/test_ipex_quant.py"
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    VLLM_USE_V1=0 pytest -s -v \
+    tests/quantization/test_ipex_quant.py"
+
+  # Run chunked-prefill and prefix-cache test
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -s -v -k cpu_model \
+    tests/basic_correctness/test_chunked_prefill.py"  
 
   # online serving
-  docker exec cpu-test-"$NUMA_NODE" bash -c '
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 &
-    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
-    python3 benchmarks/benchmark_serving.py \
+    python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --dtype half & 
+    timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
+    VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \
       --backend vllm \
       --dataset-name random \
-      --model meta-llama/Llama-3.2-3B-Instruct \
+      --model facebook/opt-125m \
       --num-prompts 20 \
-      --endpoint /v1/completions'
+      --endpoint /v1/completions \
+      --tokenizer facebook/opt-125m"
 
   # Run multi-lora tests
   docker exec cpu-test-"$NUMA_NODE" bash -c "

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

@@ -6,17 +6,19 @@ set -exuo pipefail
 
 # Try building the docker image
 cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
-FROM gaudi-base-image:latest
+FROM 1.22-413-pt2.7.1:latest
 
 COPY ./ /workspace/vllm
 
 WORKDIR /workspace/vllm
 
+RUN pip install -v -r requirements/hpu.txt
+RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+
 ENV no_proxy=localhost,127.0.0.1
 ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
 
-RUN VLLM_TARGET_DEVICE=empty pip install .
-RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+RUN VLLM_TARGET_DEVICE=hpu python3 setup.py install
 
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils

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

@@ -62,8 +62,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4 \
-    && python3 -m pip install --progress-bar off hf-transfer
+    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1
@@ -71,7 +70,7 @@ export VLLM_XLA_CACHE_PATH=
 echo "Using VLLM V1"
 
 echo "--- Hardware Information ---"
-# tpu-info
+tpu-info
 echo "--- Starting Tests ---"
 set +e
 overall_script_exit_code=0
@@ -151,7 +150,7 @@ run_and_track_test 9 "test_multimodal.py" \
 run_and_track_test 10 "test_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py"
 run_and_track_test 11 "test_struct_output_generate.py" \
-    "HF_HUB_DISABLE_XET=1 python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
 run_and_track_test 12 "test_moe_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
 run_and_track_test 13 "test_lora.py" \

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

@@ -11,8 +11,8 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head
 docker build -t ${image_name} -f docker/Dockerfile.xpu .
 
 # Setup cleanup
-remove_docker_container() {
-  docker rm -f "${container_name}" || true;
+remove_docker_container() { 
+  docker rm -f "${container_name}" || true; 
   docker image rm -f "${image_name}" || true;
   docker system prune -f || true;
 }
@@ -26,18 +26,7 @@ docker run \
     --name "${container_name}" \
     "${image_name}" \
     sh -c '
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
     VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
-    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    cd tests
-    pytest -v -s v1/core
-    pytest -v -s v1/engine
-    pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
-    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
-    pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py
-    pytest -v -s v1/test_serial_utils.py
-    pytest -v -s v1/test_utils.py
-    pytest -v -s v1/test_metrics_reader.py
 '

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -22,6 +22,16 @@ trap remove_docker_container EXIT
 # Remove the container that might not be cleaned up in the previous run.
 remove_docker_container
 
+# Build docker image.
+# TODO: build the image outside the script and share the image with other
+# tpu test if building time is too long.
+DOCKER_BUILDKIT=1 docker build \
+  --build-arg max_jobs=16 \
+  --build-arg USE_SCCACHE=1 \
+  --build-arg GIT_REPO_CHECK=0 \
+  --tag vllm/vllm-tpu-bm \
+  --progress plain -f docker/Dockerfile.tpu .
+
 LOG_ROOT=$(mktemp -d)
 # If mktemp fails, set -e will cause the script to exit.
 echo "Results will be stored in: $LOG_ROOT"

.buildkite/test-pipeline.yaml

@@ -117,7 +117,7 @@ steps:
   commands:
   - pytest -v -s core
 
-- label: Entrypoints Test (LLM) # 40min
+- label: Entrypoints Test # 40min
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -125,6 +125,8 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -133,21 +135,9 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
-
-- label: Entrypoints Test (API Server) # 40min
-  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/tests"
-  fast_check: true
-  torch_nightly: true
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   mirror_hardwares: [amdexperimental]
@@ -159,14 +149,13 @@ steps:
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
+  - tests/spec_decode/e2e/test_integration_dist_tp4
   - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/test_async_llm_dp.py
   - tests/v1/test_external_lb_dp.py
-  - tests/v1/test_internal_lb_dp.py
-  - tests/v1/test_hybrid_lb_dp.py
   - tests/v1/engine/test_engine_core_client.py
   commands:
   # test with tp=2 and external_dp=2
@@ -178,13 +167,12 @@ steps:
   - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/test_internal_lb_dp.py
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/test_hybrid_lb_dp.py
   - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
   - 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 spec_decode/e2e/test_integration_dist_tp4.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
@@ -229,7 +217,7 @@ steps:
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -268,7 +256,6 @@ steps:
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/spec_decode
     - pytest -v -s v1/kv_connector/unit
-    - pytest -v -s v1/metrics
     - pytest -v -s v1/test_serial_utils.py
     - pytest -v -s v1/test_utils.py
     - pytest -v -s v1/test_oracle.py
@@ -277,11 +264,11 @@ steps:
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
     - pytest -v -s v1/e2e
     # Integration test for streaming correctness (requires special branch).
-    - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
+    - pip install -U git+https://github.com/robertgshaw2-neuralmagic/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: Examples Test # 25min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
@@ -295,7 +282,7 @@ steps:
     - python3 offline_inference/llm_engine_example.py
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_pooling.py --seed 0
+    - python3 offline_inference/vision_language_embedding.py --seed 0
     - python3 offline_inference/vision_language_multi_image.py --seed 0
     - VLLM_USE_V1=0 python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
     - python3 offline_inference/encoder_decoder.py
@@ -315,7 +302,7 @@ steps:
 
 
 - label: Platform Tests (CUDA)
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/cuda
@@ -333,9 +320,20 @@ steps:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
-- label: LoRA Test %N # 15min each
+- label: Speculative decoding tests # 40min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
+  - vllm/spec_decode
+  - tests/spec_decode
+  - vllm/model_executor/models/eagle.py
+  commands:
+    - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
+    - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py --ignore=spec_decode/e2e/test_mtp_correctness.py
+    - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
+
+- label: LoRA Test %N # 15min each
+  mirror_hardwares: [amdexperimental, amdproduction]
+  source_file_dependencies:
   - vllm/lora
   - tests/lora
   command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
@@ -386,7 +384,7 @@ steps:
     - pytest -v -s kernels/core
 
 - label: Kernels Attention Test %N
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - csrc/attention/
   - vllm/attention
@@ -397,7 +395,7 @@ steps:
   parallelism: 2
 
 - label: Kernels Quantization Test %N
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - csrc/quantization/
   - vllm/model_executor/layers/quantization
@@ -416,7 +414,7 @@ steps:
     - pytest -v -s kernels/moe
 
 - label: Kernels Mamba Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -424,7 +422,7 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -438,6 +436,7 @@ steps:
 
 - label: Model Executor Test
   mirror_hardwares: [amdexperimental, amdproduction]
+  soft_fail: true
   source_file_dependencies:
   - vllm/model_executor
   - tests/model_executor
@@ -494,7 +493,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/
 
 - label: Encoder Decoder tests # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/encoder_decoder
@@ -502,7 +501,7 @@ steps:
     - pytest -v -s encoder_decoder
 
 - label: OpenAI-Compatible Tool Use # 20 min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: false
   source_file_dependencies:
     - vllm/
@@ -614,7 +613,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
 
 - label: Quantized Models Test
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization
@@ -631,18 +630,6 @@ steps:
     # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
     # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*
 
-- label: Transformers Nightly Models Test
-  working_dir: "/vllm-workspace/"
-  optional: true
-  commands:
-    - pip install --upgrade git+https://github.com/huggingface/transformers
-    - pytest -v -s tests/models/test_initialization.py
-    - 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
-
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -717,10 +704,10 @@ steps:
   - pytest -v -s distributed/test_sequence_parallel.py
   # this test fails consistently.
   # TODO: investigate and fix
+  # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
-  - pytest -v -s models/multimodal/generation/test_maverick.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   mirror_hardwares: [amdexperimental]

.gemini/config.yaml

@@ -1,6 +0,0 @@
-# https://developers.google.com/gemini-code-assist/docs/customize-gemini-behavior-github
-have_fun: false  # Just review the code
-code_review:
-  comment_severity_threshold: HIGH  # Reduce quantity of comments
-  pull_request_opened:
-    summary: false  # Don't summarize the PR in a separate comment

.github/CODEOWNERS

@@ -16,7 +16,6 @@
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm
 /vllm/entrypoints @aarnphm
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
@@ -43,6 +42,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/multimodal @DarkLight1337 @ywang96
 /tests/prefix_caching @comaniac @KuntaiDu
 /tests/quantization @mgoin @robertgshaw2-redhat
+/tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -46,7 +46,7 @@ body:
 - type: markdown
   attributes:
     value: >
-      Thanks for contributing 🎉! The vLLM core team hosts a biweekly RFC review session at 9:30AM Pacific Time, while most RFCs can be discussed online, you can optionally sign up for a slot to discuss your RFC online [here](https://docs.google.com/document/d/1CiLVBZeIVfR7_PNAKVSusxpceywkoOOB78qoWqHvSZc/edit).
+      Thanks for contributing 🎉!
 - type: checkboxes
   id: askllm
   attributes:

.github/mergify.yml

@@ -86,6 +86,8 @@ pull_request_rules:
     - and:
       - files~=^vllm/model_executor/models/
       - files=vllm/model_executor/models/registry.py
+      - files=tests/models/registry.py
+      - files=docs/models/supported_models.md
   actions:
     label:
       add:
@@ -164,7 +166,10 @@ pull_request_rules:
   description: Automatically apply speculative-decoding label
   conditions:
     - or:
+      - files~=^vllm/spec_decode/
       - files~=^vllm/v1/spec_decode/
+      - files=vllm/model_executor/layers/spec_decode_base_sampler.py
+      - files~=^tests/spec_decode/
       - files~=^tests/v1/spec_decode/
       - files~=^examples/.*(spec_decode|mlpspeculator|eagle|speculation).*\.py
       - files~=^vllm/model_executor/models/.*eagle.*\.py

.gitignore

@@ -146,7 +146,6 @@ venv.bak/
 
 # mkdocs documentation
 /site
-docs/argparse
 docs/examples
 
 # mypy

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   - id: ruff-format
     files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/crate-ci/typos
-  rev: v1.34.0
+  rev: v1.32.0
   hooks:
   - id: typos
 - repo: https://github.com/PyCQA/isort
@@ -166,11 +166,11 @@ repos:
     language: python
     types: [python]
     pass_filenames: true
-    files: vllm/config.py|tests/test_config.py|vllm/entrypoints/openai/cli_args.py
+    files: vllm/config.py|tests/test_config.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion
-    entry: bash -c 'echo "To bypass all the pre-commit hooks, add --no-verify to git commit. To skip a specific hook, prefix the commit command with SKIP=<hook-id>."'
+    entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
     language: system
     verbose: true
     pass_filenames: false

CMakeLists.txt

@@ -45,7 +45,7 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.7.1")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.7.0")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
 
 #
@@ -171,6 +171,7 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
+
 #
 # Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
 # setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.
@@ -231,6 +232,7 @@ endif()
 
 set(VLLM_EXT_SRC
   "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
+  "csrc/mamba/causal_conv1d/causal_conv1d.cu"
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
@@ -296,8 +298,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
     "csrc/cutlass_extensions/common.cpp"
-    "csrc/attention/mla/cutlass_mla_entry.cu"
-    "csrc/quantization/fp8/per_token_group_quant.cu")
+    "csrc/attention/mla/cutlass_mla_entry.cu")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -392,7 +393,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.0 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
        "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
@@ -408,7 +409,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
@@ -423,7 +424,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.8 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
@@ -437,7 +438,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
                      "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                      "later if you intend on running FP8 quantized models on "
@@ -452,7 +453,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
@@ -467,7 +468,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
                      "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                      "later if you intend on running FP8 quantized models on "
@@ -510,7 +511,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
   # require CUDA 12.2 or later (and only work on Hopper).
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -519,7 +520,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
     message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -531,7 +532,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # FP4 Archs and flags
   cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
       "csrc/quantization/fp4/nvfp4_experts_quant.cu"
@@ -552,10 +553,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # CUTLASS MLA Archs and flags
   cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
     set(SRCS
-      "csrc/attention/mla/cutlass_mla_kernels.cu"
-      "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
+      "csrc/attention/mla/cutlass_mla_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${MLA_ARCHS}")
@@ -578,7 +578,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -596,26 +596,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
-    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
-                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
-                     "if you intend on running FP8 quantized MoE models on Blackwell.")
-    else()
-      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
-                     "in CUDA target architectures.")
-    endif()
-  endif()
-
   # moe_data.cu is used by all CUTLASS MoE kernels.
   cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
@@ -662,7 +642,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The machete kernels only work on hopper and require CUDA 12.0 or later.
   # Only build Machete kernels if we are building for something compatible with sm90a
   cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND MACHETE_ARCHS)
     #
     # For the Machete kernels we automatically generate sources for various
     # preselected input type pairs and schedules.
@@ -714,7 +694,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
     message(STATUS "Building Machete kernels for archs: ${MACHETE_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
         AND MACHETE_ARCHS)
       message(STATUS "Not building Machete kernels as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "

README.md

@@ -63,11 +63,13 @@ vLLM is fast with:
 - Speculative decoding
 - Chunked prefill
 
+**Performance benchmark**: We include a performance benchmark at the end of [our blog post](https://blog.vllm.ai/2024/09/05/perf-update.html). It compares the performance of vLLM against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [SGLang](https://github.com/sgl-project/sglang) and [LMDeploy](https://github.com/InternLM/lmdeploy)). The implementation is under [nightly-benchmarks folder](.buildkite/nightly-benchmarks/) and you can [reproduce](https://github.com/vllm-project/vllm/issues/8176) this benchmark using our one-click runnable script.
+
 vLLM is flexible and easy to use with:
 
 - Seamless integration with popular Hugging Face models
 - High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
-- Tensor, pipeline, data and expert parallelism support for distributed inference
+- Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
 - Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron

RELEASE.md

@@ -52,36 +52,3 @@ After branch cut, we approach finalizing the release branch with clear criteria
 * Release branch specific changes (e.g. change version identifiers or CI fixes)
 
 Please note: **No feature work allowed for cherry picks**. All PRs that are considered for cherry-picks need to be merged on trunk, the only exception are Release branch specific changes.
-
-## Manual validations
-
-### E2E Performance Validation
-
-Before each release, we perform end-to-end performance validation to ensure no regressions are introduced. This validation uses the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) on PyTorch CI.
-
-**Current Coverage:**
-* Models: Llama3, Llama4, and Mixtral
-* Hardware: NVIDIA H100 and AMD MI300x
-* *Note: Coverage may change based on new model releases and hardware availability*
-
-**Performance Validation Process:**
-
-**Step 1: Get Access**
-Request write access to the [pytorch/pytorch-integration-testing](https://github.com/pytorch/pytorch-integration-testing) repository to run the benchmark workflow.
-
-**Step 2: Review Benchmark Setup**
-Familiarize yourself with the benchmark configurations:
-* [CUDA setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/cuda)
-* [ROCm setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/rocm)
-
-**Step 3: Run the Benchmark**
-Navigate to the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) and configure:
-* **vLLM branch**: Set to the release branch (e.g., `releases/v0.9.2`)
-* **vLLM commit**: Set to the RC commit hash
-
-**Step 4: Review Results**
-Once the workflow completes, benchmark results will be available on the [vLLM benchmark dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm) under the corresponding branch and commit.
-
-**Step 5: Performance Comparison**
-Compare the current results against the previous release to verify no performance regressions have occurred. Here is an
-example of [v0.9.1 vs v0.9.2](https://hud.pytorch.org/benchmark/llms?startTime=Thu%2C%2017%20Apr%202025%2021%3A43%3A50%20GMT&stopTime=Wed%2C%2016%20Jul%202025%2021%3A43%3A50%20GMT&granularity=week&lBranch=releases/v0.9.1&lCommit=b6553be1bc75f046b00046a4ad7576364d03c835&rBranch=releases/v0.9.2&rCommit=a5dd03c1ebc5e4f56f3c9d3dc0436e9c582c978f&repoName=vllm-project%2Fvllm&benchmarkName=&modelName=All%20Models&backendName=All%20Backends&modeName=All%20Modes&dtypeName=All%20DType&deviceName=All%20Devices&archName=All%20Platforms).

benchmarks/auto_tune.sh

@@ -1,7 +1,36 @@
 #!/bin/bash
 
 # This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
-# See details in README (benchmarks/auto_tune/README.md).
+# The current server parameter combination is  max_num_seqs and max_num_batched_tokens
+# It also supports additional requirement: e2e latency and prefix cache. 
+
+# Pre-requisite:
+# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. 
+# 2. If the model is customized, replace the MODEL's config with the customized config.
+# 3. Set variables (ALL REQUIRED)
+#   BASE: your directory for vllm repo
+#   MODEL: the model served by vllm
+#   SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support.
+#   TP: ways of tensor parallelism
+#   DOWNLOAD_DIR: directory to download and load model weights.
+#   INPUT_LEN: request input len
+#   OUTPUT_LEN: request output len
+#   MIN_CACHE_HIT_PCT: prefix cache rate
+#   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
+#   NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with.
+#   NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with.
+#   Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST.
+# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
+# 5. The final result will be saved in RESULT file. 
+
+
+# Example use cases 
+# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000
+# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500
+# 3. If we want to reach 60% prefix cache, what's the best server parameter? 
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500
 
 TAG=$(date +"%Y_%m_%d_%H_%M")
 BASE=""
@@ -126,12 +155,11 @@ run_benchmark() {
     # get a basic qps by using request-rate inf
     bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
     prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
-adjusted_input_len=$(( INPUT_LEN - prefix_len ))
-    python3 benchmarks/benchmark_serving.py \
+    python benchmarks/benchmark_serving.py \
         --backend vllm \
         --model $MODEL  \
         --dataset-name random \
-        --random-input-len $adjusted_input_len \
+        --random-input-len $INPUT_LEN \
         --random-output-len $OUTPUT_LEN \
         --ignore-eos \
         --disable-tqdm \
@@ -160,11 +188,11 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
             curl -X POST http://0.0.0.0:8004/reset_prefix_cache
             sleep 5
             bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
-            python3 benchmarks/benchmark_serving.py \
+            python benchmarks/benchmark_serving.py \
                 --backend vllm \
                 --model $MODEL  \
                 --dataset-name random \
-                --random-input-len $adjusted_input_len \
+                --random-input-len $INPUT_LEN \
                 --random-output-len $OUTPUT_LEN \
                 --ignore-eos \
                 --disable-tqdm \

benchmarks/auto_tune/README.md

@@ -1,137 +0,0 @@
-# Automated vLLM Server Parameter Tuning
-
-This script automates the process of finding the optimal server parameter combination (`max-num-seqs` and `max-num-batched-tokens`) to maximize throughput for a vLLM server. It also supports additional constraints such as E2E latency and prefix cache hit rate.
-
-## Table of Contents
-- [Prerequisites](#prerequisites)
-- [Configuration](#configuration)
-- [How to Run](#how-to-run)
-- [Example Use Cases](#example-use-cases)
-- [Output](#output)
-- [How It Works](#how-it-works)
-
-## Prerequisites
-
-Before running the script, please ensure the following steps are completed:
-
-1. **Clone vLLM & Set Up Branch**: Clone the vLLM repository and check out to your desired branch.
-
-```bash
-git clone https://github.com/vllm-project/vllm.git
-cd vllm
-# git checkout <your-branch>
-```
-
-1. **Install Environment**: Install or update the correct running environment. For TPU usage, activate your `conda` environment and install the corresponding `torch` and `torch_xla` versions.
-
-2. **Model Configuration**: If you are using a customized model, ensure its configuration files are correctly placed and accessible.
-
-## Configuration
-
-You must set the following variables at the top of the script before execution.
-
-| Variable | Description | Example Value |
-| --- | --- | --- |
-| `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
-| `MODEL` | **Required.** The Hugging Face model identifier to be served by vllm. | `"meta-llama/Llama-3.1-8B-Instruct"` |
-| `SYSTEM`| **Required.** The hardware you are running on. Choices: `TPU` or `GPU`. (For other systems, it might not support saving profiles) | `"TPU"` |
-| `TP` | **Required.** The tensor-parallelism size. | `1` |
-| `DOWNLOAD_DIR` | **Required.** Directory to download and load model weights from. | `""` (default download path) |
-| `INPUT_LEN` | **Required.** Request input length. | `4000` |
-| `OUTPUT_LEN` | **Required.** Request output length. | `16` |
-| `MIN_CACHE_HIT_PCT` | Prefix cache hit rate in percentage (0-100). Set to `0` to disable. | `60` |
-| `MAX_LATENCY_ALLOWED_MS` | The maximum allowed P99 end-to-end latency in milliseconds. Set to a very large number (e.g., `100000000000`) to effectively ignore the latency constraint. | `500` |
-| `NUM_SEQS_LIST` | A space-separated string of `max-num-seqs` values to test. | `"128 256"` |
-| `NUM_BATCHED_TOKENS_LIST` | A space-separated string of `max-num-batched-tokens` values to test. | `"1024 2048 4096"` |
-
-**Note**: The default `NUM_SEQS_LIST` and `NUM_BATCHED_TOKENS_LIST` are set for medium-sized inputs/outputs. For very short contexts (e.g., 20 input, 20 output tokens), you may need to test larger values for `max-num-seqs`.
-
-## How to Run
-
-1. **Configure**: Edit the script and set the variables in the [Configuration](#configuration) section.
-2. **Execute**: Run the script. Since the process can take a long time, it is highly recommended to use a terminal multiplexer like `tmux` or `screen` to prevent the script from stopping if your connection is lost.
-
-```
-cd <FOLDER_OF_THIS_SCRIPT>
-bash auto_tune.sh
-```
-
-    Please note that the `bash auto_tune.sh` command cannot contain full or partial path with keyword `vllm`, otherwise `pkill -f vllm` command will also kill this script itself.
-
-## Example Use Cases
-
-Here are a few examples of how to configure the script for different goals:
-
-### 1. Maximize Throughput (No Latency Constraint)
-- **Goal**: Find the best `max-num-seqs` and `max-num-batched-tokens` to get the highest possible throughput for 1800 input tokens and 20 output tokens.
-- **Configuration**:
-
-```bash
-INPUT_LEN=1800
-OUTPUT_LEN=20
-MIN_CACHE_HIT_PCT=0
-MAX_LATENCY_ALLOWED_MS=100000000000 # A very large number
-```
-
-#### 2. Maximize Throughput with a Latency Requirement
-- **Goal**: Find the best server parameters when P99 end-to-end latency must be below 500ms.
-- **Configuration**:
-
-```bash
-INPUT_LEN=1800
-OUTPUT_LEN=20
-MIN_CACHE_HIT_PCT=0
-MAX_LATENCY_ALLOWED_MS=500
-```
-
-#### 3. Maximize Throughput with Prefix Caching and Latency Requirements
-- **Goal**: Find the best server parameters assuming a 60% prefix cache hit rate and a latency requirement of 500ms.
-- **Configuration**:
-
-```bash
-INPUT_LEN=1800
-OUTPUT_LEN=20
-MIN_CACHE_HIT_PCT=60
-MAX_LATENCY_ALLOWED_MS=500
-```
-
-## Output
-
-After the script finishes, you will find the results in a new, timestamped directory created inside `$BASE/auto-benchmark/`.
-
-- **Log Files**: The directory (`$BASE/auto-benchmark/YYYY_MM_DD_HH_MM/`) contains detailed logs for each run:
-    - `vllm_log_...txt`: The log output from the vLLM server for each parameter combination.
-    - `bm_log_...txt`: The log output from the `benchmark_serving.py` script for each benchmark run.
-
-- **Final Result Summary**: A file named `result.txt` is created in the log directory. It contains a summary of each tested combination and concludes with the overall best parameters found.
-
-```
-# Example result.txt content
-hash:a1b2c3d4...
-max_num_seqs: 128, max_num_batched_tokens: 2048, request_rate: 10.0, e2el: 450.5, throughput: 9.8, goodput: 9.8
-max_num_seqs: 128, max_num_batched_tokens: 4096 does not meet latency requirement 500
-...
-best_max_num_seqs: 256, best_num_batched_tokens: 2048, best_throughput: 12.5, profile saved in: /home/user/vllm/auto-benchmark/2024_08_01_10_30/profile
-```
-
-  If it cannot find the best parameters, the final row will be `best_max_num_seqs: 0, best_num_batched_tokens: 0, best_throughput: 0`. This can be due to either the server not starting properly, or the latency requirement being too strict.
-
-- **Profiler Trace**: A directory named `profile` is created inside the log directory. It contains the profiler trace file (e.g., `.xplane.pb` for TPU or a `.json` trace for GPU) from the single best-performing run.
-
-## How It Works
-
-The script follows a systematic process to find the optimal parameters:
-
-1. **Find Max GPU Memory Utilization**: The script first determines the highest safe `gpu-memory-utilization` (starting from 0.98 and decreasing) that does not cause an Out-Of-Memory (OOM) error when launching the server. This ensures the benchmark runs use the maximum available memory without crashing.
-
-2. **Iterate and Benchmark**: It then enters a nested loop, iterating through every combination of `max-num-seqs` and `max-num-batched-tokens` provided in the configuration lists.
-
-3. **Latency-Aware Throughput Search**: For each parameter combination:
-    - The vLLM server is started.
-    - A benchmark is first run with an infinite request rate (`--request-rate inf`).
-    - If the resulting P99 E2E latency is within the `MAX_LATENCY_ALLOWED_MS` limit, this throughput is considered the maximum for this configuration.
-    - If the latency is too high, the script performs a search by iteratively decreasing the request rate until the latency constraint is met. This finds the highest sustainable throughput for the given parameters and latency requirement.
-
-4. **Track Best Result**: Throughout the process, the script tracks the parameter combination that has yielded the highest valid throughput so far.
-
-5. **Profile Collection**: For the best-performing run, the script saves the vLLM profiler output, which can be used for deep-dive performance analysis with tools like TensorBoard.

benchmarks/benchmark_dataset.py

@@ -324,9 +324,6 @@ class RandomDataset(BenchmarkDataset):
         input_low = int(real_input_len * (1 - range_ratio))
         input_high = int(real_input_len * (1 + range_ratio))
         output_low = int(output_len * (1 - range_ratio))
-        # Ensure the lower bound for output length is at least 1 to prevent
-        # sampling 0 tokens, which can cause request failures.
-        output_low = max(output_low, 1)
         output_high = int(output_len * (1 + range_ratio))
 
         # Add logging for debugging
@@ -704,7 +701,6 @@ class HuggingFaceDataset(BenchmarkDataset):
         self,
         dataset_path: str,
         dataset_split: str,
-        no_stream: bool = False,
         dataset_subset: Optional[str] = None,
         **kwargs,
     ) -> None:
@@ -712,7 +708,6 @@ class HuggingFaceDataset(BenchmarkDataset):
 
         self.dataset_split = dataset_split
         self.dataset_subset = dataset_subset
-        self.load_stream = not no_stream
         self.load_data()
 
     def load_data(self) -> None:
@@ -721,7 +716,7 @@ class HuggingFaceDataset(BenchmarkDataset):
             self.dataset_path,
             name=self.dataset_subset,
             split=self.dataset_split,
-            streaming=self.load_stream,
+            streaming=True,
         )
         self.data = self.data.shuffle(seed=self.random_seed)
 

benchmarks/benchmark_serving.py

@@ -30,7 +30,7 @@ import os
 import random
 import time
 import warnings
-from collections.abc import Iterable
+from collections.abc import AsyncGenerator, Iterable
 from dataclasses import dataclass
 from datetime import datetime
 from typing import Any, Literal, Optional
@@ -73,7 +73,6 @@ from benchmark_dataset import (
     VisionArenaDataset,
 )
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
-from vllm.benchmarks.serve import get_request
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
@@ -108,6 +107,101 @@ class BenchmarkMetrics:
     percentiles_e2el_ms: list[tuple[float, float]]
 
 
+def _get_current_request_rate(
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]],
+    ramp_up_start_rps: Optional[int],
+    ramp_up_end_rps: Optional[int],
+    request_index: int,
+    total_requests: int,
+    request_rate: float,
+) -> float:
+    if (
+        ramp_up_strategy
+        and ramp_up_start_rps is not None
+        and ramp_up_end_rps is not None
+    ):
+        progress = request_index / max(total_requests - 1, 1)
+        if ramp_up_strategy == "linear":
+            increase = (ramp_up_end_rps - ramp_up_start_rps) * progress
+            return ramp_up_start_rps + increase
+        elif ramp_up_strategy == "exponential":
+            ratio = ramp_up_end_rps / ramp_up_start_rps
+            return ramp_up_start_rps * (ratio**progress)
+        else:
+            raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}")
+    return request_rate
+
+
+async def get_request(
+    input_requests: list[SampleRequest],
+    request_rate: float,
+    burstiness: float = 1.0,
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
+    ramp_up_start_rps: Optional[int] = None,
+    ramp_up_end_rps: Optional[int] = None,
+) -> AsyncGenerator[tuple[SampleRequest, float], None]:
+    """
+    Asynchronously generates requests at a specified rate
+    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
+
+    Args:
+        input_requests:
+            A list of input requests, each represented as a SampleRequest.
+        request_rate:
+            The rate at which requests are generated (requests/s).
+        burstiness (optional):
+            The burstiness factor of the request generation.
+            Only takes effect when request_rate is not inf.
+            Default value is 1, which follows a Poisson process.
+            Otherwise, the request intervals follow a gamma distribution.
+            A lower burstiness value (0 < burstiness < 1) results
+            in more bursty requests, while a higher burstiness value
+            (burstiness > 1) results in a more uniform arrival of requests.
+         ramp_up_strategy (optional):
+            The ramp-up strategy. Can be "linear" or "exponential".
+            If None, uses constant request rate (specified by request_rate).
+        ramp_up_start_rps (optional):
+            The starting request rate for ramp-up.
+        ramp_up_end_rps (optional):
+            The ending request rate for ramp-up.
+    """
+    assert burstiness > 0, (
+        f"A positive burstiness factor is expected, but given {burstiness}."
+    )
+    # Convert to list to get length for ramp-up calculations
+    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
+        input_requests = list(input_requests)
+
+    total_requests = len(input_requests)
+    request_index = 0
+
+    for request in input_requests:
+        current_request_rate = _get_current_request_rate(
+            ramp_up_strategy,
+            ramp_up_start_rps,
+            ramp_up_end_rps,
+            request_index,
+            total_requests,
+            request_rate,
+        )
+
+        yield request, current_request_rate
+
+        request_index += 1
+
+        if current_request_rate == float("inf"):
+            # If the request rate is infinity, then we don't need to wait.
+            continue
+
+        theta = 1.0 / (current_request_rate * burstiness)
+
+        # Sample the request interval from the gamma distribution.
+        # If burstiness is 1, it follows exponential distribution.
+        interval = np.random.gamma(shape=burstiness, scale=theta)
+        # The next request will be sent after the interval.
+        await asyncio.sleep(interval)
+
+
 def calculate_metrics(
     input_requests: list[SampleRequest],
     outputs: list[RequestFuncOutput],
@@ -731,7 +825,6 @@ def main(args: argparse.Namespace):
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
             random_seed=args.seed,
-            no_stream=args.no_stream,
         ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
@@ -940,11 +1033,6 @@ def create_argument_parser():
         help="Path to the sharegpt/sonnet dataset. "
         "Or the huggingface dataset ID if using HF dataset.",
     )
-    parser.add_argument(
-        "--no-stream",
-        action="store_true",
-        help="Do not load the dataset in streaming mode.",
-    )
     parser.add_argument(
         "--max-concurrency",
         type=int,

benchmarks/benchmark_throughput.py

@@ -356,7 +356,6 @@ def get_requests(args, tokenizer):
     elif args.dataset_name == "burstgpt":
         dataset_cls = BurstGPTDataset
     elif args.dataset_name == "hf":
-        common_kwargs["no_stream"] = args.no_stream
         if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
             dataset_cls = VisionArenaDataset
             common_kwargs["dataset_subset"] = None
@@ -611,11 +610,6 @@ def create_argument_parser():
         help="Name of the dataset to benchmark on.",
         default="sharegpt",
     )
-    parser.add_argument(
-        "--no-stream",
-        action="store_true",
-        help="Do not load the dataset in streaming mode.",
-    )
     parser.add_argument(
         "--dataset",
         type=str,

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -1,141 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import argparse
-import copy
-import itertools
-
-import torch
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm import _custom_ops as ops
-from vllm.platforms import current_platform
-from vllm.scalar_type import scalar_types
-from vllm.triton_utils import triton
-
-if not current_platform.has_device_capability(100):
-    raise RuntimeError("NVFP4 requires compute capability of 10.0 (Blackwell)")
-
-
-FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
-FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
-
-PROVIDER_CFGS = {
-    "torch-bf16": dict(enabled=True),
-    "nvfp4": dict(no_a_quant=False, enabled=True),
-    "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
-}
-
-_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
-
-
-def _quant_weight_nvfp4(b: torch.Tensor, device: str):
-    # 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)
-    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)
-
-    # Compute global scale for activation
-    # NOTE: This is generally provided ahead-of-time by the model checkpoint.
-    a_amax = torch.abs(a).max().to(torch.float32)
-    a_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / a_amax
-
-    # Alpha for the GEMM operation
-    alpha = 1.0 / (a_global_scale * b_global_scale)
-
-    if cfg["no_a_quant"]:
-        # Pre-quantize activation
-        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
-
-        def run():
-            return ops.cutlass_scaled_fp4_mm(
-                a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
-            )
-
-        return run
-
-    # Quantize activation on-the-fly
-    def run():
-        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
-        return ops.cutlass_scaled_fp4_mm(
-            a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
-        )
-
-    return run
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
-        x_log=False,
-        line_arg="provider",
-        line_vals=_enabled,
-        line_names=_enabled,
-        ylabel="TFLOP/s (larger is better)",
-        plot_name="BF16 vs NVFP4 GEMMs",
-        args={},
-    )
-)
-def benchmark(batch_size, provider, N, K):
-    M = batch_size
-    device = "cuda"
-    dtype = torch.bfloat16
-
-    a = torch.randn((M, K), device=device, dtype=dtype)
-    b = torch.randn((N, K), device=device, dtype=dtype)
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch-bf16":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_nvfp4_runner(cfg, a, b, dtype, device)
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), quantiles=quantiles
-        )
-
-    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
-    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
-
-
-def prepare_shapes(args):
-    out = []
-    for model, tp_size in itertools.product(args.models, args.tp_sizes):
-        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
-            KN[tp_dim] //= tp_size
-            KN.append(model)
-            out.append(KN)
-    return out
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=["meta-llama/Llama-3.1-8B-Instruct"],
-        choices=list(WEIGHT_SHAPES.keys()),
-    )
-    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
-    args = parser.parse_args()
-
-    for K, N, model in prepare_shapes(args):
-        print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
-        benchmark.run(
-            print_data=True,
-            show_plots=True,
-            save_path=f"bench_nvfp4_res_n{N}_k{K}",
-            N=N,
-            K=K,
-        )
-
-    print("Benchmark finished!")

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -1,98 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import itertools
-from typing import Callable
-
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.config import CompilationConfig, VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
-from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
-from vllm.triton_utils import triton
-
-
-# TODO(luka): use standalone_compile utility
-def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
-    def inner(*args):
-        torch._dynamo.mark_dynamic(args[arg_index], dim_index)
-        return fn(*args)
-
-    return inner
-
-
-torch._dynamo.config.recompile_limit = 8888
-compilation_config = CompilationConfig(custom_ops=["none"])
-with set_current_vllm_config(VllmConfig(compilation_config=compilation_config)):
-    torch_per_token_quant_fp8 = torch.compile(
-        QuantFP8(False, GroupShape.PER_TOKEN),
-        fullgraph=True,
-        dynamic=False,  # recompile for different shapes
-    )
-
-    # First dim is explicitly dynamic to simulate vLLM usage
-    torch_per_token_quant_fp8 = with_dyn_arg(torch_per_token_quant_fp8, 0, 0)
-
-
-def cuda_per_token_quant_fp8(
-    input: torch.Tensor,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    return ops.scaled_fp8_quant(input)
-
-
-def calculate_diff(batch_size: int, seq_len: int):
-    """Calculate difference between Triton and CUDA implementations."""
-    device = torch.device("cuda")
-    x = torch.rand((batch_size * seq_len, 4096), dtype=torch.float16, device=device)
-
-    torch_out, torch_scale = torch_per_token_quant_fp8(x)
-    cuda_out, cuda_scale = cuda_per_token_quant_fp8(x)
-
-    if torch.allclose(
-        cuda_out.to(torch.float32), torch_out.to(torch.float32), rtol=1e-3, atol=1e-5
-    ) and torch.allclose(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5):
-        print("✅ All implementations match")
-    else:
-        print("❌ Implementations differ")
-
-
-batch_size_range = [1, 16, 32, 64, 128]
-seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
-
-configs = list(itertools.product(batch_size_range, seq_len_range))
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size", "seq_len"],
-        x_vals=configs,
-        line_arg="provider",
-        line_vals=["torch", "cuda"],
-        line_names=["Torch", "CUDA"],
-        styles=[("blue", "-"), ("green", "-")],
-        ylabel="us",
-        plot_name="per-token-dynamic-quant-fp8-performance",
-        args={},
-    )
-)
-def benchmark_quantization(batch_size, seq_len, provider):
-    dtype = torch.float16
-    device = torch.device("cuda")
-
-    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch":
-        fn = lambda: torch_per_token_quant_fp8(x.clone())
-    elif provider == "cuda":
-        fn = lambda: cuda_per_token_quant_fp8(x.clone())
-
-    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles)
-
-    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
-
-
-if __name__ == "__main__":
-    calculate_diff(batch_size=4, seq_len=4096)
-    benchmark_quantization.run(print_data=True)

benchmarks/kernels/benchmark_moe.py

@@ -86,9 +86,6 @@ def benchmark_config(
             (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
         )
         w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
-    if use_deep_gemm:
-        # we use the default block shape for deepgemm
-        block_quant_shape = [128, 128]
     if use_fp8_w8a8:
         if block_quant_shape:
             block_n, block_k = block_quant_shape[0], block_quant_shape[1]
@@ -576,11 +573,7 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif config.architectures[0] in (
-        "DeepseekV3ForCausalLM",
-        "DeepseekV2ForCausalLM",
-        "Glm4MoeForCausalLM",
-    ):
+    elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
@@ -590,11 +583,6 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
-        E = config.num_experts
-        topk = config.moe_topk[0]
-        intermediate_size = config.moe_intermediate_size[0]
-        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Support for llama4
         config = config.get_text_config()

benchmarks/kernels/benchmark_moe_align_block_size.py

@@ -33,13 +33,15 @@ def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8):
     sorted_ids_triton = torch.empty(
         (max_num_tokens_padded,), dtype=torch.int32, device="cuda"
     )
-    expert_ids_triton = torch.empty(
+    sorted_ids_triton.fill_(topk_ids.numel())  # fill with sentinel value
+    expert_ids_triton = torch.zeros(
         (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda"
     )
     num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda")
 
     sorted_ids_vllm = torch.empty_like(sorted_ids_triton)
-    expert_ids_vllm = torch.empty_like(expert_ids_triton)
+    sorted_ids_vllm.fill_(topk_ids.numel())
+    expert_ids_vllm = torch.zeros_like(expert_ids_triton)
     num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton)
 
     # 2. run implementations
@@ -100,6 +102,7 @@ def benchmark(num_tokens, num_experts, topk, provider):
 
     max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
     sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda")
+    sorted_ids.fill_(topk_ids.numel())
     max_num_m_blocks = max_num_tokens_padded // block_size
     expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda")
     num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda")

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -318,7 +318,6 @@ def main(args: argparse.Namespace):
     elif (
         config.architectures[0] == "DeepseekV3ForCausalLM"
         or config.architectures[0] == "DeepseekV2ForCausalLM"
-        or config.architectures[0] == "Glm4MoeForCausalLM"
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok

benchmarks/kernels/benchmark_trtllm_attention.py

@@ -1,240 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import csv
-import os
-import random
-from datetime import datetime
-
-import flashinfer
-import torch
-
-FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
-
-# KV Cache Layout for TRT-LLM
-# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
-
-
-def to_float8(x, dtype=torch.float8_e4m3fn):
-    finfo = torch.finfo(dtype)
-    min_val, max_val = x.aminmax()
-    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
-    scale = finfo.max / amax * 0.1
-    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
-    return x_scl_sat.to(dtype), scale.float().reciprocal()
-
-
-@torch.no_grad()
-def benchmark_decode(
-    num_seqs,
-    max_seq_len,
-    page_size=16,
-    dtype=torch.bfloat16,
-    kv_layout="HND",
-    num_kv_heads=8,
-    kv_cache_dtype="auto",
-    head_dim=128,
-    warmup=10,
-    trials=20,
-):
-    torch.set_default_device("cuda")
-    device = "cuda"
-    torch.manual_seed(0)
-
-    # Currently only HEAD_GRP_SIZE == 8 is supported
-    HEAD_GRP_SIZE = 8
-    MAX_SEQ_LEN = max_seq_len
-
-    # large number to reduce kv_cache reuse
-    NUM_BLOCKS = int(256000 / page_size)
-
-    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8, device=device)
-
-    # For decode, batch_size is num_decode_token
-    num_qo_heads = num_kv_heads * HEAD_GRP_SIZE
-    sm_scale = float(1.0 / (head_dim**0.5))
-    q = torch.randn(num_seqs, num_qo_heads, head_dim, device=device, dtype=dtype)
-    kv_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
-
-    max_kv_len = max(kv_lens)
-    kv_lens_tensor = torch.tensor(kv_lens, dtype=torch.int, device=device)
-    max_num_blocks_per_seq = (max_kv_len + page_size - 1) // page_size
-
-    block_tables = torch.randint(
-        0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
-    )
-
-    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
-    kv_cache = torch.randn(size=kv_cache_shape, device=device, dtype=dtype)
-    k_scale = v_scale = 1.0
-
-    if kv_cache_dtype.startswith("fp8"):
-        kv_cache, _ = to_float8(kv_cache)
-
-    # Benchmark TRT decode
-    def trt_decode():
-        return flashinfer.decode.trtllm_batch_decode_with_kv_cache(
-            q,
-            kv_cache,
-            workspace_buffer,
-            num_qo_heads,
-            num_kv_heads,
-            sm_scale,
-            block_tables,
-            kv_lens_tensor,
-            page_size,
-            max_kv_len,
-            kv_cache_dtype,
-            k_scale,
-            v_scale,
-        )
-
-    def time_fn(fn, warmup=10, trials=20):
-        torch.cuda.synchronize()
-        start = torch.cuda.Event(enable_timing=True)
-        end = torch.cuda.Event(enable_timing=True)
-        times = []
-        for i in range(warmup):
-            fn()
-        for i in range(trials):
-            start.record()
-            fn()
-            end.record()
-            torch.cuda.synchronize()
-            times.append(start.elapsed_time(end))  # ms
-        return sum(times) / len(times), torch.std(torch.tensor(times))
-
-    # TRT Decode
-    trt_mean, trt_std = time_fn(trt_decode)
-
-    kv_indptr = [0]
-    kv_indices = []
-    kv_last_page_lens = []
-    for i in range(num_seqs):
-        seq_len = kv_lens[i]
-        assert seq_len > 0
-        num_blocks = (seq_len + page_size - 1) // page_size
-        kv_indices.extend(block_tables[i, :num_blocks])
-        kv_indptr.append(kv_indptr[-1] + num_blocks)
-        kv_last_page_len = seq_len % page_size
-        if kv_last_page_len == 0:
-            kv_last_page_len = page_size
-        kv_last_page_lens.append(kv_last_page_len)
-
-    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
-    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
-    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
-
-    wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
-        workspace_buffer,
-        kv_layout,
-        use_tensor_cores=((num_qo_heads // num_kv_heads) > 4),
-    )
-
-    wrapper.plan(
-        kv_indptr,
-        kv_indices,
-        kv_last_page_lens,
-        num_qo_heads,
-        num_kv_heads,
-        head_dim,
-        page_size,
-        "NONE",
-        q_data_type=dtype,
-        kv_data_type=torch.float8_e4m3fn if kv_cache_dtype.startswith("fp8") else dtype,
-    )
-
-    def baseline_decode():
-        return wrapper.run(q, kv_cache, sm_scale, k_scale, v_scale)
-
-    baseline_mean, baseline_std = time_fn(baseline_decode)
-
-    # Calculate percentage speedup (positive means TRT is faster)
-    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
-
-    print(
-        f"\t{num_seqs}\t{max_seq_len}\t{trt_mean:.3f}\t{trt_std.item():.3f}"
-        f"\t{baseline_mean:.3f}\t{baseline_std.item():.3f}\t{speedup_percent:.3f}"
-    )
-
-    # Return results for CSV writing
-    return {
-        "num_seqs": num_seqs,
-        "trt_mean": trt_mean,
-        "trt_std": trt_std.item(),
-        "baseline_mean": baseline_mean,
-        "baseline_std": baseline_std.item(),
-        "speedup_percent": speedup_percent,
-        "q_dtype": str(dtype),
-        "kv_cache_dtype": kv_cache_dtype,
-        "page_size": page_size,
-        "num_kv_heads": num_kv_heads,
-        "head_dim": head_dim,
-        "max_seq_len": max_seq_len,
-    }
-
-
-def write_results_to_csv(results, filename=None):
-    """Write benchmark results to CSV file."""
-    if filename is None:
-        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
-        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"
-
-    fieldnames = [
-        "num_seqs",
-        "trt_mean",
-        "trt_std",
-        "baseline_mean",
-        "baseline_std",
-        "speedup_percent",
-        "q_dtype",
-        "kv_cache_dtype",
-        "page_size",
-        "num_kv_heads",
-        "head_dim",
-        "max_seq_len",
-    ]
-
-    file_exists = os.path.exists(filename)
-
-    with open(filename, "a", newline="") as csvfile:
-        writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
-
-        if not file_exists:
-            writer.writeheader()
-
-        for result in results:
-            writer.writerow(result)
-
-    print(f"Results written to {filename}")
-
-
-if __name__ == "__main__":
-    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
-    max_seq_lens = [1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
-    all_results = []
-
-    print("Running benchmark for kv_cache_dtype: bfloat16")
-    print(
-        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
-    )
-    for max_seq_len in max_seq_lens:
-        for bs in num_seqs:
-            result = benchmark_decode(
-                bs, max_seq_len, dtype=torch.bfloat16, kv_cache_dtype="auto"
-            )
-            all_results.append(result)
-
-    print("Running benchmark for q_dtype = bfloat16, kv_cache_dtype: fp8")
-    print(
-        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
-    )
-    for max_seq_len in max_seq_lens:
-        for bs in num_seqs:
-            result = benchmark_decode(
-                bs, max_seq_len, dtype=torch.bfloat16, kv_cache_dtype="fp8"
-            )
-            all_results.append(result)
-
-    # Write all results to CSV
-    write_results_to_csv(all_results)

benchmarks/kv_cache/benchmark_block_pool.py

@@ -1,108 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import gc
-import time
-from typing import Optional
-
-from tabulate import tabulate
-
-from vllm.utils import FlexibleArgumentParser
-from vllm.v1.core.block_pool import BlockPool
-
-
-class Metric:
-    def __init__(self) -> None:
-        self.cnt: int = 0
-        self.sum_v: int = 0
-        self.max_v: Optional[int] = None
-
-    def update(self, v: int) -> None:
-        self.cnt += 1
-        self.sum_v += v
-        if self.max_v is None:
-            self.max_v = v
-        else:
-            self.max_v = max(self.max_v, v)
-
-    def avg_v(self) -> float:
-        return self.sum_v * 1.0 / self.cnt
-
-
-def main(args):
-    rows = []
-    for allocate_block in args.allocate_blocks:
-        # Enforce a GC collect ahead to minimize the impact among runs
-        gc.collect()
-        block_pool = BlockPool(num_gpu_blocks=args.num_gpu_blocks, enable_caching=True)
-
-        get_blocks_metric: Metric = Metric()
-        free_blocks_metric: Metric = Metric()
-        for _ in range(args.num_iteration):
-            t1 = time.monotonic_ns()
-            blocks = block_pool.get_new_blocks(allocate_block)
-            t2 = time.monotonic_ns()
-            block_pool.free_blocks(blocks)
-            t3 = time.monotonic_ns()
-            get_blocks_metric.update(t2 - t1)
-            free_blocks_metric.update(t3 - t2)
-
-        if get_blocks_metric.max_v is not None and free_blocks_metric.max_v is not None:
-            rows.append(
-                [
-                    get_blocks_metric.cnt,
-                    args.num_gpu_blocks,
-                    allocate_block,
-                    get_blocks_metric.avg_v() / 1000000,
-                    get_blocks_metric.max_v / 1000000.0,
-                    free_blocks_metric.avg_v() / 1000000,
-                    free_blocks_metric.max_v / 1000000.0,
-                ]
-            )
-        else:
-            print(
-                "No valid metrics found."
-                f" {get_blocks_metric.max_v=} {free_blocks_metric.max_v=}"
-            )
-
-    print(
-        tabulate(
-            rows,
-            headers=[
-                "Iterations",
-                "Total\nBlocks",
-                "Allocated\nBlocks",
-                "Get Blocks\nAvg (ms)",
-                "Get Blocks\nMax (ms)",
-                "Free Blocks\nAvg (ms)",
-                "Free Blocks\nMax (ms)",
-            ],
-            tablefmt="grid",
-            floatfmt=".6f",
-        )
-    )
-
-
-def invoke_main() -> None:
-    parser = FlexibleArgumentParser(
-        description="Benchmark the performance of BlockPool for KV Cache."
-    )
-    parser.add_argument("--num-gpu-blocks", type=int, default=100000)
-    parser.add_argument(
-        "--num-iteration",
-        type=int,
-        default=1000,
-        help="Number of iterations to run to stablize final data readings",
-    )
-    parser.add_argument(
-        "--allocate-blocks",
-        type=int,
-        nargs="*",
-        default=[10, 50, 100, 500, 1000],
-        help="Number of blocks to allocate",
-    )
-    args = parser.parse_args()
-    main(args)
-
-
-if __name__ == "__main__":
-    invoke_main()  # pragma: no cover

cmake/cpu_extension.cmake

@@ -165,32 +165,17 @@ else()
 endif()
 
 #
-# Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 /ARM platforms)
-# Flag to enable ACL kernels for AARCH64 platforms
-if ( VLLM_BUILD_ACL STREQUAL "ON")
-    set(USE_ACL ON)
-else()
-    set(USE_ACL OFF)
-endif()
-
-if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
+# Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 platforms)
+#
+if (AVX512_FOUND AND NOT AVX512_DISABLED)
     FetchContent_Declare(
         oneDNN
         GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG  v3.8.1
+        GIT_TAG  v3.7.1
         GIT_PROGRESS TRUE
         GIT_SHALLOW TRUE
     )
 
-    if(USE_ACL)
-        find_library(ARM_COMPUTE_LIBRARY NAMES arm_compute PATHS $ENV{ACL_ROOT_DIR}/build/)
-        if(NOT ARM_COMPUTE_LIBRARY)
-            message(FATAL_ERROR "Could not find ARM Compute Library: please set ACL_ROOT_DIR")
-        endif()
-        set(ONEDNN_AARCH64_USE_ACL "ON")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
-        endif()
-
     set(ONEDNN_LIBRARY_TYPE "STATIC")
     set(ONEDNN_BUILD_DOC "OFF")
     set(ONEDNN_BUILD_EXAMPLES "OFF")
@@ -279,11 +264,6 @@ elseif(POWER10_FOUND)
         "csrc/cpu/quant.cpp"
         ${VLLM_EXT_SRC})
 endif()
-if (ASIMD_FOUND)
-    set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
-        ${VLLM_EXT_SRC})
-endif()
 
 message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
 

csrc/attention/attention_kernels.cuh

@@ -24,7 +24,6 @@
 
 #include "attention_dtypes.h"
 #include "attention_utils.cuh"
-#include "cuda_compat.h"
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -34,6 +33,12 @@ typedef __hip_bfloat16 __nv_bfloat16;
   #include "../quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
+
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
@@ -665,6 +670,7 @@ __global__ void paged_attention_v2_reduce_kernel(
 
 }  // namespace vllm
 
+#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

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

@@ -1,372 +0,0 @@
-/***************************************************************************************************
- * Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
- * SPDX-License-Identifier: BSD-3-Clause
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice,
- *this list of conditions and the following disclaimer.
- *
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- * this list of conditions and the following disclaimer in the documentation
- * and/or other materials provided with the distribution.
- *
- * 3. Neither the name of the copyright holder nor the names of its
- * contributors may be used to endorse or promote products derived from
- * this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
- * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
- *ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
- *LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
- *CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
- *SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
- *INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
- *CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
- *ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
- *POSSIBILITY OF SUCH DAMAGE.
- *
- **************************************************************************************************/
-/*
- * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
- * by Alcanderian JieXin Liang
- */
-
-/*!
- \file
- \brief An universal device layer for cutlass 3.x-style kernels.
-*/
-
-// clang-format off
-#pragma once
-
-// common
-#include "cutlass/cutlass.h"
-#include "cutlass/device_kernel.h"
-
-#if !defined(__CUDACC_RTC__)
-#include "cutlass/cluster_launch.hpp"
-#include "cutlass/trace.h"
-#endif // !defined(__CUDACC_RTC__)
-
-#include "../kernel/sm100_fmha_mla_tma_warpspecialized.hpp"
-#include "../kernel/sm100_fmha_mla_reduction.hpp"
-
-////////////////////////////////////////////////////////////////////////////////
-
-namespace cutlass::fmha::device {
-
-using namespace cute;
-using namespace cutlass::fmha::kernel;
-
-
-////////////////////////////////////////////////////////////////////////////////
-////////////////////////////// CUTLASS 3.x API /////////////////////////////////
-////////////////////////////////////////////////////////////////////////////////
-
-template<
-    class Kernel_
->
-class MLA {
-public:
-
-  using Kernel = Kernel_;
-
-  using ReductionKernel = cutlass::fmha::kernel::Sm100FmhaMlaReductionKernel<
-      typename Kernel::ElementOut,
-      typename Kernel::ElementAcc,
-      typename Kernel::ElementAcc,
-      Kernel::TileShapeH::value,
-      Kernel::TileShapeL::value,
-      256 /*Max split*/
-  >;
-
-  /// Argument structure: User API
-  using KernelArguments = typename Kernel::Arguments;
-  using ReductionArguments = typename ReductionKernel::Arguments;
-
-  using Arguments = KernelArguments;
-
-  /// Argument structure: Kernel API
-  using KernelParams = typename Kernel::Params;
-  using ReductionParams = typename ReductionKernel::Params;
-  struct Params {
-    KernelParams fmha_params;
-    ReductionParams reduction_params;
-  };
-
-private:
-
-  /// Kernel API parameters object
-  Params params_;
-
-  bool is_initialized(bool set = false) {
-    static bool initialized = false;
-    if (set) initialized = true;
-    return initialized;
-  }
-
-  static ReductionArguments to_reduction_args(Arguments const& args) {
-    auto [H, K, D, B] = args.problem_shape;
-    return ReductionArguments{
-      nullptr, args.epilogue.ptr_o, nullptr, args.epilogue.ptr_lse,
-      args.mainloop.softmax_scale, B, args.split_kv, K, args.mainloop.ptr_seq,
-      args.ptr_split_kv, Kernel::TileShapeS::value
-    };
-  }
-
-public:
-
-  /// Access the Params structure
-  Params const& params() const {
-    return params_;
-  }
-
-  static void set_split_kv (KernelArguments& args) {
-    // printf("set_split_kv start");
-    if (args.split_kv >= 1) return;
-    auto [H, K, D, B] = args.problem_shape;
-    // std::cout << H << " " << K << " " << D << " " << B << "\n";      
-    int sm_count = args.hw_info.sm_count;
-    // printf("    sm_count = %d\n", sm_count);
-    int max_splits = ceil_div(K, 128);
-    max_splits = min(16, max_splits);
-    // printf("    max_splits = %d\n", max_splits);
-    int sms_per_batch = max(1, sm_count / B);
-    // printf("    sms_per_batch = %d\n", sms_per_batch);
-    int split_heur = min(max_splits, sms_per_batch);
-    int waves = ceil_div(B * split_heur, sm_count);
-    int k_waves = ceil_div(max_splits, split_heur);
-    int split_wave_aware = ceil_div(max_splits, k_waves);
-    args.split_kv = split_wave_aware;
-    // printf("    args.split_kv = %d\n", args.split_kv);
-
-  }
-
-  /// Determines whether the GEMM can execute the given problem.
-  static Status
-  can_implement(Arguments const& args) {
-    if (! Kernel::can_implement(args)) {
-      return Status::kInvalid;
-    }
-    if (! ReductionKernel::can_implement(to_reduction_args(args))) {
-      return Status::kInvalid;
-    }
-    return Status::kSuccess;
-  }
-
-  /// Gets the workspace size
-  static size_t
-  get_workspace_size(Arguments const& args) {
-    size_t workspace_bytes = 0;
-    workspace_bytes += Kernel::get_workspace_size(args);
-    workspace_bytes += ReductionKernel::get_workspace_size(to_reduction_args(args));
-    return workspace_bytes;
-  }
-
-  /// Computes the maximum number of active blocks per multiprocessor
-  static int maximum_active_blocks(int /* smem_capacity */ = -1) {
-    CUTLASS_TRACE_HOST("MLA::maximum_active_blocks()");
-    int max_active_blocks = -1;
-    int smem_size = Kernel::SharedStorageSize;
-
-    // first, account for dynamic smem capacity if needed
-    cudaError_t result;
-    if (smem_size >= (48 << 10)) {
-      CUTLASS_TRACE_HOST("  Setting smem size to " << smem_size);
-      result = cudaFuncSetAttribute(
-          device_kernel<Kernel>,
-          cudaFuncAttributeMaxDynamicSharedMemorySize,
-          smem_size);
-      if (cudaSuccess != result) {
-        result = cudaGetLastError(); // to clear the error bit
-        CUTLASS_TRACE_HOST(
-          "  cudaFuncSetAttribute() returned error: "
-          << cudaGetErrorString(result));
-        return -1;
-      }
-    }
-
-    // query occupancy after setting smem size
-    result = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
-        &max_active_blocks,
-        device_kernel<Kernel>,
-        Kernel::MaxThreadsPerBlock,
-        smem_size);
-
-    if (cudaSuccess != result) {
-      result = cudaGetLastError(); // to clear the error bit
-      CUTLASS_TRACE_HOST(
-        "  cudaOccupancyMaxActiveBlocksPerMultiprocessor() returned error: "
-        << cudaGetErrorString(result));
-      return -1;
-    }
-
-    CUTLASS_TRACE_HOST("  max_active_blocks: " << max_active_blocks);
-    return max_active_blocks;
-  }
-
-  /// Initializes GEMM state from arguments.
-  Status
-  initialize(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr) {
-    CUTLASS_TRACE_HOST("MLA::initialize() - workspace "
-      << workspace << ", stream: " << (stream ? "non-null" : "null"));
-
-    // Initialize the workspace
-    Status status = Kernel::initialize_workspace(args, workspace, stream);
-    if (status != Status::kSuccess) {
-      return status;
-    }
-    status = ReductionKernel::initialize_workspace(to_reduction_args(args), workspace, stream);
-    if (status != Status::kSuccess) {
-      return status;
-    }
-    KernelParams kernel_params = Kernel::to_underlying_arguments(args, workspace);
-
-    ReductionArguments reduction_args = to_reduction_args(args);
-    if (reduction_args.split_kv > 1) {
-      reduction_args.ptr_oaccum   = kernel_params.epilogue.ptr_o_acc;
-      reduction_args.ptr_lseaccum = kernel_params.epilogue.ptr_lse_acc;
-    }
-    ReductionParams reduction_params = ReductionKernel::to_underlying_arguments(reduction_args, workspace);
-    // Initialize the Params structure
-    params_ = Params {kernel_params, reduction_params};
-
-    if (is_initialized()) return Status::kSuccess;
-
-    // account for dynamic smem capacity if needed
-    // no dynamic smem is needed for reduction kernel
-    int smem_size = Kernel::SharedStorageSize;
-    if (smem_size >= (48 << 10)) {
-      CUTLASS_TRACE_HOST("  Setting smem size to " << smem_size);
-      cudaError_t result = cudaFuncSetAttribute(
-          device_kernel<Kernel>,
-          cudaFuncAttributeMaxDynamicSharedMemorySize,
-          smem_size);
-      if (cudaSuccess != result) {
-        result = cudaGetLastError(); // to clear the error bit
-        CUTLASS_TRACE_HOST("  cudaFuncSetAttribute() returned error: " << cudaGetErrorString(result));
-        return Status::kErrorInternal;
-      }
-    }
-
-    is_initialized(true);
-
-    return Status::kSuccess;
-  }
-
-  /// Update API is preserved in 3.0, but does not guarantee a lightweight update of params.
-  Status
-  update(Arguments const& args, void* workspace = nullptr) {
-    CUTLASS_TRACE_HOST("MLA()::update() - workspace: " << workspace);
-
-    size_t workspace_bytes = get_workspace_size(args);
-    if (workspace_bytes > 0 && nullptr == workspace) {
-      return Status::kErrorWorkspaceNull;
-    }
-
-    auto fmha_params = Kernel::to_underlying_arguments(args, workspace);
-
-    ReductionArguments reduction_args = to_reduction_args(args);
-    if (reduction_args.split_kv > 1) {
-      reduction_args.ptr_oaccum   = fmha_params.epilogue.ptr_o_acc;
-      reduction_args.ptr_lseaccum = fmha_params.epilogue.ptr_lse_acc;
-    }
-    ReductionParams reduction_params = ReductionKernel::to_underlying_arguments(reduction_args, workspace);
-    // Initialize the Params structure
-    params_ = Params {fmha_params, reduction_params};
-
-    return Status::kSuccess;
-  }
-
-  /// Primary run() entry point API that is static allowing users to create and manage their own params.
-  /// Supplied params struct must be construct by calling Kernel::to_underling_arguments()
-  static Status
-  run(Params& params, cudaStream_t stream = nullptr) {
-    CUTLASS_TRACE_HOST("MLA::run()");
-    dim3 const block = Kernel::get_block_shape();
-    dim3 const grid = Kernel::get_grid_shape(params.fmha_params);
-
-    // configure smem size and carveout
-    int smem_size = Kernel::SharedStorageSize;
-
-    Status launch_result;
-    // Use extended launch API only for mainloops that use it
-    if constexpr(Kernel::ArchTag::kMinComputeCapability >= 90) {
-      dim3 cluster(cute::size<0>(typename Kernel::ClusterShape{}),
-                   cute::size<1>(typename Kernel::ClusterShape{}),
-                   cute::size<2>(typename Kernel::ClusterShape{}));
-      void const* kernel = (void const*) device_kernel<Kernel>;
-      void* kernel_params[] = {&params.fmha_params};
-      launch_result = ClusterLauncher::launch(grid, cluster, block, smem_size, stream, kernel, kernel_params);
-    }
-    else {
-      launch_result = Status::kSuccess;
-      device_kernel<Kernel><<<grid, block, smem_size, stream>>>(params.fmha_params);
-    }
-
-    cudaError_t result = cudaGetLastError();
-    if (cudaSuccess != result or Status::kSuccess != launch_result) {
-      //return Status::kSuccess;
-      CUTLASS_TRACE_HOST("  Kernel launch failed. Reason: " << result);
-      return Status::kErrorInternal;
-    }
-    if (params.reduction_params.split_kv > 1) {
-      // launch reduction kernel
-      dim3 const block = ReductionKernel::get_block_shape();
-      dim3 const grid  = ReductionKernel::get_grid_shape(params.reduction_params);
-      device_kernel<ReductionKernel><<<grid, block, 0, stream>>>(params.reduction_params);
-      cudaError_t result = cudaGetLastError();
-      if (cudaSuccess == result) {
-        return Status::kSuccess;
-      }
-      else {
-        CUTLASS_TRACE_HOST("  Kernel launch failed. Reason: " << result);
-        return Status::kErrorInternal;
-      }
-    }
-    else {
-      return Status::kSuccess;
-    }
-  }
-
-  //
-  // Non-static launch overloads that first create and set the internal params struct of this kernel handle.
-  //
-
-  /// Launches the kernel after first constructing Params internal state from supplied arguments.
-  Status
-  run(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr) {
-    Status status = initialize(args, workspace, stream);
-    if (Status::kSuccess == status) {
-      status = run(params_, stream);
-    }
-    return status;
-  }
-
-  /// Launches the kernel after first constructing Params internal state from supplied arguments.
-  Status
-  operator()(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr) {
-    return run(args, workspace, stream);
-  }
-
-  /// Overload that allows a user to re-launch the same kernel without updating internal params struct.
-  Status
-  run(cudaStream_t stream = nullptr) {
-    return run(params_, stream);
-  }
-
-  /// Overload that allows a user to re-launch the same kernel without updating internal params struct.
-  Status
-  operator()(cudaStream_t stream = nullptr) {
-    return run(params_, stream);
-  }
-};
-
-////////////////////////////////////////////////////////////////////////////////
-
-} // namespace cutlass::fmha::device
-
-////////////////////////////////////////////////////////////////////////////////

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_fmha_mla_reduction.hpp

@@ -1,203 +0,0 @@
-/***************************************************************************************************
- * Copyright (c) 2024 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights
- *reserved. SPDX-License-Identifier: BSD-3-Clause
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice,
- *this list of conditions and the following disclaimer.
- *
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- * this list of conditions and the following disclaimer in the documentation
- * and/or other materials provided with the distribution.
- *
- * 3. Neither the name of the copyright holder nor the names of its
- * contributors may be used to endorse or promote products derived from
- * this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
- * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
- *ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
- *LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
- *CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
- *SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
- *INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
- *CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
- *ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
- *POSSIBILITY OF SUCH DAMAGE.
- *
- **************************************************************************************************/
-/*
- * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
- * by Alcanderian JieXin Liang
- */
-
-// clang-format off
-#pragma once
-
-#include "cutlass/cutlass.h"
-#include "cutlass/arch/arch.h"
-#include "cute/tensor.hpp"
-
-namespace cutlass::fmha::kernel {
-
-using namespace cute;
-template<
-    class ElementOut,
-    class ElementAcc,
-    class ElementScale,
-    size_t kNumHeads,
-    size_t kHeadDimLatent,
-    int kMaxSplits
->
-struct Sm100FmhaMlaReductionKernel {
-
-  static const int SharedStorageSize = 0;
-  static const int MaxThreadsPerBlock = 128;
-  static const int MinBlocksPerMultiprocessor = 1;
-
-  using ArchTag = cutlass::arch::Sm100;
-
-  static_assert(kHeadDimLatent % MaxThreadsPerBlock == 0);
-  struct Arguments {
-    ElementAcc* ptr_oaccum = nullptr;
-    ElementOut* ptr_o = nullptr;
-    ElementAcc* ptr_lseaccum = nullptr;
-    ElementAcc* ptr_lse = nullptr;
-    ElementScale scale = 1.f;
-    int num_batches = 0;
-    int split_kv = -1;
-    int dim_k = -1;
-    int* ptr_seq = nullptr;
-    int* ptr_split_kv = nullptr;
-    int tile_shape_s = 128;
-  };
-  using Params = Arguments;
-
-  static Params to_underlying_arguments(Arguments const& args, void* workspace) {
-    return {args.ptr_oaccum, args.ptr_o, args.ptr_lseaccum, args.ptr_lse,
-	    args.scale, args.num_batches, args.split_kv, args.dim_k, args.ptr_seq,
-	    args.ptr_split_kv, args.tile_shape_s};
-  }
-
-  static size_t get_workspace_size(Arguments const& /*args*/) {
-    return 0;
-  }
-
-  static Status initialize_workspace(
-      Arguments const& /*args*/, void* /*ws*/, cudaStream_t /*stream*/) {
-    return Status::kSuccess;
-  }
-
-  static dim3 get_grid_shape(Params const& params) {
-    return dim3(kNumHeads, 1, params.num_batches);
-  }
-
-  static dim3 get_block_shape() {
-    return dim3(MaxThreadsPerBlock, 1, 1);
-  }
-
-  static bool can_implement(Arguments const& args) {
-    if (args.num_batches <= 0) return false;
-    if (args.split_kv <= 0) return false;
-    return true;
-  }
-
-  CUTLASS_DEVICE void operator() (Params const& params, char* smem_raw) {
-    if (params.split_kv <= 1) return;
-    auto blk_coord = make_coord(blockIdx.x, _0{}, blockIdx.z);
-
-    __shared__ ElementAcc sLseScale[kMaxSplits];
-    const size_t offset_lseaccum = get<0>(blk_coord) + kNumHeads * params.split_kv * get<2>(blk_coord);
-    const size_t offset_lse = get<0>(blk_coord) + kNumHeads * get<2>(blk_coord);
-
-    Tensor gLSEaccum = make_tensor(make_gmem_ptr(params.ptr_lseaccum + offset_lseaccum),
-                                   make_shape(params.split_kv), Stride<Int<kNumHeads>>{});
-
-    Tensor gLSE = make_tensor(make_gmem_ptr(params.ptr_lse + offset_lse),
-                              Shape<_1>{}, Stride<_1>{});
-
-    auto dim_k = params.ptr_seq == nullptr ?  params.dim_k : params.ptr_seq[get<2>(blk_coord)];
-    auto local_split_kv = params.ptr_split_kv == nullptr ? params.split_kv : params.ptr_split_kv[get<2>(blk_coord)];
-    auto k_tile_total = ceil_div(dim_k, params.tile_shape_s);
-    auto k_tile_per_cta = ceil_div(k_tile_total, local_split_kv);
-    local_split_kv = ceil_div(k_tile_total, k_tile_per_cta);
-
-    int warp_idx = cutlass::canonical_warp_idx_sync();
-    if (warp_idx == 0) {
-      constexpr int kNLsePerThread = cute::ceil_div(kMaxSplits, 32);
-
-      ElementAcc local_lse[kNLsePerThread];
-
-      CUTLASS_PRAGMA_UNROLL
-      for (int i = 0; i < kNLsePerThread; ++i) {
-        const int split = i * 32 + threadIdx.x;
-        local_lse[i] = split < local_split_kv ? gLSEaccum(split) : -std::numeric_limits<ElementAcc>::infinity();
-      }
-
-      ElementAcc lse_max = -std::numeric_limits<ElementAcc>::infinity();
-      CUTLASS_PRAGMA_UNROLL
-      for (int i = 0; i < kNLsePerThread; ++i) {
-        lse_max = max(lse_max, local_lse[i]);
-      }
-      CUTLASS_PRAGMA_UNROLL
-      for (int offset = 16; offset >= 1; offset /= 2) {
-        lse_max = max(lse_max, __shfl_xor_sync(0xffffffff, lse_max, offset));
-      }
-      lse_max = lse_max == -std::numeric_limits<ElementAcc>::infinity() ? 0.0f : lse_max;  // In case all local LSEs are -inf
-      lse_max = __shfl_sync(0xffffffff, lse_max, 0);
-
-      ElementAcc sum_lse = 0;
-      CUTLASS_PRAGMA_UNROLL
-      for (int i = 0; i < kNLsePerThread; ++i) {
-        sum_lse = sum_lse + expf(local_lse[i] - lse_max);
-      }
-
-      CUTLASS_PRAGMA_UNROLL
-      for (int offset = 16; offset >= 1; offset /= 2) {
-        sum_lse = sum_lse + __shfl_xor_sync(0xffffffff, sum_lse, offset);
-      }
-
-      sum_lse = __shfl_sync(0xffffffff, sum_lse, 0);
-
-      ElementAcc global_lse = (sum_lse == 0.f || sum_lse != sum_lse) ? std::numeric_limits<ElementAcc>::infinity() : logf(sum_lse) + lse_max;
-      if (threadIdx.x == 0 and params.ptr_lse != nullptr) {
-        gLSE(0) = global_lse;
-      }
-
-      CUTLASS_PRAGMA_UNROLL
-      for (int i = 0; i < kNLsePerThread; ++i) {
-        const int split = i * 32 + threadIdx.x;
-        if (split < local_split_kv) {
-          sLseScale[split] = expf(local_lse[i] - global_lse);
-        }
-      }
-    }
-    __syncthreads();
-
-    constexpr int Elements = kHeadDimLatent / MaxThreadsPerBlock;
-    const size_t offset_oaccum = kHeadDimLatent * params.split_kv * (get<0>(blk_coord) + kNumHeads * get<2>(blk_coord));
-    Tensor gOaccum = make_tensor(make_gmem_ptr(params.ptr_oaccum + offset_oaccum),
-                               Shape<Int<kHeadDimLatent>>{}, Stride<_1>{});
-    ElementAcc local_val[Elements] = {0};
-    for (int split = 0; split < local_split_kv; ++split) {
-      ElementAcc lse_scale = sLseScale[split];
-      CUTLASS_PRAGMA_UNROLL
-      for(int i = 0; i < Elements; ++i) {
-        local_val[i] += lse_scale * gOaccum(threadIdx.x + MaxThreadsPerBlock * i);
-      }
-      gOaccum.data() = gOaccum.data() + kHeadDimLatent;
-    }
-    auto ptr_o_local = params.ptr_o + (get<0>(blk_coord) + get<2>(blk_coord) * kNumHeads) * kHeadDimLatent;
-    Tensor gO = make_tensor(make_gmem_ptr(ptr_o_local), Shape<Int<kHeadDimLatent>>{}, Stride<_1>{});
-
-    CUTLASS_PRAGMA_UNROLL
-    for(int i = 0; i < Elements; ++i) {
-      gO(threadIdx.x + MaxThreadsPerBlock * i) = static_cast<ElementOut>(local_val[i]);
-    }
-  }
-};
-
-}  // namespace cutlass::fmha::kernel

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_mla_tile_scheduler.hpp

@@ -1,165 +0,0 @@
-/***************************************************************************************************
- * Copyright (c) 2024 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights
- *reserved. SPDX-License-Identifier: BSD-3-Clause
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice,
- *this list of conditions and the following disclaimer.
- *
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- * this list of conditions and the following disclaimer in the documentation
- * and/or other materials provided with the distribution.
- *
- * 3. Neither the name of the copyright holder nor the names of its
- * contributors may be used to endorse or promote products derived from
- * this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
- * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
- *ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
- *LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
- *CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
- *SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
- *INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
- *CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
- *ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
- *POSSIBILITY OF SUCH DAMAGE.
- *
- **************************************************************************************************/
-/*
- * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
- * by Alcanderian JieXin Liang
- */
-
-// clang-format off
-#pragma once
-
-#include "cutlass/cutlass.h"
-#include "cutlass/fast_math.h"
-#include "cutlass/kernel_hardware_info.h"
-
-namespace cutlass::fmha::kernel {
-
-////////////////////////////////////////////////////////////////////////////////
-
-struct Sm100MlaIndividualTileScheduler {
-
-  struct Params {
-    dim3 grid;
-  };
-
-  bool valid_ = true;
-
-  CUTLASS_DEVICE
-  Sm100MlaIndividualTileScheduler(Params const&) {}
-
-  template<class ProblemShape, class ClusterShape>
-  static Params to_underlying_arguments(
-      ProblemShape const& problem_shape, KernelHardwareInfo hw_info,
-      ClusterShape const& cluster_shape, int const& split_kv) {
-    using namespace cute;
-    dim3 grid(get<0>(cluster_shape), get<3>(problem_shape) /* Batch */, split_kv /*Maximum Split KV*/);
-    return Params{ grid };
-  }
-
-  static dim3 get_grid_shape(Params const& params) {
-    return params.grid;
-  }
-
-  CUTLASS_DEVICE
-  bool is_valid() {
-    return valid_;
-  }
-
-  CUTLASS_DEVICE
-  auto get_block_coord() {
-    using namespace cute;
-    return make_coord(blockIdx.x, _0{}, blockIdx.y, blockIdx.z);
-  }
-
-  CUTLASS_DEVICE
-  Sm100MlaIndividualTileScheduler& operator++() {
-    valid_ = false;
-    return *this;
-  }
-};
-
-////////////////////////////////////////////////////////////////////////////////
-
-struct Sm100MlaPersistentTileScheduler {
-
-  struct Params {
-    int num_blocks;
-    FastDivmod divmod_m_block;
-    FastDivmod divmod_b;
-    FastDivmod divmod_split_kv;
-    KernelHardwareInfo hw_info;
-  };
-
-  int block_idx = 0;
-  Params params;
-
-  CUTLASS_DEVICE
-  Sm100MlaPersistentTileScheduler(Params const& params) : block_idx(blockIdx.x), params(params) {}
-
-  template<class ProblemShape, class ClusterShape>
-  static Params to_underlying_arguments(
-      ProblemShape const& problem_shape, KernelHardwareInfo hw_info,
-      ClusterShape const& cluster_shape, int const& split_kv) {
-    using namespace cute;
-    // Get SM count if needed, otherwise use user supplied SM count
-    int sm_count = hw_info.sm_count;
-    if (sm_count <= 1 || sm_count % size<0>(cluster_shape) != 0) {
-      CUTLASS_TRACE_HOST("  WARNING: Arguments do not include a valid SM count.\n"
-          "  For optimal performance, populate the arguments KernelHardwareInfo struct with the SM count.");
-      sm_count = KernelHardwareInfo::query_device_multiprocessor_count(hw_info.device_id);
-    }
-
-    CUTLASS_TRACE_HOST("to_underlying_arguments(): Setting persistent grid SM count to " << sm_count);
-    hw_info.sm_count = sm_count;
-
-    int num_m_blocks = size<0>(cluster_shape);
-    int num_blocks = num_m_blocks * get<3>(problem_shape)  /* Batch */;
-    num_blocks *= split_kv; /* Maximum Split KV*/
-
-    return Params {
-      num_blocks,
-      { num_m_blocks}, { get<3>(problem_shape) }, {split_kv},
-      hw_info
-    };
-  }
-
-  static dim3 get_grid_shape(Params const& params) {
-    dim3 grid(std::min(params.num_blocks, params.hw_info.sm_count), 1, 1);
-    return grid;
-  }
-
-  CUTLASS_DEVICE
-  bool is_valid() {
-    return block_idx < params.num_blocks;
-  }
-
-  CUTLASS_DEVICE
-  auto get_block_coord() {
-    using namespace cute;
-    int block_decode = block_idx;
-    int m_block, bidb, n_split_kv;
-    params.divmod_m_block(block_decode, m_block, block_decode);
-    params.divmod_b(block_decode, bidb, block_decode);
-    params.divmod_split_kv(block_decode, n_split_kv, block_decode);
-    return make_coord(m_block, _0{}, bidb, n_split_kv);
-  }
-
-  CUTLASS_DEVICE
-  Sm100MlaPersistentTileScheduler& operator++() {
-    block_idx += gridDim.x;
-    return *this;
-  }
-};
-
-////////////////////////////////////////////////////////////////////////////////
-
-} // namespace cutlass::fmha::kernel

csrc/attention/mla/sm100_cutlass_mla_kernel.cu

@@ -1,283 +0,0 @@
-/*
-Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
-Copyright 2025 SGLang Team. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-/*
- * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
- * by Alcanderian JieXin Liang
- */
-#include "core/registration.h"
-
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-#include <cutlass/cutlass.h>
-#include <cutlass/kernel_hardware_info.h>
-#include <torch/all.h>
-
-#include <cute/tensor.hpp>
-#include <iostream>
-
-#include "cutlass_sm100_mla/device/sm100_mla.hpp"
-#include "cutlass_sm100_mla/kernel/sm100_mla_tile_scheduler.hpp"
-
-// clang-format off
-#if !defined(CUDA_VERSION) || CUDA_VERSION < 12040
-void sm100_cutlass_mla_decode(
-    torch::Tensor const& out,
-    torch::Tensor const& q_nope,
-    torch::Tensor const& q_pe,
-    torch::Tensor const& kv_c_and_k_pe_cache,
-    torch::Tensor const& seq_lens,
-    torch::Tensor const& page_table,
-    torch::Tensor const& workspace,
-    int64_t num_kv_splits) {
-  TORCH_CHECK(false, "CUDA version must be >= 12.4 for cutlass_mla_decode");
-}
-int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_batches, int64_t sm_count, int64_t num_kv_splits) {
-  TORCH_CHECK(false, "CUDA version must be >= 12.4 for cutlass_mla_get_workspace_size");
-}
-#else
-
-#define CUTLASS_CHECK(status)                                                       \
-  {                                                                                 \
-    cutlass::Status error = status;                                                 \
-    TORCH_CHECK(error == cutlass::Status::kSuccess, cutlassGetStatusString(error)); \
-  }
-
-using namespace cute;
-using namespace cutlass::fmha::kernel;
-
-template <bool v>
-struct IsPersistent {
-  static const bool value = v;
-};
-
-template <typename T, bool IsPaged128, typename PersistenceOption = IsPersistent<true>>
-struct MlaSm100 {
-  using Element = T;
-  using ElementAcc = float;
-  using ElementOut = T;
-
-  using TileShape = Shape<_128, _128, Shape<_512, _64>>;
-  using TileShapeH = cute::tuple_element_t<0, TileShape>;
-  using TileShapeD = cute::tuple_element_t<2, TileShape>;
-
-  // H K (D_latent D_rope) B
-  using ProblemShape = cute::tuple<TileShapeH, int, TileShapeD, int>;
-
-  using StrideQ = cute::tuple<int64_t, _1, int64_t>;  // H D B
-  using StrideK = cute::tuple<int64_t, _1, int64_t>;  // K D B
-  using StrideO = StrideK;                            // H D B
-  using StrideLSE = cute::tuple<_1, int>;             // H B
-
-  using TileScheduler =
-      std::conditional_t<PersistenceOption::value, Sm100MlaPersistentTileScheduler, Sm100MlaIndividualTileScheduler>;
-
-  using FmhaKernel = cutlass::fmha::kernel::Sm100FmhaMlaKernelTmaWarpspecialized<
-      TileShape,
-      Element,
-      ElementAcc,
-      ElementOut,
-      ElementAcc,
-      TileScheduler,
-      /*kIsCpAsync=*/!IsPaged128>;
-  using Fmha = cutlass::fmha::device::MLA<FmhaKernel>;
-};
-
-template <typename T>
-typename T::Fmha::Arguments args_from_options(
-    at::Tensor const& out,
-    at::Tensor const& q_nope,
-    at::Tensor const& q_pe,
-    at::Tensor const& kv_c_and_k_pe_cache,
-    at::Tensor const& seq_lens,
-    at::Tensor const& page_table,
-    double sm_scale,
-    int64_t num_kv_splits) {
-  cutlass::KernelHardwareInfo hw_info;
-  hw_info.device_id = q_nope.device().index();
-  hw_info.sm_count = cutlass::KernelHardwareInfo::query_device_multiprocessor_count(hw_info.device_id);
-
-  int batches = q_nope.sizes()[0];
-  int page_count_per_seq = page_table.sizes()[1];
-  int page_count_total = kv_c_and_k_pe_cache.sizes()[0];
-  int page_size = kv_c_and_k_pe_cache.sizes()[1];
-  int max_seq_len = page_size * page_count_per_seq;
-  using TileShapeH = typename T::TileShapeH;
-  using TileShapeD = typename T::TileShapeD;
-  auto problem_shape = cute::make_tuple(TileShapeH{}, max_seq_len, TileShapeD{}, batches);
-
-  auto [H, K, D, B] = problem_shape;
-  auto [D_latent, D_rope] = D;
-
-  float scale = float(sm_scale);
-
-  using StrideQ = typename T::StrideQ;
-  using StrideK = typename T::StrideK;
-  using StrideO = typename T::StrideO;
-  using StrideLSE = typename T::StrideLSE;
-
-  StrideQ stride_Q_nope = cute::make_tuple(
-      static_cast<int64_t>(q_nope.stride(1)), _1{}, static_cast<int64_t>(q_nope.stride(0)));
-  StrideQ stride_Q_pe = cute::make_tuple(
-      static_cast<int64_t>(q_pe.stride(1)), _1{}, static_cast<int64_t>(q_pe.stride(0)));
-
-  StrideK stride_C = cute::make_tuple(
-      static_cast<int64_t>(0 + D_latent + D_rope), _1{}, static_cast<int64_t>(page_size * (D_latent + D_rope)));
-  StrideLSE stride_PT = cute::make_stride(_1{}, page_count_per_seq);
-  StrideLSE stride_LSE = cute::make_tuple(_1{}, 0 + H);
-  StrideO stride_O = cute::make_tuple(static_cast<int64_t>(0 + D_latent), _1{}, static_cast<int64_t>(0 + H * D_latent));
-
-  using Element = typename T::Element;
-  using ElementOut = typename T::ElementOut;
-  using ElementAcc = typename T::ElementAcc;
-  auto Q_nope_ptr = static_cast<Element*>(q_nope.data_ptr());
-  auto Q_pe_ptr = static_cast<Element*>(q_pe.data_ptr());
-  auto C_ptr = static_cast<Element*>(kv_c_and_k_pe_cache.data_ptr());
-  typename T::Fmha::Arguments arguments{
-      problem_shape,
-      {scale,
-       Q_nope_ptr,
-       stride_Q_nope,
-       Q_pe_ptr,
-       stride_Q_pe,
-       C_ptr,
-       stride_C,
-       C_ptr + D_latent,
-       stride_C,
-       static_cast<int*>(seq_lens.data_ptr()),
-       static_cast<int*>(page_table.data_ptr()),
-       stride_PT,
-       page_count_total,
-       page_size},
-      {static_cast<ElementOut*>(out.data_ptr()), stride_O, static_cast<ElementAcc*>(nullptr), stride_LSE},
-      hw_info,
-      // TODO(trevor-m): Change split_kv back to -1 when
-      // https://github.com/NVIDIA/cutlass/issues/2274 is fixed. Split_kv=1 will
-      // perform worse with larger context length and smaller batch sizes.
-      num_kv_splits, // split_kv
-      nullptr,       // is_var_split_kv
-  };
-  // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
-  // split_kv automatically based on batch size and sequence length to balance
-  // workload across available SMs. Consider using var_split_kv for manual
-  // control if needed.
-  T::Fmha::set_split_kv(arguments);
-  return arguments;
-}
-
-template <typename Element, bool IsPaged128, typename PersistenceOption>
-void runMla(
-    at::Tensor const& out,
-    at::Tensor const& q_nope,
-    at::Tensor const& q_pe,
-    at::Tensor const& kv_c_and_k_pe_cache,
-    at::Tensor const& seq_lens,
-    at::Tensor const& page_table,
-    at::Tensor const& workspace,
-    double sm_scale,
-    int64_t num_kv_splits,
-    cudaStream_t stream) {
-  using MlaSm100Type = MlaSm100<Element, IsPaged128, PersistenceOption>;
-  typename MlaSm100Type::Fmha fmha;
-  auto arguments = args_from_options<MlaSm100Type>(out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, sm_scale, num_kv_splits);
-
-  CUTLASS_CHECK(fmha.can_implement(arguments));
-
-  CUTLASS_CHECK(fmha.initialize(arguments, workspace.data_ptr(), stream));
-
-  CUTLASS_CHECK(fmha.run(arguments, workspace.data_ptr(), stream));
-}
-
-#define DISPATCH_BOOL(expr, const_expr, ...) \
-  [&]() -> bool {                            \
-    if (expr) {                              \
-      constexpr bool const_expr = true;      \
-      return __VA_ARGS__();                  \
-    } else {                                 \
-      constexpr bool const_expr = false;     \
-      return __VA_ARGS__();                  \
-    }                                        \
-  }()
-
-void sm100_cutlass_mla_decode(
-    torch::Tensor const& out,
-    torch::Tensor const& q_nope,
-    torch::Tensor const& q_pe,
-    torch::Tensor const& kv_c_and_k_pe_cache,
-    torch::Tensor const& seq_lens,
-    torch::Tensor const& page_table,
-    torch::Tensor const& workspace,
-    double sm_scale,
-    int64_t num_kv_splits) {
-  auto in_dtype = q_nope.dtype();
-  at::cuda::CUDAGuard device_guard{(char)q_nope.get_device()};
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream(q_nope.get_device());
-  const int page_size = kv_c_and_k_pe_cache.sizes()[1];
-  
-  // NOTE(alcanderian): IsPersistent has bug with manual split_kv.
-  // Kernel will hang if batch is too large with large num_kv_splits. (for example bs=8, num_kv_splits=8)
-  // Maybe per batch split kv will fix this.
-  DISPATCH_BOOL(page_size == 128, IsPaged128, [&] {
-    DISPATCH_BOOL(num_kv_splits <= 1, NotManualSplitKV, [&] {
-      if (in_dtype == at::ScalarType::Half) {
-        runMla<cutlass::half_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
-          out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
-      } else if (in_dtype == at::ScalarType::BFloat16) {
-        runMla<cutlass::bfloat16_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
-          out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
-      } else if (in_dtype == at::ScalarType::Float8_e4m3fn) {
-        runMla<cutlass::float_e4m3_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
-          out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
-      } else {
-        TORCH_CHECK(false, "Unsupported input data type of MLA");
-      }
-      return true;
-    });
-    return true;
-  });
-}
-
-int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_batches, int64_t sm_count, int64_t num_kv_splits) {
-  // Workspace size depends on ElementAcc and ElementLSE (same as ElementAcc)
-  // which are float, so Element type here doesn't matter.
-  using MlaSm100Type = MlaSm100<cutlass::half_t, true>;
-
-  // Get split kv. Requires problem shape and sm_count only.
-  typename MlaSm100Type::Fmha::Arguments arguments;
-  using TileShapeH = typename MlaSm100Type::TileShapeH;
-  using TileShapeD = typename MlaSm100Type::TileShapeD;
-  arguments.problem_shape =
-      cute::make_tuple(TileShapeH{}, static_cast<int>(max_seq_len), TileShapeD{}, static_cast<int>(num_batches));
-  // Assumes device 0 when getting sm_count.
-  arguments.hw_info.sm_count =
-      sm_count <= 0 ? cutlass::KernelHardwareInfo::query_device_multiprocessor_count(/*device_id=*/0) : sm_count;
-  arguments.split_kv = num_kv_splits;
-  MlaSm100Type::Fmha::set_split_kv(arguments);
-
-  return MlaSm100Type::Fmha::get_workspace_size(arguments);
-}
-
-#endif
-
-TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
-  m.impl("sm100_cutlass_mla_decode", &sm100_cutlass_mla_decode);
-}
-
-TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CatchAll, m) {
-  m.impl("sm100_cutlass_mla_get_workspace_size", &sm100_cutlass_mla_get_workspace_size);
-}
-
-// clang-format on

csrc/attention/paged_attention_v1.cu

@@ -18,7 +18,12 @@
  */
 
 #include "attention_kernels.cuh"
-#include "cuda_compat.h"
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -182,6 +187,7 @@ void paged_attention_v1(
                              CALL_V1_LAUNCHER_BLOCK_SIZE)
 }
 
+#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/attention/paged_attention_v2.cu

@@ -18,7 +18,12 @@
  */
 
 #include "attention_kernels.cuh"
-#include "cuda_compat.h"
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -192,6 +197,7 @@ void paged_attention_v2(
                              CALL_V2_LAUNCHER_BLOCK_SIZE)
 }
 
+#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/cpu/cpu_types_arm.hpp

@@ -33,8 +33,6 @@ namespace vec_op {
 #endif
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
-// Number of elements in single ASIMD vector of given Datatype
-#define NUM_ELEMENTS_REG(vec) (sizeof(vec) / sizeof(vec[0]))
 
 namespace {
 template <typename T, T... indexes, typename F>
@@ -88,8 +86,8 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
   }
 
   void save(void* ptr, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+    int full_blocks = elem_num / 8;
+    int remainder = elem_num % 8;
 
     if (full_blocks > 0) {
       vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);
@@ -199,25 +197,6 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
              vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3])}) {};
 
   void save(void* ptr) const { *reinterpret_cast<bfloat16x8x2_t*>(ptr) = reg; };
-  void save(void* ptr, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
-    for (int i = 0; i < full_blocks; i++)
-      vst1q_bf16(
-          reinterpret_cast<__bf16*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
-          reg.val[i]);
-    if (remainder > 0) {
-      bfloat16x8_t temp = reg.val[full_blocks];
-      bfloat16_t* base = reinterpret_cast<bfloat16_t*>(ptr) + full_blocks * 8;
-      if (remainder > 0) base[0] = vgetq_lane_bf16(temp, 0);
-      if (remainder > 1) base[1] = vgetq_lane_bf16(temp, 1);
-      if (remainder > 2) base[2] = vgetq_lane_bf16(temp, 2);
-      if (remainder > 3) base[3] = vgetq_lane_bf16(temp, 3);
-      if (remainder > 4) base[4] = vgetq_lane_bf16(temp, 4);
-      if (remainder > 5) base[5] = vgetq_lane_bf16(temp, 5);
-      if (remainder > 6) base[6] = vgetq_lane_bf16(temp, 6);
-    }
-  };
 };
 
 struct BF16Vec32 : public Vec<BF16Vec32> {
@@ -234,25 +213,6 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
       : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {};
 
   void save(void* ptr) const { *reinterpret_cast<bfloat16x8x4_t*>(ptr) = reg; };
-  void save(void* ptr, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
-    for (int i = 0; i < full_blocks; i++)
-      vst1q_bf16(
-          reinterpret_cast<__bf16*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
-          reg.val[i]);
-    if (remainder > 0) {
-      bfloat16x8_t temp = reg.val[full_blocks];
-      bfloat16_t* base = reinterpret_cast<bfloat16_t*>(ptr) + full_blocks * 8;
-      base[0] = vgetq_lane_bf16(temp, 0);
-      if (remainder > 1) base[1] = vgetq_lane_bf16(temp, 1);
-      if (remainder > 2) base[2] = vgetq_lane_bf16(temp, 2);
-      if (remainder > 3) base[3] = vgetq_lane_bf16(temp, 3);
-      if (remainder > 4) base[4] = vgetq_lane_bf16(temp, 4);
-      if (remainder > 5) base[5] = vgetq_lane_bf16(temp, 5);
-      if (remainder > 6) base[6] = vgetq_lane_bf16(temp, 6);
-    }
-  };
 };
 #endif
 
@@ -412,48 +372,6 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   }
 };
 
-struct INT32Vec16 : public Vec<INT32Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  union AliasReg {
-    int32x4x4_t reg;
-    int32_t values[VEC_ELEM_NUM];
-  };
-  int32x4x4_t reg;
-
-  explicit INT32Vec16(const void* ptr) {
-    reg.val[0] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr));
-    reg.val[1] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr) + 4);
-    reg.val[2] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr) + 8);
-    reg.val[3] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr) + 12);
-  }
-
-  void save(int32_t* ptr) const {
-    vst1q_s32(ptr, reg.val[0]);
-    vst1q_s32(ptr + 4, reg.val[1]);
-    vst1q_s32(ptr + 8, reg.val[2]);
-    vst1q_s32(ptr + 12, reg.val[3]);
-  };
-
-  void save(int32_t* ptr, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
-
-    for (int i = 0; i < full_blocks; i++)
-      vst1q_s32(
-          reinterpret_cast<__int32_t*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
-          reg.val[i]);
-
-    if (remainder > 0) {
-      int32x4_t temp = reg.val[full_blocks];
-      int32_t* base = reinterpret_cast<int32_t*>(ptr) + full_blocks * 4;
-      if (remainder > 0) base[0] = vgetq_lane_s32(temp, 0);
-      if (remainder > 1) base[1] = vgetq_lane_s32(temp, 1);
-      if (remainder > 2) base[2] = vgetq_lane_s32(temp, 2);
-      if (remainder > 3) base[3] = vgetq_lane_s32(temp, 3);
-    }
-  }
-};
-
 struct FP32Vec16 : public Vec<FP32Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
   union AliasReg {
@@ -516,12 +434,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1]));
     reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1]));
   };
-  explicit FP32Vec16(const INT32Vec16& v) {
-    reg.val[0] = vcvtq_f32_s32(v.reg.val[0]);
-    reg.val[1] = vcvtq_f32_s32(v.reg.val[1]);
-    reg.val[2] = vcvtq_f32_s32(v.reg.val[2]);
-    reg.val[3] = vcvtq_f32_s32(v.reg.val[3]);
-  };
+
   FP32Vec16 operator+(const FP32Vec16& b) const {
     return FP32Vec16(float32x4x4_t({vaddq_f32(reg.val[0], b.reg.val[0]),
                                     vaddq_f32(reg.val[1], b.reg.val[1]),
@@ -550,85 +463,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                                     vdivq_f32(reg.val[3], b.reg.val[3])}));
   };
 
-  FP32Vec16 clamp(const FP32Vec16& min, const FP32Vec16& max) const {
-    return FP32Vec16(float32x4x4_t(
-        {vminq_f32(max.reg.val[0], vmaxq_f32(min.reg.val[0], reg.val[0])),
-         vminq_f32(max.reg.val[1], vmaxq_f32(min.reg.val[1], reg.val[1])),
-         vminq_f32(max.reg.val[2], vmaxq_f32(min.reg.val[2], reg.val[2])),
-         vminq_f32(max.reg.val[3], vmaxq_f32(min.reg.val[3], reg.val[3]))}));
-  };
-
-  FP32Vec16 max(const FP32Vec16& b) const {
-    return FP32Vec16(float32x4x4_t({vmaxq_f32(b.reg.val[0], reg.val[0]),
-                                    vmaxq_f32(b.reg.val[1], reg.val[1]),
-                                    vmaxq_f32(b.reg.val[2], reg.val[2]),
-                                    vmaxq_f32(b.reg.val[3], reg.val[3])}));
-  };
-
-  FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
-    float32x4x4_t temp;
-
-    for (int i = 0; i < full_blocks; i++)
-      temp.val[i] = vmaxq_f32(b.reg.val[i], reg.val[i]);
-
-    if (remainder > 0) {
-      float max_v = std::max(vgetq_lane_f32(reg.val[full_blocks], 0),
-                             vgetq_lane_f32(b.reg.val[full_blocks], 0));
-      temp.val[full_blocks] = vsetq_lane_f32(max_v, temp.val[full_blocks], 0);
-    }
-    if (remainder > 1) {
-      float max_v = std::max(vgetq_lane_f32(reg.val[full_blocks], 1),
-                             vgetq_lane_f32(b.reg.val[full_blocks], 1));
-      temp.val[full_blocks] = vsetq_lane_f32(max_v, temp.val[full_blocks], 1);
-    }
-    if (remainder > 2) {
-      float max_v = std::max(vgetq_lane_f32(reg.val[full_blocks], 2),
-                             vgetq_lane_f32(b.reg.val[full_blocks], 2));
-      temp.val[full_blocks] = vsetq_lane_f32(max_v, temp.val[full_blocks], 2);
-    }
-    return FP32Vec16(temp);
-  };
-
-  FP32Vec16 min(const FP32Vec16& b) const {
-    return FP32Vec16(float32x4x4_t({
-        vminq_f32(b.reg.val[0], reg.val[0]),
-        vminq_f32(b.reg.val[1], reg.val[1]),
-        vminq_f32(b.reg.val[2], reg.val[2]),
-        vminq_f32(b.reg.val[3], reg.val[3]),
-    }));
-  };
-  FP32Vec16 min(const FP32Vec16& b, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    const int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
-    float32x4x4_t temp;
-    for (int i = 0; i < full_blocks; i++)
-      temp.val[i] = vminq_f32(b.reg.val[i], reg.val[i]);
-
-    if (remainder > 0) {
-      float min_v = std::min(vgetq_lane_f32(reg.val[full_blocks], 0),
-                             vgetq_lane_f32(b.reg.val[full_blocks], 0));
-      temp.val[full_blocks] = vsetq_lane_f32(min_v, temp.val[full_blocks], 0);
-    }
-    if (remainder > 1) {
-      float min_v = std::min(vgetq_lane_f32(reg.val[full_blocks], 1),
-                             vgetq_lane_f32(b.reg.val[full_blocks], 1));
-      temp.val[full_blocks] = vsetq_lane_f32(min_v, temp.val[full_blocks], 1);
-    }
-    if (remainder > 2) {
-      float min_v = std::min(vgetq_lane_f32(reg.val[full_blocks], 2),
-                             vgetq_lane_f32(b.reg.val[full_blocks], 2));
-      temp.val[full_blocks] = vsetq_lane_f32(min_v, temp.val[full_blocks], 2);
-    }
-
-    return FP32Vec16(temp);
-  };
-  FP32Vec16 abs() const {
-    return FP32Vec16(
-        float32x4x4_t({vabsq_f32(reg.val[0]), vabsq_f32(reg.val[1]),
-                       vabsq_f32(reg.val[2]), vabsq_f32(reg.val[3])}));
-  }
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
@@ -639,24 +473,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return answer;
   };
 
-  float reduce_max() const {
-    AliasReg ar;
-    ar.reg = reg;
-    float max_v = std::numeric_limits<float>::lowest();
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&max_v, &ar](int i) { max_v = std::max(max_v, ar.values[i]); });
-    return max_v;
-  }
-
-  float reduce_min() const {
-    AliasReg ar;
-    ar.reg = reg;
-    float min_v = std::numeric_limits<float>::max();
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&min_v, &ar](int i) { min_v = std::min(min_v, ar.values[i]); });
-    return min_v;
-  }
-
   template <int group_size>
   float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
@@ -677,83 +493,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     vst1q_f32(ptr + 8, reg.val[2]);
     vst1q_f32(ptr + 12, reg.val[3]);
   };
-
-  void save(float* ptr, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
-
-    for (int i = 0; i < full_blocks; i++)
-      vst1q_f32(
-          reinterpret_cast<float32_t*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
-          reg.val[i]);
-
-    if (remainder > 0) {
-      float32x4_t temp = reg.val[full_blocks];
-      float* base = reinterpret_cast<float32_t*>(ptr) +
-                    full_blocks * NUM_ELEMENTS_REG(reg.val[0]);
-      if (remainder > 0) base[0] = vgetq_lane_f32(temp, 0);
-      if (remainder > 1) base[1] = vgetq_lane_f32(temp, 1);
-      if (remainder > 2) base[2] = vgetq_lane_f32(temp, 2);
-    }
-  }
-};
-
-struct INT8Vec16 : public Vec<INT8Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  union AliasReg {
-    int8x16_t reg;
-    int8_t values[VEC_ELEM_NUM];
-  };
-  int8x16_t reg;
-
-  explicit INT8Vec16(const FP32Vec16& vec) {
-    // Convert each 128-bit float32 vector to int32
-    int32x4_t part0 =
-        vcvtq_s32_f32(vec.reg.val[0]);  // Convert first 128-bit block
-    int32x4_t part1 =
-        vcvtq_s32_f32(vec.reg.val[1]);  // Convert second 128-bit block
-    int32x4_t part2 =
-        vcvtq_s32_f32(vec.reg.val[2]);  // Convert third 128-bit block
-    int32x4_t part3 =
-        vcvtq_s32_f32(vec.reg.val[3]);  // Convert fourth 128-bit block
-
-    // Narrow each 32-bit vector to 8 bits and combine
-    int8x8_t lower =
-        vqmovn_s16(vcombine_s16(vqmovn_s32(part0), vqmovn_s32(part1)));
-    int8x8_t upper =
-        vqmovn_s16(vcombine_s16(vqmovn_s32(part2), vqmovn_s32(part3)));
-    reg = vcombine_s8(lower, upper);  // Combine to form a single 128-bit vector
-  }
-
-  void save(int8_t* ptr) const { vst1q_s8(ptr, reg); };
-
-  void save(int8_t* ptr, const int elem_num) const {
-    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg);
-    int remainder = elem_num % NUM_ELEMENTS_REG(reg);
-
-    for (int i = 0; i < full_blocks; i++)
-      vst1q_s8(reinterpret_cast<int8_t*>(ptr) + NUM_ELEMENTS_REG(reg) * i, reg);
-    if (remainder > 0) {
-      int8x16_t temp = reg;
-      int8_t* base =
-          reinterpret_cast<int8_t*>(ptr) + full_blocks * NUM_ELEMENTS_REG(reg);
-      if (remainder > 0) base[0] = vgetq_lane_s8(temp, 0);
-      if (remainder > 1) base[1] = vgetq_lane_s8(temp, 1);
-      if (remainder > 2) base[2] = vgetq_lane_s8(temp, 2);
-      if (remainder > 3) base[3] = vgetq_lane_s8(temp, 3);
-      if (remainder > 4) base[4] = vgetq_lane_s8(temp, 4);
-      if (remainder > 5) base[5] = vgetq_lane_s8(temp, 5);
-      if (remainder > 6) base[6] = vgetq_lane_s8(temp, 6);
-      if (remainder > 7) base[7] = vgetq_lane_s8(temp, 7);
-      if (remainder > 8) base[8] = vgetq_lane_s8(temp, 8);
-      if (remainder > 9) base[9] = vgetq_lane_s8(temp, 9);
-      if (remainder > 10) base[10] = vgetq_lane_s8(temp, 10);
-      if (remainder > 11) base[11] = vgetq_lane_s8(temp, 11);
-      if (remainder > 12) base[12] = vgetq_lane_s8(temp, 12);
-      if (remainder > 13) base[13] = vgetq_lane_s8(temp, 13);
-      if (remainder > 14) base[14] = vgetq_lane_s8(temp, 14);
-    }
-  };
 };
 
 template <typename T>

csrc/cpu/dnnl_helper.hpp

@@ -57,7 +57,6 @@ class DNNLPrimitiveHelper {
   // Note: Due to the limitation of oneDNN
   // (https://github.com/oneapi-src/oneDNN/issues/1636), the quantized bias is
   // not supported.
-
   template <typename OutputT, typename BiasT>
   static void gemm_s8s8_jit(const int8_t* a, const int8_t* b, OutputT* c,
                             const BiasT* bias, dnnl_dim_t M, dnnl_dim_t N,
@@ -91,27 +90,6 @@ class DNNLPrimitiveHelper {
     }
 
     dnnl::matmul::primitive_desc matmul_pd;
-// Create memory descriptors with format_tag::any for the primitive. This
-// enables the matmul primitive to choose memory layouts for an
-// optimized primitive implementation, and these layouts may differ from the
-// ones provided by the user.
-#ifdef __aarch64__
-    auto mat_src_md = dnnl::memory::desc({M, K}, dnnl::memory::data_type::s8,
-                                         dnnl::memory::format_tag::any);
-    auto mat_weights_md = dnnl::memory::desc(
-        {K, N}, dnnl::memory::data_type::s8, dnnl::memory::format_tag::any);
-    auto mat_dst_md =
-        dnnl::memory::desc({M, N}, OutputType, dnnl::memory::format_tag::any);
-    if (bias) {
-      dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), mat_src_md,
-                                               mat_weights_md, bias_md,
-                                               mat_dst_md, attr);
-    } else {
-      matmul_pd = dnnl::matmul::primitive_desc(
-          default_engine(), mat_src_md, mat_weights_md, mat_dst_md, attr);
-    }
-#else
     if (bias) {
       dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
       matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
@@ -120,7 +98,6 @@ class DNNLPrimitiveHelper {
       matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
                                                c_md, attr);
     }
-#endif
     dnnl::matmul matmul(matmul_pd);
 
     auto& engine = default_engine();
@@ -134,34 +111,24 @@ class DNNLPrimitiveHelper {
                             (void*)b_scales);
 
     auto& stream = default_stream();
-
-    auto mat_src_mem = a_m;
-    auto mat_weights_mem = b_m;
-    auto mat_dst_mem = c_m;
-#ifdef __aarch64__
-    if (matmul_pd.weights_desc() != b_m.get_desc()) {
-      mat_weights_mem = dnnl::memory(matmul_pd.weights_desc(), engine);
-      dnnl::reorder(b_m, mat_weights_mem).execute(stream, b_m, mat_weights_mem);
-    }
-#endif
     if constexpr (InputNoScale) {
       if (bias) {
         dnnl::memory::desc bias_md({N}, BiasType, {1});
         dnnl::memory bias_m(bias_md, engine, (void*)bias);
         matmul.execute(
             stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
                         {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, mat_dst_mem},
+                        {DNNL_ARG_DST, c_m},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
                     });
       } else {
         matmul.execute(
             stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_DST, mat_dst_mem},
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_DST, c_m},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
                     });
       }
@@ -171,19 +138,19 @@ class DNNLPrimitiveHelper {
         dnnl::memory bias_m(bias_md, engine, (void*)bias);
         matmul.execute(
             stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
                         {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, mat_dst_mem},
+                        {DNNL_ARG_DST, c_m},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
                     });
       } else {
         matmul.execute(
             stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_DST, mat_dst_mem},
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_DST, c_m},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
                     });
@@ -203,4 +170,5 @@ class DNNLPrimitiveHelper {
     return stream;
   }
 };
+
 #endif

csrc/cpu/quant.cpp

@@ -36,7 +36,7 @@ struct KernelVecType<c10::Half> {
   using cvt_vec_type = vec_op::FP32Vec16;
 };
 
-#if defined(__AVX512F__) || defined(__aarch64__)
+#ifdef __AVX512F__
 template <bool AZP, typename scalar_t>
 void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
                                    const float* scale, const int32_t* azp,
@@ -598,9 +598,8 @@ void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
                                    const float* scale, const int32_t* azp,
                                    const int num_tokens,
                                    const int hidden_size) {
-  TORCH_CHECK(false,
-              "static_scaled_int8_quant_impl requires AVX512/powerpc64/AArch64 "
-              "support.")
+  TORCH_CHECK(
+      false, "static_scaled_int8_quant_impl requires AVX512/powerpc64 support.")
 }
 
 template <typename scalar_t>
@@ -608,9 +607,9 @@ void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
                                     float* scale, int32_t* azp,
                                     const int num_tokens,
                                     const int hidden_size) {
-  TORCH_CHECK(false,
-              "dynamic_scaled_int8_quant_impl requires "
-              "AVX512/powerpc64/AArch64 support.")
+  TORCH_CHECK(
+      false,
+      "dynamic_scaled_int8_quant_impl requires AVX512/powerpc64 support.")
 }
 
 template <bool PerChannel, typename scalar_t>
@@ -618,8 +617,7 @@ void static_quant_epilogue(const float* input, scalar_t* output,
                            const float a_scale, const float* b_scale,
                            const int32_t* azp_with_adj, const int num_tokens,
                            const int hidden_size) {
-  TORCH_CHECK(
-      false, "static_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
+  TORCH_CHECK(false, "static_quant_epilogue requires AVX512/powerpc64 support.")
 }
 
 template <typename scalar_t>
@@ -628,9 +626,8 @@ void dynamic_quant_epilogue(const float* input, scalar_t* output,
                             const int32_t* azp, const int32_t* azp_with_adj,
                             const scalar_t* bias, const int num_tokens,
                             const int hidden_size) {
-  TORCH_CHECK(
-      false,
-      "dynamic_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
+  TORCH_CHECK(false,
+              "dynamic_quant_epilogue requires AVX512/powerpc64 support.")
 }
 #endif
 }  // namespace

csrc/cpu/sgl-kernels/common.h

@@ -58,7 +58,7 @@ namespace {
 
 #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
 #define CHECK_LAST_DIM_CONTIGUOUS(x) \
-  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimension")
+  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimention")
 
 #define CHECK_INPUT(x) \
   CHECK_CPU(x);        \

csrc/cpu/sgl-kernels/gemm.h

@@ -126,7 +126,7 @@ void fused_experts_int4_w4a16_kernel_impl(
     int64_t topk,
     int64_t num_tokens_post_pad);
 
-// shared expert implementation for int8 w8a8
+// shared expert implememntation for int8 w8a8
 template <typename scalar_t>
 void shared_expert_int8_kernel_impl(
     scalar_t* __restrict__ output,

csrc/cpu/sgl-kernels/gemm_int8.cpp

@@ -41,7 +41,7 @@ struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
     __m512  vd0;
     __m512  vd1[COLS];
 
-    // oops! 4x4 spills but luckily we use 4x2
+    // oops! 4x4 spills but luckly we use 4x2
     __m512 vbias[COLS];
 
     // [NOTE]: s8s8 igemm compensation in avx512-vnni

csrc/cpu/sgl-kernels/vec.h

@@ -37,7 +37,7 @@ inline Vectorized<at::BFloat16> convert_from_float_ext<at::BFloat16>(const Vecto
 #define CVT_FP16_TO_FP32(a) \
     _mm512_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
 
-// this doesn't handle NaN.
+// this doesn't hanel NaN.
 inline __m512bh cvt_e4m3_bf16_intrinsic_no_nan(__m256i fp8_vec) {
   const __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
 

csrc/cpu/shm.cpp

@@ -7,7 +7,7 @@
 
 namespace {
 #define MAX_SHM_RANK_NUM 8
-#define PER_THREAD_SHM_BUFFER_BYTES (4 * 1024 * 1024)
+#define PER_THREAD_SHM_BUFFER_BYTES (2 * 1024 * 1024)
 static_assert(PER_THREAD_SHM_BUFFER_BYTES % 2 == 0);
 #define PER_THREAD_SHM_BUFFER_OFFSET (PER_THREAD_SHM_BUFFER_BYTES >> 1)
 #define MIN_THREAD_PROCESS_SIZE (256)
@@ -34,10 +34,9 @@ struct KernelVecType<c10::Half> {
 };
 
 struct ThreadSHMContext {
-  volatile char _curr_thread_stamp[2];
-  volatile char _ready_thread_stamp[2];
-  int local_stamp_buffer_idx;
-  int remote_stamp_buffer_idx;
+  volatile char _curr_thread_stamp;
+  volatile char _ready_thread_stamp;
+  char _padding1[6];
   int thread_id;
   int thread_num;
   int rank;
@@ -46,28 +45,23 @@ struct ThreadSHMContext {
   int swizzled_ranks[MAX_SHM_RANK_NUM];
   void* thread_shm_ptrs[MAX_SHM_RANK_NUM];
   ThreadSHMContext* shm_contexts[MAX_SHM_RANK_NUM];
-  size_t _thread_buffer_mask[2];
-  char _padding2[40];
+  size_t _thread_buffer_mask;
+  char _padding2[56];
 
   ThreadSHMContext(const int thread_id, const int thread_num, const int rank,
                    const int group_size, void* thread_shm_ptr)
-      : local_stamp_buffer_idx(0),
-        remote_stamp_buffer_idx(0),
+      : _curr_thread_stamp(1),
+        _ready_thread_stamp(0),
         thread_id(thread_id),
         thread_num(thread_num),
         rank(rank),
         group_size(group_size),
-        _spinning_count(0) {
+        _spinning_count(0),
+        _thread_buffer_mask(0) {
     static_assert(sizeof(ThreadSHMContext) % 64 == 0);
     TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
     TORCH_CHECK((size_t)this % 64 == 0);
     TORCH_CHECK((size_t)thread_shm_ptr % 64 == 0);
-    _curr_thread_stamp[0] = 1;
-    _curr_thread_stamp[1] = 1;
-    _ready_thread_stamp[0] = 0;
-    _ready_thread_stamp[1] = 0;
-    _thread_buffer_mask[0] = 0;
-    _thread_buffer_mask[1] = 0;
     for (int i = 0; i < MAX_SHM_RANK_NUM; ++i) {
       shm_contexts[i] = nullptr;
       thread_shm_ptrs[i] = nullptr;
@@ -76,11 +70,6 @@ struct ThreadSHMContext {
     set_context(rank, this, thread_shm_ptr);
   }
 
-  void set_stamp_buffer_idx(int local, int remote) {
-    local_stamp_buffer_idx = local;
-    remote_stamp_buffer_idx = remote;
-  }
-
   void set_context(int rank, ThreadSHMContext* ptr, void* thread_shm_ptr) {
     TORCH_CHECK(rank < MAX_SHM_RANK_NUM);
     TORCH_CHECK(ptr);
@@ -95,27 +84,23 @@ struct ThreadSHMContext {
   T* get_thread_shm_ptr(int rank) {
     return reinterpret_cast<T*>(
         reinterpret_cast<int8_t*>(thread_shm_ptrs[rank]) +
-        (PER_THREAD_SHM_BUFFER_OFFSET &
-         _thread_buffer_mask[local_stamp_buffer_idx]));
+        (PER_THREAD_SHM_BUFFER_OFFSET & _thread_buffer_mask));
   }
 
-  void next_buffer() {
-    _thread_buffer_mask[local_stamp_buffer_idx] ^= 0xFFFFFFFFFFFFFFFF;
-  }
+  void next_buffer() { _thread_buffer_mask ^= 0xFFFFFFFFFFFFFFFF; }
 
-  char get_curr_stamp(int idx) const { return _curr_thread_stamp[idx]; }
+  char get_curr_stamp() const { return _curr_thread_stamp; }
 
-  char get_ready_stamp(int idx) const { return _ready_thread_stamp[idx]; }
+  char get_ready_stamp() const { return _ready_thread_stamp; }
 
   void next_stamp() {
     _mm_mfence();
-    _curr_thread_stamp[local_stamp_buffer_idx] += 1;
+    _curr_thread_stamp += 1;
   }
 
   void commit_ready_stamp() {
     _mm_mfence();
-    _ready_thread_stamp[local_stamp_buffer_idx] =
-        _curr_thread_stamp[local_stamp_buffer_idx];
+    _ready_thread_stamp = _curr_thread_stamp;
   }
 
   int get_swizzled_rank(int idx) { return swizzled_ranks[idx]; }
@@ -132,11 +117,10 @@ struct ThreadSHMContext {
   void wait_for_one(int rank, Cond&& cond) {
     ThreadSHMContext* rank_ctx = shm_contexts[rank];
     for (;;) {
-      char local_curr_stamp = get_curr_stamp(local_stamp_buffer_idx);
-      char local_ready_stamp = get_ready_stamp(local_stamp_buffer_idx);
-      char rank_curr_stamp = rank_ctx->get_curr_stamp(remote_stamp_buffer_idx);
-      char rank_ready_stamp =
-          rank_ctx->get_ready_stamp(remote_stamp_buffer_idx);
+      char local_curr_stamp = get_curr_stamp();
+      char local_ready_stamp = get_ready_stamp();
+      char rank_curr_stamp = rank_ctx->get_curr_stamp();
+      char rank_ready_stamp = rank_ctx->get_ready_stamp();
       if (cond(local_curr_stamp, local_ready_stamp, rank_curr_stamp,
                rank_ready_stamp)) {
         break;
@@ -377,15 +361,6 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
     }
   }
 }
-
-void reset_threads_stamp_buffer_idx(ThreadSHMContext* ctx, int local,
-                                    int remote) {
-  int thread_num = ctx->thread_num;
-  for (int i = 0; i < thread_num; ++i) {
-    ThreadSHMContext* thread_ctx = ctx + i;
-    thread_ctx->set_stamp_buffer_idx(local, remote);
-  }
-}
 };  // namespace shm_cc_ops
 
 namespace shm_cc_ops {
@@ -657,7 +632,6 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx, int64_t dst,
   TensorListMeta* metadata = new (metadata_tensor.data_ptr()) TensorListMeta();
   metadata->bind_tensor_list(tensor_list_with_metadata);
 
-  shm_cc_ops::reset_threads_stamp_buffer_idx(ctx, 0, 1);
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata->total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
@@ -685,7 +659,6 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   torch::Tensor metadata_tensor =
       torch::empty({sizeof(TensorListMeta)}, options);
 
-  shm_cc_ops::reset_threads_stamp_buffer_idx(ctx, 1, 0);
   ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
   shm_cc_ops::memcpy(metadata_tensor.data_ptr(),
                      ctx->get_thread_shm_ptr<void>(src),
@@ -704,7 +677,7 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
       ctx, metadata.total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
           int64_t data_elem_num, bool fast_mode) {
-        thread_ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
+        ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
         int64_t curr_shm_offset = 0;
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata.get_data(data_offset + curr_shm_offset);

csrc/cpu/torch_bindings.cpp

@@ -151,9 +151,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
 
   // Quantization
-#if defined(__AVX512F__) || defined(__aarch64__)
+#ifdef __AVX512F__
   at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
-
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
       "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"

csrc/cuda_compat.h

@@ -4,10 +4,10 @@
   #include <hip/hip_runtime.h>
 #endif
 
-#if defined(USE_ROCM) && defined(__GFX9__)
-  #define WARP_SIZE 64
-#else
+#ifndef USE_ROCM
   #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
 #endif
 
 #ifndef USE_ROCM

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c2x.hpp

@@ -153,7 +153,7 @@ struct ScaledEpilogueBias
       cutlass::epilogue::threadblock::Sm80EVT<Compute0, ScaleB, Accum>;
 
   using Compute1 = cutlass::epilogue::threadblock::VisitorCompute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -210,7 +210,7 @@ struct ScaledEpilogueBiasAzp
                                               EVTComputeAzp>;
 
   using ComputeScaleBiasA = cutlass::epilogue::threadblock::VisitorCompute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -288,7 +288,7 @@ struct ScaledEpilogueBiasAzpToken
                                               EVTComputeAcc>;
 
   using ComputeScaleBiasA = cutlass::epilogue::threadblock::VisitorCompute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp

@@ -195,7 +195,7 @@ struct ScaledEpilogueBias
       cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
 
   using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -238,7 +238,7 @@ struct ScaledEpilogueColumnBias
       cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
 
   using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -295,7 +295,7 @@ struct ScaledEpilogueBiasAzp
       cutlass::epilogue::fusion::Sm90EVT<ComputeScaleB, ScaleB, EVTComputeAzp>;
 
   using ComputeScaleBiasA = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -371,7 +371,7 @@ struct ScaledEpilogueBiasAzpToken
       cutlass::epilogue::fusion::Sm90EVT<ComputeScaleB, ScaleB, EVTComputeAcc>;
 
   using ComputeScaleBiasA = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::homogeneous_multiply_add, ElementD, float,
+      cutlass::multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:

csrc/layernorm_kernels.cu

@@ -15,16 +15,15 @@ namespace vllm {
 // TODO(woosuk): Further optimize this kernel.
 template <typename scalar_t>
 __global__ void rms_norm_kernel(
-    scalar_t* __restrict__ out,          // [..., hidden_size]
-    const scalar_t* __restrict__ input,  // [..., hidden_size]
-    const int64_t input_stride,
+    scalar_t* __restrict__ out,           // [..., hidden_size]
+    const scalar_t* __restrict__ input,   // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float)input[blockIdx.x * input_stride + idx];
+    const float x = (float)input[blockIdx.x * hidden_size + idx];
     variance += x * x;
   }
 
@@ -38,7 +37,7 @@ __global__ void rms_norm_kernel(
   __syncthreads();
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * input_stride + idx];
+    float x = (float)input[blockIdx.x * hidden_size + idx];
     out[blockIdx.x * hidden_size + idx] =
         ((scalar_t)(x * s_variance)) * weight[idx];
   }
@@ -51,8 +50,7 @@ __global__ void rms_norm_kernel(
 template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
-    scalar_t* __restrict__ input,  // [..., hidden_size]
-    const int64_t input_stride,
+    scalar_t* __restrict__ input,         // [..., hidden_size]
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -61,7 +59,6 @@ fused_add_rms_norm_kernel(
   static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
 
   const int vec_hidden_size = hidden_size / width;
-  const int64_t vec_input_stride = input_stride / width;
   __shared__ float s_variance;
   float variance = 0.0f;
   /* These and the argument pointers are all declared `restrict` as they are
@@ -76,8 +73,7 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
-    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
-    _f16Vec<scalar_t, width> temp = input_v[strided_id];
+    _f16Vec<scalar_t, width> temp = input_v[id];
     temp += residual_v[id];
     variance += temp.sum_squares();
     residual_v[id] = temp;
@@ -94,11 +90,10 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
-    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
     _f16Vec<scalar_t, width> temp = residual_v[id];
     temp *= s_variance;
     temp *= weight_v[idx];
-    input_v[strided_id] = temp;
+    input_v[id] = temp;
   }
 }
 
@@ -108,8 +103,7 @@ fused_add_rms_norm_kernel(
 template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
-    scalar_t* __restrict__ input,  // [..., hidden_size]
-    const int64_t input_stride,
+    scalar_t* __restrict__ input,         // [..., hidden_size]
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -117,7 +111,7 @@ fused_add_rms_norm_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    scalar_t z = input[blockIdx.x * input_stride + idx];
+    scalar_t z = input[blockIdx.x * hidden_size + idx];
     z += residual[blockIdx.x * hidden_size + idx];
     float x = (float)z;
     variance += x * x;
@@ -135,7 +129,7 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
     float x = (float)residual[blockIdx.x * hidden_size + idx];
-    input[blockIdx.x * input_stride + idx] =
+    input[blockIdx.x * hidden_size + idx] =
         ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
@@ -147,12 +141,11 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
               torch::Tensor& weight,  // [hidden_size]
               double epsilon) {
   TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(input.stride(-1) == 1);
+  TORCH_CHECK(input.is_contiguous());
   TORCH_CHECK(weight.is_contiguous());
 
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
-  int64_t input_stride = input.stride(-2);
 
   dim3 grid(num_tokens);
   dim3 block(std::min(hidden_size, 1024));
@@ -160,29 +153,26 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_kernel", [&] {
     vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), input_stride,
+        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
         weight.data_ptr<scalar_t>(), epsilon, num_tokens, hidden_size);
   });
 }
 
-#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                    \
-  VLLM_DISPATCH_FLOATING_TYPES(                                             \
-      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {               \
-        vllm::fused_add_rms_norm_kernel<scalar_t, width>                    \
-            <<<grid, block, 0, stream>>>(                                   \
-                input.data_ptr<scalar_t>(), input_stride,                   \
-                residual.data_ptr<scalar_t>(), weight.data_ptr<scalar_t>(), \
-                epsilon, num_tokens, hidden_size);                          \
+#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                       \
+  VLLM_DISPATCH_FLOATING_TYPES(                                                \
+      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {                  \
+        vllm::fused_add_rms_norm_kernel<scalar_t, width>                       \
+            <<<grid, block, 0, stream>>>(input.data_ptr<scalar_t>(),           \
+                                         residual.data_ptr<scalar_t>(),        \
+                                         weight.data_ptr<scalar_t>(), epsilon, \
+                                         num_tokens, hidden_size);             \
       });
 
 void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
                         torch::Tensor& residual,  // [..., hidden_size]
                         torch::Tensor& weight,    // [hidden_size]
                         double epsilon) {
-  TORCH_CHECK(residual.is_contiguous());
-  TORCH_CHECK(weight.is_contiguous());
   int hidden_size = input.size(-1);
-  int64_t input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -204,16 +194,9 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
   auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
   auto res_ptr = reinterpret_cast<std::uintptr_t>(residual.data_ptr());
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
-  constexpr int vector_width = 8;
-  constexpr int req_alignment_bytes =
-      vector_width * 2;  // vector_width * sizeof(bfloat16 or float16) (float32
-                         // falls back to non-vectorized version anyway)
-  bool ptrs_are_aligned = inp_ptr % req_alignment_bytes == 0 &&
-                          res_ptr % req_alignment_bytes == 0 &&
-                          wt_ptr % req_alignment_bytes == 0;
-  bool offsets_are_multiple_of_vector_width =
-      hidden_size % vector_width == 0 && input_stride % vector_width == 0;
-  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width) {
+  bool ptrs_are_aligned =
+      inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
+  if (ptrs_are_aligned && hidden_size % 8 == 0) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/layernorm_quant_kernels.cu

@@ -23,9 +23,8 @@ namespace vllm {
 // TODO(woosuk): Further optimize this kernel.
 template <typename scalar_t, typename fp8_type>
 __global__ void rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,          // [..., hidden_size]
-    const scalar_t* __restrict__ input,  // [..., hidden_size]
-    const int input_stride,
+    fp8_type* __restrict__ out,           // [..., hidden_size]
+    const scalar_t* __restrict__ input,   // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -33,7 +32,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float)input[blockIdx.x * input_stride + idx];
+    const float x = (float)input[blockIdx.x * hidden_size + idx];
     variance += x * x;
   }
 
@@ -50,7 +49,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   float const scale_inv = 1.0f / *scale;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * input_stride + idx];
+    float x = (float)input[blockIdx.x * hidden_size + idx];
     float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
     out[blockIdx.x * hidden_size + idx] =
         scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
@@ -64,9 +63,8 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 template <typename scalar_t, int width, typename fp8_type>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,    // [..., hidden_size]
-    scalar_t* __restrict__ input,  // [..., hidden_size]
-    const int input_stride,
+    fp8_type* __restrict__ out,           // [..., hidden_size]
+    scalar_t* __restrict__ input,         // [..., hidden_size]
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
@@ -76,7 +74,6 @@ fused_add_rms_norm_static_fp8_quant_kernel(
   static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
 
   const int vec_hidden_size = hidden_size / width;
-  const int vec_input_stride = input_stride / width;
   __shared__ float s_variance;
   float variance = 0.0f;
   /* These and the argument pointers are all declared `restrict` as they are
@@ -90,9 +87,8 @@ fused_add_rms_norm_static_fp8_quant_kernel(
       reinterpret_cast<const _f16Vec<scalar_t, width>*>(weight);
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
-    int stride_id = blockIdx.x * vec_input_stride + idx;
     int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> temp = input_v[stride_id];
+    _f16Vec<scalar_t, width> temp = input_v[id];
     temp += residual_v[id];
     variance += temp.sum_squares();
     residual_v[id] = temp;
@@ -129,9 +125,8 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 template <typename scalar_t, int width, typename fp8_type>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,    // [..., hidden_size]
-    scalar_t* __restrict__ input,  // [..., hidden_size]
-    const int input_stride,
+    fp8_type* __restrict__ out,           // [..., hidden_size]
+    scalar_t* __restrict__ input,         // [..., hidden_size]
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
@@ -140,7 +135,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    scalar_t z = input[blockIdx.x * input_stride + idx];
+    scalar_t z = input[blockIdx.x * hidden_size + idx];
     z += residual[blockIdx.x * hidden_size + idx];
     float x = (float)z;
     variance += x * x;
@@ -174,9 +169,7 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
                                torch::Tensor& weight,  // [hidden_size]
                                torch::Tensor& scale,   // [1]
                                double epsilon) {
-  TORCH_CHECK(out.is_contiguous());
   int hidden_size = input.size(-1);
-  int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -190,9 +183,8 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
               vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t>
                   <<<grid, block, 0, stream>>>(
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                      input_stride, weight.data_ptr<scalar_t>(),
-                      scale.data_ptr<float>(), epsilon, num_tokens,
-                      hidden_size);
+                      weight.data_ptr<scalar_t>(), scale.data_ptr<float>(),
+                      epsilon, num_tokens, hidden_size);
             });
       });
 }
@@ -206,7 +198,7 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
                                                                width, fp8_t> \
                   <<<grid, block, 0, stream>>>(                              \
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),     \
-                      input_stride, residual.data_ptr<scalar_t>(),           \
+                      residual.data_ptr<scalar_t>(),                         \
                       weight.data_ptr<scalar_t>(), scale.data_ptr<float>(),  \
                       epsilon, num_tokens, hidden_size);                     \
             });                                                              \
@@ -218,10 +210,7 @@ void fused_add_rms_norm_static_fp8_quant(
     torch::Tensor& weight,    // [hidden_size]
     torch::Tensor& scale,     // [1]
     double epsilon) {
-  TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(residual.is_contiguous());
   int hidden_size = input.size(-1);
-  int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -245,7 +234,7 @@ void fused_add_rms_norm_static_fp8_quant(
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
   bool ptrs_are_aligned =
       inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 0) {
+  if (ptrs_are_aligned && hidden_size % 8 == 0) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/mamba/causal_conv1d/causal_conv1d.cu

@@ -0,0 +1,656 @@
+// clang-format off
+// adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/causal_conv1d_fwd.cu 
+// and https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/causal_conv1d_update.cu
+#include <torch/all.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include "causal_conv1d.h"
+#include <c10/util/BFloat16.h>
+#include <c10/util/Half.h>
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_store.cuh>
+
+#ifdef USE_ROCM
+    namespace cub = hipcub;
+#endif
+
+#include "static_switch.h"
+
+
+
+#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")
+
+#define DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(ITYPE, NAME, ...)              \
+    if (ITYPE == at::ScalarType::Half) {                                            \
+        using input_t = at::Half;                                                   \
+        using weight_t = at::Half;                                                  \
+        __VA_ARGS__();                                                              \
+    } else if (ITYPE == at::ScalarType::BFloat16) {                                 \
+        using input_t = at::BFloat16;                                               \
+        using weight_t = at::BFloat16;                                              \
+        __VA_ARGS__();                                                              \
+    } else if (ITYPE == at::ScalarType::Float)  {                                   \
+        using input_t = float;                                                      \
+        using weight_t = float;                                                     \
+        __VA_ARGS__();                                                              \
+    } else {                                                                        \
+        AT_ERROR(#NAME, " not implemented for input type '", toString(ITYPE), "'"); \
+    }
+
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_fwd_cuda(ConvParamsBase &params, cudaStream_t stream);
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_update_cuda(ConvParamsBase &params, cudaStream_t stream);
+
+void set_conv_params_fwd(ConvParamsBase &params,
+                         // sizes
+                         const size_t batch,
+                         const size_t dim,
+                         const size_t seqlen,
+                         const size_t width,
+                         // device pointers
+                         const at::Tensor x,
+                         const at::Tensor weight,
+                         const at::Tensor out,
+                         const std::optional<at::Tensor>& bias,
+                         bool silu_activation,
+                         int64_t pad_slot_id,
+                         const std::optional<at::Tensor>& query_start_loc = std::nullopt,
+                         const std::optional<at::Tensor>& cache_indices = std::nullopt,
+                         const std::optional<at::Tensor>& has_initial_state = std::nullopt) {
+
+    // Reset the parameters
+    memset(&params, 0, sizeof(params));
+
+    params.batch = batch;
+    params.dim = dim;
+    params.seqlen = seqlen;
+    params.width = width;
+    params.pad_slot_id = pad_slot_id;
+
+    params.silu_activation = silu_activation;
+
+    // Set the pointers and strides.
+    params.x_ptr = x.data_ptr();
+    params.weight_ptr = weight.data_ptr();
+    params.bias_ptr = bias.has_value() ? bias.value().data_ptr() : nullptr;
+    params.out_ptr = out.data_ptr();
+    // All stride are in elements, not bytes.
+    params.query_start_loc_ptr = query_start_loc.has_value() ? query_start_loc.value().data_ptr() : nullptr;
+    params.cache_indices_ptr = cache_indices.has_value() ? cache_indices.value().data_ptr() : nullptr;
+    params.has_initial_state_ptr = has_initial_state.has_value() ? has_initial_state.value().data_ptr() : nullptr;
+    const bool varlen = params.query_start_loc_ptr != nullptr;
+    params.x_batch_stride = x.stride(varlen ? 1 : 0);
+    params.x_c_stride = x.stride(varlen ? 0 : 1);
+    params.x_l_stride = x.stride(varlen ? 1 : -1);
+    params.weight_c_stride = weight.stride(0);
+    params.weight_width_stride = weight.stride(1);
+    params.out_batch_stride = out.stride(varlen ? 1 : 0);
+    params.out_c_stride = out.stride(varlen ? 0 : 1);
+    params.out_l_stride = out.stride(varlen ? 1 : -1);
+}
+
+
+void causal_conv1d_fwd(const at::Tensor &x, const at::Tensor &weight,
+                  const std::optional<at::Tensor> &bias_,
+                  const std::optional<at::Tensor> &conv_states,
+                  const std::optional<at::Tensor> &query_start_loc,
+                  const std::optional<at::Tensor> &cache_indices,
+                  const std::optional<at::Tensor> &has_initial_state,
+                  bool silu_activation,
+                 // used to identify padding entries if cache_indices provided
+                 // in case of padding, the kernel will return early
+                  int64_t pad_slot_id) {
+    auto input_type = x.scalar_type();
+    auto weight_type = weight.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
+
+    TORCH_CHECK(x.is_cuda());
+    TORCH_CHECK(weight.is_cuda());
+    
+    const bool varlen = query_start_loc.has_value() ? true : false;
+    const auto sizes = x.sizes();
+    const int batch_size = varlen ? query_start_loc.value().sizes()[0] - 1 : sizes[0];
+    const int dim = varlen ? sizes[0] : sizes[1];
+    const int seqlen = varlen ? sizes[1] : sizes[2];
+    const int width = weight.size(-1);
+    if (varlen){
+        CHECK_SHAPE(x, dim, seqlen);
+    }
+    else {
+        CHECK_SHAPE(x, batch_size, dim, seqlen);
+    }
+    CHECK_SHAPE(weight, dim, width);
+
+
+
+    if (bias_.has_value()) {
+        auto bias = bias_.value();
+        TORCH_CHECK(bias.scalar_type() == weight_type);
+        TORCH_CHECK(bias.is_cuda());
+        TORCH_CHECK(bias.stride(-1) == 1);
+        CHECK_SHAPE(bias, dim);
+    }
+
+
+    if (has_initial_state.has_value()) {
+        auto has_initial_state_ = has_initial_state.value();
+        TORCH_CHECK(has_initial_state_.scalar_type() == at::ScalarType::Bool);
+        TORCH_CHECK(has_initial_state_.is_cuda());
+        CHECK_SHAPE(has_initial_state_, batch_size);
+    }
+
+
+    if (query_start_loc.has_value()) {
+        auto query_start_loc_ = query_start_loc.value();
+        TORCH_CHECK(query_start_loc_.scalar_type() == at::ScalarType::Int);
+        TORCH_CHECK(query_start_loc_.is_cuda());
+    }
+
+
+    if (cache_indices.has_value()) {
+        auto cache_indices_ = cache_indices.value();
+        TORCH_CHECK(cache_indices_.scalar_type() == at::ScalarType::Int);
+        TORCH_CHECK(cache_indices_.is_cuda());
+        CHECK_SHAPE(cache_indices_, batch_size);
+    }
+
+    at::Tensor out = x;
+
+    ConvParamsBase params;
+    set_conv_params_fwd(params, batch_size, dim, seqlen, width, x, weight, out,
+                        bias_,
+                        silu_activation, 
+                        pad_slot_id,
+                        query_start_loc,
+                        cache_indices,
+                        has_initial_state
+                        );
+
+    if (conv_states.has_value()) {
+        auto conv_states_ = conv_states.value();
+        TORCH_CHECK(conv_states_.scalar_type() == input_type);
+        TORCH_CHECK(conv_states_.is_cuda());
+        params.conv_states_ptr = conv_states_.data_ptr();
+        params.conv_states_batch_stride = conv_states_.stride(0);
+        params.conv_states_c_stride = conv_states_.stride(1);
+        params.conv_states_l_stride = conv_states_.stride(2);
+    } else {
+        params.conv_states_ptr = nullptr;
+    }
+
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_fwd", [&] {
+            causal_conv1d_fwd_cuda<input_t, weight_t>(params, stream);
+    });
+}
+
+
+void causal_conv1d_update(const at::Tensor &x,
+                     const at::Tensor &conv_state,
+                     const at::Tensor &weight,
+                     const std::optional<at::Tensor> &bias_,
+                     bool silu_activation,
+                     const std::optional<at::Tensor> &cache_seqlens_,
+                     const std::optional<at::Tensor> &conv_state_indices_,
+                     // used to identify padding entries if cache_indices provided
+                     // in case of padding, the kernel will return early
+                     int64_t pad_slot_id) {
+    auto input_type = x.scalar_type();
+    auto weight_type = weight.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == input_type, "weight type must equal to input type, other variations are disabled due to binary size limitations");
+    TORCH_CHECK(conv_state.scalar_type() == input_type);
+
+    TORCH_CHECK(x.is_cuda());
+    TORCH_CHECK(conv_state.is_cuda());
+    TORCH_CHECK(weight.is_cuda());
+
+    const auto sizes = x.sizes();
+    const int batch_size = sizes[0];
+    const int dim = sizes[1];
+    const int seqlen = sizes[2];
+    const int width = weight.size(-1);
+    const int conv_state_len = conv_state.size(2);
+    TORCH_CHECK(conv_state_len >= width - 1);
+
+    CHECK_SHAPE(x, batch_size, dim, seqlen);
+    CHECK_SHAPE(weight, dim, width);
+
+    TORCH_CHECK(width >= 2 && width <= 4, "causal_conv1d only supports width between 2 and 4");
+
+    if (bias_.has_value()) {
+        auto bias = bias_.value();
+        TORCH_CHECK(bias.scalar_type() == weight_type);
+        TORCH_CHECK(bias.is_cuda());
+        TORCH_CHECK(bias.stride(-1) == 1);
+        CHECK_SHAPE(bias, dim);
+    }
+
+    at::Tensor out = x;
+
+    ConvParamsBase params;
+    set_conv_params_fwd(params, batch_size, dim, seqlen, width, x, weight, out,
+                        bias_,
+                        silu_activation,
+                        pad_slot_id);
+    params.conv_state_ptr = conv_state.data_ptr();
+    params.conv_state_len = conv_state_len;
+    // All stride are in elements, not bytes.
+    params.conv_state_batch_stride = conv_state.stride(0);
+    params.conv_state_c_stride = conv_state.stride(1);
+    params.conv_state_l_stride = conv_state.stride(2);
+
+    if (cache_seqlens_.has_value()) {
+        auto cache_seqlens = cache_seqlens_.value();
+        TORCH_CHECK(cache_seqlens.scalar_type() == torch::kInt32);
+        TORCH_CHECK(cache_seqlens.is_cuda());
+        TORCH_CHECK(cache_seqlens.stride(-1) == 1);
+        CHECK_SHAPE(cache_seqlens, batch_size);
+        params.cache_seqlens = cache_seqlens.data_ptr<int32_t>();
+    } else {
+        params.cache_seqlens = nullptr;
+    }
+
+    if (conv_state_indices_.has_value()) {
+        auto conv_state_indices = conv_state_indices_.value();
+        TORCH_CHECK(conv_state_indices.scalar_type() == torch::kInt32)
+        TORCH_CHECK(conv_state_indices.is_cuda());
+        TORCH_CHECK(conv_state_indices.stride(0) == 1)
+        CHECK_SHAPE(conv_state_indices, batch_size);
+
+        int conv_state_entries = conv_state.size(0);
+        CHECK_SHAPE(conv_state, conv_state_entries, dim, conv_state_len);
+
+        params.conv_state_indices_ptr = conv_state_indices.data_ptr<int32_t>();
+    } else {
+        CHECK_SHAPE(conv_state, batch_size, dim, conv_state_len);
+        params.conv_state_indices_ptr = nullptr;
+    }
+
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_update", [&] {
+            causal_conv1d_update_cuda<input_t, weight_t>(params, stream);
+    });
+}
+
+template<int kNThreads_, int kWidth_, bool kIsVecLoad_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_fwd_kernel_traits {
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static constexpr int kWidth = kWidth_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+    static constexpr int kNElts = kNBytes == 4 ? 4 : 8;
+    static_assert(kWidth <= kNElts);
+    static constexpr bool kIsVecLoad = kIsVecLoad_;
+    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
+    using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNElts, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockLoadVecT = cub::BlockLoad<vec_t, kNThreads, 1, cub::BLOCK_LOAD_DIRECT>;
+    using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNElts, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    using BlockStoreVecT = cub::BlockStore<vec_t, kNThreads, 1, cub::BLOCK_STORE_DIRECT>;
+    static constexpr int kSmemIOSize = kIsVecLoad
+        ? 0
+        : custom_max({sizeof(typename BlockLoadT::TempStorage), sizeof(typename BlockStoreT::TempStorage)});
+    static constexpr int kSmemExchangeSize = kNThreads * kNBytes * kNElts;
+    static constexpr int kSmemSize = kSmemIOSize + kSmemExchangeSize;
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_fwd_kernel(ConvParamsBase params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    constexpr int kNElts = Ktraits::kNElts;
+    constexpr bool kIsVecLoad = Ktraits::kIsVecLoad;
+    using input_t = typename Ktraits::input_t;
+    using vec_t = typename Ktraits::vec_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    // Shared memory.
+    extern __shared__ char smem_[];
+    auto& smem_load = reinterpret_cast<typename Ktraits::BlockLoadT::TempStorage&>(smem_);
+    auto& smem_load_vec = reinterpret_cast<typename Ktraits::BlockLoadVecT::TempStorage&>(smem_);
+    auto& smem_store = reinterpret_cast<typename Ktraits::BlockStoreT::TempStorage&>(smem_);
+    auto& smem_store_vec = reinterpret_cast<typename Ktraits::BlockStoreVecT::TempStorage&>(smem_);
+    vec_t *smem_exchange = reinterpret_cast<vec_t *>(smem_ + Ktraits::kSmemIOSize);
+
+    const bool kVarlen = params.query_start_loc_ptr != nullptr;
+    const int tidx = threadIdx.x;
+    const int batch_id = blockIdx.x;
+    const int channel_id = blockIdx.y;
+    const int *query_start_loc = kVarlen ? reinterpret_cast<int *>(params.query_start_loc_ptr) : nullptr;
+    const int sequence_start_index = kVarlen ? query_start_loc[batch_id] : batch_id;
+    const int seqlen = kVarlen ? query_start_loc[batch_id + 1] - sequence_start_index : params.seqlen;
+
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + sequence_start_index * params.x_batch_stride
+        + channel_id * params.x_c_stride;
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + channel_id * params.weight_c_stride;
+    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + sequence_start_index * params.out_batch_stride
+        + channel_id * params.out_c_stride;
+    float bias_val = params.bias_ptr == nullptr ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[channel_id]);
+
+    bool has_initial_state = params.has_initial_state_ptr == nullptr ? false
+        : reinterpret_cast<bool *>(params.has_initial_state_ptr)[batch_id];
+
+    int* cache_indices = params.cache_indices_ptr == nullptr ? nullptr
+        : reinterpret_cast<int *>(params.cache_indices_ptr);
+    int cache_index = cache_indices == nullptr ? batch_id : cache_indices[batch_id];
+    // cache_index == params.pad_slot_id is defined as padding, so we exit early
+    if (cache_index == params.pad_slot_id){
+        return;
+    }
+    input_t *conv_states = params.conv_states_ptr == nullptr ? nullptr
+        : reinterpret_cast<input_t *>(params.conv_states_ptr) + cache_index * params.conv_states_batch_stride + channel_id * params.conv_states_c_stride;
+
+    // Thread 0 will load the last elements of the previous chunk, so we initialize those to 0.
+    if (tidx == 0) {
+        input_t initial_state[kNElts] = {0};
+        if (has_initial_state) {
+            #pragma unroll
+            for (int w = 0; w < kWidth - 1; ++w){ initial_state[kNElts - 1 - (kWidth - 2) + w ] = conv_states[w]; }
+        }
+        smem_exchange[kNThreads - 1] = reinterpret_cast<vec_t *>(initial_state)[0];
+    }
+
+    float weight_vals[kWidth];
+    #pragma unroll
+    for (int i = 0; i < kWidth; ++i) { weight_vals[i] = float(weight[i * params.weight_width_stride]); }
+
+    constexpr int kChunkSize = kNThreads * kNElts;
+    const int n_chunks = (seqlen + kChunkSize - 1) / kChunkSize;
+    for (int chunk = 0; chunk < n_chunks; ++chunk) {
+        input_t x_vals_load[2 * kNElts] = {0};
+        if constexpr(kIsVecLoad) {
+            typename Ktraits::BlockLoadVecT(smem_load_vec).Load(reinterpret_cast<vec_t*>(x), *reinterpret_cast<vec_t (*)[1]>(&x_vals_load[kNElts]), (seqlen - chunk * kChunkSize) / kNElts);
+        } else {
+            __syncthreads();
+            typename Ktraits::BlockLoadT(smem_load).Load(x, *reinterpret_cast<input_t (*)[kNElts]>(&x_vals_load[kNElts]), seqlen - chunk * kChunkSize);
+        }
+        x += kChunkSize;
+        __syncthreads();
+        // Thread kNThreads - 1 don't write yet, so that thread 0 can read
+        // the last elements of the previous chunk.
+        if (tidx < kNThreads - 1) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1]; }
+        __syncthreads();
+        reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[tidx > 0 ? tidx - 1 : kNThreads - 1];
+        __syncthreads();
+        // Now thread kNThreads - 1 can write the last elements of the current chunk.
+        if (tidx == kNThreads - 1) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1]; }
+
+        float x_vals[2 * kNElts];
+        #pragma unroll
+        for (int i = 0; i < 2 * kNElts; ++i) { x_vals[i] = float(x_vals_load[i]); }
+
+        float out_vals[kNElts];
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) {
+            out_vals[i] = bias_val;
+            #pragma unroll
+            for (int w = 0; w < kWidth; ++w) {
+                out_vals[i] += weight_vals[w] * x_vals[kNElts + i - (kWidth - w - 1)];
+            }
+        }
+
+        if (params.silu_activation) {
+            #pragma unroll
+            for (int i = 0; i < kNElts; ++i) {
+                out_vals[i] = out_vals[i] / (1 + expf(-out_vals[i]));
+            }
+        }
+
+        input_t out_vals_store[kNElts];
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) { out_vals_store[i] = out_vals[i]; }
+        if constexpr(kIsVecLoad) {
+            typename Ktraits::BlockStoreVecT(smem_store_vec).Store(reinterpret_cast<vec_t*>(out), reinterpret_cast<vec_t (&)[1]>(out_vals_store), (seqlen - chunk * kChunkSize) / kNElts);
+        } else {
+            typename Ktraits::BlockStoreT(smem_store).Store(out, out_vals_store, seqlen - chunk * kChunkSize);
+        }
+        out += kChunkSize;
+
+        int final_state_position =  ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize);
+        // in case the final state is separated between the last "smem_exchange" and 
+        // and the one before it (chunk = n_chunks - 1 and chunk = n_chunks - 2), 
+        // (which occurs when `final_state_position` is a non-positive index)
+        // we load the correct data from smem_exchange from both chunks, the last chunk iteration and the one before it
+        if (conv_states != nullptr && final_state_position < 0 && seqlen > kWidth){
+            input_t vals_load[kNElts] = {0};
+            if ((chunk == n_chunks - 2) && (tidx == kNThreads - 1)){
+                // chunk = n_chunks - 2, a segment of the final state sits in the last index
+                reinterpret_cast<vec_t *>(vals_load)[0] = smem_exchange[kNThreads - 1];
+                #pragma unroll
+                for (int w = 0; w < -final_state_position; ++w){
+                    conv_states[w] = vals_load[kNElts + final_state_position + w];
+                }
+            }
+            if ((chunk == n_chunks - 1) && tidx == 0){
+                // chunk = n_chunks - 1, the second segment of the final state first positions
+                reinterpret_cast<vec_t *>(vals_load)[0] = smem_exchange[0];
+                for (int w = -final_state_position; w < kWidth - 1; ++w){
+                    conv_states[w] = vals_load[w + final_state_position];
+                }
+                return;
+            }
+        }
+    }
+    // Final state is stored in the smem_exchange last token slot,
+    // in case seqlen < kWidth, we would need to take the final state from the 
+    // initial state which is stored in conv_states
+    // in case seqlen > kWidth, we would need to load the last kWidth - 1 data
+    // and load it into conv_state accordingly
+    int last_thread =  ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize) / kNElts;
+    if (conv_states != nullptr && tidx == last_thread) { 
+        input_t x_vals_load[kNElts * 2] = {0};
+        // in case we are on the first kWidth tokens
+        if (last_thread == 0 && seqlen < kWidth){
+            // Need to take the initial state
+            reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[0];
+            const int offset = seqlen - (kWidth - 1);
+            #pragma unroll
+            for (int w = 0; w < kWidth - 1; ++w){
+                // pad the existing state
+                if ((w - seqlen) >= 0 && has_initial_state) { conv_states[w - seqlen] = conv_states[w]; }
+                else if ((w - seqlen) >= 0 && !has_initial_state) { conv_states[w - seqlen] = input_t(0.0f); }
+            }
+            #pragma unroll
+            for (int w = 0; w < kWidth - 1; ++w){
+                if (offset + w >= 0) 
+                    conv_states[w] = x_vals_load[offset + w ];
+            }
+        }
+        else {
+            // in case the final state is in between the threads data
+            const int offset = ((seqlen - (kWidth - 1)) % (kNElts));
+            if ((offset + kWidth - 2) >= kNElts && (last_thread + 1 < kNThreads)){
+                // In case last_thread == kNThreads - 1, accessing last_thread + 1 will result in a 
+                // illegal access error on H100.
+                // Therefore, we access last_thread + 1, only if the final state data sits there
+                reinterpret_cast<vec_t *>(x_vals_load)[1] = smem_exchange[last_thread + 1];
+            }
+            reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[last_thread];
+            #pragma unroll
+            for (int w = 0; w < kWidth - 1; ++w){
+                conv_states[w] = x_vals_load[offset + w ];
+            }
+        }
+        
+    }
+}
+
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_fwd_launch(ConvParamsBase &params, cudaStream_t stream) {
+    static constexpr int kNElts = sizeof(input_t) == 4 ? 4 : 8;
+    const bool kVarlen = params.query_start_loc_ptr != nullptr;
+    BOOL_SWITCH(params.seqlen % kNElts == 0 && !kVarlen, kIsVecLoad, [&] {
+        using Ktraits = Causal_conv1d_fwd_kernel_traits<kNThreads, kWidth, kIsVecLoad, input_t, weight_t>;
+        constexpr int kSmemSize = Ktraits::kSmemSize;
+        dim3 grid(params.batch, params.dim);
+
+        auto kernel = &causal_conv1d_fwd_kernel<Ktraits>;
+
+        if (kSmemSize >= 48 * 1024) {
+            C10_CUDA_CHECK(cudaFuncSetAttribute(
+                (void *) kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+            std::cerr << "Warning (causal_conv1d fwd launch): attempting to set maxDynamicSharedMemorySize on an AMD GPU which is currently a non-op (in ROCm versions <= 6.1). This might lead to undefined behavior. \n" << std::endl;
+        }
+        kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+
+        C10_CUDA_KERNEL_LAUNCH_CHECK();
+    });
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_fwd_cuda(ConvParamsBase &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_fwd_launch<128, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_fwd_launch<128, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_fwd_launch<128, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+
+template void causal_conv1d_fwd_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+
+
+
+
+template<int kNThreads_, int kWidth_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_update_kernel_traits {
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static constexpr int kWidth = kWidth_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+};
+
+template<typename Ktraits, bool kIsCircularBuffer>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_update_kernel(ConvParamsBase params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    using input_t = typename Ktraits::input_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    const int tidx = threadIdx.x;
+    const int batch_id = blockIdx.x;
+    const int channel_id = blockIdx.y * kNThreads + tidx;
+    if (channel_id >= params.dim) return;
+
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
+        + channel_id * params.x_c_stride;
+
+    // If params.conv_state_batch_indices is set, then the conv state is gathered from the conv state tensor
+    // along the batch axis. Otherwise, the conv state coordinate is the same as the batch id.
+    const int conv_state_batch_coord = params.conv_state_indices_ptr == nullptr
+        ? batch_id
+        : params.conv_state_indices_ptr[batch_id];
+    // conv_state_batch_coord == params.pad_slot_id is defined as padding so we exit early
+    if (conv_state_batch_coord == params.pad_slot_id){
+        return;
+    }
+    input_t *conv_state = reinterpret_cast<input_t *>(params.conv_state_ptr) 
+        + conv_state_batch_coord * params.conv_state_batch_stride
+        + channel_id * params.conv_state_c_stride;
+
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + channel_id * params.weight_c_stride;
+    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + batch_id * params.out_batch_stride
+        + channel_id * params.out_c_stride;
+    float bias_val = params.bias_ptr == nullptr ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[channel_id]);
+
+    int state_len = params.conv_state_len;
+    int advance_len = params.seqlen;
+    int cache_seqlen = kIsCircularBuffer ? params.cache_seqlens[batch_id] % state_len : 0;
+    int update_idx = cache_seqlen - (kWidth - 1);
+    update_idx = update_idx < 0 ? update_idx + state_len : update_idx;
+
+    float weight_vals[kWidth] = {0};
+    #pragma unroll
+    for (int i = 0; i < kWidth; ++i) { weight_vals[i] = float(weight[i * params.weight_width_stride]); }
+
+    float x_vals[kWidth] = {0};
+    if constexpr (!kIsCircularBuffer) {
+        #pragma unroll 2
+        for (int i = 0; i < state_len - advance_len - (kWidth - 1); ++i) {
+            conv_state[i * params.conv_state_l_stride] = conv_state[(i + advance_len) * params.conv_state_l_stride];
+        }
+        #pragma unroll
+        for (int i = 0; i < kWidth - 1; ++i) {
+            input_t state_val = conv_state[(state_len - (kWidth - 1) + i) * params.conv_state_l_stride];
+            if (i < advance_len + (kWidth - 1) && state_len - advance_len - (kWidth - 1) + i >= 0) {
+                conv_state[(state_len - advance_len - (kWidth - 1) + i) * params.conv_state_l_stride] = state_val;
+            }
+            x_vals[i] = float(state_val);
+        }
+    } else {
+        #pragma unroll
+        for (int i = 0; i < kWidth - 1; ++i, update_idx = update_idx + 1 >= state_len ? update_idx + 1 - state_len : update_idx + 1) {
+            input_t state_val = conv_state[update_idx * params.conv_state_l_stride];
+            x_vals[i] = float(state_val);
+        }
+    }
+    #pragma unroll 2
+    for (int i = 0; i < params.seqlen; ++i) {
+        input_t x_val = x[i * params.x_l_stride];
+        if constexpr (!kIsCircularBuffer) {
+            if (i < advance_len && state_len - advance_len + i >= 0) {
+                conv_state[(state_len - advance_len + i) * params.conv_state_l_stride] = x_val;
+            }
+        } else {
+            conv_state[update_idx * params.conv_state_l_stride] = x_val;
+            ++update_idx;
+            update_idx = update_idx >= state_len ? update_idx - state_len : update_idx;
+        }
+        x_vals[kWidth - 1] = float(x_val);
+        float out_val = bias_val;
+        #pragma unroll
+        for (int j = 0; j < kWidth; ++j) { out_val += weight_vals[j] * x_vals[j]; }
+        if (params.silu_activation) { out_val = out_val / (1 + expf(-out_val)); }
+        out[i * params.out_l_stride] = input_t(out_val);
+        // Shift the input buffer by 1
+        #pragma unroll
+        for (int i = 0; i < kWidth - 1; ++i) { x_vals[i] = x_vals[i + 1]; }
+    }
+}
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_update_launch(ConvParamsBase &params, cudaStream_t stream) {
+    using Ktraits = Causal_conv1d_update_kernel_traits<kNThreads, kWidth, input_t, weight_t>;
+    dim3 grid(params.batch, (params.dim + kNThreads - 1) / kNThreads);
+    auto kernel = params.cache_seqlens == nullptr
+        ? &causal_conv1d_update_kernel<Ktraits, false>
+        : &causal_conv1d_update_kernel<Ktraits, true>;
+    kernel<<<grid, Ktraits::kNThreads, 0, stream>>>(params);
+    C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_update_cuda(ConvParamsBase &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_update_launch<64, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_update_launch<64, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_update_launch<64, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+template void causal_conv1d_update_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);

csrc/mamba/causal_conv1d/causal_conv1d.h

@@ -0,0 +1,159 @@
+/******************************************************************************
+ * Copyright (c) 2024, Tri Dao.
+ ******************************************************************************/
+// clang-format off
+// adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/causal_conv1d.h
+#pragma once
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+struct ConvParamsBase {
+    using index_t = uint32_t;
+
+    int batch, dim, seqlen, width;
+    int64_t pad_slot_id;
+    bool silu_activation;
+
+    index_t x_batch_stride;
+    index_t x_c_stride;
+    index_t x_l_stride;
+    index_t weight_c_stride;
+    index_t weight_width_stride;
+    index_t out_batch_stride;
+    index_t out_c_stride;
+    index_t out_l_stride;
+
+    int conv_state_len;
+    index_t conv_state_batch_stride;
+    index_t conv_state_c_stride;
+    index_t conv_state_l_stride;
+
+    // Common data pointers.
+    void *__restrict__ x_ptr;
+    void *__restrict__ weight_ptr;
+    void *__restrict__ bias_ptr;
+    void *__restrict__ out_ptr;
+
+    void *__restrict__ conv_state_ptr;
+    void *__restrict__ query_start_loc_ptr;
+    void *__restrict__ has_initial_state_ptr;
+    void *__restrict__ cache_indices_ptr;
+    int32_t *__restrict__ cache_seqlens;
+
+    // For the continuous batching case. Makes it so that the mamba state for 
+    // the current batch doesn't need to be a contiguous tensor.
+    int32_t *__restrict__ conv_state_indices_ptr;
+
+    void *__restrict__ seq_idx_ptr;
+
+    // No __restrict__ since initial_states could be the same as final_states.
+    void * initial_states_ptr;
+    index_t initial_states_batch_stride;
+    index_t initial_states_l_stride;
+    index_t initial_states_c_stride;
+
+    void * final_states_ptr;
+    index_t final_states_batch_stride;
+    index_t final_states_l_stride;
+    index_t final_states_c_stride;
+
+    void *  conv_states_ptr;
+    index_t conv_states_batch_stride;
+    index_t conv_states_l_stride;
+    index_t conv_states_c_stride;
+};
+
+
+#ifndef USE_ROCM
+    #include <cuda_bf16.h>
+
+    template<typename T>
+    __device__ inline T shuffle_xor(T val, int offset) {
+        return __shfl_xor_sync(uint32_t(-1), val, offset);
+    }
+
+    constexpr size_t custom_max(std::initializer_list<size_t> ilist) 
+    {
+        return std::max(ilist);
+    }
+
+    template<typename T>
+    constexpr T constexpr_min(T a, T b) {
+        return std::min(a, b);
+    }
+
+#else
+    #include <hip/hip_bf16.h>
+
+    template<typename T>
+    __device__ inline T shuffle_xor(T val, int offset) {
+        return __shfl_xor(val, offset);
+    }
+    constexpr size_t custom_max(std::initializer_list<size_t> ilist) 
+    {
+        return *std::max_element(ilist.begin(), ilist.end());
+    }
+
+    template<typename T>
+    constexpr T constexpr_min(T a, T b) {
+        return a < b ? a : b;
+    }
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<int BYTES> struct BytesToType {};
+
+template<> struct BytesToType<16> {
+    using Type = uint4;
+    static_assert(sizeof(Type) == 16);
+};
+
+template<> struct BytesToType<8> {
+    using Type = uint64_t;
+    static_assert(sizeof(Type) == 8);
+};
+
+template<> struct BytesToType<4> {
+    using Type = uint32_t;
+    static_assert(sizeof(Type) == 4);
+};
+
+template<> struct BytesToType<2> {
+    using Type = uint16_t;
+    static_assert(sizeof(Type) == 2);
+};
+
+template<> struct BytesToType<1> {
+    using Type = uint8_t;
+    static_assert(sizeof(Type) == 1);
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<typename T>
+struct SumOp {
+__device__ inline T operator()(T const & x, T const & y) { return x + y; }
+};
+
+template<int THREADS>
+struct Allreduce {
+    static_assert(THREADS == 32 || THREADS == 16 || THREADS == 8 || THREADS == 4);
+    template<typename T, typename Operator>
+    static __device__ inline T run(T x, Operator &op) {
+        constexpr int OFFSET = THREADS / 2;
+        x = op(x, __shfl_xor_sync(uint32_t(-1), x, OFFSET));
+        return Allreduce<OFFSET>::run(x, op);
+    }
+};
+
+template<>
+struct Allreduce<2> {
+template<typename T, typename Operator>
+static __device__ inline T run(T x, Operator &op) {
+    x = op(x, __shfl_xor_sync(uint32_t(-1), x, 1));
+    return x;
+}
+};

csrc/mamba/causal_conv1d/static_switch.h

@@ -0,0 +1,28 @@
+// Inspired by
+// https://github.com/NVIDIA/DALI/blob/main/include/dali/core/static_switch.h
+// and https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Dispatch.h
+// clang-format off
+// adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/static_switch.h
+
+#pragma once
+
+/// @param COND       - a boolean expression to switch by
+/// @param CONST_NAME - a name given for the constexpr bool variable.
+/// @param ...       - code to execute for true and false
+///
+/// Usage:
+/// ```
+/// BOOL_SWITCH(flag, BoolConst, [&] {
+///     some_function<BoolConst>(...);
+/// });
+/// ```
+#define BOOL_SWITCH(COND, CONST_NAME, ...)                                           \
+    [&] {                                                                            \
+        if (COND) {                                                                  \
+            static constexpr bool CONST_NAME = true;                                 \
+            return __VA_ARGS__();                                                    \
+        } else {                                                                     \
+            static constexpr bool CONST_NAME = false;                                \
+            return __VA_ARGS__();                                                    \
+        }                                                                            \
+    }()

csrc/mamba/mamba_ssm/selective_scan_fwd.cu

@@ -7,11 +7,7 @@
 
 #include <c10/util/BFloat16.h>
 #include <c10/util/Half.h>
-#ifdef USE_ROCM
-    #include <c10/hip/HIPException.h>  // For C10_HIP_CHECK and C10_HIP_KERNEL_LAUNCH_CHECK
-#else
-    #include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
-#endif
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
 
 #ifndef USE_ROCM
     #include <cub/block/block_load.cuh>
@@ -316,25 +312,19 @@ void selective_scan_fwd_launch(SSMParamsBase &params, cudaStream_t stream) {
     // kIsVariableB, kIsVariableC and kHasZ are all set to True to reduce binary size
     constexpr bool kIsVariableB = true;
     constexpr bool kIsVariableC = true;
+    constexpr bool kHasZ = true;
     BOOL_SWITCH(params.seqlen % (kNThreads * kNItems) == 0, kIsEvenLen, [&] {
-        BOOL_SWITCH(params.z_ptr != nullptr , kHasZ, [&] {
-            BOOL_SWITCH(params.query_start_loc_ptr != nullptr , kVarlen, [&] {
-                using Ktraits = Selective_Scan_fwd_kernel_traits<kNThreads, kNItems, kNRows, kIsEvenLen, kIsVariableB, kIsVariableC, kHasZ,  kVarlen, input_t, weight_t>;
-                constexpr int kSmemSize = Ktraits::kSmemSize + kNRows * MAX_DSTATE * sizeof(typename Ktraits::scan_t);
-                dim3 grid(params.batch, params.dim / kNRows);
-                auto kernel = &selective_scan_fwd_kernel<Ktraits>;
-                if (kSmemSize >= 48 * 1024) {
-#ifdef USE_ROCM
-                    C10_HIP_CHECK(hipFuncSetAttribute(
-                        reinterpret_cast<const void*>(kernel), hipFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
-#else
-                    C10_CUDA_CHECK(cudaFuncSetAttribute(
-                        kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
-#endif
-                }
-                kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
-                C10_CUDA_KERNEL_LAUNCH_CHECK();
-            });
+        BOOL_SWITCH(params.query_start_loc_ptr != nullptr , kVarlen, [&] {
+            using Ktraits = Selective_Scan_fwd_kernel_traits<kNThreads, kNItems, kNRows, kIsEvenLen, kIsVariableB, kIsVariableC, kHasZ,  kVarlen, input_t, weight_t>;
+            constexpr int kSmemSize = Ktraits::kSmemSize + kNRows * MAX_DSTATE * sizeof(typename Ktraits::scan_t);
+            dim3 grid(params.batch, params.dim / kNRows);
+            auto kernel = &selective_scan_fwd_kernel<Ktraits>;
+            if (kSmemSize >= 48 * 1024) {
+                C10_CUDA_CHECK(cudaFuncSetAttribute(
+                    (void *) kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+            }
+            kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+            C10_CUDA_KERNEL_LAUNCH_CHECK();
         });
     });
 }
@@ -622,20 +612,19 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
 
     at::Tensor z, out_z;
     const bool has_z = z_.has_value();
-    if (has_z) {
-        z = z_.value();
-        TORCH_CHECK(z.scalar_type() == input_type);
-        TORCH_CHECK(z.is_cuda());
-        TORCH_CHECK(z.stride(-1) == 1 || z.size(-1) == 1);
-        if (varlen){
-            CHECK_SHAPE(z, dim, seqlen);
-        } else {
-            CHECK_SHAPE(z, batch_size, dim, seqlen);
-        }
-        
-        out_z = z;
+    TORCH_CHECK(has_z, "has_z = False is disabled in favor of reduced binary size")
+    z = z_.value();
+    TORCH_CHECK(z.scalar_type() == input_type);
+    TORCH_CHECK(z.is_cuda());
+    TORCH_CHECK(z.stride(-1) == 1 || z.size(-1) == 1);
+    if (varlen){
+        CHECK_SHAPE(z, dim, seqlen);
+    } else {
+        CHECK_SHAPE(z, batch_size, dim, seqlen);
     }
 
+    out_z = z;
+
     // Right now u has BHL layout and delta has HBL layout, and we want out to have HBL layout
     at::Tensor out = delta;
     TORCH_CHECK(ssm_states.scalar_type() == input_type);
@@ -664,3 +653,4 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
         selective_scan_fwd_cuda<input_t, weight_t>(params, stream);
     });
 }
+

csrc/moe/moe_align_sum_kernels.cu

@@ -1,7 +1,6 @@
 #include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
-#include <cub/cub.cuh>
 
 #include <ATen/ATen.h>
 #include <ATen/cuda/Atomic.cuh>
@@ -20,14 +19,9 @@ __global__ void moe_align_block_size_kernel(
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
     int32_t padded_num_experts, int32_t experts_per_warp, int32_t block_size,
-    size_t numel, int32_t* __restrict__ cumsum, int32_t max_num_tokens_padded) {
+    size_t numel, int32_t* __restrict__ cumsum) {
   extern __shared__ int32_t shared_counts[];
 
-  // Initialize sorted_token_ids with numel
-  for (size_t it = threadIdx.x; it < max_num_tokens_padded; it += blockDim.x) {
-    sorted_token_ids[it] = numel;
-  }
-
   const int warp_id = threadIdx.x / WARP_SIZE;
   const int my_expert_start = warp_id * experts_per_warp;
 
@@ -51,27 +45,18 @@ __global__ void moe_align_block_size_kernel(
 
   __syncthreads();
 
-  // Compute prefix sum over token counts per expert
-  using BlockScan = cub::BlockScan<int32_t, 1024>;
-  __shared__ typename BlockScan::TempStorage temp_storage;
+  if (threadIdx.x == 0) {
+    cumsum[0] = 0;
+    for (int i = 1; i <= num_experts; ++i) {
+      int expert_count = 0;
+      int warp_idx = (i - 1) / experts_per_warp;
+      int expert_offset = (i - 1) % experts_per_warp;
+      expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
 
-  int expert_count = 0;
-  int expert_id = threadIdx.x;
-  if (expert_id < num_experts) {
-    int warp_idx = expert_id / experts_per_warp;
-    int expert_offset = expert_id % experts_per_warp;
-    expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
-    expert_count = CEILDIV(expert_count, block_size) * block_size;
-  }
-
-  int cumsum_val;
-  BlockScan(temp_storage).ExclusiveSum(expert_count, cumsum_val);
-  if (expert_id <= num_experts) {
-    cumsum[expert_id] = cumsum_val;
-  }
-
-  if (expert_id == num_experts) {
-    *total_tokens_post_pad = cumsum_val;
+      cumsum[i] =
+          cumsum[i - 1] + CEILDIV(expert_count, block_size) * block_size;
+    }
+    *total_tokens_post_pad = cumsum[num_experts];
   }
 
   __syncthreads();
@@ -82,13 +67,6 @@ __global__ void moe_align_block_size_kernel(
       expert_ids[i / block_size] = threadIdx.x;
     }
   }
-
-  // Fill remaining expert_ids with 0
-  const size_t fill_start_idx = cumsum[num_experts] / block_size + threadIdx.x;
-  const size_t expert_ids_size = CEILDIV(max_num_tokens_padded, block_size);
-  for (size_t i = fill_start_idx; i < expert_ids_size; i += blockDim.x) {
-    expert_ids[i] = 0;
-  }
 }
 
 template <typename scalar_t>
@@ -127,12 +105,7 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
     const scalar_t* __restrict__ topk_ids,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
-    int32_t block_size, size_t numel, int32_t max_num_tokens_padded) {
-  // Initialize sorted_token_ids with numel
-  for (size_t it = threadIdx.x; it < max_num_tokens_padded; it += blockDim.x) {
-    sorted_token_ids[it] = numel;
-  }
-
+    int32_t block_size, size_t numel) {
   const size_t tid = threadIdx.x;
   const size_t stride = blockDim.x;
 
@@ -180,13 +153,6 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
     }
   }
 
-  // Fill remaining expert_ids with 0
-  const size_t fill_start_idx = cumsum[num_experts] / block_size + threadIdx.x;
-  const size_t expert_ids_size = CEILDIV(max_num_tokens_padded, block_size);
-  for (size_t i = fill_start_idx; i < expert_ids_size; i += blockDim.x) {
-    expert_ids[i] = 0;
-  }
-
   for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
     int32_t rank_post_pad =
@@ -213,17 +179,13 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
   int threads = 1024;
   threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
 
-  // BlockScan uses 1024 threads and assigns one thread per expert.
-  TORCH_CHECK(padded_num_experts < 1024,
-              "padded_num_experts must be less than 1024");
-
   VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
       topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
         // calc needed amount of shared mem for `cumsum` tensors
         auto options_int =
             torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
         torch::Tensor cumsum_buffer =
-            torch::empty({num_experts + 1}, options_int);
+            torch::zeros({num_experts + 1}, options_int);
         bool small_batch_expert_mode =
             (topk_ids.numel() < 1024) && (num_experts <= 64);
 
@@ -241,7 +203,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               sorted_token_ids.data_ptr<int32_t>(),
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
-              topk_ids.numel(), sorted_token_ids.size(0));
+              topk_ids.numel());
         } else {
           auto align_kernel = vllm::moe::moe_align_block_size_kernel<scalar_t>;
 
@@ -255,8 +217,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts,
               padded_num_experts, experts_per_warp, block_size,
-              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>(),
-              sorted_token_ids.size(0));
+              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
 
           const int block_threads = std::min(256, (int)threads);
           const int num_blocks =

csrc/ops.h

@@ -287,11 +287,6 @@ void scaled_fp4_experts_quant(
     torch::Tensor const& input, torch::Tensor const& input_global_scale,
     torch::Tensor const& input_offset_by_experts,
     torch::Tensor const& output_scale_offset_by_experts);
-
-void per_token_group_quant_fp8(const torch::Tensor& input,
-                               torch::Tensor& output_q, torch::Tensor& output_s,
-                               int64_t group_size, double eps, double fp8_min,
-                               double fp8_max, bool scale_ue8m0);
 #endif
 
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
@@ -331,6 +326,22 @@ 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);
 
+void causal_conv1d_update(const at::Tensor& x, const at::Tensor& conv_state,
+                          const at::Tensor& weight,
+                          const std::optional<at::Tensor>& bias_,
+                          bool silu_activation,
+                          const std::optional<at::Tensor>& cache_seqlens_,
+                          const std::optional<at::Tensor>& conv_state_indices_,
+                          int64_t pad_slot_id);
+
+void causal_conv1d_fwd(const at::Tensor& x, const at::Tensor& weight,
+                       const std::optional<at::Tensor>& bias_,
+                       const std::optional<at::Tensor>& conv_states,
+                       const std::optional<at::Tensor>& query_start_loc,
+                       const std::optional<at::Tensor>& cache_indices,
+                       const std::optional<at::Tensor>& has_initial_state,
+                       bool silu_activation, int64_t pad_slot_id);
+
 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/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu

@@ -201,10 +201,11 @@ void run_blockwise_scaled_group_mm(
       reinterpret_cast<typename ScheduleConfig::LayoutSFB*>(
           layout_sfb.data_ptr())};
 
-  int device_id = a_ptrs.device().index();
-  static const cutlass::KernelHardwareInfo hw_info{
-      device_id, cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
-                     device_id)};
+  cutlass::KernelHardwareInfo hw_info;
+  hw_info.device_id = a_ptrs.get_device();
+  hw_info.sm_count =
+      cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
+          hw_info.device_id);
 
   // Epilogue Arguments
   typename GemmKernel::EpilogueArguments epilogue_args{

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu

@@ -21,49 +21,27 @@ struct sm90_fp8_config_default {
       cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
   using TileShape = cute::Shape<cute::_64, cute::_256, cute::_128>;
   using ClusterShape = cute::Shape<cute::_1, cute::_2, cute::_1>;
-  using ArchTag = cutlass::arch::Sm90;
 
   using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule>;
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
-struct sm90_fp8_config_M4 {
-  // M in [1, 4]
+struct sm90_fp8_config_M16 {
+  // M in [1, 16]
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule =
       cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
   using EpilogueSchedule =
       cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
-  using TileShape = cute::Shape<cute::_128, cute::_16, cute::_128>;
-  using ClusterShape = cute::Shape<cute::_1, cute::_1, cute::_1>;
-  using ArchTag = cutlass::arch::Sm90;
+  using TileShape = cute::Shape<cute::_64, cute::_64, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_4, cute::_1>;
 
   using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule,
-                            true>;
-};
-
-template <typename InType, typename OutType,
-          template <typename, typename, typename> typename Epilogue>
-struct sm90_fp8_config_M64 {
-  // M in (4, 64]
-  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
-  using KernelSchedule =
-      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
-  using EpilogueSchedule =
-      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
-  using TileShape = cute::Shape<cute::_128, cute::_16, cute::_256>;
-  using ClusterShape = cute::Shape<cute::_2, cute::_1, cute::_1>;
-  using ArchTag = cutlass::arch::Sm90;
-
-  using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule,
-                            true>;
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -77,11 +55,10 @@ struct sm90_fp8_config_K8192 {
       cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
   using TileShape = cute::Shape<cute::_128, cute::_128, cute::_128>;
   using ClusterShape = cute::Shape<cute::_1, cute::_8, cute::_1>;
-  using ArchTag = cutlass::arch::Sm90;
 
   using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule>;
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -95,11 +72,10 @@ struct sm90_fp8_config_N8192 {
       cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
   using TileShape = cute::Shape<cute::_64, cute::_128, cute::_256>;
   using ClusterShape = cute::Shape<cute::_1, cute::_8, cute::_1>;
-  using ArchTag = cutlass::arch::Sm90;
 
   using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule>;
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType>
@@ -119,13 +95,14 @@ void run_cutlass_moe_mm_sm90(
   TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn,
               "B tensors must be of type float8_e4m3fn.");
 
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn);
+
   using Cutlass3xGemmN8192 = typename sm90_fp8_config_N8192<
       InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
   using Cutlass3xGemmK8192 = typename sm90_fp8_config_K8192<
       InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
-  using Cutlass3xGemmM4 = typename sm90_fp8_config_M4<
-      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
-  using Cutlass3xGemmM64 = typename sm90_fp8_config_M64<
+  using Cutlass3xGemmM16 = typename sm90_fp8_config_M16<
       InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
   using Cutlass3xGemmDefault = typename sm90_fp8_config_default<
       InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
@@ -134,18 +111,7 @@ void run_cutlass_moe_mm_sm90(
   uint32_t const n = out_tensors.size(1);
   uint32_t const k = a_tensors.size(1);
 
-  // Use swap_ab for M <= 64 by default to reduce padding
-  if (m <= 4) {
-    cutlass_group_gemm_caller<Cutlass3xGemmM4>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  } else if (m <= 64) {
-    cutlass_group_gemm_caller<Cutlass3xGemmM64>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  } else if (n >= 8192) {
+  if (n >= 8192) {
     cutlass_group_gemm_caller<Cutlass3xGemmN8192>(
         out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
         problem_sizes, a_strides, b_strides, c_strides, per_act_token,
@@ -155,6 +121,11 @@ void run_cutlass_moe_mm_sm90(
         out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
         problem_sizes, a_strides, b_strides, c_strides, per_act_token,
         per_out_ch);
+  } else if (m <= 16) {
+    cutlass_group_gemm_caller<Cutlass3xGemmM16>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
+        per_out_ch);
   } else {
     cutlass_group_gemm_caller<Cutlass3xGemmDefault>(
         out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cuh

@@ -18,34 +18,28 @@ using ProblemShape =
     cutlass::gemm::GroupProblemShape<cute::Shape<int, int, int>>;
 
 using ElementAccumulator = float;
+using ArchTag = cutlass::arch::Sm90;
 using OperatorClass = cutlass::arch::OpClassTensorOp;
 
 using LayoutA = cutlass::layout::RowMajor;
-using LayoutA_Transpose =
-    typename cutlass::layout::LayoutTranspose<LayoutA>::type;
 using LayoutB = cutlass::layout::ColumnMajor;
-using LayoutB_Transpose =
-    typename cutlass::layout::LayoutTranspose<LayoutB>::type;
-using LayoutD = cutlass::layout::RowMajor;
-using LayoutD_Transpose =
-    typename cutlass::layout::LayoutTranspose<LayoutD>::type;
-using LayoutC = LayoutD;
-using LayoutC_Transpose = LayoutD_Transpose;
+using LayoutC = cutlass::layout::RowMajor;
 
-template <typename ElementAB_, typename ElementC_, typename ArchTag_,
+template <typename ElementAB_, typename ElementC_,
           template <typename, typename, typename> typename Epilogue_,
           typename TileShape, typename ClusterShape, typename KernelSchedule,
-          typename EpilogueSchedule, bool swap_ab_ = false>
+          typename EpilogueSchedule>
 struct cutlass_3x_group_gemm {
-  static constexpr bool swap_ab = swap_ab_;
   using ElementAB = ElementAB_;
   using ElementC = void;
   using ElementD = ElementC_;
   using ElementAccumulator = float;
-  using ArchTag = ArchTag_;
 
   using Epilogue = Epilogue_<ElementAccumulator, ElementD, TileShape>;
 
+  using StrideC =
+      cute::remove_pointer_t<cute::Stride<int64_t, cute::Int<1>, cute::Int<0>>>;
+
   static constexpr int AlignmentAB =
       128 / cutlass::sizeof_bits<ElementAB>::value;
   static constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementD>::value;
@@ -56,28 +50,21 @@ struct cutlass_3x_group_gemm {
       typename cutlass::epilogue::collective::CollectiveBuilder<
           ArchTag, OperatorClass, TileShape, ClusterShape,
           cutlass::epilogue::collective::EpilogueTileAuto, ElementAccumulator,
-          ElementAccumulator, ElementC,
-          conditional_t<swap_ab, LayoutC_Transpose*, LayoutC*>, AlignmentC,
-          ElementD, conditional_t<swap_ab, LayoutD_Transpose*, LayoutD*>,
-          AlignmentC, EpilogueSchedule, EVTCompute>::CollectiveOp;
+          ElementAccumulator, ElementC, LayoutC*, AlignmentC, ElementD,
+          LayoutC*, AlignmentC, EpilogueSchedule, EVTCompute>::CollectiveOp;
 
   static constexpr size_t CEStorageSize =
       sizeof(typename CollectiveEpilogue::SharedStorage);
   using Stages = typename cutlass::gemm::collective::StageCountAutoCarveout<
       static_cast<int>(CEStorageSize)>;
 
-  using CollectiveMainloop = conditional_t<
-      swap_ab,
-      typename cutlass::gemm::collective::CollectiveBuilder<
-          ArchTag, OperatorClass, ElementAB, LayoutB_Transpose*, AlignmentAB,
-          ElementAB, LayoutA_Transpose*, AlignmentAB, ElementAccumulator,
-          TileShape, ClusterShape, Stages, KernelSchedule>::CollectiveOp,
+  using CollectiveMainloop =
       typename cutlass::gemm::collective::CollectiveBuilder<
           ArchTag, OperatorClass, ElementAB, LayoutA*, AlignmentAB, ElementAB,
           LayoutB*, AlignmentAB, ElementAccumulator, TileShape, ClusterShape,
-          Stages, KernelSchedule>::CollectiveOp>;
+          Stages, KernelSchedule>::CollectiveOp;
 
-  using KernelType = enable_sm90_or_later<cutlass::gemm::kernel::GemmUniversal<
+  using KernelType = enable_sm90_only<cutlass::gemm::kernel::GemmUniversal<
       ProblemShape, CollectiveMainloop, CollectiveEpilogue>>;
 
   struct GemmKernel : public KernelType {};
@@ -91,12 +78,12 @@ void cutlass_group_gemm_caller(
     torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
     torch::Tensor const& b_strides, torch::Tensor const& c_strides,
     bool per_act_token, bool per_out_ch) {
-  static constexpr bool swap_ab = Gemm::swap_ab;
-
   using ElementAB = typename Gemm::ElementAB;
   using ElementD = typename Gemm::ElementD;
 
   int num_experts = static_cast<int>(expert_offsets.size(0));
+  int k_size = a_tensors.size(1);
+  int n_size = out_tensors.size(1);
 
   auto stream = at::cuda::getCurrentCUDAStream(a_tensors.device().index());
 
@@ -123,47 +110,26 @@ void cutlass_group_gemm_caller(
           problem_sizes.data_ptr());
   ProblemShape prob_shape{num_experts, problem_sizes_as_shapes, nullptr};
 
-  typename GemmKernel::MainloopArguments mainloop_args;
-  if constexpr (swap_ab) {
-    mainloop_args = typename GemmKernel::MainloopArguments{
-        static_cast<const ElementAB**>(b_ptrs.data_ptr()),
-        static_cast<StrideB*>(b_strides.data_ptr()),
-        static_cast<const ElementAB**>(a_ptrs.data_ptr()),
-        static_cast<StrideA*>(a_strides.data_ptr())};
-  } else {
-    mainloop_args = typename GemmKernel::MainloopArguments{
-        static_cast<const ElementAB**>(a_ptrs.data_ptr()),
-        static_cast<StrideA*>(a_strides.data_ptr()),
-        static_cast<const ElementAB**>(b_ptrs.data_ptr()),
-        static_cast<StrideB*>(b_strides.data_ptr())};
-  }
+  typename GemmKernel::MainloopArguments mainloop_args{
+      static_cast<const ElementAB**>(a_ptrs.data_ptr()),
+      static_cast<StrideA*>(a_strides.data_ptr()),
+      static_cast<const ElementAB**>(b_ptrs.data_ptr()),
+      static_cast<StrideB*>(b_strides.data_ptr())};
 
   // Currently, we are only able to do broadcast on either all or none a_scales
   // and on either all or none b_scales
   typename GemmKernel::EpilogueArguments epilogue_args{
       Gemm::Epilogue::prepare_args(
-          swap_ab ? static_cast<const ElementAccumulator**>(
-                        b_scales_ptrs.data_ptr())
-                  : static_cast<const ElementAccumulator**>(
-                        a_scales_ptrs.data_ptr()),
-          swap_ab ? static_cast<const ElementAccumulator**>(
-                        a_scales_ptrs.data_ptr())
-                  : static_cast<const ElementAccumulator**>(
-                        b_scales_ptrs.data_ptr()),
-          swap_ab ? per_out_ch : per_act_token,
-          swap_ab ? per_act_token : per_out_ch),
+          static_cast<const ElementAccumulator**>(a_scales_ptrs.data_ptr()),
+          static_cast<const ElementAccumulator**>(b_scales_ptrs.data_ptr()),
+          per_act_token, per_out_ch),
       nullptr, static_cast<StrideC*>(c_strides.data_ptr()),
       static_cast<ElementD**>(out_ptrs.data_ptr()),
       static_cast<StrideC*>(c_strides.data_ptr())};
 
-  int device_id = a_tensors.device().index();
-  static const cutlass::KernelHardwareInfo hw_info{
-      device_id, cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
-                     device_id)};
-
   typename GemmKernel::Arguments args{
       cutlass::gemm::GemmUniversalMode::kGrouped, prob_shape, mainloop_args,
-      epilogue_args, hw_info};
+      epilogue_args};
 
   using GemmOp = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
   GemmOp gemm_op;

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu

@@ -1,140 +0,0 @@
-#include <cudaTypedefs.h>
-
-#include <c10/cuda/CUDAGuard.h>
-#include <torch/all.h>
-
-#include "cutlass/cutlass.h"
-#include "grouped_mm_c3x.cuh"
-
-using namespace cute;
-
-namespace {
-
-template <typename InType, typename OutType,
-          template <typename, typename, typename> typename Epilogue>
-struct sm100_fp8_config_default {
-  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
-  using KernelSchedule =
-      cutlass::gemm::KernelPtrArrayTmaWarpSpecialized1SmSm100;
-  using EpilogueSchedule = cutlass::epilogue::PtrArrayTmaWarpSpecialized1Sm;
-  using TileShape = cute::Shape<cute::_128, cute::_256, cute::_128>;
-  using ClusterShape = cute::Shape<cute::_1, cute::_1, cute::_1>;
-  using ArchTag = cutlass::arch::Sm100;
-
-  using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule>;
-};
-
-template <typename InType, typename OutType,
-          template <typename, typename, typename> typename Epilogue>
-struct sm100_fp8_config_M64 {
-  // M in [1,64]
-  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
-  using KernelSchedule =
-      cutlass::gemm::KernelPtrArrayTmaWarpSpecialized1SmSm100;
-  using EpilogueSchedule = cutlass::epilogue::PtrArrayTmaWarpSpecialized1Sm;
-  using TileShape = cute::Shape<cute::_128, cute::_16, cute::_128>;
-  using ClusterShape = cute::Shape<cute::_1, cute::_1, cute::_1>;
-  using ArchTag = cutlass::arch::Sm100;
-
-  using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule,
-                            true>;
-};
-
-template <typename InType, typename OutType,
-          template <typename, typename, typename> typename Epilogue>
-struct sm100_fp8_config_N8192 {
-  // N in [8192, inf)
-  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
-  using KernelSchedule =
-      cutlass::gemm::KernelPtrArrayTmaWarpSpecialized2SmSm100;
-  using EpilogueSchedule = cutlass::epilogue::PtrArrayTmaWarpSpecialized2Sm;
-  using TileShape = cute::Shape<cute::_128, cute::_256, cute::_128>;
-  using ClusterShape = cute::Shape<cute::_2, cute::_1, cute::_1>;
-  using ArchTag = cutlass::arch::Sm100;
-
-  using Cutlass3xGemm =
-      cutlass_3x_group_gemm<InType, OutType, ArchTag, Epilogue, TileShape,
-                            ClusterShape, KernelSchedule, EpilogueSchedule>;
-};
-
-template <typename InType, typename OutType>
-void run_cutlass_moe_mm_sm100(
-    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
-    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
-    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
-    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
-    torch::Tensor const& b_strides, torch::Tensor const& c_strides,
-    bool per_act_token, bool per_out_ch) {
-  TORCH_CHECK(a_tensors.size(0) > 0, "No input A tensors provided.");
-  TORCH_CHECK(b_tensors.size(0) > 0, "No input B tensors provided.");
-  TORCH_CHECK(out_tensors.size(0) > 0, "No output tensors provided.");
-
-  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn,
-              "A tensors must be of type float8_e4m3fn.");
-  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn,
-              "B tensors must be of type float8_e4m3fn.");
-
-  using Cutlass3xGemmDefault = typename sm100_fp8_config_default<
-      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
-  using Cutlass3xGemmN8192 = typename sm100_fp8_config_N8192<
-      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
-  using Cutlass3xGemmM64 = typename sm100_fp8_config_M64<
-      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
-
-  uint32_t const m = a_tensors.size(0);
-  uint32_t const n = out_tensors.size(1);
-
-  if (m <= 64) {
-    cutlass_group_gemm_caller<Cutlass3xGemmM64>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  } else if (n >= 8192) {
-    cutlass_group_gemm_caller<Cutlass3xGemmN8192>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  } else {
-    cutlass_group_gemm_caller<Cutlass3xGemmDefault>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  }
-}
-}  // namespace
-
-void dispatch_moe_mm_sm100(
-    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
-    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
-    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
-    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
-    torch::Tensor const& b_strides, torch::Tensor const& c_strides,
-    bool per_act_token, bool per_out_ch) {
-  if (out_tensors.dtype() == torch::kBFloat16) {
-    run_cutlass_moe_mm_sm100<cutlass::float_e4m3_t, cutlass::bfloat16_t>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  } else {
-    run_cutlass_moe_mm_sm100<cutlass::float_e4m3_t, cutlass::half_t>(
-        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
-        problem_sizes, a_strides, b_strides, c_strides, per_act_token,
-        per_out_ch);
-  }
-}
-
-void cutlass_moe_mm_sm100(
-    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
-    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
-    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
-    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
-    torch::Tensor const& b_strides, torch::Tensor const& c_strides,
-    bool per_act_token, bool per_out_ch) {
-  dispatch_moe_mm_sm100(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
-                        expert_offsets, problem_sizes, a_strides, b_strides,
-                        c_strides, per_act_token, per_out_ch);
-}

csrc/quantization/cutlass_w8a8/moe/moe_data.cu

@@ -6,11 +6,8 @@
 #include <iostream>
 
 constexpr uint64_t THREADS_PER_EXPERT = 512;
-// threshold must match the dispatch logic in run_cutlass_moe_mm_sm90()
-constexpr int SWAP_AB_THRESHOLD = 64;
 
-template <bool SWAP_AB>
-__global__ void compute_problem_sizes(const int32_t* __restrict__ topk_ids,
+__global__ void compute_problem_sizes(const uint32_t* __restrict__ topk_ids,
                                       int32_t* problem_sizes1,
                                       int32_t* problem_sizes2,
                                       int32_t* atomic_buffer,
@@ -27,58 +24,45 @@ __global__ void compute_problem_sizes(const int32_t* __restrict__ topk_ids,
 
   if (threadIdx.x == 0) {
     int final_occurrences = atomic_buffer[expert_id];
-    if constexpr (!SWAP_AB) {
-      problem_sizes1[expert_id * 3] = final_occurrences;
-      problem_sizes1[expert_id * 3 + 1] = 2 * n;
-      problem_sizes1[expert_id * 3 + 2] = k;
-      problem_sizes2[expert_id * 3] = final_occurrences;
-      problem_sizes2[expert_id * 3 + 1] = k;
-      problem_sizes2[expert_id * 3 + 2] = n;
-    } else {
-      problem_sizes1[expert_id * 3] = 2 * n;
-      problem_sizes1[expert_id * 3 + 1] = final_occurrences;
-      problem_sizes1[expert_id * 3 + 2] = k;
-      problem_sizes2[expert_id * 3] = k;
-      problem_sizes2[expert_id * 3 + 1] = final_occurrences;
-      problem_sizes2[expert_id * 3 + 2] = n;
-    }
+    problem_sizes1[expert_id * 3] = final_occurrences;
+    problem_sizes1[expert_id * 3 + 1] = 2 * n;
+    problem_sizes1[expert_id * 3 + 2] = k;
+    problem_sizes2[expert_id * 3] = final_occurrences;
+    problem_sizes2[expert_id * 3 + 1] = k;
+    problem_sizes2[expert_id * 3 + 2] = n;
   }
 }
 
 __global__ void compute_expert_offsets(
     const int32_t* __restrict__ problem_sizes1, int32_t* expert_offsets,
-    int32_t* atomic_buffer, const int num_experts, const int topk_length) {
+    int32_t* atomic_buffer, const int num_experts) {
   int32_t tot_offset = 0;
   expert_offsets[0] = 0;
   for (int i = 0; i < num_experts; ++i) {
     atomic_buffer[i] = tot_offset;
-    tot_offset += topk_length > SWAP_AB_THRESHOLD ? problem_sizes1[i * 3]
-                                                  : problem_sizes1[i * 3 + 1];
+    tot_offset += problem_sizes1[i * 3];
     expert_offsets[i + 1] = tot_offset;
   }
 }
 
 __global__ void compute_expert_blockscale_offsets(
     const int32_t* __restrict__ problem_sizes1, int32_t* expert_offsets,
-    int32_t* blockscale_offsets, int32_t* atomic_buffer, const int num_experts,
-    const int topk_length) {
+    int32_t* blockscale_offsets, int32_t* atomic_buffer,
+    const int num_experts) {
   int32_t tot_offset = 0;
   int32_t tot_offset_round = 0;
   expert_offsets[0] = 0;
   blockscale_offsets[0] = 0;
   for (int i = 0; i < num_experts; ++i) {
-    int32_t cur_offset = topk_length > SWAP_AB_THRESHOLD
-                             ? problem_sizes1[i * 3]
-                             : problem_sizes1[i * 3 + 1];
     atomic_buffer[i] = tot_offset;
-    tot_offset += cur_offset;
+    tot_offset += problem_sizes1[i * 3];
     expert_offsets[i + 1] = tot_offset;
-    tot_offset_round += (cur_offset + (128 - 1)) / 128 * 128;
+    tot_offset_round += (problem_sizes1[i * 3] + (128 - 1)) / 128 * 128;
     blockscale_offsets[i + 1] = tot_offset_round;
   }
 }
 
-__global__ void compute_arg_sorts(const int32_t* __restrict__ topk_ids,
+__global__ void compute_arg_sorts(const uint32_t* __restrict__ topk_ids,
                                   const int32_t* __restrict__ expert_offsets,
                                   int32_t* input_permutation,
                                   int32_t* output_permutation,
@@ -118,39 +102,25 @@ void get_cutlass_moe_mm_data_caller(
   torch::Tensor atomic_buffer = torch::zeros(num_experts, options_int32);
 
   int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
-
-  if (topk_ids.numel() > SWAP_AB_THRESHOLD) {
-    compute_problem_sizes<false><<<num_experts, num_threads, 0, stream>>>(
-        static_cast<const int32_t*>(topk_ids.data_ptr()),
-        static_cast<int32_t*>(problem_sizes1.data_ptr()),
-        static_cast<int32_t*>(problem_sizes2.data_ptr()),
-        static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n,
-        k);
-  } else {
-    compute_problem_sizes<true><<<num_experts, num_threads, 0, stream>>>(
-        static_cast<const int32_t*>(topk_ids.data_ptr()),
-        static_cast<int32_t*>(problem_sizes1.data_ptr()),
-        static_cast<int32_t*>(problem_sizes2.data_ptr()),
-        static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n,
-        k);
-  }
-
+  compute_problem_sizes<<<num_experts, num_threads, 0, stream>>>(
+      static_cast<const uint32_t*>(topk_ids.data_ptr()),
+      static_cast<int32_t*>(problem_sizes1.data_ptr()),
+      static_cast<int32_t*>(problem_sizes2.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n, k);
   if (blockscale_offsets.has_value()) {
     compute_expert_blockscale_offsets<<<1, 1, 0, stream>>>(
         static_cast<const int32_t*>(problem_sizes1.data_ptr()),
         static_cast<int32_t*>(expert_offsets.data_ptr()),
         static_cast<int32_t*>(blockscale_offsets.value().data_ptr()),
-        static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts,
-        topk_ids.numel());
+        static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
   } else {
     compute_expert_offsets<<<1, 1, 0, stream>>>(
         static_cast<const int32_t*>(problem_sizes1.data_ptr()),
         static_cast<int32_t*>(expert_offsets.data_ptr()),
-        static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts,
-        topk_ids.numel());
+        static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
   }
   compute_arg_sorts<<<num_experts, num_threads, 0, stream>>>(
-      static_cast<const int32_t*>(topk_ids.data_ptr()),
+      static_cast<const uint32_t*>(topk_ids.data_ptr()),
       static_cast<const int32_t*>(expert_offsets.data_ptr()),
       static_cast<int32_t*>(input_permutation.data_ptr()),
       static_cast<int32_t*>(output_permutation.data_ptr()),
@@ -190,4 +160,4 @@ void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
       static_cast<int32_t*>(problem_sizes2.data_ptr()),
       static_cast<const int32_t*>(expert_num_tokens.data_ptr()), padded_m, n,
       k);
-}
+}

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -41,16 +41,6 @@ void cutlass_moe_mm_sm90(
 
 #endif
 
-#if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100
-void cutlass_moe_mm_sm100(
-    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
-    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
-    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
-    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
-    torch::Tensor const& b_strides, torch::Tensor const& c_strides,
-    bool per_act_token, bool per_out_ch);
-#endif
-
 #if defined ENABLE_SCALED_MM_SM120 && ENABLE_SCALED_MM_SM120
 void cutlass_scaled_mm_sm120(torch::Tensor& c, torch::Tensor const& a,
                              torch::Tensor const& b,
@@ -140,10 +130,10 @@ bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) {
   // and at least SM90 (Hopper)
 
 #if defined CUDA_VERSION
-  if (cuda_device_capability >= 100) {
-    return CUDA_VERSION >= 12080;
-  } else if (cuda_device_capability >= 90) {
+  if (cuda_device_capability >= 90 && cuda_device_capability < 100) {
     return CUDA_VERSION >= 12000;
+  } else if (cuda_device_capability >= 100) {
+    return CUDA_VERSION >= 12080;
   }
 #endif
 
@@ -151,14 +141,11 @@ bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) {
 }
 
 bool cutlass_group_gemm_supported(int64_t cuda_device_capability) {
-  // CUTLASS grouped FP8 kernels need at least CUDA 12.3 and SM90 (Hopper)
-  // or CUDA 12.8 and SM100 (Blackwell)
+  // CUTLASS grouped FP8 kernels need at least CUDA 12.3
+  // and SM90 (Hopper)
 
 #if defined CUDA_VERSION
-  if (cuda_device_capability >= 100) {
-    return CUDA_VERSION >= 12080;
-  }
-  if (cuda_device_capability >= 90) {
+  if (cuda_device_capability == 90) {
     return CUDA_VERSION >= 12030;
   }
 #endif
@@ -247,26 +234,16 @@ void cutlass_moe_mm(
     torch::Tensor const& b_strides, torch::Tensor const& c_strides,
     bool per_act_token, bool per_out_ch) {
   int32_t version_num = get_sm_version_num();
-#if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100
-  if (version_num >= 100) {
-    cutlass_moe_mm_sm100(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
-                         expert_offsets, problem_sizes, a_strides, b_strides,
-                         c_strides, per_act_token, per_out_ch);
-    return;
-  }
-#endif
 #if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
-  if (version_num >= 90) {
-    cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
-                        expert_offsets, problem_sizes, a_strides, b_strides,
-                        c_strides, per_act_token, per_out_ch);
-    return;
-  }
+  cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
+                      expert_offsets, problem_sizes, a_strides, b_strides,
+                      c_strides, per_act_token, per_out_ch);
+  return;
 #endif
   TORCH_CHECK_NOT_IMPLEMENTED(
       false,
       "No compiled cutlass_scaled_mm for CUDA device capability: ", version_num,
-      ". Required capability: 90 or 100");
+      ". Required capability: 90");
 }
 
 void get_cutlass_moe_mm_data(

csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu

@@ -30,40 +30,35 @@
 
 #include "cutlass/util/packed_stride.hpp"
 
-#include "core/math.hpp"
-
 using namespace cute;
 
 #if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+// Kernel Perf config
+template <typename T>
+struct KernelTraits;
 
-// Configuration for M in (256, inf)
-struct sm100_fp4_config_default {
-  using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
-  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
-  using TileShape = Shape<_256, _256, _256>;
-  using ClusterShape = Shape<_2, _1, _1>;
-  using PerSmTileShape_MNK = Shape<_128, _256, _256>;
-};
-
-// Configuration for M in (16, 256]
-struct sm100_fp4_config_M256 {
-  using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
-  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
-  using TileShape = Shape<_256, _128, _256>;
-  using ClusterShape = Shape<_2, _1, _1>;
-  using PerSmTileShape_MNK = Shape<_128, _128, _256>;
-};
-
-// Configuration for M in [1, 16]
-struct sm100_fp4_config_M16 {
-  using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
-  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
-  using TileShape = Shape<_128, _128, _256>;
+template <>
+struct KernelTraits<float> {
+  using MmaTileShape = Shape<_128, _128, _256>;
   using ClusterShape = Shape<_1, _1, _1>;
   using PerSmTileShape_MNK = Shape<_128, _128, _256>;
 };
 
-template <typename Config, typename OutType>
+template <>
+struct KernelTraits<cutlass::half_t> {
+  using MmaTileShape = Shape<_256, _256, _256>;
+  using ClusterShape = Shape<_4, _4, _1>;
+  using PerSmTileShape_MNK = Shape<_128, _256, _256>;
+};
+
+template <>
+struct KernelTraits<cutlass::bfloat16_t> {
+  using MmaTileShape = Shape<_256, _256, _256>;
+  using ClusterShape = Shape<_4, _4, _1>;
+  using PerSmTileShape_MNK = Shape<_128, _256, _256>;
+};
+
+template <typename T>
 struct Fp4GemmSm100 {
   // A matrix configuration
   using ElementA = cutlass::nv_float4_t<cutlass::float_e2m1_t>;
@@ -76,22 +71,21 @@ struct Fp4GemmSm100 {
   static constexpr int AlignmentB = 32;
 
   // C/D matrix configuration
-  using ElementD = OutType;
-  using ElementC = OutType;
+  using ElementD = T;
+  using ElementC = T;
   using LayoutCTag = cutlass::layout::RowMajor;
   using LayoutDTag = cutlass::layout::RowMajor;
   static constexpr int AlignmentD = 128 / cutlass::sizeof_bits<ElementD>::value;
   static constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementC>::value;
-
   // Kernel functional config
   using ElementAccumulator = float;
   using ArchTag = cutlass::arch::Sm100;
   using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp;
 
-  // Use config's tile shapes
-  using MmaTileShape = typename Config::TileShape;
-  using ClusterShape = typename Config::ClusterShape;
-  using PerSmTileShape_MNK = typename Config::PerSmTileShape_MNK;
+  // Kernel Perf config
+  using MmaTileShape = typename KernelTraits<T>::MmaTileShape;
+  using ClusterShape = typename KernelTraits<T>::ClusterShape;
+  using PerSmTileShape_MNK = typename KernelTraits<T>::PerSmTileShape_MNK;
 
   using CollectiveEpilogue =
       typename cutlass::epilogue::collective::CollectiveBuilder<
@@ -125,22 +119,22 @@ struct Fp4GemmSm100 {
   using LayoutD = decltype(cute::make_layout(make_shape(0, 0, 0), StrideD{}));
 };
 
-template <typename Config>
-typename Config::Gemm::Arguments args_from_options(
+template <typename T>
+typename T::Gemm::Arguments args_from_options(
     at::Tensor& D, at::Tensor const& A, at::Tensor const& B,
     at::Tensor const& A_sf, at::Tensor const& B_sf, at::Tensor const& alpha,
     int64_t M, int64_t N, int64_t K) {
-  using ElementA = typename Config::Gemm::ElementA;
-  using ElementB = typename Config::Gemm::ElementB;
+  using ElementA = typename T::Gemm::ElementA;
+  using ElementB = typename T::Gemm::ElementB;
   using ElementSFA = cutlass::float_ue4m3_t;
   using ElementSFB = cutlass::float_ue4m3_t;
-  using ElementD = typename Config::Gemm::ElementD;
+  using ElementD = typename T::Gemm::ElementD;
   using ElementCompute = float;
-  using StrideA = typename Config::StrideA;
-  using StrideB = typename Config::StrideB;
-  using StrideD = typename Config::StrideD;
-  using Sm100BlkScaledConfig = typename Config::Gemm::GemmKernel::
-      CollectiveMainloop::Sm1xxBlkScaledConfig;
+  using StrideA = typename T::StrideA;
+  using StrideB = typename T::StrideB;
+  using StrideD = typename T::StrideD;
+  using Sm100BlkScaledConfig =
+      typename T::Gemm::GemmKernel::CollectiveMainloop::Sm1xxBlkScaledConfig;
 
   int m = static_cast<int>(M);
   int n = static_cast<int>(N);
@@ -154,7 +148,7 @@ typename Config::Gemm::Arguments args_from_options(
   auto layout_SFB = Sm100BlkScaledConfig::tile_atom_to_shape_SFB(
       cute::make_shape(m, n, k, 1));
 
-  typename Config::Gemm::Arguments arguments{
+  typename T::Gemm::Arguments arguments{
       cutlass::gemm::GemmUniversalMode::kGemm,
       {m, n, k, 1},
       {// Mainloop arguments
@@ -173,17 +167,17 @@ typename Config::Gemm::Arguments args_from_options(
   return arguments;
 }
 
-template <typename Config>
+template <typename T>
 void runGemm(at::Tensor& D, at::Tensor const& A, at::Tensor const& B,
              at::Tensor const& A_sf, at::Tensor const& B_sf,
              at::Tensor const& alpha, int64_t m, int64_t n, int64_t k,
              cudaStream_t stream) {
-  typename Config::Gemm gemm;
+  typename Fp4GemmSm100<T>::Gemm gemm;
 
   auto arguments =
-      args_from_options<Config>(D, A, B, A_sf, B_sf, alpha, m, n, k);
+      args_from_options<Fp4GemmSm100<T>>(D, A, B, A_sf, B_sf, alpha, m, n, k);
 
-  size_t workspace_size = Config::Gemm::get_workspace_size(arguments);
+  size_t workspace_size = Fp4GemmSm100<T>::Gemm::get_workspace_size(arguments);
   auto const workspace_options =
       torch::TensorOptions().dtype(torch::kUInt8).device(A.device());
   auto workspace = torch::empty(workspace_size, workspace_options);
@@ -194,40 +188,12 @@ void runGemm(at::Tensor& D, at::Tensor const& A, at::Tensor const& B,
 
   CUTLASS_CHECK(gemm.run(arguments, workspace.data_ptr(), stream));
 }
-
-// Dispatch function to select appropriate config based on M
-template <typename OutType>
-void cutlass_fp4_gemm_dispatch(torch::Tensor& D, torch::Tensor const& A,
-                               torch::Tensor const& B,
-                               torch::Tensor const& A_sf,
-                               torch::Tensor const& B_sf,
-                               torch::Tensor const& alpha, int64_t m, int64_t n,
-                               int64_t k, cudaStream_t stream) {
-  uint32_t const mp2 = std::max(static_cast<uint32_t>(16), next_pow_2(m));
-
-  if (mp2 <= 16) {
-    // m in [1, 16]
-    runGemm<Fp4GemmSm100<sm100_fp4_config_M16, OutType>>(
-        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
-  } else if (mp2 <= 256) {
-    // m in (16, 256]
-    runGemm<Fp4GemmSm100<sm100_fp4_config_M256, OutType>>(
-        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
-  } else {
-    // m in (256, inf)
-    runGemm<Fp4GemmSm100<sm100_fp4_config_default, OutType>>(
-        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
-  }
-}
-
 #else
-template <typename OutType>
-void cutlass_fp4_gemm_dispatch(torch::Tensor& D, torch::Tensor const& A,
-                               torch::Tensor const& B,
-                               torch::Tensor const& A_sf,
-                               torch::Tensor const& B_sf,
-                               torch::Tensor const& alpha, int64_t m, int64_t n,
-                               int64_t k, cudaStream_t stream) {
+template <typename T>
+void runGemm(at::Tensor& D, at::Tensor const& A, at::Tensor const& B,
+             at::Tensor const& A_sf, at::Tensor const& B_sf,
+             at::Tensor const& alpha, int64_t m, int64_t n, int64_t k,
+             cudaStream_t stream) {
   TORCH_CHECK(false,
               "Unsupported CUTLASS version. Set VLLM_CUTLASS_SRC_DIR to "
               "a CUTLASS 3.8 source directory to enable support.");
@@ -305,13 +271,12 @@ void cutlass_scaled_fp4_mm_sm100a(torch::Tensor& D, torch::Tensor const& A,
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream(A.get_device());
 
   if (out_dtype == at::ScalarType::Half) {
-    cutlass_fp4_gemm_dispatch<cutlass::half_t>(D, A, B, A_sf, B_sf, alpha, m, n,
-                                               k, stream);
+    runGemm<cutlass::half_t>(D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
   } else if (out_dtype == at::ScalarType::BFloat16) {
-    cutlass_fp4_gemm_dispatch<cutlass::bfloat16_t>(D, A, B, A_sf, B_sf, alpha,
-                                                   m, n, k, stream);
+    runGemm<cutlass::bfloat16_t>(D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
+  } else if (out_dtype == at::ScalarType::Float) {
+    runGemm<float>(D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
   } else {
-    TORCH_CHECK(false, "Unsupported output data type of nvfp4 mm (", out_dtype,
-                ")");
+    TORCH_CHECK(false, "Unsupported output data type of nvfp4 mm");
   }
 }

csrc/quantization/fp8/common.cu

@@ -88,8 +88,6 @@ void static_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
                              torch::Tensor const& input,  // [..., d]
                              torch::Tensor const& scale)  // [1]
 {
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
   int const block_size = 256;
   int const num_tokens = input.numel() / input.size(-1);
   int const num_elems = input.numel();
@@ -113,8 +111,6 @@ void dynamic_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
                               torch::Tensor const& input,  // [..., d]
                               torch::Tensor& scale)        // [1]
 {
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
   int const block_size = 256;
   int const num_tokens = input.numel() / input.size(-1);
   int const num_elems = input.numel();

csrc/quantization/fp8/per_token_group_quant.cu

@@ -1,213 +0,0 @@
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/util/Float8_e4m3fn.h>
-
-#include <cmath>
-
-#include <cuda_fp16.h>
-#include <cuda_bf16.h>
-
-#include <torch/all.h>
-
-#include "../vectorization.cuh"
-#include "../vectorization_utils.cuh"
-#include "../../dispatch_utils.h"
-
-__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
-  unsigned mask = 0xffff;
-
-  val = fmaxf(val, __shfl_xor_sync(mask, val, 8));
-  val = fmaxf(val, __shfl_xor_sync(mask, val, 4));
-  val = fmaxf(val, __shfl_xor_sync(mask, val, 2));
-  val = fmaxf(val, __shfl_xor_sync(mask, val, 1));
-  return val;
-}
-
-template <typename T, typename DST_DTYPE, bool IS_COLUMN_MAJOR = false,
-          bool SCALE_UE8M0 = false, typename scale_packed_t = float>
-__global__ void per_token_group_quant_8bit_kernel(
-    const T* __restrict__ input, void* __restrict__ output_q,
-    scale_packed_t* __restrict__ output_s, const int group_size,
-    const int num_groups, const int groups_per_block, const float eps,
-    const float min_8bit, const float max_8bit, const int scale_num_rows = 0,
-    const int scale_stride = 0) {
-  const int threads_per_group = 16;
-  const int64_t local_group_id = threadIdx.x / threads_per_group;
-  const int lane_id = threadIdx.x % threads_per_group;
-
-  const int64_t block_group_id = blockIdx.x * groups_per_block;
-  const int64_t global_group_id = block_group_id + local_group_id;
-  const int64_t block_group_offset = global_group_id * group_size;
-
-  float local_absmax = eps;
-
-  using scale_element_t = float;
-  static_assert(sizeof(scale_packed_t) % sizeof(scale_element_t) == 0);
-
-  const T* group_input = input + block_group_offset;
-  DST_DTYPE* group_output =
-      static_cast<DST_DTYPE*>(output_q) + block_group_offset;
-  scale_element_t* scale_output;
-
-  if constexpr (IS_COLUMN_MAJOR) {
-    const int num_elems_per_pack =
-        static_cast<int>(sizeof(scale_packed_t) / sizeof(scale_element_t));
-    const int scale_num_rows_element = scale_num_rows * num_elems_per_pack;
-    const int row_idx = global_group_id / scale_num_rows_element;
-    const int col_idx_raw = global_group_id % scale_num_rows_element;
-    const int col_idx = col_idx_raw / num_elems_per_pack;
-    const int pack_idx = col_idx_raw % num_elems_per_pack;
-    scale_output = reinterpret_cast<scale_element_t*>(output_s) +
-                   (col_idx * scale_stride * num_elems_per_pack +
-                    row_idx * num_elems_per_pack + pack_idx);
-  } else {
-    scale_output = output_s + global_group_id;
-  }
-
-  // shared memory to cache each group's data to avoid double DRAM reads.
-  extern __shared__ __align__(16) char smem_raw[];
-  T* smem = reinterpret_cast<T*>(smem_raw);
-  T* smem_group = smem + local_group_id * group_size;
-
-  constexpr int vec_size = 16 / sizeof(T);
-  using vec_t = vllm::vec_n_t<T, vec_size>;
-
-  // copy global -> shared & compute absmax
-  auto scalar_op_cache = [&] __device__(T & dst, const T& src) {
-    float abs_v = fabsf(static_cast<float>(src));
-    local_absmax = fmaxf(local_absmax, abs_v);
-    dst = src;
-  };
-
-  vllm::vectorize_with_alignment<vec_size>(
-      group_input,        // in
-      smem_group,         // out (shared)
-      group_size,         // elements per group
-      lane_id,            // thread id
-      threads_per_group,  // stride in group
-      scalar_op_cache);   // scalar handler
-
-  local_absmax = GroupReduceMax(local_absmax, lane_id);
-
-  float y_s = local_absmax / max_8bit;
-  if constexpr (SCALE_UE8M0) {
-    y_s = exp2f(ceilf(log2f(fmaxf(fabsf(y_s), 1e-10f))));
-  }
-
-  scale_element_t y_s_quant = y_s;
-
-  if (lane_id == 0) {
-    *scale_output = y_s_quant;
-  }
-
-  __syncthreads();
-
-  // quantize shared -> global 8-bit
-  auto scalar_op_quant = [&] __device__(DST_DTYPE & dst, const T& src) {
-    float q = fminf(fmaxf(static_cast<float>(src) / y_s, min_8bit), max_8bit);
-    dst = DST_DTYPE(q);
-  };
-
-  vllm::vectorize_with_alignment<vec_size>(
-      smem_group,         // in (shared)
-      group_output,       // out (global quant tensor)
-      group_size,         // elements
-      lane_id,            // tid
-      threads_per_group,  // stride
-      scalar_op_quant);   // scalar handler
-}
-
-void per_token_group_quant_8bit(const torch::Tensor& input,
-                                torch::Tensor& output_q,
-                                torch::Tensor& output_s, int64_t group_size,
-                                double eps, double min_8bit, double max_8bit,
-                                bool scale_ue8m0 = false) {
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(output_q.is_contiguous());
-
-  const int num_groups = input.numel() / group_size;
-
-  TORCH_CHECK(input.numel() % group_size == 0);
-  TORCH_CHECK(output_s.dim() == 2);
-
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  constexpr int THREADS_PER_GROUP = 16;
-
-  int groups_per_block = 1;
-
-  if (num_groups % 16 == 0) {
-    groups_per_block = 16;
-  } else if (num_groups % 8 == 0) {
-    groups_per_block = 8;
-  } else if (num_groups % 4 == 0) {
-    groups_per_block = 4;
-  } else if (num_groups % 2 == 0) {
-    groups_per_block = 2;
-  }
-
-  auto dst_type = output_q.scalar_type();
-  const int num_blocks = num_groups / groups_per_block;
-  const int num_threads = groups_per_block * THREADS_PER_GROUP;
-
-  const bool is_column_major = output_s.stride(0) < output_s.stride(1);
-  const int scale_num_rows = output_s.size(1);
-  const int scale_stride = output_s.stride(1);
-
-#define LAUNCH_KERNEL(T, DST_DTYPE)                                        \
-  do {                                                                     \
-    dim3 grid(num_blocks);                                                 \
-    dim3 block(num_threads);                                               \
-    size_t smem_bytes =                                                    \
-        static_cast<size_t>(groups_per_block) * group_size * sizeof(T);    \
-    if (is_column_major) {                                                 \
-      if (scale_ue8m0) {                                                   \
-        per_token_group_quant_8bit_kernel<T, DST_DTYPE, true, true>        \
-            <<<grid, block, smem_bytes, stream>>>(                         \
-                static_cast<T*>(input.data_ptr()), output_q.data_ptr(),    \
-                static_cast<float*>(output_s.data_ptr()), group_size,      \
-                num_groups, groups_per_block, (float)eps, (float)min_8bit, \
-                (float)max_8bit, scale_num_rows, scale_stride);            \
-      } else {                                                             \
-        per_token_group_quant_8bit_kernel<T, DST_DTYPE, true, false>       \
-            <<<grid, block, smem_bytes, stream>>>(                         \
-                static_cast<T*>(input.data_ptr()), output_q.data_ptr(),    \
-                static_cast<float*>(output_s.data_ptr()), group_size,      \
-                num_groups, groups_per_block, (float)eps, (float)min_8bit, \
-                (float)max_8bit, scale_num_rows, scale_stride);            \
-      }                                                                    \
-    } else {                                                               \
-      if (scale_ue8m0) {                                                   \
-        per_token_group_quant_8bit_kernel<T, DST_DTYPE, false, true>       \
-            <<<grid, block, smem_bytes, stream>>>(                         \
-                static_cast<T*>(input.data_ptr()), output_q.data_ptr(),    \
-                static_cast<float*>(output_s.data_ptr()), group_size,      \
-                num_groups, groups_per_block, (float)eps, (float)min_8bit, \
-                (float)max_8bit);                                          \
-      } else {                                                             \
-        per_token_group_quant_8bit_kernel<T, DST_DTYPE, false, false>      \
-            <<<grid, block, smem_bytes, stream>>>(                         \
-                static_cast<T*>(input.data_ptr()), output_q.data_ptr(),    \
-                static_cast<float*>(output_s.data_ptr()), group_size,      \
-                num_groups, groups_per_block, (float)eps, (float)min_8bit, \
-                (float)max_8bit);                                          \
-      }                                                                    \
-    }                                                                      \
-  } while (0)
-
-  VLLM_DISPATCH_FLOATING_TYPES(
-      input.scalar_type(), "per_token_group_quant_8bit", ([&] {
-        if (dst_type == at::ScalarType::Float8_e4m3fn) {
-          LAUNCH_KERNEL(scalar_t, c10::Float8_e4m3fn);
-        }
-      }));
-
-#undef LAUNCH_KERNEL
-}
-
-void per_token_group_quant_fp8(const torch::Tensor& input,
-                               torch::Tensor& output_q, torch::Tensor& output_s,
-                               int64_t group_size, double eps, double fp8_min,
-                               double fp8_max, bool scale_ue8m0) {
-  per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
-                             fp8_min, fp8_max, scale_ue8m0);
-}

csrc/torch_bindings.cpp

@@ -20,17 +20,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
   //
 
-  // The default behavior in PyTorch 2.6 was changed to "requires_contiguous",
-  // so we need
+  // The default behavior in PyTorch 2.6 is "requires_contiguous", so we need
   // to override this for many GEMMs with the following tag. Otherwise,
   // torch.compile will force all input tensors to be contiguous(), which
   // will break many custom ops that require column-major weight matrices.
-  // This was a bug and PyTorch 2.7 has since fixed this.
-#if TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 6
-  #define stride_tag at::Tag::needs_fixed_stride_order
-#else
-  #define stride_tag
-#endif
+  // TODO: remove this for PyTorch 2.8, when the default is planned to switch
+  // to match exact eager-mode strides.
+  at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
 
   ops.def("weak_ref_tensor(Tensor input) -> Tensor");
   ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);
@@ -518,22 +514,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "                   Tensor page_table, float scale) -> ()");
   ops.impl("cutlass_mla_decode", torch::kCUDA, &cutlass_mla_decode);
 
-  // SM100 CUTLASS MLA decode
-  ops.def(
-      "sm100_cutlass_mla_decode(Tensor! out, Tensor q_nope, Tensor q_pe,"
-      "                         Tensor kv_c_and_k_pe_cache, Tensor seq_lens,"
-      "                         Tensor page_table, Tensor workspace, float "
-      "scale,"
-      "                         int num_kv_splits) -> ()");
-  // conditionally compiled so impl in source file
-
-  // SM100 CUTLASS MLA workspace
-  ops.def(
-      "sm100_cutlass_mla_get_workspace_size(int max_seq_len, int num_batches,"
-      "                                     int sm_count, int num_kv_splits) "
-      "-> int");
-  // conditionally compiled so impl in source file
-
   // Compute NVFP4 block quantized tensor.
   ops.def(
       "scaled_fp4_quant(Tensor! output, Tensor input,"
@@ -614,16 +594,29 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "int pad_slot_id) -> ()");
   ops.impl("selective_scan_fwd", torch::kCUDA, &selective_scan_fwd);
 
-#ifndef USE_ROCM
-  // Compute per-token-group FP8 quantized tensor and scaling factor.
   ops.def(
-      "per_token_group_fp8_quant(Tensor input, Tensor! output_q, Tensor! "
-      "output_s, "
-      "int group_size, float eps, float fp8_min, float fp8_max, bool "
-      "scale_ue8m0) -> ()");
-  ops.impl("per_token_group_fp8_quant", torch::kCUDA,
-           &per_token_group_quant_fp8);
+      "causal_conv1d_update(Tensor! x,"
+      "Tensor! conv_state,"
+      "Tensor! weight,"
+      "Tensor? bias_,"
+      "bool silu_activation,"
+      "Tensor? cache_seqlens_,"
+      "Tensor? conv_state_indices,"
+      "int pad_slot_id) -> ()");
+  ops.impl("causal_conv1d_update", torch::kCUDA, &causal_conv1d_update);
 
+  ops.def(
+      "causal_conv1d_fwd(Tensor! x, Tensor! weight,"
+      "Tensor? bias_,"
+      "Tensor!? conv_states,"
+      "Tensor? query_start_loc,"
+      "Tensor? cache_indices,"
+      "Tensor? has_initial_state,"
+      "bool silu_activation,"
+      "int pad_slot_id) -> ()");
+  ops.impl("causal_conv1d_fwd", torch::kCUDA, &causal_conv1d_fwd);
+
+#ifndef USE_ROCM
   // reorder weight for AllSpark Ampere W8A16 Fused Gemm kernel
   ops.def(
       "rearrange_kn_weight_as_n32k16_order(Tensor b_qweight, Tensor b_scales, "

docker/Dockerfile

@@ -63,7 +63,7 @@ ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL=https://download.pytorch.org/whl/nightly
 ARG PIP_KEYRING_PROVIDER=disabled
 ARG UV_KEYRING_PROVIDER=${PIP_KEYRING_PROVIDER}
 
-# Flag enables built-in KV-connector dependency libs into docker images
+# Flag enables build-in KV-connector dependency libs into docker images
 ARG INSTALL_KV_CONNECTORS=false
 
 #################### BASE BUILD IMAGE ####################
@@ -207,19 +207,6 @@ ARG SCCACHE_ENDPOINT
 ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
-
-# Flag to control whether to use pre-built vLLM wheels
-ARG VLLM_USE_PRECOMPILED
-# TODO: in setup.py VLLM_USE_PRECOMPILED is sensitive to truthiness, it will take =0 as "true", this should be fixed
-ENV VLLM_USE_PRECOMPILED=""
-RUN if [ "${VLLM_USE_PRECOMPILED}" = "1" ]; then \
-        export VLLM_USE_PRECOMPILED=1 && \
-        echo "Using precompiled wheels"; \
-    else \
-        unset VLLM_USE_PRECOMPILED && \
-        echo "Leaving VLLM_USE_PRECOMPILED unset to build wheels from source"; \
-    fi
-
 # if USE_SCCACHE is set, use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
@@ -265,7 +252,7 @@ RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
 #################### EXTENSION Build IMAGE ####################
 
 #################### DEV IMAGE ####################
-FROM base AS dev
+FROM base as dev
 
 ARG PIP_INDEX_URL UV_INDEX_URL
 ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
@@ -388,33 +375,48 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
 # -rw-rw-r-- 1 mgoin mgoin 205M Jun  9 18:03 flashinfer_python-0.2.6.post1-cp39-abi3-linux_x86_64.whl
 # $ # upload the wheel to a public location, e.g. https://wheels.vllm.ai/flashinfer/v0.2.6.post1/flashinfer_python-0.2.6.post1-cp39-abi3-linux_x86_64.whl
 
-# Install FlashInfer from source
+# Allow specifying a version, Git revision or local .whl file
+ARG FLASHINFER_CUDA128_INDEX_URL="https://download.pytorch.org/whl/cu128/flashinfer"
+ARG FLASHINFER_CUDA128_WHEEL="flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl"
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-ARG FLASHINFER_GIT_REF="v0.2.8rc1"
+ARG FLASHINFER_GIT_REF="v0.2.6.post1"
 RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
   . /etc/environment
-    git clone --depth 1 --recursive --shallow-submodules \
-        --branch ${FLASHINFER_GIT_REF} \
-        ${FLASHINFER_GIT_REPO} flashinfer
-    # Exclude CUDA arches for older versions (11.x and 12.0-12.7)
-    # TODO: Update this to allow setting TORCH_CUDA_ARCH_LIST as a build arg.
-    if [[ "${CUDA_VERSION}" == 11.* ]]; then
-        FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9"
-    elif [[ "${CUDA_VERSION}" == 12.[0-7]* ]]; then
-        FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a"
-    else
-        # CUDA 12.8+ supports 10.0a and 12.0
-        FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0a 12.0"
-    fi
-    echo "🏗️  Building FlashInfer for arches: ${FI_TORCH_CUDA_ARCH_LIST}"
-    # Needed to build AOT kernels
-    pushd flashinfer
-        TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
+  if [ "$TARGETPLATFORM" != "linux/arm64" ]; then
+      # FlashInfer already has a wheel for PyTorch 2.7.0 and CUDA 12.8. This is enough for CI use
+      if [[ "$CUDA_VERSION" == 12.8* ]]; then
+          uv pip install --system ${FLASHINFER_CUDA128_INDEX_URL}/${FLASHINFER_CUDA128_WHEEL}
+      else
+          export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0a 10.0a 12.0'
+          git clone ${FLASHINFER_GIT_REPO} --single-branch --branch ${FLASHINFER_GIT_REF} --recursive
+          # Needed to build AOT kernels
+          (cd flashinfer && \
+              python3 -m flashinfer.aot && \
+              uv pip install --system --no-build-isolation . \
+          )
+          rm -rf flashinfer
+
+          # Default arches (skipping 10.0a and 12.0 since these need 12.8)
+          # TODO: Update this to allow setting TORCH_CUDA_ARCH_LIST as a build arg.
+          TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a"
+          if [[ "${CUDA_VERSION}" == 11.* ]]; then
+              TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9"
+          fi
+          echo "🏗️  Building FlashInfer for arches: ${TORCH_CUDA_ARCH_LIST}"
+
+          git clone --depth 1 --recursive --shallow-submodules \
+            --branch v0.2.6.post1 \
+            https://github.com/flashinfer-ai/flashinfer.git flashinfer
+
+          pushd flashinfer
             python3 -m flashinfer.aot
-        TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
-            uv pip install --system --no-build-isolation .
-    popd
-    rm -rf flashinfer
+            TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}" \
+              uv pip install --system --no-build-isolation .
+          popd
+
+          rm -rf flashinfer
+      fi \
+  fi
 BASH
 COPY examples examples
 COPY benchmarks benchmarks
@@ -506,11 +508,10 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         uv pip install --system -r requirements/kv_connectors.txt; \
     fi; \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        BITSANDBYTES_VERSION="0.42.0"; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
     else \
-        BITSANDBYTES_VERSION="0.46.1"; \
-    fi; \
-    uv pip install --system accelerate hf_transfer modelscope "bitsandbytes>=${BITSANDBYTES_VERSION}" 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.46.1' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+    fi
 
 ENV VLLM_USAGE_SOURCE production-docker-image
 

docker/Dockerfile.cpu

@@ -95,7 +95,7 @@ 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/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

docker/Dockerfile.hpu

@@ -0,0 +1,21 @@
+FROM vault.habana.ai/gaudi-docker/1.20.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+RUN pip install -v -r requirements/hpu.txt
+
+ENV no_proxy=localhost,127.0.0.1
+ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
+
+RUN VLLM_TARGET_DEVICE=hpu python3 setup.py install
+
+# install development dependencies (for testing)
+RUN python3 -m pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.rocm_base

@@ -12,7 +12,7 @@ ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
 ARG FA_BRANCH="1a7f4dfa"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="916bf3c"
+ARG AITER_BRANCH="6487649"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 
 FROM ${BASE_IMAGE} AS base

docker/Dockerfile.tpu

@@ -1,5 +1,5 @@
-ARG NIGHTLY_DATE="20250714"
-ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.12_tpuvm_$NIGHTLY_DATE"
+ARG NIGHTLY_DATE="20250124"
+ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
 
 FROM $BASE_IMAGE
 WORKDIR /workspace/vllm

docker/Dockerfile.xpu

@@ -47,7 +47,7 @@ FROM vllm-base AS vllm-openai
 
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer pytest pytest_asyncio lm_eval[api] modelscope
+    pip install accelerate hf_transfer 'modelscope!=1.15.0'
 
 ENV VLLM_USAGE_SOURCE production-docker-image \
     TRITON_XPU_PROFILE 1

docs/.nav.yml

@@ -55,7 +55,6 @@ nav:
       - contributing/model/registration.md
       - contributing/model/tests.md
       - contributing/model/multimodal.md
-    - CI: contributing/ci
     - Design Documents:
       - V0: design
       - V1: design/v1

docs/README.md

@@ -36,7 +36,7 @@ vLLM is flexible and easy to use with:
 
 - Seamless integration with popular HuggingFace models
 - High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
-- Tensor, pipeline, data and expert parallelism support for distributed inference
+- Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
 - Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs, Gaudi® accelerators and GPUs, IBM Power CPUs, TPU, and AWS Trainium and Inferentia Accelerators.
@@ -48,4 +48,4 @@ For more information, check out the following:
 - [vLLM announcing blog post](https://vllm.ai) (intro to PagedAttention)
 - [vLLM paper](https://arxiv.org/abs/2309.06180) (SOSP 2023)
 - [How continuous batching enables 23x throughput in LLM inference while reducing p50 latency](https://www.anyscale.com/blog/continuous-batching-llm-inference) by Cade Daniel et al.
-- [vLLM Meetups](community/meetups.md)
+- [vLLM Meetups][meetups]

docs/api/README.md

@@ -8,12 +8,14 @@ API documentation for vLLM's configuration classes.
 
 - [vllm.config.ModelConfig][]
 - [vllm.config.CacheConfig][]
+- [vllm.config.TokenizerPoolConfig][]
 - [vllm.config.LoadConfig][]
 - [vllm.config.ParallelConfig][]
 - [vllm.config.SchedulerConfig][]
 - [vllm.config.DeviceConfig][]
 - [vllm.config.SpeculativeConfig][]
 - [vllm.config.LoRAConfig][]
+- [vllm.config.PromptAdapterConfig][]
 - [vllm.config.MultiModalConfig][]
 - [vllm.config.PoolerConfig][]
 - [vllm.config.DecodingConfig][]
@@ -62,7 +64,7 @@ vLLM provides experimental support for multi-modal models through the [vllm.mult
 Multi-modal inputs can be passed alongside text and token prompts to [supported models][supported-mm-models]
 via the `multi_modal_data` field in [vllm.inputs.PromptType][].
 
-Looking to add your own multi-modal model? Please follow the instructions listed [here](../contributing/model/multimodal.md).
+Looking to add your own multi-modal model? Please follow the instructions listed [here][supports-multimodal].
 
 - [vllm.multimodal.MULTIMODAL_REGISTRY][]
 

docs/ci/update_pytorch_version.md

@@ -1,12 +1,15 @@
-# Update PyTorch version on vLLM OSS CI/CD
+---
+title: Update PyTorch version on vLLM OSS CI/CD
+---
 
 vLLM's current policy is to always use the latest PyTorch stable
 release in CI/CD. It is standard practice to submit a PR to update the
 PyTorch version as early as possible when a new [PyTorch stable
 release](https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-cadence) becomes available.
 This process is non-trivial due to the gap between PyTorch
-releases. Using <gh-pr:16859> as an example, this document outlines common steps to achieve this
-update along with a list of potential issues and how to address them.
+releases. Using [#16859](https://github.com/vllm-project/vllm/pull/16859) as
+an example, this document outlines common steps to achieve this update along with
+a list of potential issues and how to address them.
 
 ## Test PyTorch release candidates (RCs)
 
@@ -16,12 +19,11 @@ by waiting for the next release or by implementing hacky workarounds in vLLM.
 The better solution is to test vLLM with PyTorch release candidates (RC) to ensure
 compatibility before each release.
 
-PyTorch release candidates can be downloaded from [PyTorch test index](https://download.pytorch.org/whl/test).
-For example, `torch2.7.0+cu12.8` RC can be installed using the following command:
+PyTorch release candidates can be downloaded from PyTorch test index at https://download.pytorch.org/whl/test.
+For example, torch2.7.0+cu12.8 RC can be installed using the following command:
 
-```bash
-uv pip install torch torchvision torchaudio \
-    --index-url https://download.pytorch.org/whl/test/cu128
+```
+uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu128
 ```
 
 When the final RC is ready for testing, it will be announced to the community
@@ -29,28 +31,13 @@ on the [PyTorch dev-discuss forum](https://dev-discuss.pytorch.org/c/release-ann
 After this announcement, we can begin testing vLLM integration by drafting a pull request
 following this 3-step process:
 
-1. Update [requirements files](https://github.com/vllm-project/vllm/tree/main/requirements)
-to point to the new releases for `torch`, `torchvision`, and `torchaudio`.
-
-2. Use the following option to get the final release candidates' wheels. Some common platforms are `cpu`, `cu128`, and `rocm6.2.4`.
-
-    ```bash
-    --extra-index-url https://download.pytorch.org/whl/test/<PLATFORM>
-    ```
-
-3. Since vLLM uses `uv`, ensure the following index strategy is applied:
-
-    - Via environment variable:
-
-    ```bash
-    export UV_INDEX_STRATEGY=unsafe-best-match
-    ```
-
-    - Or via CLI flag:
-
-    ```bash
-    --index-strategy unsafe-best-match
-    ```
+1. Update requirements files in https://github.com/vllm-project/vllm/tree/main/requirements
+to point to the new releases for torch, torchvision, and torchaudio.
+2. Use `--extra-index-url https://download.pytorch.org/whl/test/<PLATFORM>` to
+get the final release candidates' wheels.  Some common platforms are `cpu`, `cu128`,
+and `rocm6.2.4`.
+3. As vLLM uses uv, make sure that `unsafe-best-match` strategy is set either
+via `UV_INDEX_STRATEGY` env variable or via `--index-strategy unsafe-best-match`.
 
 If failures are found in the pull request, raise them as issues on vLLM and
 cc the PyTorch release team to initiate discussion on how to address them.
@@ -58,25 +45,20 @@ cc the PyTorch release team to initiate discussion on how to address them.
 ## Update CUDA version
 
 The PyTorch release matrix includes both stable and experimental [CUDA versions](https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix). Due to limitations, only the latest stable CUDA version (for example,
-`torch2.7.0+cu12.6`) is uploaded to PyPI. However, vLLM may require a different CUDA version,
+torch2.7.0+cu12.6) is uploaded to PyPI. However, vLLM may require a different CUDA version,
 such as 12.8 for Blackwell support.
 This complicates the process as we cannot use the out-of-the-box
 `pip install torch torchvision torchaudio` command. The solution is to use
 `--extra-index-url` in vLLM's Dockerfiles.
 
-- Important indexes at the moment include:
-
-| Platform | `--extra-index-url` |
-|----------|-----------------|
-| CUDA 12.8| [https://download.pytorch.org/whl/cu128](https://download.pytorch.org/whl/cu128)|
-| CPU      | [https://download.pytorch.org/whl/cpu](https://download.pytorch.org/whl/cpu)|
-| ROCm 6.2 | [https://download.pytorch.org/whl/rocm6.2.4](https://download.pytorch.org/whl/rocm6.2.4) |
-| ROCm 6.3 | [https://download.pytorch.org/whl/rocm6.3](https://download.pytorch.org/whl/rocm6.3) |
-| XPU      | [https://download.pytorch.org/whl/xpu](https://download.pytorch.org/whl/xpu) |
-
-- Update the below files to match the CUDA version from step 1. This makes sure that the release vLLM wheel is tested on CI.
-    - `.buildkite/release-pipeline.yaml`
-    - `.buildkite/scripts/upload-wheels.sh`
+1. Use `--extra-index-url https://download.pytorch.org/whl/cu128` to install torch+cu128.
+2. Other important indexes at the moment include:
+    1. CPU ‒ https://download.pytorch.org/whl/cpu
+    2. ROCm ‒ https://download.pytorch.org/whl/rocm6.2.4 and https://download.pytorch.org/whl/rocm6.3
+    3. XPU ‒ https://download.pytorch.org/whl/xpu
+3. Update .buildkite/release-pipeline.yaml and .buildkite/scripts/upload-wheels.sh to
+match the CUDA version from step 1.  This makes sure that the release vLLM wheel is tested
+on CI.
 
 ## Address long vLLM build time
 
@@ -86,8 +68,8 @@ and timeout. Additionally, since vLLM's fastcheck pipeline runs in read-only mod
 it doesn't populate the cache, so re-running it to warm up the cache
 is ineffective.
 
-While ongoing efforts like [#17419](gh-issue:17419)
-address the long build time at its source, the current workaround is to set `VLLM_CI_BRANCH`
+While ongoing efforts like [#17419](https://github.com/vllm-project/vllm/issues/17419)
+address the long build time at its source, the current workaround is to set VLLM_CI_BRANCH
 to a custom branch provided by @khluu (`VLLM_CI_BRANCH=khluu/use_postmerge_q`)
 when manually triggering a build on Buildkite. This branch accomplishes two things:
 
@@ -107,18 +89,17 @@ releases (which would take too much time), they can be built from
 source to unblock the update process.
 
 ### FlashInfer
-Here is how to build and install it from source with `torch2.7.0+cu128` in vLLM [Dockerfile](https://github.com/vllm-project/vllm/blob/27bebcd89792d5c4b08af7a65095759526f2f9e1/docker/Dockerfile#L259-L271):
+Here is how to build and install it from source with torch2.7.0+cu128 in vLLM [Dockerfile](https://github.com/vllm-project/vllm/blob/27bebcd89792d5c4b08af7a65095759526f2f9e1/docker/Dockerfile#L259-L271):
 
 ```bash
 export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0 10.0+PTX'
 export FLASHINFER_ENABLE_SM90=1
-uv pip install --system \
-    --no-build-isolation "git+https://github.com/flashinfer-ai/flashinfer@v0.2.6.post1"
+uv pip install --system --no-build-isolation "git+https://github.com/flashinfer-ai/flashinfer@v0.2.6.post1"
 ```
 
 One caveat is that building FlashInfer from source adds approximately 30
 minutes to the vLLM build time. Therefore, it's preferable to cache the wheel in a
-public location for immediate installation, such as [this FlashInfer wheel link](https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl). For future releases, contact the PyTorch release
+public location for immediate installation, such as https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl. For future releases, contact the PyTorch release
 team if you want to get the package published there.
 
 ### xFormers
@@ -126,15 +107,13 @@ Similar to FlashInfer, here is how to build and install xFormers from source:
 
 ```bash
 export TORCH_CUDA_ARCH_LIST='7.0 7.5 8.0 8.9 9.0 10.0+PTX'
-MAX_JOBS=16 uv pip install --system \
-    --no-build-isolation "git+https://github.com/facebookresearch/xformers@v0.0.30"
+MAX_JOBS=16 uv pip install --system --no-build-isolation "git+https://github.com/facebookresearch/xformers@v0.0.30"
 ```
 
 ### Mamba
 
 ```bash
-uv pip install --system \
-    --no-build-isolation "git+https://github.com/state-spaces/mamba@v2.2.4"
+uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@v2.2.4"
 ```
 
 ### causal-conv1d
@@ -149,6 +128,7 @@ Rather than attempting to update all vLLM platforms in a single pull request, it
 to handle some platforms separately. The separation of requirements and Dockerfiles
 for different platforms in vLLM CI/CD allows us to selectively choose
 which platforms to update. For instance, updating XPU requires the corresponding
-release from [Intel Extension for PyTorch](https://github.com/intel/intel-extension-for-pytorch) by Intel.
-While <gh-pr:16859> updated vLLM to PyTorch 2.7.0 on CPU, CUDA, and ROCm,
-<gh-pr:17444> completed the update for XPU.
+release from https://github.com/intel/intel-extension-for-pytorch by Intel.
+While https://github.com/vllm-project/vllm/pull/16859 updated vLLM to PyTorch
+2.7.0 on CPU, CUDA, and ROCm, https://github.com/vllm-project/vllm/pull/17444
+completed the update for XPU.

docs/cli/README.md

@@ -1,7 +1,3 @@
----
-toc_depth: 4
----
-
 # vLLM CLI Guide
 
 The vllm command-line tool is used to run and manage vLLM models. You can start by viewing the help message with:
@@ -20,7 +16,7 @@ vllm {chat,complete,serve,bench,collect-env,run-batch}
 
 Start the vLLM OpenAI Compatible API server.
 
-??? console "Examples"
+??? Examples
 
     ```bash
     # Start with a model
@@ -41,15 +37,8 @@ Start the vLLM OpenAI Compatible API server.
 
     # To search by keyword
     vllm serve --help=max
-
-    # To view full help with pager (less/more)
-    vllm serve --help=page
     ```
 
-### Options
-
---8<-- "docs/argparse/serve.md"
-
 ## chat
 
 Generate chat completions via the running API server.

docs/community/contact_us.md

@@ -1,3 +1,6 @@
-# Contact Us
+---
+title: Contact Us
+---
+[](){ #contactus }
 
 --8<-- "README.md:contact-us"

docs/community/meetups.md

@@ -1,4 +1,7 @@
-# Meetups
+---
+title: Meetups
+---
+[](){ #meetups }
 
 We host regular meetups in San Francisco Bay Area every 2 months. We will share the project updates from the vLLM team and have guest speakers from the industry to share their experience and insights. Please find the materials of our previous meetups below:
 

docs/configuration/conserving_memory.md

@@ -33,7 +33,7 @@ Quantized models take less memory at the cost of lower precision.
 Statically quantized models can be downloaded from HF Hub (some popular ones are available at [Red Hat AI](https://huggingface.co/RedHatAI))
 and used directly without extra configuration.
 
-Dynamic quantization is also supported via the `quantization` option -- see [here](../features/quantization/README.md) for more details.
+Dynamic quantization is also supported via the `quantization` option -- see [here][quantization-index] for more details.
 
 ## Context length and batch size
 
@@ -57,7 +57,7 @@ By default, we optimize model inference using CUDA graphs which take up extra me
 
 You can adjust `compilation_config` to achieve a better balance between inference speed and memory usage:
 
-??? code
+??? Code
 
     ```python
     from vllm import LLM
@@ -129,7 +129,7 @@ reduce the size of the processed multi-modal inputs, which in turn saves memory.
 
 Here are some examples:
 
-??? code
+??? Code
 
     ```python
     from vllm import LLM

docs/configuration/engine_args.md

@@ -1,20 +1,18 @@
 ---
-toc_depth: 3
+title: Engine Arguments
 ---
-
-# Engine Arguments
+[](){ #engine-args }
 
 Engine arguments control the behavior of the vLLM engine.
 
-- For [offline inference](../serving/offline_inference.md), they are part of the arguments to [LLM][vllm.LLM] class.
-- For [online serving](../serving/openai_compatible_server.md), they are part of the arguments to `vllm serve`.
+- For [offline inference][offline-inference], they are part of the arguments to [LLM][vllm.LLM] class.
+- For [online serving][serving-openai-compatible-server], they are part of the arguments to `vllm serve`.
 
-The engine argument classes, [EngineArgs][vllm.engine.arg_utils.EngineArgs] and [AsyncEngineArgs][vllm.engine.arg_utils.AsyncEngineArgs], are a combination of the configuration classes defined in [vllm.config][]. Therefore, if you are interested in developer documentation, we recommend looking at these configuration classes as they are the source of truth for types, defaults and docstrings.
+You can look at [EngineArgs][vllm.engine.arg_utils.EngineArgs] and [AsyncEngineArgs][vllm.engine.arg_utils.AsyncEngineArgs] to see the available engine arguments.
 
-## `EngineArgs`
+However, these classes are a combination of the configuration classes defined in [vllm.config][]. Therefore, we would recommend you read about them there where they are best documented.
 
---8<-- "docs/argparse/engine_args.md"
+For offline inference you will have access to these configuration classes and for online serving you can cross-reference the configs with `vllm serve --help`, which has its arguments grouped by config.
 
-## `AsyncEngineArgs`
-
---8<-- "docs/argparse/async_engine_args.md"
+!!! note
+    Additional arguments are available to the [AsyncLLMEngine][vllm.engine.async_llm_engine.AsyncLLMEngine] which is used for online serving. These can be found by running `vllm serve --help`

docs/configuration/env_vars.md

@@ -7,7 +7,7 @@ vLLM uses the following environment variables to configure the system:
 
     All environment variables used by vLLM are prefixed with `VLLM_`. **Special care should be taken for Kubernetes users**: please do not name the service as `vllm`, otherwise environment variables set by Kubernetes might conflict with vLLM's environment variables, because [Kubernetes sets environment variables for each service with the capitalized service name as the prefix](https://kubernetes.io/docs/concepts/services-networking/service/#environment-variables).
 
-??? code
+??? Code
 
     ```python
     --8<-- "vllm/envs.py:env-vars-definition"

docs/configuration/model_resolution.md

@@ -14,10 +14,10 @@ For example:
 ```python
 from vllm import LLM
 
-llm = LLM(
+model = LLM(
     model="cerebras/Cerebras-GPT-1.3B",
     hf_overrides={"architectures": ["GPT2LMHeadModel"]},  # GPT-2
 )
 ```
 
-Our [list of supported models](../models/supported_models.md) shows the model architectures that are recognized by vLLM.
+Our [list of supported models][supported-models] shows the model architectures that are recognized by vLLM.

docs/configuration/serve_args.md

@@ -1,16 +1,19 @@
-# Server Arguments
+---
+title: Server Arguments
+---
+[](){ #serve-args }
 
 The `vllm serve` command is used to launch the OpenAI-compatible server.
 
 ## CLI Arguments
 
 The `vllm serve` command is used to launch the OpenAI-compatible server.
-To see the available options, take a look at the [CLI Reference](../cli/README.md#options)!
+To see the available CLI arguments, run `vllm serve --help`!
 
 ## Configuration file
 
 You can load CLI arguments via a [YAML](https://yaml.org/) config file.
-The argument names must be the long form of those outlined [above](serve_args.md).
+The argument names must be the long form of those outlined [above][serve-args].
 
 For example:
 

docs/contributing/README.md

@@ -95,7 +95,7 @@ For additional features and advanced configurations, refer to the official [MkDo
 
 ## Testing
 
-??? console "Commands"
+??? note "Commands"
 
     ```bash
     pip install -r requirements/dev.txt

docs/contributing/benchmarks.md

@@ -1,4 +1,7 @@
-# Benchmark Suites
+---
+title: Benchmark Suites
+---
+[](){ #benchmarks }
 
 vLLM contains two sets of benchmarks:
 

docs/contributing/ci-failures.md

@@ -6,9 +6,9 @@ the failure?
 - Check the dashboard of current CI test failures:  
   👉 [CI Failures Dashboard](https://github.com/orgs/vllm-project/projects/20)
 
-- If your failure **is already listed**, it's likely unrelated to your PR.
-  Help fixing it is always welcome!
-    - Leave comments with links to additional instances of the failure.
+- If your failure **is already listed**, it's likely unrelated to your PR.  
+  Help fixing it is always welcome!  
+    - Leave comments with links to additional instances of the failure.  
     - React with a 👍 to signal how many are affected.
 
 - If your failure **is not listed**, you should **file an issue**.
@@ -19,25 +19,25 @@ the failure?
     👉 [New CI Failure Report](https://github.com/vllm-project/vllm/issues/new?template=450-ci-failure.yml)
 
 - **Use this title format:**
-
+  
     ```
     [CI Failure]: failing-test-job - regex/matching/failing:test
     ```
 
 - **For the environment field:**
-
+  
     ```
  Still failing on main as of commit abcdef123
     ```
 
 - **In the description, include failing tests:**
-
+  
     ```
-    FAILED failing/test.py:failing_test1 - Failure description
-    FAILED failing/test.py:failing_test2 - Failure description
-    https://github.com/orgs/vllm-project/projects/20
-    https://github.com/vllm-project/vllm/issues/new?template=400-bug-report.yml
-    FAILED failing/test.py:failing_test3 - Failure description
+    FAILED failing/test.py:failing_test1 - Failure description  
+     FAILED failing/test.py:failing_test2 - Failure description  
+    https://github.com/orgs/vllm-project/projects/20  
+    https://github.com/vllm-project/vllm/issues/new?template=400-bug-report.yml  
+    FAILED failing/test.py:failing_test3 - Failure description  
     ```
 
 - **Attach logs** (collapsible section example):
@@ -45,17 +45,17 @@ the failure?
     <summary>Logs:</summary>
 
     ```text
-    ERROR 05-20 03:26:38 [dump_input.py:68] Dumping input data
+    ERROR 05-20 03:26:38 [dump_input.py:68] Dumping input data  
     --- Logging error ---  
     Traceback (most recent call last):  
       File "/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core.py", line 203, in execute_model  
-        return self.model_executor.execute_model(scheduler_output)
+        return self.model_executor.execute_model(scheduler_output)  
     ...
-    FAILED failing/test.py:failing_test1 - Failure description
-    FAILED failing/test.py:failing_test2 - Failure description
-    FAILED failing/test.py:failing_test3 - Failure description
+    FAILED failing/test.py:failing_test1 - Failure description  
+    FAILED failing/test.py:failing_test2 - Failure description  
+    FAILED failing/test.py:failing_test3 - Failure description  
     ```
-
+  
     </details>
 
 ## Logs Wrangling
@@ -78,7 +78,7 @@ tail -525 ci_build.log | wl-copy
 
 ## Investigating a CI Test Failure
 
-1. Go to 👉 [Buildkite main branch](https://buildkite.com/vllm/ci/builds?branch=main)
+1. Go to 👉 [Buildkite main branch](https://buildkite.com/vllm/ci/builds?branch=main)  
 2. Bisect to find the first build that shows the issue.  
 3. Add your findings to the GitHub issue.  
 4. If you find a strong candidate PR, mention it in the issue and ping contributors.
@@ -97,9 +97,9 @@ CI test failures may be flaky. Use a bash loop to run repeatedly:
 
 If you submit a PR to fix a CI failure:
 
-- Link the PR to the issue:
+- Link the PR to the issue:  
   Add `Closes #12345` to the PR description.
-- Add the `ci-failure` label:
+- Add the `ci-failure` label:  
   This helps track it in the [CI Failures GitHub Project](https://github.com/orgs/vllm-project/projects/20).
 
 ## Other Resources

docs/contributing/dockerfile/dockerfile.md

@@ -1,7 +1,7 @@
 # Dockerfile
 
 We provide a <gh-file:docker/Dockerfile> to construct the image for running an OpenAI compatible server with vLLM.
-More information about deploying with Docker can be found [here](../../deployment/docker.md).
+More information about deploying with Docker can be found [here][deployment-docker].
 
 Below is a visual representation of the multi-stage Dockerfile. The build graph contains the following nodes:
 

docs/contributing/incremental_build.md

@@ -84,7 +84,6 @@ Below is an example of what the generated `CMakeUserPresets.json` might look lik
 ```
 
 **What do the various configurations mean?**
-
 - `CMAKE_CUDA_COMPILER`: Path to your `nvcc` binary. The script attempts to find this automatically.
 - `CMAKE_C_COMPILER_LAUNCHER`, `CMAKE_CXX_COMPILER_LAUNCHER`, `CMAKE_CUDA_COMPILER_LAUNCHER`: Setting these to `ccache` (or `sccache`) significantly speeds up rebuilds by caching compilation results. Ensure `ccache` is installed (e.g., `sudo apt install ccache` or `conda install ccache`). The script sets these by default.
 - `VLLM_PYTHON_EXECUTABLE`: Path to the Python executable in your vLLM development environment. The script will prompt for this, defaulting to the current Python environment if suitable.
@@ -99,16 +98,16 @@ Once your `CMakeUserPresets.json` is configured:
 1. **Initialize the CMake build environment:**
    This step configures the build system according to your chosen preset (e.g., `release`) and creates the build directory at `binaryDir`
 
-    ```console
-    cmake --preset release
-    ```
+   ```console
+   cmake --preset release
+   ```
 
 2. **Build and install the vLLM components:**
    This command compiles the code and installs the resulting binaries into your vLLM source directory, making them available to your editable Python installation.
 
-    ```console
-    cmake --build --preset release --target install
-    ```
+   ```console
+   cmake --build --preset release --target install
+   ```
 
 3. **Make changes and repeat!**
     Now you start using your editable install of vLLM, testing and making changes as needed. If you need to build again to update based on changes, simply run the CMake command again to build only the affected files.

docs/contributing/model/README.md

@@ -1,9 +1,12 @@
-# Summary
+---
+title: Summary
+---
+[](){ #new-model }
 
 !!! important
     Many decoder language models can now be automatically loaded using the [Transformers backend][transformers-backend] without having to implement them in vLLM. See if `vllm serve <model>` works first!
 
-vLLM models are specialized [PyTorch](https://pytorch.org/) models that take advantage of various [features](../../features/compatibility_matrix.md) to optimize their performance.
+vLLM models are specialized [PyTorch](https://pytorch.org/) models that take advantage of various [features][compatibility-matrix] to optimize their performance.
 
 The complexity of integrating a model into vLLM depends heavily on the model's architecture.
 The process is considerably straightforward if the model shares a similar architecture with an existing model in vLLM.

docs/contributing/model/basic.md

@@ -1,4 +1,7 @@
-# Basic Model
+---
+title: Basic Model
+---
+[](){ #new-model-basic }
 
 This guide walks you through the steps to implement a basic vLLM model.
 
@@ -24,7 +27,7 @@ All vLLM modules within the model must include a `prefix` argument in their cons
 
 The initialization code should look like this:
 
-??? code
+??? Code
 
     ```python
     from torch import nn
@@ -73,8 +76,6 @@ def forward(
     self,
     input_ids: torch.Tensor,
     positions: torch.Tensor,
-    intermediate_tensors: Optional[IntermediateTensors] = None,
-    inputs_embeds: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
     ...
 ```
@@ -107,7 +108,7 @@ This method should load the weights from the HuggingFace's checkpoint file and a
 
 ## 5. Register your model
 
-See [this page](registration.md) for instructions on how to register your new model to be used by vLLM.
+See [this page][new-model-registration] for instructions on how to register your new model to be used by vLLM.
 
 ## Frequently Asked Questions
 

docs/contributing/model/multimodal.md

@@ -1,15 +1,18 @@
-# Multi-Modal Support
+---
+title: Multi-Modal Support
+---
+[](){ #supports-multimodal }
 
-This document walks you through the steps to extend a basic model so that it accepts [multi-modal inputs](../../features/multimodal_inputs.md).
+This document walks you through the steps to extend a basic model so that it accepts [multi-modal inputs][multimodal-inputs].
 
 ## 1. Update the base vLLM model
 
-It is assumed that you have already implemented the model in vLLM according to [these steps](basic.md).
+It is assumed that you have already implemented the model in vLLM according to [these steps][new-model-basic].
 Further update the model as follows:
 
 - Implement [get_placeholder_str][vllm.model_executor.models.interfaces.SupportsMultiModal.get_placeholder_str] to define the placeholder string which is used to represent the multi-modal item in the text prompt. This should be consistent with the chat template of the model.
 
-    ??? code
+    ??? Code
 
         ```python
         class YourModelForImage2Seq(nn.Module):
@@ -38,7 +41,7 @@ Further update the model as follows:
 
 - Implement [get_multimodal_embeddings][vllm.model_executor.models.interfaces.SupportsMultiModal.get_multimodal_embeddings] that returns the embeddings from running the multimodal inputs through the multimodal tokenizer of the model. Below we provide a boilerplate of a typical implementation pattern, but feel free to adjust it to your own needs.
 
-    ??? code
+    ??? Code
 
         ```python
         class YourModelForImage2Seq(nn.Module):
@@ -68,7 +71,7 @@ Further update the model as follows:
 
 - Implement [get_input_embeddings][vllm.model_executor.models.interfaces.SupportsMultiModal.get_input_embeddings] to merge `multimodal_embeddings` with text embeddings from the `input_ids`. If input processing for the model is implemented correctly (see sections below), then you can leverage the utility function we provide to easily merge the embeddings.
 
-    ??? code
+    ??? Code
 
         ```python
         from .utils import merge_multimodal_embeddings
@@ -152,7 +155,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     Looking at the code of HF's `LlavaForConditionalGeneration`:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L530-L544
@@ -176,7 +179,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
     The number of placeholder feature tokens per image is `image_features.shape[1]`.
     `image_features` is calculated inside the `get_image_features` method:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L290-L300
@@ -214,7 +217,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     To find the sequence length, we turn to the code of `CLIPVisionEmbeddings`:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L247-L257
@@ -241,7 +244,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     Overall, the number of placeholder feature tokens for an image can be calculated as:
 
-    ??? code
+    ??? Code
 
         ```python
         def get_num_image_tokens(
@@ -266,7 +269,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
     Notice that the number of image tokens doesn't depend on the image width and height.
     We can simply use a dummy `image_size` to calculate the multimodal profiling data:
 
-    ??? code
+    ??? Code
 
         ```python
         # NOTE: In actuality, this is usually implemented as part of the
@@ -311,7 +314,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     Looking at the code of HF's `FuyuForCausalLM`:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/modeling_fuyu.py#L311-L322
@@ -341,7 +344,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
     In `FuyuImageProcessor.preprocess`, the images are resized and padded to the target `FuyuImageProcessor.size`,
     returning the dimensions after resizing (but before padding) as metadata.
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/processing_fuyu.py#L541-L544
@@ -379,7 +382,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     In `FuyuImageProcessor.preprocess_with_tokenizer_info`, the images are split into patches based on this metadata:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/processing_fuyu.py#L417-L425
@@ -417,7 +420,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     The number of patches is in turn defined by `FuyuImageProcessor.get_num_patches`:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L552-L562
@@ -454,7 +457,7 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
 
     For the multimodal image profiling data, the logic is very similar to LLaVA:
 
-    ??? code
+    ??? Code
 
         ```python
         def get_dummy_mm_data(
@@ -480,7 +483,7 @@ Afterwards, create a subclass of [BaseMultiModalProcessor][vllm.multimodal.proce
 to fill in the missing details about HF processing.
 
 !!! info
-    [Multi-Modal Data Processing](../../design/mm_processing.md)
+    [Multi-Modal Data Processing][mm-processing]
 
 ### Multi-modal fields
 
@@ -543,7 +546,7 @@ return a schema of the tensors outputted by the HF processor that are related to
     In order to support the use of [MultiModalFieldConfig.batched][] like in LLaVA,
     we remove the extra batch dimension by overriding [BaseMultiModalProcessor._call_hf_processor][]:
 
-    ??? code
+    ??? Code
 
         ```python
         def _call_hf_processor(
@@ -620,7 +623,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
     It simply repeats each input `image_token` a number of times equal to the number of placeholder feature tokens (`num_image_tokens`).
     Based on this, we override [_get_prompt_updates][vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates] as follows:
 
-    ??? code
+    ??? Code
 
         ```python
         def _get_prompt_updates(
@@ -665,7 +668,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
 
     We define a helper function to return `ncols` and `nrows` directly:
 
-    ??? code
+    ??? Code
 
         ```python
         def get_image_feature_grid_size(
@@ -695,7 +698,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
 
     Based on this, we can initially define our replacement tokens as:
 
-    ??? code
+    ??? Code
 
         ```python
         def get_replacement(item_idx: int):
@@ -715,7 +718,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
     However, this is not entirely correct. After `FuyuImageProcessor.preprocess_with_tokenizer_info` is called,
     a BOS token (`<s>`) is also added to the promopt:
 
-    ??? code
+    ??? Code
 
         ```python
         # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/processing_fuyu.py#L417-L435
@@ -742,7 +745,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
     To assign the vision embeddings to only the image tokens, instead of a string
     you can return an instance of [PromptUpdateDetails][vllm.multimodal.processing.PromptUpdateDetails]:
 
-    ??? code
+    ??? Code
 
         ```python
         hf_config = self.info.get_hf_config()
@@ -769,7 +772,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
     Finally, noticing that the HF processor removes the `|ENDOFTEXT|` token from the tokenized prompt,
     we can search for it to conduct the replacement at the start of the string:
 
-    ??? code
+    ??? Code
 
         ```python
         def _get_prompt_updates(
@@ -816,7 +819,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
 After you have defined [BaseProcessingInfo][vllm.multimodal.processing.BaseProcessingInfo] (Step 2),
 [BaseDummyInputsBuilder][vllm.multimodal.profiling.BaseDummyInputsBuilder] (Step 3),
 and [BaseMultiModalProcessor][vllm.multimodal.processing.BaseMultiModalProcessor] (Step 4),
-decorate the model class with [MULTIMODAL_REGISTRY.register_processor][vllm.multimodal.processing.MultiModalRegistry.register_processor]
+decorate the model class with {meth}`MULTIMODAL_REGISTRY.register_processor <vllm.multimodal.registry.MultiModalRegistry.register_processor>`
 to register them to the multi-modal registry:
 
 ```diff
@@ -843,7 +846,7 @@ Examples:
 
 ### Handling prompt updates unrelated to multi-modal data
 
-[_get_prompt_updates][vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates] assumes that each application of prompt update corresponds to one multi-modal item. If the HF processor performs additional processing regardless of how many multi-modal items there are, you should override [_apply_hf_processor_tokens_only][vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_tokens_only] so that the processed token inputs are consistent with the result of applying the HF processor on text inputs. This is because token inputs bypass the HF processor according to [our design](../../design/mm_processing.md).
+[_get_prompt_updates][vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates] assumes that each application of prompt update corresponds to one multi-modal item. If the HF processor performs additional processing regardless of how many multi-modal items there are, you should override [_apply_hf_processor_tokens_only][vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_tokens_only] so that the processed token inputs are consistent with the result of applying the HF processor on text inputs. This is because token inputs bypass the HF processor according to [our design][mm-processing].
 
 Examples:
 

docs/contributing/model/registration.md

@@ -1,7 +1,10 @@
-# Registering a Model
+---
+title: Registering a Model
+---
+[](){ #new-model-registration }
 
 vLLM relies on a model registry to determine how to run each model.
-A list of pre-registered architectures can be found [here](../../models/supported_models.md).
+A list of pre-registered architectures can be found [here][supported-models].
 
 If your model is not on this list, you must register it to vLLM.
 This page provides detailed instructions on how to do so.
@@ -11,16 +14,16 @@ This page provides detailed instructions on how to do so.
 To add a model directly to the vLLM library, start by forking our [GitHub repository](https://github.com/vllm-project/vllm) and then [build it from source][build-from-source].
 This gives you the ability to modify the codebase and test your model.
 
-After you have implemented your model (see [tutorial](basic.md)), put it into the <gh-dir:vllm/model_executor/models> directory.
+After you have implemented your model (see [tutorial][new-model-basic]), put it into the <gh-dir:vllm/model_executor/models> directory.
 Then, add your model class to `_VLLM_MODELS` in <gh-file:vllm/model_executor/models/registry.py> so that it is automatically registered upon importing vLLM.
-Finally, update our [list of supported models](../../models/supported_models.md) to promote your model!
+Finally, update our [list of supported models][supported-models] to promote your model!
 
 !!! important
     The list of models in each section should be maintained in alphabetical order.
 
 ## Out-of-tree models
 
-You can load an external model [using a plugin](../../design/plugin_system.md) without modifying the vLLM codebase.
+You can load an external model [using a plugin][plugin-system] without modifying the vLLM codebase.
 
 To register the model, use the following code:
 
@@ -48,4 +51,4 @@ def register():
 
 !!! important
     If your model is a multimodal model, ensure the model class implements the [SupportsMultiModal][vllm.model_executor.models.interfaces.SupportsMultiModal] interface.
-    Read more about that [here](multimodal.md).
+    Read more about that [here][supports-multimodal].

docs/contributing/model/tests.md

@@ -1,4 +1,7 @@
-# Unit Testing
+---
+title: Unit Testing
+---
+[](){ #new-model-tests }
 
 This page explains how to write unit tests to verify the implementation of your model.