Reduce Scatter Plumbing

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

.buildkite/nightly-benchmarks/README.md

@@ -1,15 +1,13 @@
 # vLLM benchmark suite
 
-
 ## Introduction
 
 This directory contains two sets of benchmark for vllm.
+
 - Performance benchmark: benchmark vllm's performance under various workload, for **developers** to gain clarity on whether their PR improves/degrades vllm's performance
 - Nightly benchmark: compare vllm's performance against alternatives (tgi, trt-llm and lmdeploy), for **the public** to know when to choose vllm.
 
-
-See  [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
-
+See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
 
 ## Performance benchmark quick overview
 
@@ -19,17 +17,14 @@ See  [vLLM performance dashboard](https://perf.vllm.ai) for the latest performan
 
 **For benchmarking developers**: please try your best to constraint the duration of benchmarking to about 1 hr so that it won't take forever to run.
 
-
 ## Nightly benchmark quick overview
 
-**Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B. 
+**Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B.
 
 **Benchmarking engines**: vllm, TGI, trt-llm and lmdeploy.
 
 **Benchmarking Duration**: about 3.5hrs.
 
-
-
 ## Trigger the benchmark
 
 Performance benchmark will be triggered when:
@@ -39,16 +34,11 @@ Performance benchmark will be triggered when:
 Nightly benchmark will be triggered when:
 - Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label.
 
-
-
-
 ## Performance benchmark details
 
-
 See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
 
-
-#### Latency test
+### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
 
@@ -68,23 +58,25 @@ Here is an example of one test inside `latency-tests.json`:
 ```
 
 In this example:
--  The `test_name` attributes is a unique identifier for the test. In `latency-tests.json`, it must start with `latency_`.
--  The `parameters` attribute control the command line arguments to be used for `benchmark_latency.py`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-performance-benchmarks.sh` will convert the underline to dash when feeding the arguments to `benchmark_latency.py`. For example, the corresponding command line arguments for `benchmark_latency.py` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15`
+
+- The `test_name` attributes is a unique identifier for the test. In `latency-tests.json`, it must start with `latency_`.
+- The `parameters` attribute control the command line arguments to be used for `benchmark_latency.py`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-performance-benchmarks.sh` will convert the underline to dash when feeding the arguments to `benchmark_latency.py`. For example, the corresponding command line arguments for `benchmark_latency.py` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15`
 
 Note that the performance numbers are highly sensitive to the value of the parameters. Please make sure the parameters are set correctly.
 
 WARNING: The benchmarking script will save json results by itself, so please do not configure `--output-json` parameter in the json file.
 
+### Throughput test
 
-#### Throughput test
 The tests are specified in `throughput-tests.json`. The syntax is similar to `latency-tests.json`, except for that the parameters will be fed forward to `benchmark_throughput.py`.
 
 The number of this test is also stable -- a slight change on the value of this number might vary the performance numbers by a lot.
 
-#### Serving test
+### Serving test
+
 We test the throughput by using `benchmark_serving.py` with request rate = inf to cover the online serving overhead. The corresponding parameters are in `serving-tests.json`, and here is an example:
 
-```
+```json
 [
     {
         "test_name": "serving_llama8B_tp1_sharegpt",
@@ -109,6 +101,7 @@ We test the throughput by using `benchmark_serving.py` with request rate = inf t
 ```
 
 Inside this example:
+
 - The `test_name` attribute is also a unique identifier for the test. It must start with `serving_`.
 - The `server-parameters` includes the command line arguments for vLLM server.
 - The `client-parameters` includes the command line arguments for `benchmark_serving.py`.
@@ -118,36 +111,33 @@ The number of this test is less stable compared to the delay and latency benchma
 
 WARNING: The benchmarking script will save json results by itself, so please do not configure `--save-results` or other results-saving-related parameters in `serving-tests.json`.
 
-#### Visualizing the results
+### Visualizing the results
+
 The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results.
 You can find the result presented as a table inside the `buildkite/performance-benchmark` job page.
 If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.
 The raw benchmarking results (in the format of json files) are in the `Artifacts` tab of the benchmarking.
 
-
-
 ## Nightly test details
 
 See [nightly-descriptions.md](nightly-descriptions.md) for the detailed description on test workload, models and docker containers of benchmarking other llm engines.
 
+### Workflow
 
-#### Workflow
-
-- The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines. 
+- The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines.
 - Inside each container, we run [run-nightly-suite.sh](run-nightly-suite.sh), which will probe the serving engine of the current container.
 - The `run-nightly-suite.sh` will redirect the request to `tests/run-[llm serving engine name]-nightly.sh`, which parses the workload described in [nightly-tests.json](tests/nightly-tests.json) and performs the benchmark.
 - At last, we run [scripts/plot-nightly-results.py](scripts/plot-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite.
 
-#### Nightly tests
+### Nightly tests
 
 In [nightly-tests.json](tests/nightly-tests.json), we include the command line arguments for benchmarking commands, together with the benchmarking test cases. The format is highly similar to performance benchmark.
 
-#### Docker containers
+### Docker containers
 
 The docker containers for benchmarking are specified in `nightly-pipeline.yaml`.
 
 WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `tests/run-[llm serving engine name]-nightly.sh`.
 
 WARNING: populating `trt-llm` to latest version is not easy, as it requires updating several protobuf files in [tensorrt-demo](https://github.com/neuralmagic/tensorrt-demo.git).
-

.buildkite/nightly-benchmarks/benchmark-pipeline.yaml

@@ -10,12 +10,18 @@ steps:
           - image: badouralix/curl-jq
             command:
             - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
-
+  - label: "Cleanup H100"
+    agents:
+      queue: H100
+    depends_on: ~
+    command: docker system prune -a --volumes --force
+  
   - label: "A100"
     # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
     agents:
       queue: A100
     depends_on: wait-for-container-image
+    if: build.branch == "main"
     plugins:
     - kubernetes:
         podSpec:
@@ -50,6 +56,7 @@ steps:
     agents:
       queue: H200
     depends_on: wait-for-container-image
+    if: build.branch == "main"
     plugins:
     - docker#v5.12.0:
         image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
@@ -75,6 +82,7 @@ steps:
     agents:
       queue: H100
     depends_on: wait-for-container-image
+    if: build.branch == "main"
     plugins:
     - docker#v5.12.0:
         image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
@@ -90,3 +98,87 @@ steps:
         environment:
         - VLLM_USAGE_SOURCE
         - HF_TOKEN
+
+  # Premerge benchmark
+  - label: "A100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: A100
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - kubernetes:
+        podSpec:
+          priorityClassName: perf-benchmark
+          containers:
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+            command:
+            - bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+            env:
+            - name: VLLM_USAGE_SOURCE
+              value: ci-test
+            - name: HF_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: hf-token-secret
+                  key: token
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+
+  - label: "H200"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H200
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: 4,5,6,7
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
+
+  #- block: "Run H100 Benchmark"
+    #key: block-h100
+    #depends_on: ~
+
+  - label: "H100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H100
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: all # see CUDA_VISIBLE_DEVICES for actual GPUs used
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN

.buildkite/nightly-benchmarks/nightly-annotation.md

@@ -9,20 +9,19 @@ This file contains the downloading link for benchmarking results.
 
 Please download the visualization scripts in the post
 
-
 ## Results reproduction
 
 - Find the docker we use in `benchmarking pipeline`
 - Deploy the docker, and inside the docker:
-  - Download `nightly-benchmarks.zip`. 
-  - In the same folder, run the following code
-```
-export HF_TOKEN=<your HF token>
-apt update
-apt install -y git
-unzip nightly-benchmarks.zip
-VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
-```
+  - Download `nightly-benchmarks.zip`.
+  - In the same folder, run the following code:
+
+  ```console
+  export HF_TOKEN=<your HF token>
+  apt update
+  apt install -y git
+  unzip nightly-benchmarks.zip
+  VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+  ```
 
 And the results will be inside `./benchmarks/results`.
-

.buildkite/nightly-benchmarks/nightly-descriptions.md

@@ -2,6 +2,7 @@
 # Nightly benchmark
 
 This benchmark aims to:
+
 - Provide performance clarity: Provide clarity on which one (vllm, tensorrt-llm, lmdeploy and SGLang) leads in performance in what workload.
 - Be reproducible: one can run the exact same set of benchmarking commands inside the exact same docker by following reproducing instructions.
 
@@ -9,7 +10,6 @@ Latest results: [results link](https://blog.vllm.ai/2024/09/05/perf-update.html)
 
 Latest reproduction guilde: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
 
-
 ## Setup
 
 - Docker images:
@@ -33,7 +33,7 @@ Latest reproduction guilde: [github issue link](https://github.com/vllm-project/
     - Queries are randomly sampled, and arrival patterns are determined via Poisson process, but all with fixed random seed.
   - Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).
 
-# Known issues
+## Known issues
 
 - TRT-LLM crashes with Llama 3.1 8B [issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105).
-- TGI does not support `ignore-eos` flag.
+- TGI does not support `ignore-eos` flag.

.buildkite/nightly-benchmarks/performance-benchmarks-descriptions.md

@@ -7,10 +7,8 @@
 - Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - Evaluation metrics: end-to-end latency (mean, median, p99).
 
-
 {latency_tests_markdown_table}
 
-
 ## Throughput tests
 
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
@@ -19,10 +17,8 @@
 - Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - Evaluation metrics: throughput.
 
-
 {throughput_tests_markdown_table}
 
-
 ## Serving tests
 
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
@@ -33,13 +29,11 @@
 - We also added a speculative decoding test for llama-3 70B, under QPS 2
 - Evaluation metrics: throughput, TTFT (time to the first token, with mean, median and p99), ITL (inter-token latency, with mean, median and p99).
 
-
 {serving_tests_markdown_table}
 
-
 ## json version of the benchmarking tables
 
-This section contains the data of the markdown tables above in JSON format. 
+This section contains the data of the markdown tables above in JSON format.
 You can load the benchmarking tables into pandas dataframes as follows:
 
 ```python
@@ -54,9 +48,9 @@ serving_results = pd.DataFrame.from_dict(benchmarking_results["serving"])
 ```
 
 The json string for all benchmarking tables:
+
 ```json
 {benchmarking_results_in_json_string}
 ```
 
 You can also check the raw experiment data in the Artifact tab of the Buildkite page.
-

.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -84,8 +84,13 @@ if __name__ == "__main__":
             # this result is generated via `benchmark_serving.py`
 
             # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
             raw_result.update(command)
 
             # update the test name of this result
@@ -99,8 +104,13 @@ if __name__ == "__main__":
             # this result is generated via `benchmark_latency.py`
 
             # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
             raw_result.update(command)
 
             # update the test name of this result
@@ -121,8 +131,13 @@ if __name__ == "__main__":
             # this result is generated via `benchmark_throughput.py`
 
             # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
             raw_result.update(command)
 
             # update the test name of this result

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

@@ -309,11 +309,14 @@ run_serving_tests() {
 
       new_test_name=$test_name"_qps_"$qps
 
+      # pass the tensor parallel size to the client so that it can be displayed
+      # on the benchmark dashboard
       client_command="python3 benchmark_serving.py \
         --save-result \
         --result-dir $RESULTS_FOLDER \
         --result-filename ${new_test_name}.json \
         --request-rate $qps \
+        --metadata "tensor_parallel_size=$tp" \
         $client_args"
 
       echo "Running test case $test_name with qps $qps"
@@ -345,6 +348,11 @@ main() {
   check_gpus
   check_hf_token
 
+  # Set to v1 to run v1 benchmark
+  if [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
+    export VLLM_USE_V1=1
+  fi
+
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
   (which jq) || (apt-get update && apt-get -y install jq)

.buildkite/nightly-benchmarks/scripts/wait-for-image.sh

@@ -1,6 +1,10 @@
 #!/bin/sh
 TOKEN=$(curl -s -L "https://public.ecr.aws/token?service=public.ecr.aws&scope=repository:q9t5s3a7/vllm-ci-postmerge-repo:pull" | jq -r .token)
-URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-postmerge-repo/manifests/$BUILDKITE_COMMIT"
+if [[ "$BUILDKITE_BRANCH" == "main" ]]; then
+    URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-postmerge-repo/manifests/$BUILDKITE_COMMIT"
+else
+    URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-test-repo/manifests/$BUILDKITE_COMMIT"
+fi
 
 TIMEOUT_SECONDS=10
 

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

@@ -29,4 +29,4 @@
             "num-iters": 15
         }
     }
-]
+]

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

@@ -66,8 +66,7 @@
             "swap_space": 16, 
             "speculative_model": "turboderp/Qwama-0.5B-Instruct",
             "num_speculative_tokens": 4,
-            "speculative_draft_tensor_parallel_size": 1,
-            "use_v2_block_manager": ""
+            "speculative_draft_tensor_parallel_size": 1
         },
         "client_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",

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

@@ -32,4 +32,4 @@
             "backend": "vllm"
         }
     }
-]
+]

.buildkite/release-pipeline.yaml

@@ -1,4 +1,15 @@
 steps:
+  - label: "Build wheel - CUDA 12.4"
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "mkdir artifacts"
+      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+      - "bash .buildkite/upload-wheels.sh"
+    env:
+      DOCKER_BUILDKIT: "1"
+
   - label: "Build wheel - CUDA 12.1"
     agents:
       queue: cpu_queue_postmerge
@@ -37,7 +48,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
   - label: "Build and publish TPU release image"

.buildkite/run-amd-test.sh

@@ -92,7 +92,9 @@ if [[ $commands == *" kernels "* ]]; then
   --ignore=kernels/test_moe.py \
   --ignore=kernels/test_prefix_prefill.py \
   --ignore=kernels/test_rand.py \
-  --ignore=kernels/test_sampler.py"
+  --ignore=kernels/test_sampler.py \
+  --ignore=kernels/test_cascade_flash_attn.py \
+  --ignore=kernels/test_mamba_mixer2.py"
 fi
 
 #ignore certain Entrypoints tests
@@ -121,6 +123,8 @@ if [[ $commands == *"--shard-id="* ]]; then
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
         -e HF_TOKEN \
+        -e AWS_ACCESS_KEY_ID \
+        -e AWS_SECRET_ACCESS_KEY \
         -v "${HF_CACHE}:${HF_MOUNT}" \
         -e "HF_HOME=${HF_MOUNT}" \
         --name "${container_name}_${GPU}" \
@@ -148,6 +152,8 @@ else
           --rm \
           -e HIP_VISIBLE_DEVICES=0 \
           -e HF_TOKEN \
+          -e AWS_ACCESS_KEY_ID \
+          -e AWS_SECRET_ACCESS_KEY \
           -v "${HF_CACHE}:${HF_MOUNT}" \
           -e "HF_HOME=${HF_MOUNT}" \
           --name "${container_name}" \

.buildkite/run-cpu-test.sh

@@ -30,7 +30,7 @@ function cpu_tests() {
   # offline inference
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
     set -e
-    python3 examples/offline_inference/basic.py"
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "

.buildkite/run-gh200-test.sh

@@ -24,5 +24,5 @@ remove_docker_container
 
 # Run the image and test offline inference
 docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
-    python3 examples/offline_inference/cli.py --model meta-llama/Llama-3.2-1B
+    python3 examples/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
 '

.buildkite/run-hpu-test.sh

@@ -20,5 +20,5 @@ trap remove_docker_container_and_exit EXIT
 remove_docker_container
 
 # Run the image and launch offline inference
-docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic.py
+docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
 EXITCODE=$?

.buildkite/run-neuron-test.sh

@@ -29,9 +29,6 @@ if [ -f /tmp/neuron-docker-build-timestamp ]; then
         docker image prune -f
         # Remove unused volumes / force the system prune for old images as well.
         docker volume prune -f && docker system prune -f
-        # Remove huggingface model artifacts and compiler cache
-        rm -rf "${HF_MOUNT:?}/*"
-        rm -rf "${NEURON_COMPILE_CACHE_MOUNT:?}/*"
         echo "$current_time" > /tmp/neuron-docker-build-timestamp
     fi
 else

.buildkite/run-openvino-test.sh

@@ -13,4 +13,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic.py
+docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic/generate.py --model facebook/opt-125m

.buildkite/run-xpu-test.sh

@@ -14,6 +14,6 @@ remove_docker_container
 
 # Run the image and test offline inference/tensor parallel
 docker run --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test sh -c '
-    python3 examples/offline_inference/basic.py
-    python3 examples/offline_inference/cli.py -tp 2
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
 '

.buildkite/test-pipeline.yaml

@@ -2,7 +2,7 @@
 # adding a new command to an existing step. See different options here for examples.
 
 # This script will be feed into Jinja template in `test-template-aws.j2` at
-# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2 
+# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2
 # to generate the final pipeline yaml file.
 
 # Documentation
@@ -15,7 +15,7 @@
 # mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
 # gpu(str): override the GPU selection for the test. default is on L4 GPUs. currently only supports a100
 # num_gpus(int): override the number of GPUs for the test. default to 1 GPU. currently support 2,4.
-# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host, 
+# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host,
 #     in this case, commands must be specified. the first command runs on first host, the second
 #     command runs on the second host.
 # working_dir(str): specify the place where command should execute, default to /vllm-workspace/tests
@@ -24,8 +24,8 @@
 # When adding a test
 # - If the test belong to an existing group, add it there
 # - If the test is short, add to any existing step
-# - If the test takes more than 10min, then it is okay to create a new step. 
-#   Note that all steps execute in parallel. 
+# - If the test takes more than 10min, then it is okay to create a new step.
+#   Note that all steps execute in parallel.
 
 steps:
 ##### fast check tests  #####
@@ -107,13 +107,17 @@ steps:
   mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
+  - tests/entrypoints/llm
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  - tests/entrypoints/offline_mode
   commands:
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - 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
   - pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
@@ -124,12 +128,15 @@ steps:
   source_file_dependencies:
   - vllm/distributed/
   - vllm/core/
-  - tests/distributed
+  - tests/distributed/test_utils
+  - tests/distributed/test_pynccl
   - tests/spec_decode/e2e/test_integration_dist_tp4
-  - tests/compile
+  - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
-  - examples/offline_inference/ray_placement.py
+  - examples/offline_inference/rlhf_colocate.py
+  - tests/examples/offline_inference/data_parallel.py
   commands:
+  - VLLM_USE_V1=1 python3 ../examples/offline_inference/data_parallel.py
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
@@ -137,17 +144,17 @@ steps:
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - python3 ../examples/offline_inference/rlhf.py
-  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/ray_placement.py
+  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/rlhf_colocate.py
 
 - label: Metrics, Tracing Test # 10min
-  num_gpus: 2 
+  num_gpus: 2
   fast_check: true
   source_file_dependencies:
   - vllm/
   - tests/metrics
   - tests/tracing
   commands:
-  - pytest -v -s metrics 
+  - pytest -v -s metrics
   - "pip install \
       'opentelemetry-sdk>=1.26.0,<1.27.0' \
       'opentelemetry-api>=1.26.0,<1.27.0' \
@@ -174,6 +181,9 @@ steps:
   - vllm/
   - tests/engine
   - tests/tokenization
+  - tests/test_sequence
+  - tests/test_config
+  - tests/test_logger
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py
   # OOM in the CI unless we run this separately
@@ -195,6 +205,9 @@ steps:
     # TODO: accuracy does not match, whether setting
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
     - VLLM_USE_V1=1 pytest -v -s v1/e2e
+    # Integration test for streaming correctness (requires special branch).
+    - 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
   working_dir: "/vllm-workspace/examples"
@@ -204,18 +217,18 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-    - python3 offline_inference/basic.py
-    - python3 offline_inference/cpu_offload.py
-    - python3 offline_inference/chat.py
+    - python3 offline_inference/basic/generate.py --model facebook/opt-125m
+    - python3 offline_inference/basic/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
+    - python3 offline_inference/basic/chat.py
     - python3 offline_inference/prefix_caching.py
     - python3 offline_inference/llm_engine_example.py
     - python3 offline_inference/vision_language.py
     - python3 offline_inference/vision_language_multi_image.py
     - python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/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
-    - python3 offline_inference/classification.py
-    - python3 offline_inference/embedding.py
-    - python3 offline_inference/scoring.py
+    - python3 offline_inference/basic/classify.py
+    - python3 offline_inference/basic/embed.py
+    - python3 offline_inference/basic/score.py
     - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
@@ -243,7 +256,7 @@ steps:
   - vllm/model_executor/guided_decoding
   - tests/test_logits_processor
   - tests/model_executor/test_guided_processors
-  commands: 
+  commands:
     - pytest -v -s test_logits_processor.py
     - pytest -v -s model_executor/test_guided_processors.py
 
@@ -254,7 +267,7 @@ steps:
   - 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
+    - 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
@@ -265,7 +278,7 @@ steps:
   command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py
   parallelism: 4
 
-- label: "PyTorch Fullgraph Smoke Test" # 9min
+- label: PyTorch Fullgraph Smoke Test # 9min
   fast_check: true
   source_file_dependencies:
   - vllm/
@@ -276,7 +289,7 @@ steps:
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
 
-- label: "PyTorch Fullgraph Test" # 18min
+- label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -328,6 +341,14 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - bash ./run-tests.sh -c configs/models-small.txt -t 1
 
+- label: OpenAI API correctness
+  source_file_dependencies:
+  - csrc/
+  - vllm/entrypoints/openai/
+  - vllm/model_executor/models/whisper.py
+  commands: # LMEval+Transcription WER check
+  - pytest -s entrypoints/openai/correctness/
+
 - label: Encoder Decoder tests # 5min
   source_file_dependencies:
   - vllm/
@@ -482,6 +503,7 @@ steps:
   - entrypoints/llm/test_collective_rpc.py
   commands:
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - VLLM_USE_V1=1 torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
@@ -512,6 +534,7 @@ steps:
   - pip uninstall vllm_add_dummy_platform -y
   # end platform plugin tests
   # other tests continue here:
+  - pytest -v -s plugins_tests/test_scheduler_plugins.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
   - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
@@ -561,7 +584,7 @@ steps:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # This test runs llama 13B, so it is required to run on 4 GPUs.
     - pytest -v -s -x lora/test_long_context.py
-    # There is some Tensor Parallelism related processing logic in LoRA that 
+    # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
@@ -586,7 +609,7 @@ steps:
   - vllm/
   - tests/weight_loading
   commands:
-    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt 
+    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
 
 
 ##### multi gpus test #####
@@ -598,7 +621,7 @@ steps:
   num_gpus: 4
   source_file_dependencies:
   - vllm/
-  commands: 
+  commands:
   # NOTE: don't test llama model here, it seems hf implementation is buggy
   # see https://github.com/vllm-project/vllm/pull/5689 for details
   - pytest -v -s distributed/test_custom_all_reduce.py

.buildkite/upload-wheels.sh

@@ -50,8 +50,11 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
+elif [[ $normal_wheel == *"cu121"* ]]; then
+    # if $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
 else
-    # only upload index.html for cu12 wheels (default wheels)
+    # only upload index.html for cu124 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
     aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
@@ -63,8 +66,11 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
+elif [[ $normal_wheel == *"cu121"* ]]; then
+    # if $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
 else
-    # only upload index.html for cu12 wheels (default wheels)
+    # only upload index.html for cu124 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi
 

.github/PULL_REQUEST_TEMPLATE.md

@@ -2,4 +2,5 @@ FILL IN THE PR DESCRIPTION HERE
 
 FIX #xxxx (*link existing issues this PR will resolve*)
 
-**BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html **
+<!--- pyml disable-next-line no-emphasis-as-heading -->
+**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>**

.github/mergify.yml

@@ -5,6 +5,7 @@ pull_request_rules:
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
+      - files~=^examples/
   actions:
     label:
       add:

.github/workflows/cleanup_pr_body.yml

@@ -16,7 +16,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: Set up Python
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+        uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
         with:
           python-version: '3.12'
 

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

@@ -12,17 +12,17 @@ jobs:
           fetch-depth: 0
 
       - name: Set up Helm
-        uses: azure/setup-helm@fe7b79cd5ee1e45176fcad797de68ecaf3ca4814 # v4.2.0
+        uses: azure/setup-helm@b9e51907a09c216f16ebe8536097933489208112 # v4.3.0
         with:
           version: v3.14.4
 
        #Python is required because ct lint runs Yamale and yamllint which require Python.
-      - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+      - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
         with:
           python-version: '3.13'
 
       - name: Set up chart-testing
-        uses: helm/chart-testing-action@e6669bcd63d7cb57cb4380c33043eebe5d111992 # v2.6.1
+        uses: helm/chart-testing-action@0d28d3144d3a25ea2cc349d6e59901c4ff469b3b # v2.7.0
         with:
           version: v3.10.1
 
@@ -47,7 +47,7 @@ jobs:
           aws --endpoint-url http://127.0.0.1:9000/ s3 cp opt-125m/ s3://testbucket/opt-125m --recursive
 
       - name: Create kind cluster
-        uses: helm/kind-action@0025e74a8c7512023d06dc019c617aa3cf561fde # v1.10.0
+        uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0
 
       - name: Build the Docker image vllm cpu
         run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .

.github/workflows/pre-commit.yml

@@ -10,10 +10,11 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+    - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
       with:
         python-version: "3.12"
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
+    - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"
     - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1
       with:
         extra_args: --all-files --hook-stage manual

.github/workflows/stale.yml

@@ -13,7 +13,7 @@ jobs:
       actions: write
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@28ca1036281a5e5922ead5184a1bbf96e5fc984e # v9.0.0
+      - uses: actions/stale@5bef64f19d7facfb25b37b414482c7164d639639 # v9.1.0
         with:
           # Increasing this value ensures that changes to this workflow
           # propagate to all issues and PRs in days rather than months

.pre-commit-config.yaml

@@ -1,6 +1,7 @@
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
+exclude: 'vllm/third_party/.*'
 repos:
 - repo: https://github.com/google/yapf
   rev: v0.43.0
@@ -12,32 +13,39 @@ repos:
   rev: v0.9.3
   hooks:
   - id: ruff
-    args: [--output-format, github]
+    args: [--output-format, github, --fix]
 - repo: https://github.com/codespell-project/codespell
   rev: v2.4.0
   hooks:
   - id: codespell
-    exclude: 'benchmarks/sonnet.txt|(build|tests/(lora/data|models/fixtures|prompts))/.*'
+    additional_dependencies: ['tomli']
+    args: ['--toml', 'pyproject.toml']
 - repo: https://github.com/PyCQA/isort
-  rev: 5.13.2
+  rev: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
   hooks:
   - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
   rev: v19.1.7
   hooks:
   - id: clang-format
-    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))'
+    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/jackdewinter/pymarkdown
   rev: v0.9.27
   hooks:
   - id: pymarkdown
-    files: docs/.*
+    args: [fix]
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
   - id: actionlint
+- repo: https://github.com/astral-sh/uv-pre-commit
+  rev: 0.6.2
+  hooks:
+    - id: pip-compile
+      args: [requirements-test.in, -o, requirements-test.txt]
+      files: ^requirements-test\.(in|txt)$
 - repo: local
   hooks:
   - id: mypy-local
@@ -102,9 +110,20 @@ repos:
     entry: python tools/check_spdx_header.py
     language: python
     types: [python]
+  - id: check-filenames
+    name: Check for spaces in all filenames
+    entry: bash
+    args:
+      - -c
+      - 'git ls-files | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
+    language: system
+    always_run: true
+    pass_filenames: false
+  # Keep `suggestion` last
   - id: suggestion
     name: Suggestion
     entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
     language: system
     verbose: true
     pass_filenames: false
+  # Insert new entries above the `suggestion` entry

CMakeLists.txt

@@ -34,7 +34,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -174,6 +174,25 @@ include(FetchContent)
 file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 
+#
+# Set rocm version dev int.
+#
+if(VLLM_GPU_LANG STREQUAL "HIP")
+  #
+  # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
+  #
+  set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
+  set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
+
+
+  #
+  # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
+  # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
+  #
+  set(CMAKE_${VLLM_GPU_LANG}_FLAGS "${CMAKE_${VLLM_GPU_LANG}_FLAGS} -Wno-unused-result")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-result")
+endif()
+
 #
 # Define other extension targets
 #
@@ -192,7 +211,7 @@ set_gencode_flags_for_srcs(
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   message(STATUS "Enabling cumem allocator extension.")
   # link against cuda driver library
-  list(APPEND CUMEM_LIBS cuda)
+  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
   define_gpu_extension_target(
     cumem_allocator
     DESTINATION vllm
@@ -228,7 +247,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  set(CUTLASS_REVISION "v3.6.0" CACHE STRING "CUTLASS revision to use")
+  # Please keep this in sync with FetchContent_Declare line below.
+  set(CUTLASS_REVISION "v3.8.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -245,7 +265,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     FetchContent_Declare(
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        GIT_TAG v3.7.0
+        # Please keep this in sync with CUTLASS_REVISION line above.
+        GIT_TAG v3.8.0
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -264,8 +285,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/custom_all_reduce.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
+    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
+    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
-    "csrc/sparse/cutlass/sparse_compressor_entry.cu"
     "csrc/cutlass_extensions/common.cpp")
 
   set_gencode_flags_for_srcs(
@@ -299,7 +321,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
   cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-    set(SRCS 
+    set(SRCS
        "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
@@ -358,8 +380,7 @@ 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, 9.0a for now).
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
-    set(SRCS "csrc/sparse/cutlass/sparse_compressor_c3x.cu"
-             "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
+    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
@@ -377,6 +398,24 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # FP4 Archs and flags
+  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
+    set(SRCS
+      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
+      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${FP4_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
+  else()
+    message(STATUS "Not building NVFP4 as no compatible archs were found.")
+    # clear FP4_ARCHS
+    set(FP4_ARCHS)
+  endif()
 
   #
   # Machete kernels
@@ -458,7 +497,7 @@ define_gpu_extension_target(
   SOURCES ${VLLM_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
-  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
@@ -536,77 +575,8 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     WITH_SOABI)
 endif()
 
-# vllm-flash-attn currently only supported on CUDA
-if (NOT VLLM_GPU_LANG STREQUAL "CUDA")
-  return()
+# For CUDA we also build and ship some external projects.
+if (VLLM_GPU_LANG STREQUAL "CUDA")
+    include(cmake/external_projects/flashmla.cmake)
+    include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()
-
-# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
-# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
-# arches in the CUDA case (and instead set the gencodes on a per file basis)
-# we need to manually set VLLM_GPU_ARCHES here.
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  foreach(_ARCH ${CUDA_ARCHS})
-    string(REPLACE "." "" _ARCH "${_ARCH}")
-    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
-  endforeach()
-endif()
-
-#
-# Build vLLM flash attention from source
-#
-# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
-# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
-# They should be identical but if they aren't, this is a massive footgun.
-#
-# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
-# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
-# If no component is specified, vllm-flash-attn is still installed.
-
-# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
-# This is to enable local development of vllm-flash-attn within vLLM.
-# It can be set as an environment variable or passed as a cmake argument.
-# The environment variable takes precedence.
-if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
-  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
-endif()
-
-if(VLLM_FLASH_ATTN_SRC_DIR)
-  FetchContent_Declare(
-          vllm-flash-attn SOURCE_DIR 
-          ${VLLM_FLASH_ATTN_SRC_DIR}
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-else()
-  FetchContent_Declare(
-          vllm-flash-attn
-          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG d4e09037abf588af1ec47d0e966b237ee376876c
-          GIT_PROGRESS TRUE
-          # Don't share the vllm-flash-attn build between build types
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-endif()
-
-
-# Fetch the vllm-flash-attn library
-FetchContent_MakeAvailable(vllm-flash-attn)
-message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
-
-# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
-# case only one is built, in the case both are built redundant work is done)
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa2_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa3_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-# Nothing after vllm-flash-attn, see comment about macros above

CODE_OF_CONDUCT.md

@@ -125,4 +125,3 @@ Community Impact Guidelines were inspired by
 For answers to common questions about this code of conduct, see the
 [Contributor Covenant FAQ](https://www.contributor-covenant.org/faq). Translations are available at
 [Contributor Covenant translations](https://www.contributor-covenant.org/translations).
-

Dockerfile

@@ -27,6 +27,9 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
     && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
     && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
     && python3 --version && python3 -m pip --version
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
@@ -50,15 +53,15 @@ WORKDIR /workspace
 # we need to install torch and torchvision from the nightly builds first,
 # pytorch will not appear as a vLLM dependency in all of the following steps
 # after this step
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
     fi
 
 COPY requirements-common.txt requirements-common.txt
 COPY requirements-cuda.txt requirements-cuda.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-cuda.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-cuda.txt
 
 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
@@ -78,8 +81,8 @@ ARG TARGETPLATFORM
 # install build dependencies
 COPY requirements-build.txt requirements-build.txt
 
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-build.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-build.txt
 
 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -98,7 +101,7 @@ ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
 # if USE_SCCACHE is set, use sccache to speed up compilation
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
         echo "Installing sccache..." \
@@ -118,7 +121,7 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git  \
     if [ "$USE_SCCACHE" != "1" ]; then \
         python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
@@ -143,8 +146,8 @@ FROM base as dev
 COPY requirements-lint.txt requirements-lint.txt
 COPY requirements-test.txt requirements-test.txt
 COPY requirements-dev.txt requirements-dev.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-dev.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-dev.txt
 #################### DEV IMAGE ####################
 
 #################### vLLM installation IMAGE ####################
@@ -174,6 +177,9 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
     && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
     && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
     && python3 --version && python3 -m pip --version
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -185,29 +191,32 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 # we need to install torch and torchvision from the nightly builds first,
 # pytorch will not appear as a vLLM dependency in all of the following steps
 # after this step
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215";  \
     fi
 
 # Install vllm wheel first, so that torch etc will be installed.
 RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
-    --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install dist/*.whl --verbose
+    --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system dist/*.whl --verbose
 
-# How to build this FlashInfer wheel:
+# If we need to build FlashInfer wheel before its release:
 # $ export FLASHINFER_ENABLE_AOT=1
 # $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+
 # $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX'
 # $ git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
 # $ cd flashinfer
 # $ git checkout 524304395bd1d8cd7d07db083859523fcaa246a4
+# $ rm -rf build
 # $ python3 setup.py bdist_wheel --dist-dir=dist --verbose
+# $ ls dist
+# $ # upload the wheel to a public location, e.g. https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.1.post1+cu124torch2.5-cp38-abi3-linux_x86_64.whl
 
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
 . /etc/environment && \
 if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
-    python3 -m pip install https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.0.post1-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
+    uv pip install --system https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.1.post1/flashinfer_python-0.2.1.post1+cu124torch2.5-cp38-abi3-linux_x86_64.whl ; \
 fi
 COPY examples examples
 
@@ -216,8 +225,8 @@ COPY examples examples
 # install build dependencies for JIT compilation.
 # TODO: Remove this once FlashInfer AOT wheel is fixed
 COPY requirements-build.txt requirements-build.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-build.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-build.txt
 
 #################### vLLM installation IMAGE ####################
 
@@ -229,16 +238,16 @@ FROM vllm-base AS test
 ADD . /vllm-workspace/
 
 # install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-dev.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-dev.txt
 
 # install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -e tests/vllm_test_utils
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -e tests/vllm_test_utils
 
 # enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install hf_transfer
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system hf_transfer
 ENV HF_HUB_ENABLE_HF_TRANSFER 1
 
 # Copy in the v1 package for testing (it isn't distributed yet)
@@ -257,11 +266,11 @@ RUN mv vllm test_docs/
 FROM vllm-base AS vllm-openai-base
 
 # install additional dependencies for openai api server
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
     else \
-        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
     fi
 
 ENV VLLM_USAGE_SOURCE production-docker-image

Dockerfile.neuron

@@ -23,10 +23,12 @@ WORKDIR ${APP_MOUNT}/vllm
 RUN python3 -m pip install --upgrade pip
 RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
 RUN python3 -m pip install sentencepiece transformers==4.45.2 -U
-RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
 RUN python3 -m pip install neuronx-cc==2.16.345.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
 RUN python3 -m pip install pytest
 
+# uninstall transformers-neuronx package explicitly to avoid version conflict
+RUN python3 -m pip uninstall -y transformers-neuronx
+
 COPY . .
 ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
@@ -43,6 +45,10 @@ RUN --mount=type=bind,source=.git,target=.git \
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 
+# install transformers-neuronx package as an optional dependencies (for V0)
+# FIXME: `--no-deps` argument is temporarily added to resolve transformers package version conflict
+RUN python3 -m pip install transformers-neuronx==0.13.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U --no-deps
+
 # overwrite entrypoint to run bash script
 RUN echo "import subprocess; import sys; subprocess.check_call(sys.argv[1:])" > /usr/local/bin/dockerd-entrypoint.py
 

Dockerfile.rocm_base

@@ -6,7 +6,7 @@ ARG RCCL_BRANCH="648a58d"
 ARG RCCL_REPO="https://github.com/ROCm/rccl"
 ARG TRITON_BRANCH="e5be006"
 ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
-ARG PYTORCH_BRANCH="8d4926e"
+ARG PYTORCH_BRANCH="3a585126"
 ARG PYTORCH_VISION_BRANCH="v0.19.1"
 ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"

README.md

@@ -15,9 +15,14 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
+We are excited to invite you to our Menlo Park meetup with Meta, evening of Thursday, February 27! Meta engineers will discuss the improvements on top of vLLM, and vLLM contributors will share updates from the v0.7.x series of releases. [Register Now](https://lu.ma/h7g3kuj9)
+
+---
+
 *Latest News* 🔥
+
 - [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
-- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing).
+- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
@@ -33,7 +38,9 @@ Easy, fast, and cheap LLM serving for everyone
 - [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
 
 ---
+
 ## About
+
 vLLM is a fast and easy-to-use library for LLM inference and serving.
 
 Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evolved into a community-driven project with contributions from both academia and industry.
@@ -127,6 +134,7 @@ We also have an official fundraising venue through [OpenCollective](https://open
 ## Citation
 
 If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):
+
 ```bibtex
 @inproceedings{kwon2023efficient,
   title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
@@ -138,11 +146,11 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
-* For technical questions and feature requests, please use Github issues or discussions.
-* For discussing with fellow users and coordinating contributions and development, please use Slack.
-* For security disclosures, please use Github's security advisory feature.
-* For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+- For technical questions and feature requests, please use Github issues or discussions.
+- For discussing with fellow users and coordinating contributions and development, please use Slack.
+- For security disclosures, please use Github's security advisory feature.
+- For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
 
 ## Media Kit
 
-* If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
+- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).

benchmarks/README.md

@@ -3,6 +3,7 @@
 ## Downloading the ShareGPT dataset
 
 You can download the dataset by running:
+
 ```bash
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 ```
@@ -11,9 +12,18 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
 
 The json file refers to several image datasets (coco, llava, etc.). The benchmark scripts
 will ignore a datapoint if the referred image is missing.
+
 ```bash
 wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json
 mkdir coco -p
 wget http://images.cocodataset.org/zips/train2017.zip -O coco/train2017.zip
 unzip coco/train2017.zip -d coco/
 ```
+
+# Downloading the BurstGPT dataset
+
+You can download the BurstGPT v1.1 dataset by running:
+
+```bash
+wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
+```

benchmarks/benchmark_guided.py

@@ -46,6 +46,12 @@ def run_vllm(requests: List[SampleRequest],
              warmup: bool = False) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(**vars(engine_args))
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests.")
 
     # Add the requests to the engine.
     prompts: List[str] = []
@@ -115,6 +121,13 @@ async def run_vllm_async(
     async with build_async_engine_client_from_engine_args(
             engine_args, disable_frontend_multiprocessing) as llm:
 
+        assert all(
+            llm.model_config.max_model_len >= (request.prompt_len +
+                                               request.expected_output_len)
+            for request in requests), (
+                "Please ensure that max_model_len is greater than the sum of"
+                " prompt_len and expected_output_len for all requests.")
+
         # Add the requests to the engine.
         prompts: List[str] = []
         sampling_params: List[SamplingParams] = []

benchmarks/benchmark_latency.py

@@ -1,14 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 """Benchmark the latency of processing a single batch of requests."""
+
 import argparse
 import dataclasses
 import json
+import os
 import time
 from pathlib import Path
-from typing import List, Optional
+from typing import Any, Dict, List, Optional
 
 import numpy as np
 import torch
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
@@ -18,6 +21,18 @@ from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser
 
 
+def save_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                     results: Dict[str, Any]) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={"latency": results["latencies"]},
+        extra_info={k: results[k]
+                    for k in ["avg_latency", "percentiles"]})
+    if pt_records:
+        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
 def main(args: argparse.Namespace):
     print(args)
 
@@ -26,6 +41,10 @@ def main(args: argparse.Namespace):
     # NOTE(woosuk): If the request cannot be processed in a single batch,
     # the engine will automatically process the request in multiple batches.
     llm = LLM(**dataclasses.asdict(engine_args))
+    assert llm.llm_engine.model_config.max_model_len >= (
+        args.input_len +
+        args.output_len), ("Please ensure that max_model_len is greater than"
+                           " the sum of input_len and output_len.")
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -54,7 +73,8 @@ def main(args: argparse.Namespace):
                     beam_width=args.n,
                     max_tokens=args.output_len,
                     ignore_eos=True,
-                ))
+                ),
+            )
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
@@ -64,7 +84,8 @@ def main(args: argparse.Namespace):
                         torch.profiler.ProfilerActivity.CUDA,
                     ],
                     on_trace_ready=torch.profiler.tensorboard_trace_handler(
-                        str(profile_dir))) as p:
+                        str(profile_dir)),
+            ) as p:
                 llm_generate()
             print(p.key_averages().table(sort_by="self_cuda_time_total"))
         else:
@@ -81,9 +102,8 @@ def main(args: argparse.Namespace):
     if args.profile:
         profile_dir = args.profile_result_dir
         if not profile_dir:
-            profile_dir = Path(
-                "."
-            ) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
+            profile_dir = (Path(".") / "vllm_benchmark_result" /
+                           f"latency_result_{time.time()}")
         print(f"Profiling (results will be saved to '{profile_dir}')...")
         run_to_completion(profile_dir=profile_dir)
         return
@@ -95,9 +115,9 @@ def main(args: argparse.Namespace):
     latencies = np.array(latencies)
     percentages = [10, 25, 50, 75, 90, 99]
     percentiles = np.percentile(latencies, percentages)
-    print(f'Avg latency: {np.mean(latencies)} seconds')
+    print(f"Avg latency: {np.mean(latencies)} seconds")
     for percentage, percentile in zip(percentages, percentiles):
-        print(f'{percentage}% percentile latency: {percentile} seconds')
+        print(f"{percentage}% percentile latency: {percentile} seconds")
 
     # Output JSON results if specified
     if args.output_json:
@@ -108,43 +128,51 @@ def main(args: argparse.Namespace):
         }
         with open(args.output_json, "w") as f:
             json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description='Benchmark the latency of processing a single batch of '
-        'requests till completion.')
-    parser.add_argument('--input-len', type=int, default=32)
-    parser.add_argument('--output-len', type=int, default=128)
-    parser.add_argument('--batch-size', type=int, default=8)
-    parser.add_argument('--n',
-                        type=int,
-                        default=1,
-                        help='Number of generated sequences per prompt.')
-    parser.add_argument('--use-beam-search', action='store_true')
-    parser.add_argument('--num-iters-warmup',
-                        type=int,
-                        default=10,
-                        help='Number of iterations to run for warmup.')
-    parser.add_argument('--num-iters',
+        description="Benchmark the latency of processing a single batch of "
+        "requests till completion.")
+    parser.add_argument("--input-len", type=int, default=32)
+    parser.add_argument("--output-len", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, default=8)
+    parser.add_argument(
+        "--n",
+        type=int,
+        default=1,
+        help="Number of generated sequences per prompt.",
+    )
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument(
+        "--num-iters-warmup",
+        type=int,
+        default=10,
+        help="Number of iterations to run for warmup.",
+    )
+    parser.add_argument("--num-iters",
                         type=int,
                         default=30,
-                        help='Number of iterations to run.')
+                        help="Number of iterations to run.")
     parser.add_argument(
-        '--profile',
-        action='store_true',
-        help='profile the generation process of a single batch')
+        "--profile",
+        action="store_true",
+        help="profile the generation process of a single batch",
+    )
     parser.add_argument(
-        '--profile-result-dir',
+        "--profile-result-dir",
         type=str,
         default=None,
-        help=('path to save the pytorch profiler output. Can be visualized '
-              'with ui.perfetto.dev or Tensorboard.'))
+        help=("path to save the pytorch profiler output. Can be visualized "
+              "with ui.perfetto.dev or Tensorboard."),
+    )
     parser.add_argument(
-        '--output-json',
+        "--output-json",
         type=str,
         default=None,
-        help='Path to save the latency results in JSON format.')
+        help="Path to save the latency results in JSON format.",
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_prioritization.py

@@ -13,6 +13,11 @@ from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser
 
 
+#Select a equi-probable random priority
+def get_random_flag():
+    return 0 if random.random() < 0.5 else 1
+
+
 def sample_requests(
     dataset_path: str,
     num_requests: int,
@@ -55,8 +60,7 @@ def sample_requests(
             # Prune too long sequences.
             continue
 
-        #Select a equi-probable random priority
-        priority = 0 if random.random() < 0.5 else 1
+        priority = get_random_flag()
 
         filtered_dataset.append((prompt, prompt_len, output_len, priority))
 
@@ -71,6 +75,12 @@ def run_vllm(
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
 
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " input_len and output_len for all requests.")
+
     # Add the requests to the engine.
     prompts = []
     sampling_params = []
@@ -103,8 +113,8 @@ def main(args: argparse.Namespace):
     if args.dataset is None:
         # Synthesize a prompt with the given input length.
         prompt = "hi" * (args.input_len - 1)
-        requests = [(prompt, args.input_len, args.output_len)
-                    for _ in range(args.num_prompts)]
+        requests = [(prompt, args.input_len, args.output_len,
+                     get_random_flag()) for _ in range(args.num_prompts)]
     else:
         requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
                                    args.output_len)

benchmarks/benchmark_serving.py

@@ -38,6 +38,7 @@ from datetime import datetime
 from typing import Any, AsyncGenerator, Collection, Dict, List, Optional, Tuple
 
 import numpy as np
+import pandas as pd
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
                                   RequestFuncOutput)
 from datasets import load_dataset
@@ -55,6 +56,8 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
+
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
 
@@ -131,6 +134,35 @@ def sample_sharegpt_requests(
     return filtered_dataset
 
 
+def sample_burstgpt_requests(
+    dataset_path: str,
+    num_requests: int,
+    random_seed: int,
+    tokenizer: PreTrainedTokenizerBase,
+) -> List[Tuple[str, int, int, None]]:
+    df = pd.read_csv(dataset_path)
+    gpt4_df = df[df["Model"] == "GPT-4"]
+    # Remove the failed requests (i.e., response length is 0)
+    gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
+    # Randomly sample num_requests from the dataset
+    if num_requests <= len(gpt4_df):
+        gpt4_df = gpt4_df.sample(n=num_requests, random_state=random_seed)
+    else:
+        gpt4_df = gpt4_df.sample(n=num_requests,
+                                 random_state=random_seed,
+                                 replace=True)
+    # Convert the dataframe to a list of tuples
+    dataset = gpt4_df.values.tolist()
+    input_requests = []
+    for i in range(num_requests):
+        input_len = int(dataset[i][2])
+        output_len = int(dataset[i][3])
+        prompt = tokenizer.decode([(i + j) % tokenizer.vocab_size
+                                   for j in range(input_len)])
+        input_requests.append((prompt, input_len, output_len, None))
+    return input_requests
+
+
 def sample_sonnet_requests(
     dataset_path: str,
     num_requests: int,
@@ -372,21 +404,21 @@ async def get_request(
     burstiness: float = 1.0,
 ) -> AsyncGenerator[Tuple[str, int, int], None]:
     """
-    Asynchronously generates requests at a specified rate 
+    Asynchronously generates requests at a specified rate
     with OPTIONAL burstiness.
-    
+
     Args:
-        input_requests: 
+        input_requests:
             A list of input requests, each represented as a tuple.
-        request_rate: 
+        request_rate:
             The rate at which requests are generated (requests/s).
-        burstiness (optional): 
-            The burstiness factor of the request generation. 
+        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 
+            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.
     """
     input_requests = iter(input_requests)
@@ -537,6 +569,7 @@ async def benchmark(
     ignore_eos: bool,
     goodput_config_dict: Dict[str, float],
     max_concurrency: Optional[int],
+    lora_modules: Optional[List[str]],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -562,6 +595,7 @@ async def benchmark(
         multi_modal_content=test_mm_content,
         ignore_eos=ignore_eos,
     )
+
     test_output = await request_func(request_func_input=test_input)
     if not test_output.success:
         raise ValueError(
@@ -570,6 +604,11 @@ async def benchmark(
     else:
         print("Initial test run completed. Starting main benchmark run...")
 
+    if lora_modules:
+        # For each input request, choose a LoRA module at random.
+        lora_modules = iter(
+            [random.choice(lora_modules) for _ in range(len(input_requests))])
+
     if profile:
         print("Starting profiler...")
         profile_input = RequestFuncInput(model=model_id,
@@ -616,8 +655,13 @@ async def benchmark(
     tasks: List[asyncio.Task] = []
     async for request in get_request(input_requests, request_rate, burstiness):
         prompt, prompt_len, output_len, mm_content = request
-        request_func_input = RequestFuncInput(model=model_id,
-                                              model_name=model_name,
+        req_model_id, req_model_name = model_id, model_name
+        if lora_modules:
+            req_lora_module = next(lora_modules)
+            req_model_id, req_model_name = req_lora_module, req_lora_module
+
+        request_func_input = RequestFuncInput(model=req_model_id,
+                                              model_name=req_model_name,
                                               prompt=prompt,
                                               api_url=api_url,
                                               prompt_len=prompt_len,
@@ -775,6 +819,31 @@ def parse_goodput(slo_pairs):
     return goodput_config_dict
 
 
+def save_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                     results: Dict[str, Any],
+                                     file_name: str) -> None:
+    metrics = [
+        "median_ttft_ms", "mean_ttft_ms", "std_ttft_ms", "p99_ttft_ms",
+        "mean_tpot_ms", "median_tpot_ms", "std_tpot_ms", "p99_tpot_ms",
+        "median_itl_ms", "mean_itl_ms", "std_itl_ms", "p99_itl_ms"
+    ]
+    # These raw data might be useful, but they are rather big. They can be added
+    # later if needed
+    ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"]
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={k: [results[k]]
+                 for k in metrics},
+        extra_info={
+            k: results[k]
+            for k in results if k not in metrics and k not in ignored_metrics
+        })
+    if pt_records:
+        # Don't use json suffix here as we don't want CI to pick it up
+        pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
 def main(args: argparse.Namespace):
     print(args)
     random.seed(args.seed)
@@ -797,18 +866,10 @@ def main(args: argparse.Namespace):
                               tokenizer_mode=tokenizer_mode,
                               trust_remote_code=args.trust_remote_code)
 
-    if args.dataset is not None:
-        warnings.warn(
-            "The '--dataset' argument will be deprecated in the next "
-            "release. Please use '--dataset-name' and "
-            "'--dataset-path' in the future runs.",
-            stacklevel=2)
-        input_requests = sample_sharegpt_requests(
-            dataset_path=args.dataset,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
+    if args.dataset_name is None:
+        raise ValueError(
+            "Please specify '--dataset-name' and the corresponding "
+            "'--dataset-path' if required.")
 
     elif args.dataset_name == "sharegpt":
         input_requests = sample_sharegpt_requests(
@@ -818,6 +879,14 @@ def main(args: argparse.Namespace):
             fixed_output_len=args.sharegpt_output_len,
         )
 
+    elif args.dataset_name == "burstgpt":
+        input_requests = sample_burstgpt_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            random_seed=args.seed,
+            tokenizer=tokenizer,
+        )
+
     elif args.dataset_name == "sonnet":
         # Do not format the prompt, pass to message directly
         if args.backend == "openai-chat":
@@ -900,6 +969,7 @@ def main(args: argparse.Namespace):
             ignore_eos=args.ignore_eos,
             goodput_config_dict=goodput_config_dict,
             max_concurrency=args.max_concurrency,
+            lora_modules=args.lora_modules,
         ))
 
     # Save config and results to json
@@ -946,6 +1016,7 @@ def main(args: argparse.Namespace):
             file_name = os.path.join(args.result_dir, file_name)
         with open(file_name, "w", encoding='utf-8') as outfile:
             json.dump(result_json, outfile)
+        save_to_pytorch_benchmark_format(args, result_json, file_name)
 
 
 if __name__ == "__main__":
@@ -963,7 +1034,8 @@ if __name__ == "__main__":
         default=None,
         help="Server or API base url if not using http host and port.",
     )
-    parser.add_argument("--host", type=str, default="localhost")
+    # Use 127.0.0.1 here instead of localhost to force the use of ipv4
+    parser.add_argument("--host", type=str, default="127.0.0.1")
     parser.add_argument("--port", type=int, default=8000)
     parser.add_argument(
         "--endpoint",
@@ -971,18 +1043,11 @@ if __name__ == "__main__":
         default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument(
-        "--dataset",
-        type=str,
-        default=None,
-        help="Path to the ShareGPT dataset, will be deprecated in the "
-        "next release.",
-    )
     parser.add_argument(
         "--dataset-name",
         type=str,
         default="sharegpt",
-        choices=["sharegpt", "sonnet", "random", "hf"],
+        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf"],
         help="Name of the dataset to benchmark on.",
     )
     parser.add_argument("--dataset-path",
@@ -1224,11 +1289,12 @@ if __name__ == "__main__":
         '--tokenizer-mode',
         type=str,
         default="auto",
-        choices=['auto', 'slow', 'mistral'],
+        choices=['auto', 'slow', 'mistral', 'custom'],
         help='The tokenizer mode.\n\n* "auto" will use the '
         'fast tokenizer if available.\n* "slow" will '
         'always use the slow tokenizer. \n* '
-        '"mistral" will always use the `mistral_common` tokenizer.')
+        '"mistral" will always use the `mistral_common` tokenizer. \n*'
+        '"custom" will use --tokenizer to select the preregistered tokenizer.')
 
     parser.add_argument("--served-model-name",
                         type=str,
@@ -1237,5 +1303,12 @@ if __name__ == "__main__":
                         "If not specified, the model name will be the "
                         "same as the ``--model`` argument. ")
 
+    parser.add_argument("--lora-modules",
+                        nargs='+',
+                        default=None,
+                        help="A subset of LoRA module names passed in when "
+                        "launching the server. For each request, the "
+                        "script chooses a LoRA module at random.")
+
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving_guided.py

@@ -9,7 +9,7 @@ On the server side, run one of the following commands:
     ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
 
 On the client side, run:
-    python benchmarks/benchmark_serving.py \
+    python benchmarks/benchmark_serving_guided.py \
         --backend <backend> \
         --model <your_model> \
         --dataset json \
@@ -31,7 +31,7 @@ import random
 import time
 import warnings
 from dataclasses import dataclass
-from typing import AsyncGenerator, List, Optional, Tuple
+from typing import AsyncGenerator, Dict, List, Optional, Tuple
 
 import datasets
 import numpy as np
@@ -264,6 +264,7 @@ def calculate_metrics(
     tokenizer: PreTrainedTokenizerBase,
     selected_percentile_metrics: List[str],
     selected_percentiles: List[float],
+    goodput_config_dict: Optional[Dict[str, float]] = None,
 ) -> Tuple[BenchmarkMetrics, List[int]]:
     actual_output_lens: List[int] = []
     total_input = 0
@@ -287,10 +288,10 @@ def calculate_metrics(
             total_input += input_requests[i].prompt_len
             tpot = 0
             if output_len > 1:
-                tpot = (outputs[i].latency - outputs[i].ttft) / (output_len -
-                                                                 1)
+                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
+                tpot = latency_minus_ttft / (output_len - 1)
                 tpots.append(tpot)
-            outputs[i].tpot = sum(tpots) / len(tpots) if len(tpots) else 0
+            outputs[i].tpot = tpot
             # Note: if output_len <= 1, we regard tpot as 0 for goodput
             all_tpots.append(tpot)
             itls += outputs[i].itl
@@ -300,6 +301,28 @@ def calculate_metrics(
         else:
             actual_output_lens.append(0)
 
+    if goodput_config_dict:
+        valid_metrics = []
+        slo_values = []
+
+        if "ttft" in goodput_config_dict:
+            valid_metrics.append(ttfts)
+            slo_values.append(goodput_config_dict["ttft"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+        if "tpot" in goodput_config_dict:
+            valid_metrics.append(all_tpots)
+            slo_values.append(goodput_config_dict["tpot"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+        if "e2el" in goodput_config_dict:
+            valid_metrics.append(e2els)
+            slo_values.append(goodput_config_dict["e2el"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+
+        for req_metric in zip(*valid_metrics):
+            is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)])
+            if is_good_req:
+                good_completed += 1
+
     if completed == 0:
         warnings.warn(
             "All requests failed. This is likely due to a misconfiguration "
@@ -356,6 +379,7 @@ async def benchmark(
     max_concurrency: Optional[int],
     guided_decoding_ratio: float,
     guided_decoding_backend: str,
+    goodput_config_dict: Optional[Dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -483,6 +507,7 @@ async def benchmark(
         tokenizer=tokenizer,
         selected_percentile_metrics=selected_percentile_metrics,
         selected_percentiles=selected_percentiles,
+        goodput_config_dict=goodput_config_dict,
     )
 
     print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -494,6 +519,9 @@ async def benchmark(
                                  metrics.total_output))
     print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                     metrics.request_throughput))
+    if goodput_config_dict:
+        print("{:<40} {:<10.2f}".format("Request goodput (req/s):",
+                                        metrics.request_goodput))
     print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
                                     metrics.output_throughput))
     print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
@@ -617,6 +645,40 @@ def evaluate(ret, args):
             100) if len(not_none_scores) > 0 else None
 
 
+def parse_goodput(slo_pairs):
+    goodput_config_dict = {}
+    try:
+        for slo_pair in slo_pairs:
+            slo_name, slo_val = slo_pair.split(":")
+            goodput_config_dict[slo_name] = float(slo_val)
+    except ValueError as err:
+        raise argparse.ArgumentTypeError(
+            "Invalid format found for service level objectives. "
+            "Specify service level objectives for goodput as \"KEY:VALUE\" "
+            "pairs, where the key is a metric name, and the value is a "
+            "number in milliseconds.") from err
+    return goodput_config_dict
+
+
+def check_goodput_args(args):
+    goodput_config_dict = {}
+    VALID_NAMES = ["ttft", "tpot", "e2el"]
+    if args.goodput:
+        goodput_config_dict = parse_goodput(args.goodput)
+        for slo_name, slo_val in goodput_config_dict.items():
+            if slo_name not in VALID_NAMES:
+                raise ValueError(
+                    f"Invalid metric name found, {slo_name}: {slo_val}. "
+                    "The service level objective name should be one of "
+                    f"{str(VALID_NAMES)}. ")
+            if slo_val < 0:
+                raise ValueError(
+                    f"Invalid value found, {slo_name}: {slo_val}. "
+                    "The service level objective value should be "
+                    "non-negative.")
+    return goodput_config_dict
+
+
 def main(args: argparse.Namespace):
     print(args)
     random.seed(args.seed)
@@ -661,6 +723,8 @@ def main(args: argparse.Namespace):
 
     input_requests = sample_requests(tokenizer, args)
 
+    goodput_config_dict = check_goodput_args(args)
+
     benchmark_result, ret = asyncio.run(
         benchmark(
             backend=backend,
@@ -681,6 +745,7 @@ def main(args: argparse.Namespace):
             max_concurrency=args.max_concurrency,
             guided_decoding_ratio=args.guided_decoding_ratio,
             guided_decoding_backend=args.guided_decoding_backend,
+            goodput_config_dict=goodput_config_dict,
         ))
 
     # Save config and results to json
@@ -731,7 +796,8 @@ if __name__ == "__main__":
         default=None,
         help="Server or API base url if not using http host and port.",
     )
-    parser.add_argument("--host", type=str, default="localhost")
+    # Use 127.0.0.1 here instead of localhost to force the use of ipv4
+    parser.add_argument("--host", type=str, default="127.0.0.1")
     parser.add_argument("--port", type=int, default=8000)
     parser.add_argument(
         "--endpoint",
@@ -864,6 +930,18 @@ if __name__ == "__main__":
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
     )
+    parser.add_argument(
+        "--goodput",
+        nargs="+",
+        required=False,
+        help="Specify service level objectives for goodput as \"KEY:VALUE\" "
+        "pairs, where the key is a metric name, and the value is in "
+        "milliseconds. Multiple \"KEY:VALUE\" pairs can be provided, "
+        "separated by spaces. Allowed request level metric names are "
+        "\"ttft\", \"tpot\", \"e2el\". For more context on the definition of "
+        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
+        "and the blog: https://hao-ai-lab.github.io/blogs/distserve")
+
     parser.add_argument("--no-guided-decoding",
                         action='store_true',
                         default=False,

benchmarks/benchmark_throughput.py

@@ -3,13 +3,15 @@
 import argparse
 import dataclasses
 import json
+import os
 import random
 import time
 from functools import cache
-from typing import Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 import uvloop
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from PIL import Image
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
@@ -169,7 +171,12 @@ def run_vllm(
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
-
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests.")
     # Add the requests to the engine.
     prompts: List[TextPrompt] = []
     sampling_params: List[SamplingParams] = []
@@ -227,6 +234,12 @@ async def run_vllm_async(
 
     async with build_async_engine_client_from_engine_args(
             engine_args, disable_frontend_multiprocessing) as llm:
+        assert all(
+            llm.model_config.max_model_len >= (request.prompt_len +
+                                               request.expected_output_len)
+            for request in requests), (
+                "Please ensure that max_model_len is greater than the sum of"
+                " prompt_len and expected_output_len for all requests.")
 
         # Add the requests to the engine.
         prompts: List[TextPrompt] = []
@@ -338,6 +351,24 @@ def run_mii(
     return end - start
 
 
+def save_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                     results: Dict[str, Any]) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={
+            "requests_per_second": [results["requests_per_second"]],
+            "tokens_per_second": [results["tokens_per_second"]],
+        },
+        extra_info={
+            k: results[k]
+            for k in ["elapsed_time", "num_requests", "total_num_tokens"]
+        })
+    if pt_records:
+        # Don't use json suffix here as we don't want CI to pick it up
+        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
 def main(args: argparse.Namespace):
     print(args)
     random.seed(args.seed)
@@ -435,6 +466,7 @@ def main(args: argparse.Namespace):
         }
         with open(args.output_json, "w") as f:
             json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)
 
 
 if __name__ == "__main__":

benchmarks/benchmark_utils.py

@@ -0,0 +1,69 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+import json
+import math
+import os
+from typing import Any, Dict, List
+
+
+def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                        metrics: Dict[str, List],
+                                        extra_info: Dict[str, Any]) -> List:
+    """
+    Save the benchmark results in the format used by PyTorch OSS benchmark with
+    on metric per record
+    https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
+    """
+    records = []
+    if not os.environ.get("SAVE_TO_PYTORCH_BENCHMARK_FORMAT", False):
+        return records
+
+    for name, benchmark_values in metrics.items():
+        record = {
+            "benchmark": {
+                "name": "vLLM benchmark",
+                "extra_info": {
+                    "args": vars(args),
+                },
+            },
+            "model": {
+                "name": args.model,
+            },
+            "metric": {
+                "name": name,
+                "benchmark_values": benchmark_values,
+                "extra_info": extra_info,
+            },
+        }
+
+        tp = record["benchmark"]["extra_info"]["args"].get(
+            "tensor_parallel_size")
+        # Save tensor_parallel_size parameter if it's part of the metadata
+        if not tp and "tensor_parallel_size" in extra_info:
+            record["benchmark"]["extra_info"]["args"][
+                "tensor_parallel_size"] = extra_info["tensor_parallel_size"]
+
+        records.append(record)
+
+    return records
+
+
+class InfEncoder(json.JSONEncoder):
+
+    def clear_inf(self, o: Any):
+        if isinstance(o, dict):
+            return {k: self.clear_inf(v) for k, v in o.items()}
+        elif isinstance(o, list):
+            return [self.clear_inf(v) for v in o]
+        elif isinstance(o, float) and math.isinf(o):
+            return "inf"
+        return o
+
+    def iterencode(self, o: Any, *args, **kwargs) -> Any:
+        return super().iterencode(self.clear_inf(o), *args, **kwargs)
+
+
+def write_to_json(filename: str, records: List) -> None:
+    with open(filename, "w") as f:
+        json.dump(records, f, cls=InfEncoder)

benchmarks/kernels/benchmark_lora.py

@@ -89,7 +89,7 @@ def make_prompt_lora_mapping(num_prompts: int, num_active_loras: int,
                              sort_by_lora_id: bool,
                              device: str) -> torch.Tensor:
     """
-    All prompts are mapped to a Lora ID in range [0, num_active_loras).
+    All prompts are mapped to a LoRA ID in range [0, num_active_loras).
     where 0 refers to first lora, 1 refers to second lora and so on.
     """
     assert num_active_loras > 0

benchmarks/kernels/benchmark_moe.py

@@ -468,7 +468,8 @@ 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] == "DeepseekV3ForCausalLM":
+    elif (config.architectures[0] == "DeepseekV3ForCausalLM"
+          or config.architectures[0] == "DeepseekV2ForCausalLM"):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size

cmake/external_projects/flashmla.cmake

@@ -0,0 +1,66 @@
+include(FetchContent)
+
+# If FLASH_MLA_SRC_DIR is set, flash-mla is installed from that directory 
+# instead of downloading.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{FLASH_MLA_SRC_DIR})
+  set(FLASH_MLA_SRC_DIR $ENV{FLASH_MLA_SRC_DIR})
+endif()
+
+if(FLASH_MLA_SRC_DIR)
+  FetchContent_Declare(
+        flashmla 
+        SOURCE_DIR ${FLASH_MLA_SRC_DIR}
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND ""
+  )
+else()
+  FetchContent_Declare(
+        flashmla
+        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
+        GIT_TAG 575f7724b9762f265bbee5889df9c7d630801845
+        GIT_PROGRESS TRUE
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND ""
+  )
+endif()
+
+
+FetchContent_MakeAvailable(flashmla)
+message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
+
+# The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
+# Only build FlashMLA kernels if we are building for something compatible with 
+# sm90a
+cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
+    set(FlashMLA_SOURCES
+        ${flashmla_SOURCE_DIR}/csrc/flash_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_bf16_sm90.cu
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_fp16_sm90.cu
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_metadata.cu)
+
+    set(FlashMLA_INCLUDES
+        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
+        ${flashmla_SOURCE_DIR}/csrc/include)
+
+    set_gencode_flags_for_srcs(
+        SRCS "${FlashMLA_SOURCES}"
+        CUDA_ARCHS "${FLASH_MLA_ARCHS}")
+
+    define_gpu_extension_target(
+        _flashmla_C
+        DESTINATION vllm
+        LANGUAGE ${VLLM_GPU_LANG}
+        SOURCES ${FlashMLA_SOURCES}
+        COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+        ARCHITECTURES ${VLLM_GPU_ARCHES}
+        INCLUDE_DIRECTORIES ${FlashMLA_INCLUDES}
+        USE_SABI 3
+        WITH_SOABI)
+else()
+    # Create an empty target for setup.py when not targeting sm90a systems
+    add_custom_target(_flashmla_C)
+endif()
+

cmake/external_projects/vllm_flash_attn.cmake

@@ -0,0 +1,67 @@
+# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
+# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
+# arches in the CUDA case (and instead set the gencodes on a per file basis)
+# we need to manually set VLLM_GPU_ARCHES here.
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  foreach(_ARCH ${CUDA_ARCHS})
+    string(REPLACE "." "" _ARCH "${_ARCH}")
+    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
+  endforeach()
+endif()
+
+#
+# Build vLLM flash attention from source
+#
+# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
+# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
+# They should be identical but if they aren't, this is a massive footgun.
+#
+# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
+# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
+# If no component is specified, vllm-flash-attn is still installed.
+
+# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
+# This is to enable local development of vllm-flash-attn within vLLM.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
+  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
+endif()
+
+if(VLLM_FLASH_ATTN_SRC_DIR)
+  FetchContent_Declare(
+          vllm-flash-attn SOURCE_DIR 
+          ${VLLM_FLASH_ATTN_SRC_DIR}
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
+else()
+  FetchContent_Declare(
+          vllm-flash-attn
+          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
+          GIT_TAG 720c94869cf2e0ff5a706e9c7f1dce0939686ade
+          GIT_PROGRESS TRUE
+          # Don't share the vllm-flash-attn build between build types
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
+endif()
+
+
+# Fetch the vllm-flash-attn library
+FetchContent_MakeAvailable(vllm-flash-attn)
+message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
+
+# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
+# case only one is built, in the case both are built redundant work is done)
+install(
+  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  DESTINATION vllm_flash_attn
+  COMPONENT _vllm_fa2_C
+  FILES_MATCHING PATTERN "*.py"
+)
+
+install(
+  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  DESTINATION vllm_flash_attn
+  COMPONENT _vllm_fa3_C
+  FILES_MATCHING PATTERN "*.py"
+)

cmake/utils.cmake

@@ -257,9 +257,9 @@ endmacro()
 #  where `<=` is the version comparison operator.
 # In other words, for each version in `TGT_CUDA_ARCHS` find the highest version
 #  in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`.
-# We have special handling for 9.0a, if 9.0a is in `SRC_CUDA_ARCHS` and 9.0 is
-#  in `TGT_CUDA_ARCHS` then we should remove 9.0a from `SRC_CUDA_ARCHS` and add
-#  9.0a to the result (and remove 9.0 from TGT_CUDA_ARCHS). 
+# We have special handling for x.0a, if x.0a is in `SRC_CUDA_ARCHS` and x.0 is
+#  in `TGT_CUDA_ARCHS` then we should remove x.0a from `SRC_CUDA_ARCHS` and add
+#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS). 
 # The result is stored in `OUT_CUDA_ARCHS`.
 #
 # Example:
@@ -272,8 +272,8 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   list(REMOVE_DUPLICATES SRC_CUDA_ARCHS)
   set(TGT_CUDA_ARCHS_ ${TGT_CUDA_ARCHS})
 
-  # if 9.0a is in SRC_CUDA_ARCHS and 9.0 is in CUDA_ARCHS then we should
-  # remove 9.0a from SRC_CUDA_ARCHS and add 9.0a to _CUDA_ARCHS
+  # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
+  # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
   if ("9.0a" IN_LIST SRC_CUDA_ARCHS)
     list(REMOVE_ITEM SRC_CUDA_ARCHS "9.0a")
@@ -283,6 +283,14 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
     endif()
   endif()
 
+  if ("10.0a" IN_LIST SRC_CUDA_ARCHS)
+    list(REMOVE_ITEM SRC_CUDA_ARCHS "10.0a")
+    if ("10.0" IN_LIST TGT_CUDA_ARCHS)
+      list(REMOVE_ITEM TGT_CUDA_ARCHS_ "10.0")
+      set(_CUDA_ARCHS "10.0a")
+    endif()
+  endif()
+
   list(SORT SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
   # for each ARCH in TGT_CUDA_ARCHS find the highest arch in SRC_CUDA_ARCHS that

csrc/cache.h

@@ -39,3 +39,10 @@ void concat_and_cache_mla(torch::Tensor& kv_c, torch::Tensor& k_pe,
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
                  const double scale, const std::string& kv_cache_dtype);
+
+void gather_cache(
+    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
+    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
+    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
+    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
+    int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);

csrc/cache_kernels.cu

@@ -2,6 +2,7 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 
+#include "cuda_utils.h"
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
 
@@ -374,7 +375,7 @@ void reshape_and_cache(
     torch::Tensor& slot_mapping,  // [num_tokens]
     const std::string& kv_cache_dtype, torch::Tensor& k_scale,
     torch::Tensor& v_scale) {
-  int num_tokens = key.size(0);
+  int num_tokens = slot_mapping.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
   int block_size = key_cache.size(3);
@@ -570,3 +571,161 @@ void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
     TORCH_CHECK(false, "Unsupported data type: ", kv_cache_dtype);
   }
 }
+
+namespace vllm {
+
+// grid is launched with dimensions (batch, num_splits)
+template <typename scalar_t>
+__global__ void gather_cache(
+    const scalar_t* __restrict__ src_cache,   // [NUM_BLOCKS, BLOCK_SIZE,
+                                              // ENTRIES...]
+    scalar_t* __restrict__ dst,               // [TOT_TOKENS, ENTRIES...]
+    const int32_t* __restrict__ block_table,  // [BATCH, BLOCK_INDICES]
+    const int32_t* __restrict__ cu_seq_lens,  // [BATCH+1]
+    const int32_t block_size, const int32_t entry_size,
+    const int64_t block_table_stride, const int64_t cache_block_stride,
+    const int64_t cache_entry_stride, const int64_t dst_entry_stride,
+    const int32_t* __restrict__ seq_starts) {  // Optional: starting offsets per
+                                               // batch
+
+  const int64_t bid = blockIdx.x;  // Batch ID
+  const int32_t num_splits = gridDim.y;
+  const int32_t split = blockIdx.y;
+  const int32_t seq_start = cu_seq_lens[bid];
+  const int32_t seq_end = cu_seq_lens[bid + 1];
+  const int32_t seq_len = seq_end - seq_start;
+  const int32_t tot_blocks = cuda_utils::ceil_div(seq_len, block_size);
+  const int32_t split_blocks = cuda_utils::ceil_div(tot_blocks, num_splits);
+
+  const int32_t split_start = split * split_blocks;
+  const int32_t split_end = min((split + 1) * split_blocks, tot_blocks);
+
+  const bool is_active_split = (split_start < tot_blocks);
+  const bool is_last_split = (split_end == tot_blocks);
+
+  if (!is_active_split) return;
+
+  int32_t full_blocks_end = split_end;
+  int32_t partial_block_size = 0;
+
+  // Adjust the pointer for the block_table for this batch.
+  // If seq_starts is provided, compute an offset based on (seq_starts[bid] /
+  // page_size)
+  const int32_t batch_offset = bid * block_table_stride;
+  int32_t offset = 0;
+  if (seq_starts != nullptr) {
+    offset = seq_starts[bid] / block_size;
+  }
+  const int32_t* batch_block_table = block_table + batch_offset + offset;
+
+  // Adjust dst pointer based on the cumulative sequence lengths.
+  dst += seq_start * dst_entry_stride;
+
+  if (is_last_split) {
+    partial_block_size = seq_len % block_size;
+    if (partial_block_size) full_blocks_end -= 1;
+  }
+
+  auto copy_entry = [&](const scalar_t* __restrict__ _src,
+                        scalar_t* __restrict__ _dst) {
+    for (int i = threadIdx.x; i < entry_size; i += blockDim.x)
+      _dst[i] = _src[i];
+  };
+
+  for (int pid = split_start; pid < full_blocks_end; ++pid) {
+    auto block_id = batch_block_table[pid];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + pid * block_size * dst_entry_stride;
+    for (int eid = 0; eid < block_size; ++eid) {
+      copy_entry(block_start_ptr + eid * cache_entry_stride,
+                 block_dst_ptr + eid * dst_entry_stride);
+    }
+  }
+
+  if (partial_block_size) {
+    auto block_id = batch_block_table[full_blocks_end];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + full_blocks_end * block_size * dst_entry_stride;
+    for (int eid = 0; eid < partial_block_size; ++eid) {
+      copy_entry(block_start_ptr + eid * cache_entry_stride,
+                 block_dst_ptr + eid * dst_entry_stride);
+    }
+  }
+}
+
+}  // namespace vllm
+
+// Macro to dispatch the kernel based on the data type.
+#define CALL_GATHER_CACHE(CPY_DTYPE)                                    \
+  vllm::gather_cache<CPY_DTYPE><<<grid, block, 0, stream>>>(            \
+      reinterpret_cast<CPY_DTYPE*>(src_cache.data_ptr()),               \
+      reinterpret_cast<CPY_DTYPE*>(dst.data_ptr()),                     \
+      block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
+      block_size, entry_size, block_table_stride, cache_block_stride,   \
+      cache_entry_stride, dst_entry_stride, seq_starts_ptr);
+
+// Gather sequences from the cache into the destination tensor.
+//  - cu_seq_lens contains the cumulative sequence lengths for each batch
+//  - block_table contains the cache block indices for each sequence
+//  - Optionally, seq_starts (if provided) offsets the starting block index by
+//  (seq_starts[bid] / page_size)
+void gather_cache(
+    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
+    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
+    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
+    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
+    int64_t batch_size,
+    std::optional<torch::Tensor> seq_starts = std::nullopt) {
+  at::cuda::OptionalCUDAGuard device_guard(src_cache.device());
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  int32_t block_size = src_cache.size(1);
+  int32_t entry_size = src_cache.flatten(2, -1).size(2);
+
+  TORCH_CHECK(block_table.dtype() == torch::kInt32,
+              "block_table must be int32");
+  TORCH_CHECK(cu_seq_lens.dtype() == torch::kInt32,
+              "cu_seq_lens must be int32");
+  if (seq_starts.has_value()) {
+    TORCH_CHECK(seq_starts.value().dtype() == torch::kInt32,
+                "seq_starts must be int32");
+  }
+
+  TORCH_CHECK(src_cache.device() == dst.device(),
+              "src_cache and dst must be on the same device");
+  TORCH_CHECK(src_cache.device() == block_table.device(),
+              "src_cache and block_table must be on the same device");
+  TORCH_CHECK(src_cache.device() == cu_seq_lens.device(),
+              "src_cache and cu_seq_lens must be on the same device");
+  if (seq_starts.has_value()) {
+    TORCH_CHECK(src_cache.device() == seq_starts.value().device(),
+                "src_cache and seq_starts must be on the same device");
+  }
+
+  int64_t block_table_stride = block_table.stride(0);
+  int64_t cache_block_stride = src_cache.stride(0);
+  int64_t cache_entry_stride = src_cache.stride(1);
+  int64_t dst_entry_stride = dst.stride(0);
+
+  // Decide on the number of splits based on the batch size.
+  int num_splits = batch_size > 128 ? 2 : batch_size > 64 ? 4 : 16;
+  dim3 grid(batch_size, num_splits);
+  dim3 block(1024);
+
+  TORCH_CHECK(src_cache.dtype() == dst.dtype(),
+              "src_cache and dst must have the same dtype");
+
+  const int dtype_bits = src_cache.element_size() * 8;
+  const int32_t* seq_starts_ptr =
+      seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;
+
+  if (dtype_bits == 32) {
+    CALL_GATHER_CACHE(uint32_t);
+  } else if (dtype_bits == 16) {
+    CALL_GATHER_CACHE(uint16_t);
+  } else if (dtype_bits == 8) {
+    CALL_GATHER_CACHE(uint8_t);
+  } else {
+    TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
+  }
+}

csrc/core/math.hpp

@@ -7,8 +7,3 @@ inline constexpr uint32_t next_pow_2(uint32_t const num) {
   if (num <= 1) return num;
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
-
-template <typename T>
-inline constexpr std::enable_if_t<std::is_integral_v<T>, T> ceil_div(T a, T b) {
-  return (a + b - 1) / b;
-}

csrc/cuda_utils.h

@@ -1,15 +1,41 @@
 #pragma once
 
-#if defined(__CUDACC__) || defined(_NVHPC_CUDA)
-  #define HOST_DEVICE_INLINE __forceinline__ __host__ __device__
-  #define DEVICE_INLINE __forceinline__ __device__
-  #define HOST_INLINE __forceinline__ __host__
+#include <stdio.h>
+
+#if defined(__HIPCC__)
+  #define HOST_DEVICE_INLINE __host__ __device__
+  #define DEVICE_INLINE __device__
+  #define HOST_INLINE __host__
+#elif defined(__CUDACC__) || defined(_NVHPC_CUDA)
+  #define HOST_DEVICE_INLINE __host__ __device__ __forceinline__
+  #define DEVICE_INLINE __device__ __forceinline__
+  #define HOST_INLINE __host__ __forceinline__
 #else
   #define HOST_DEVICE_INLINE inline
   #define DEVICE_INLINE inline
   #define HOST_INLINE inline
 #endif
 
+#define CUDA_CHECK(cmd)                                             \
+  do {                                                              \
+    cudaError_t e = cmd;                                            \
+    if (e != cudaSuccess) {                                         \
+      printf("Failed: Cuda error %s:%d '%s'\n", __FILE__, __LINE__, \
+             cudaGetErrorString(e));                                \
+      exit(EXIT_FAILURE);                                           \
+    }                                                               \
+  } while (0)
+
 int64_t get_device_attribute(int64_t attribute, int64_t device_id);
 
 int64_t get_max_shared_memory_per_block_device_attribute(int64_t device_id);
+
+namespace cuda_utils {
+
+template <typename T>
+HOST_DEVICE_INLINE constexpr std::enable_if_t<std::is_integral_v<T>, T>
+ceil_div(T a, T b) {
+  return (a + b - 1) / b;
+}
+
+};  // namespace cuda_utils

csrc/cuda_utils_kernels.cu

@@ -1,16 +1,22 @@
+#include "cuda_utils.h"
 #ifdef USE_ROCM
   #include <hip/hip_runtime.h>
   #include <hip/hip_runtime_api.h>
 #endif
+
 int64_t get_device_attribute(int64_t attribute, int64_t device_id) {
-  int device, value;
-  if (device_id < 0) {
-    cudaGetDevice(&device);
-  } else {
-    device = device_id;
-  }
-  cudaDeviceGetAttribute(&value, static_cast<cudaDeviceAttr>(attribute),
-                         device);
+  // Return the cached value on subsequent calls
+  static int value = [=]() {
+    int device = static_cast<int>(device_id);
+    if (device < 0) {
+      CUDA_CHECK(cudaGetDevice(&device));
+    }
+    int value;
+    CUDA_CHECK(cudaDeviceGetAttribute(
+        &value, static_cast<cudaDeviceAttr>(attribute), device));
+    return static_cast<int>(value);
+  }();
+
   return value;
 }
 

csrc/cumem_allocator.cpp

@@ -12,15 +12,21 @@ extern "C" {
 #include <cuda_runtime_api.h>
 #include <cuda.h>
 
-#define CUDA_CHECK(condition)                                                  \
-  do {                                                                         \
-    CUresult error = condition;                                                \
-    if (error != 0) {                                                          \
-      char* error_string;                                                      \
-      cuGetErrorString(error, (const char**)&error_string);                    \
-      std::cerr << "CUDA Error: " << error_string << " at " << __FILE__ << ":" \
-                << __LINE__ << std::endl;                                      \
-    }                                                                          \
+char error_msg[10240];  // 10KB buffer to store error messages
+CUresult no_error = CUresult(0);
+CUresult error_code = no_error;  // store error code
+
+#define CUDA_CHECK(condition)                                           \
+  do {                                                                  \
+    CUresult error = condition;                                         \
+    if (error != 0) {                                                   \
+      error_code = error;                                               \
+      char* error_string;                                               \
+      cuGetErrorString(error, (const char**)&error_string);             \
+      snprintf(error_msg, sizeof(error_msg), "CUDA Error: %s at %s:%d", \
+               error_string, __FILE__, __LINE__);                       \
+      std::cerr << error_msg << std::endl;                              \
+    }                                                                   \
   } while (0)
 
 // Global references to Python callables
@@ -54,14 +60,22 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
 
   // Allocate memory using cuMemCreate
   CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0));
+  if (error_code != 0) {
+    return;
+  }
   CUDA_CHECK(cuMemMap(d_mem, size, 0, *p_memHandle, 0));
-
+  if (error_code != 0) {
+    return;
+  }
   CUmemAccessDesc accessDesc = {};
   accessDesc.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
   accessDesc.location.id = device;
   accessDesc.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
 
   CUDA_CHECK(cuMemSetAccess(d_mem, size, &accessDesc, 1));
+  if (error_code != 0) {
+    return;
+  }
   // std::cout << "create_and_map: device=" << device << ", size=" << size << ",
   // d_mem=" << d_mem << ", p_memHandle=" << p_memHandle << std::endl;
 }
@@ -73,7 +87,13 @@ void unmap_and_release(unsigned long long device, ssize_t size,
   // ", d_mem=" << d_mem << ", p_memHandle=" << p_memHandle << std::endl;
   ensure_context(device);
   CUDA_CHECK(cuMemUnmap(d_mem, size));
+  if (error_code != 0) {
+    return;
+  }
   CUDA_CHECK(cuMemRelease(*p_memHandle));
+  if (error_code != 0) {
+    return;
+  }
 }
 
 PyObject* create_tuple_from_c_integers(unsigned long long a,
@@ -121,12 +141,16 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   size_t granularity;
   CUDA_CHECK(cuMemGetAllocationGranularity(&granularity, &prop,
                                            CU_MEM_ALLOC_GRANULARITY_MINIMUM));
-
+  if (error_code != 0) {
+    return nullptr;
+  }
   size_t alignedSize = ((size + granularity - 1) / granularity) * granularity;
 
   CUdeviceptr d_mem;
   CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, 0, 0, 0));
-
+  if (error_code != 0) {
+    return nullptr;
+  }
   // allocate the CUmemGenericAllocationHandle
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)malloc(
@@ -208,6 +232,9 @@ void my_free(void* ptr, ssize_t size, int device, CUstream stream) {
 
   // free address and the handle
   CUDA_CHECK(cuMemAddressFree(d_mem, size));
+  if (error_code != 0) {
+    return;
+  }
   free(p_memHandle);
 }
 
@@ -258,6 +285,12 @@ static PyObject* python_unmap_and_release(PyObject* self, PyObject* args) {
 
   unmap_and_release(recv_device, recv_size, d_mem_ptr, p_memHandle);
 
+  if (error_code != 0) {
+    error_code = no_error;
+    PyErr_SetString(PyExc_RuntimeError, error_msg);
+    return nullptr;
+  }
+
   Py_RETURN_NONE;
 }
 
@@ -282,6 +315,12 @@ static PyObject* python_create_and_map(PyObject* self, PyObject* args) {
 
   create_and_map(recv_device, recv_size, d_mem_ptr, p_memHandle);
 
+  if (error_code != 0) {
+    error_code = no_error;
+    PyErr_SetString(PyExc_RuntimeError, error_msg);
+    return nullptr;
+  }
+
   Py_RETURN_NONE;
 }
 

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c2x.hpp

@@ -122,8 +122,8 @@ struct ScaledEpilogue
     auto a_args = SUPER::template args_from_tensor<ScaleA, float>(a_scales);
     auto b_args = SUPER::template args_from_tensor<ScaleB, float>(b_scales);
 
-    typename EVTCompute0::Arguments evt0_args{b_args};
-    return ArgumentType{a_args, evt0_args};
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, {}};
   }
 };
 
@@ -167,8 +167,8 @@ struct ScaledEpilogueBias
     auto b_args = SUPER::template args_from_tensor<ScaleB, float>(b_scales);
     auto bias_args = SUPER::template args_from_tensor<Bias, ElementD>(bias);
 
-    typename EVTCompute0::Arguments evt0_args{b_args};
-    return ArgumentType{a_args, evt0_args, bias_args};
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, bias_args, {}};
   }
 };
 
@@ -230,9 +230,10 @@ struct ScaledEpilogueBiasAzp
     auto azp_adj_args =
         SUPER::template args_from_tensor<AzpWithAdj, int32_t>(azp_adj);
 
-    typename EVTComputeAzp::Arguments evt_azp_args{{}, azp_adj_args};
-    typename EVTComputeScaleB::Arguments evt_scale_b_args{b_args, evt_azp_args};
-    return ArgumentType{a_args, evt_scale_b_args, bias_args};
+    typename EVTComputeAzp::Arguments evt_azp_args{{}, azp_adj_args, {}};
+    typename EVTComputeScaleB::Arguments evt_scale_b_args{
+        b_args, evt_azp_args, {}};
+    return ArgumentType{a_args, evt_scale_b_args, bias_args, {}};
   }
 };
 
@@ -309,11 +310,12 @@ struct ScaledEpilogueBiasAzpToken
     auto azp_adj_args =
         SUPER::template args_from_tensor<AzpAdj, int32_t>(azp_adj);
 
-    typename EVTComputeAzp::Arguments evt_azp_args{azp_args, azp_adj_args};
-    typename EVTComputeAcc::Arguments evt_acc_args{{}, evt_azp_args};
-    typename EVTComputeScaleB::Arguments evt_scale_b_args{b_args, evt_acc_args};
-    return ArgumentType{a_args, evt_scale_b_args, bias_args};
+    typename EVTComputeAzp::Arguments evt_azp_args{azp_args, azp_adj_args, {}};
+    typename EVTComputeAcc::Arguments evt_acc_args{{}, evt_azp_args, {}};
+    typename EVTComputeScaleB::Arguments evt_scale_b_args{
+        b_args, evt_acc_args, {}};
+    return ArgumentType{a_args, evt_scale_b_args, bias_args, {}};
   }
 };
 
-};  // namespace vllm::c2x
+};  // namespace vllm::c2x

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp

@@ -16,6 +16,30 @@ namespace vllm::c3x {
 
 using namespace cute;
 
+template <typename T>
+struct identity {
+  CUTLASS_HOST_DEVICE
+  T operator()(T lhs) const { return lhs; }
+};
+
+template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
+struct TrivialEpilogue {
+ private:
+  using Accum = cutlass::epilogue::fusion::Sm90AccFetch;
+  using Compute = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::epilogue::thread::Identity, ElementD, ElementAcc,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+ public:
+  using EVTCompute = cutlass::epilogue::fusion::Sm90EVT<Compute, Accum>;
+  using ArgumentType = typename EVTCompute::Arguments;
+
+  template <typename... Args>
+  static ArgumentType prepare_args(Args... args) {
+    return {};
+  }
+};
+
 /*
  * This class provides the common load descriptors for the
  * ScaledEpilogue[...] classes
@@ -122,8 +146,8 @@ struct ScaledEpilogue
     auto a_args = SUPER::template args_from_tensor<ScaleA, float>(a_scales);
     auto b_args = SUPER::template args_from_tensor<ScaleB, float>(b_scales);
 
-    typename EVTCompute0::Arguments evt0_args{b_args};
-    return ArgumentType{a_args, evt0_args};
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, {}};
   }
 };
 
@@ -169,8 +193,51 @@ struct ScaledEpilogueBias
     auto b_args = SUPER::template args_from_tensor<ScaleB, float>(b_scales);
     auto bias_args = SUPER::template args_from_tensor<Bias, ElementD>(bias);
 
-    typename EVTCompute0::Arguments evt0_args{b_args};
-    return ArgumentType{a_args, evt0_args, bias_args};
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, bias_args, {}};
+  }
+};
+
+/*
+ * This epilogue performs the same operation as ScaledEpilogueBias, but the
+ * bias is a column vector instead of a row vector. Useful e.g. if we are
+ * computing a GEMM via C^T += B^T A^T. This happens in the 2:4 sparse kernels.
+ */
+template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
+struct ScaledEpilogueColumnBias
+    : private ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor> {
+ private:
+  using SUPER = ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::template ColOrScalarLoad<float>;
+  using ScaleB = typename SUPER::template RowOrScalarLoad<float>;
+  using Bias = typename SUPER::template ColLoad<ElementD>;
+
+  using Compute0 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiplies, float, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using EVTCompute0 =
+      cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
+
+  using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiply_add, ElementD, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+ public:
+  using EVTCompute =
+      cutlass::epilogue::fusion::Sm90EVT<Compute1, ScaleA, EVTCompute0, Bias>;
+
+  using ArgumentType = typename EVTCompute::Arguments;
+  static ArgumentType prepare_args(torch::Tensor const& a_scales,
+                                   torch::Tensor const& b_scales,
+                                   torch::Tensor const& bias) {
+    auto a_args = SUPER::template args_from_tensor<ScaleA, float>(a_scales);
+    auto b_args = SUPER::template args_from_tensor<ScaleB, float>(b_scales);
+    auto bias_args = SUPER::template args_from_tensor<Bias, ElementD>(bias);
+
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, bias_args, {}};
   }
 };
 
@@ -230,9 +297,10 @@ struct ScaledEpilogueBiasAzp
     auto azp_adj_args =
         SUPER::template args_from_tensor<AzpWithAdj, int32_t>(azp_adj);
 
-    typename EVTComputeAzp::Arguments evt_azp_args{{}, azp_adj_args};
-    typename EVTComputeScaleB::Arguments evt_scale_b_args{b_args, evt_azp_args};
-    return ArgumentType{a_args, evt_scale_b_args, bias_args};
+    typename EVTComputeAzp::Arguments evt_azp_args{{}, azp_adj_args, {}};
+    typename EVTComputeScaleB::Arguments evt_scale_b_args{
+        b_args, evt_azp_args, {}};
+    return ArgumentType{a_args, evt_scale_b_args, bias_args, {}};
   }
 };
 
@@ -307,11 +375,12 @@ struct ScaledEpilogueBiasAzpToken
     auto azp_adj_args =
         SUPER::template args_from_tensor<AzpAdj, int32_t>(azp_adj);
 
-    typename EVTComputeAzp::Arguments evt_azp_args{azp_args, azp_adj_args};
-    typename EVTComputeAcc::Arguments evt_acc_args{{}, evt_azp_args};
-    typename EVTComputeScaleB::Arguments evt_scale_b_args{b_args, evt_acc_args};
-    return ArgumentType{a_args, evt_scale_b_args, bias_args};
+    typename EVTComputeAzp::Arguments evt_azp_args{azp_args, azp_adj_args, {}};
+    typename EVTComputeAcc::Arguments evt_acc_args{{}, evt_azp_args, {}};
+    typename EVTComputeScaleB::Arguments evt_scale_b_args{
+        b_args, evt_acc_args, {}};
+    return ArgumentType{a_args, evt_scale_b_args, bias_args, {}};
   }
 };
 
-};  // namespace vllm::c3x
+};  // namespace vllm::c3x

csrc/moe/moe_align_sum_kernels.cu

@@ -3,7 +3,7 @@
 #include <c10/cuda/CUDAGuard.h>
 
 #include <ATen/ATen.h>
-#include <THC/THCAtomics.cuh>
+#include <ATen/cuda/Atomic.cuh>
 
 #include "../cuda_compat.h"
 #include "../dispatch_utils.h"
@@ -198,26 +198,27 @@ __global__ void moe_align_block_size_global_mem_kernel(
 }
 
 // taken from
-// https://github.com/sgl-project/sglang/commit/ded9fcd09a43d5e7d5bb31a2bc3e9fc21bf65d2a
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
 template <typename scalar_t>
 __global__ void sgl_moe_align_block_size_kernel(
     scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
     int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
     int32_t block_size, size_t numel, int32_t* cumsum) {
   __shared__ int32_t shared_counts[32][8];
-  __shared__ int32_t local_offsets[256];
 
   const int warp_id = threadIdx.x / 32;
-  const int lane_id = threadIdx.x % 32;
   const int experts_per_warp = 8;
   const int my_expert_start = warp_id * experts_per_warp;
 
+  // Initialize shared_counts for this warp's experts
   for (int i = 0; i < experts_per_warp; ++i) {
     if (my_expert_start + i < num_experts) {
       shared_counts[warp_id][i] = 0;
     }
   }
 
+  __syncthreads();
+
   const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
   const size_t start_idx = threadIdx.x * tokens_per_thread;
 
@@ -230,6 +231,7 @@ __global__ void sgl_moe_align_block_size_kernel(
 
   __syncthreads();
 
+  // Single thread computes cumulative sum and total tokens
   if (threadIdx.x == 0) {
     cumsum[0] = 0;
     for (int i = 1; i <= num_experts; ++i) {
@@ -246,19 +248,28 @@ __global__ void sgl_moe_align_block_size_kernel(
 
   __syncthreads();
 
+  // Assign expert IDs to blocks
   if (threadIdx.x < num_experts) {
     for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
          i += block_size) {
       expert_ids[i / block_size] = threadIdx.x;
     }
-    local_offsets[threadIdx.x] = cumsum[threadIdx.x];
   }
+}
 
-  __syncthreads();
+// taken from
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
+template <typename scalar_t>
+__global__ void sgl_moe_token_sort_kernel(scalar_t* __restrict__ topk_ids,
+                                          int32_t* sorted_token_ids,
+                                          int32_t* cumsum_buffer,
+                                          size_t numel) {
+  const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const size_t stride = blockDim.x * gridDim.x;
 
-  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+  for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
-    int32_t rank_post_pad = atomicAdd(&local_offsets[expert_id], 1);
+    int32_t rank_post_pad = atomicAdd(&cumsum_buffer[expert_id], 1);
     sorted_token_ids[rank_post_pad] = i;
   }
 }
@@ -377,23 +388,34 @@ void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
                               torch::Tensor experts_ids,
                               torch::Tensor num_tokens_post_pad) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  TORCH_CHECK(num_experts == 256,
+              "sgl_moe_align_block_size kernel only supports deepseek v3.");
+
   VLLM_DISPATCH_INTEGRAL_TYPES(
       topk_ids.scalar_type(), "sgl_moe_align_block_size_kernel", [&] {
-        // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
-        // tensors
+        // calc needed amount of shared mem for `cumsum` tensors
         auto options_int =
             torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
-        // torch::Tensor token_cnts_buffer =
-        //     torch::empty({(num_experts + 1) * num_experts}, options_int);
         torch::Tensor cumsum_buffer =
-            torch::empty({num_experts + 1}, options_int);
+            torch::zeros({num_experts + 1}, options_int);
 
-        auto kernel = vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
-        kernel<<<1, 1024, 0, stream>>>(
+        auto align_kernel =
+            vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
+        align_kernel<<<1, 1024, 0, stream>>>(
             topk_ids.data_ptr<scalar_t>(), 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(), cumsum_buffer.data_ptr<int32_t>());
+
+        const int block_threads = 256;
+        const int num_blocks =
+            (topk_ids.numel() + block_threads - 1) / block_threads;
+        const int max_blocks = 65535;
+        const int actual_blocks = std::min(num_blocks, max_blocks);
+        auto sort_kernel = vllm::moe::sgl_moe_token_sort_kernel<scalar_t>;
+        sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
+            cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
       });
 }
 

csrc/ops.h

@@ -152,6 +152,11 @@ torch::Tensor ggml_mul_mat_a8(torch::Tensor W, torch::Tensor X, int64_t type,
                               int64_t row);
 
 #ifndef USE_ROCM
+void cutlass_scaled_fp4_mm(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);
+
 bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
 bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability);
 
@@ -176,8 +181,11 @@ void cutlass_scaled_sparse_mm(torch::Tensor& out, torch::Tensor const& a,
                               torch::Tensor const& b_scales,
                               std::optional<torch::Tensor> const& bias);
 
-bool cutlass_sparse_compress_entry(torch::Tensor& a_compressed,
-                                   torch::Tensor& e, torch::Tensor const& a);
+std::vector<torch::Tensor> cutlass_sparse_compress(torch::Tensor const& a);
+
+void scaled_fp4_quant(torch::Tensor& output, torch::Tensor const& input,
+                      torch::Tensor& output_scale,
+                      torch::Tensor const& input_scale);
 #endif
 
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,

csrc/pos_encoding_kernels.cu

@@ -124,18 +124,54 @@ __global__ void batched_rotary_embedding_kernel(
 void rotary_embedding(
     torch::Tensor& positions,  // [batch_size, seq_len] or [num_tokens]
     torch::Tensor& query,  // [batch_size, seq_len, num_heads * head_size] or
-                           // [num_tokens, num_heads * head_size]
+                           // [num_tokens, num_heads * head_size] or
+                           // [batch_size, seq_len, num_heads, head_size] or
+                           // [num_tokens, num_heads, head_size]
     torch::Tensor& key,    // [batch_size, seq_len, num_kv_heads * head_size] or
-                           // [num_tokens, num_kv_heads * head_size]
+                           // [num_tokens, num_kv_heads * head_size] or
+                           // [batch_size, seq_len, num_heads, head_size] or
+                           // [num_tokens, num_heads, head_size]
     int64_t head_size,
     torch::Tensor& cos_sin_cache,  // [max_position, rot_dim]
     bool is_neox) {
-  int64_t num_tokens = query.numel() / query.size(-1);
+  // num_tokens = batch_size * seq_len
+  int64_t num_tokens = positions.numel();
+  int positions_ndim = positions.dim();
+
+  // Make sure num_tokens dim is consistent across positions, query, and key.
+  TORCH_CHECK(
+      positions_ndim == 1 || positions_ndim == 2,
+      "positions must have shape [num_tokens] or [batch_size, seq_len]");
+  if (positions_ndim == 1) {
+    TORCH_CHECK(
+        query.size(0) == positions.size(0) && key.size(0) == positions.size(0),
+        "query, key and positions must have the same number of tokens");
+  }
+  if (positions_ndim == 2) {
+    TORCH_CHECK(
+        query.size(0) == positions.size(0) &&
+            key.size(0) == positions.size(0) &&
+            query.size(1) == positions.size(1) &&
+            key.size(1) == positions.size(1),
+        "query, key and positions must have the same batch_size and seq_len");
+  }
+
+  // Make sure head_size is valid for query and key
+  // hidden_size = num_heads * head_size
+  int query_hidden_size = query.numel() / num_tokens;
+  int key_hidden_size = key.numel() / num_tokens;
+  TORCH_CHECK(query_hidden_size % head_size == 0);
+  TORCH_CHECK(key_hidden_size % head_size == 0);
+
+  // Make sure query and key have consistent number of heads
+  int num_heads = query_hidden_size / head_size;
+  int num_kv_heads = key_hidden_size / head_size;
+  TORCH_CHECK(num_heads % num_kv_heads == 0);
+
   int rot_dim = cos_sin_cache.size(1);
-  int num_heads = query.size(-1) / head_size;
-  int num_kv_heads = key.size(-1) / head_size;
-  int64_t query_stride = query.stride(-2);
-  int64_t key_stride = key.stride(-2);
+  int seq_dim_idx = positions_ndim - 1;
+  int64_t query_stride = query.stride(seq_dim_idx);
+  int64_t key_stride = key.stride(seq_dim_idx);
 
   dim3 grid(num_tokens);
   dim3 block(std::min<int64_t>(num_heads * rot_dim / 2, 512));
@@ -165,19 +201,58 @@ and process in batched manner.
 void batched_rotary_embedding(
     torch::Tensor& positions,  // [batch_size, seq_len] or [num_tokens]
     torch::Tensor& query,  // [batch_size, seq_len, num_heads * head_size] or
-                           // [num_tokens, num_heads * head_size]
+                           // [num_tokens, num_heads * head_size] or
+                           // [batch_size, seq_len, num_heads, head_size] or
+                           // [num_tokens, num_heads, head_size]
     torch::Tensor& key,    // [batch_size, seq_len, num_kv_heads * head_size] or
-                           // [num_tokens, num_kv_heads * head_size]
+                           // [num_tokens, num_kv_heads * head_size] or
+                           // [batch_size, seq_len, num_heads, head_size] or
+                           // [num_tokens, num_heads, head_size]
     int64_t head_size,
     torch::Tensor& cos_sin_cache,  // [max_position, rot_dim]
     bool is_neox, int64_t rot_dim,
-    torch::Tensor& cos_sin_cache_offsets  // [num_tokens]
+    torch::Tensor& cos_sin_cache_offsets  // [num_tokens] or [batch_size]
 ) {
+  // num_tokens = batch_size * seq_len
   int64_t num_tokens = cos_sin_cache_offsets.size(0);
-  int num_heads = query.size(-1) / head_size;
-  int num_kv_heads = key.size(-1) / head_size;
-  int64_t query_stride = query.stride(-2);
-  int64_t key_stride = key.stride(-2);
+  TORCH_CHECK(
+      positions.size(0) == num_tokens || positions.numel() == num_tokens,
+      "positions must have the same num_tokens or batch_size as "
+      "cos_sin_cache_offsets");
+
+  int positions_ndim = positions.dim();
+  // Make sure num_tokens dim is consistent across positions, query, and key.
+  TORCH_CHECK(
+      positions_ndim == 1 || positions_ndim == 2,
+      "positions must have shape [num_tokens] or [batch_size, seq_len]");
+  if (positions_ndim == 1) {
+    TORCH_CHECK(
+        query.size(0) == positions.size(0) && key.size(0) == positions.size(0),
+        "query, key and positions must have the same number of tokens");
+  }
+  if (positions_ndim == 2) {
+    TORCH_CHECK(
+        query.size(0) == positions.size(0) &&
+            key.size(0) == positions.size(0) &&
+            query.size(1) == positions.size(1) &&
+            key.size(1) == positions.size(1),
+        "query, key and positions must have the same batch_size and seq_len");
+  }
+
+  // Make sure head_size is valid for query and key
+  int query_hidden_size = query.numel() / num_tokens;
+  int key_hidden_size = key.numel() / num_tokens;
+  TORCH_CHECK(query_hidden_size % head_size == 0);
+  TORCH_CHECK(key_hidden_size % head_size == 0);
+
+  // Make sure query and key have concistent number of heads
+  int num_heads = query_hidden_size / head_size;
+  int num_kv_heads = key_hidden_size / head_size;
+  TORCH_CHECK(num_heads % num_kv_heads == 0);
+
+  int seq_dim_idx = positions_ndim - 1;
+  int64_t query_stride = query.stride(seq_dim_idx);
+  int64_t key_stride = key.stride(seq_dim_idx);
 
   dim3 grid(num_tokens);
   dim3 block(std::min<int64_t>(num_heads * rot_dim / 2, 512));

csrc/quantization/awq/gemm_kernels.cu

@@ -334,7 +334,7 @@ __global__ void __launch_bounds__(64)
   }
 
   // TODO: Shang: Hoist loop invariance.
-  for (int ax1_0_1 = 0; ax1_0_1 < 4; ++ax1_0_1) {
+  for (int ax1_0_1 = 0; ax1_0_1 < (N / 32); ++ax1_0_1) {
     for (int local_id = 0; local_id < 8; ++local_id) {
       int row_offset = (((int)blockIdx_y) / j_factors1) * 16 +
                        ((int)threadIdx.x) / 4 + (local_id % 4) / 2 * 8;

csrc/quantization/cutlass_w8a8/Epilogues.md

@@ -1,17 +1,19 @@
 # CUTLASS Epilogues
 
 ## Introduction
-This document describes the various CUTLASS epilogues implemented for fusing de-quantization operations onto GEMMs. 
+
+This document describes the various CUTLASS epilogues implemented for fusing de-quantization operations onto GEMMs.
 
 Currently, we only support symmetric quantization for weights,
 and symmetric and asymmetric quantization for activations.
 Both can be quantized per-tensor or per-channel (weights) / per-token (activations).
 
 There are 4 epilogues:
-1. ScaledEpilogue: symmetric quantization for activations, no bias.
-1. ScaledEpilogueBias: symmetric quantization for activations, supports bias.
-1. ScaledEpilogueAzp: asymmetric per-tensor quantization for activations, supports bias.
-1. ScaledEpilogueAzpPerToken: asymmetric per-token quantization for activations, supports bias.
+
+1. `ScaledEpilogue`: symmetric quantization for activations, no bias.
+1. `ScaledEpilogueBias`: symmetric quantization for activations, supports bias.
+1. `ScaledEpilogueAzp`: asymmetric per-tensor quantization for activations, supports bias.
+1. `ScaledEpilogueAzpPerToken`: asymmetric per-token quantization for activations, supports bias.
 
 We do not have epilogues for asymmetric quantization of activations without bias in order to reduce final binary size.
 Instead, if no bias is passed, the epilogue will use 0 as the bias.
@@ -26,12 +28,15 @@ If $` \widehat X `$ is the quantized $` X `$, our matrices become the following
 ```math
 A = s_a (\widehat A - J_a z_a)
 ```
+
 ```math
 B = s_b \widehat B
 ```
+
 ```math
 D = A B + C
 ```
+
 ```math
 D = s_a s_b \widehat D + C
 ```
@@ -48,9 +53,11 @@ Expanding further, we can calculate $` \widehat D `$ as follows:
 ```math
 A B = s_a ( \widehat A - J_a z_a ) s_b \widehat B
 ```
+
 ```math
 A B = s_a s_b \left( \widehat A \widehat B - J_a z_a \widehat B \right)
 ```
+
 ```math
 \widehat D = \widehat A \widehat B - z_a J_a \widehat B
 ```
@@ -61,16 +68,19 @@ Each row of it is equal to $` \mathbf 1 \widehat B `$, which is a row-vector of
 
 ## Epilogues
 
-### ScaledEpilogue
+### `ScaledEpilogue`
+
 This epilogue computes the symmetric quantization for activations without bias, meaning $` C = 0 `$ and $` z_a = 0 `$.
 The output of the GEMM is:
 
 ```math
 \widehat D = \widehat A \widehat B
 ```
+
 ```math
 D = s_a s_b \widehat D
 ```
+
 ```math
 D = s_a s_b \widehat A \widehat B
 ```
@@ -79,44 +89,51 @@ Epilogue parameters:
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
 
-### ScaledEpilogueBias
+### `ScaledEpilogueBias`
+
 This epilogue computes the symmetric quantization for activations with bias, meaning $` z_a = 0 `$.
 The output of the GEMM is:
 
 ```math
 \widehat D = \widehat A \widehat B
 ```
+
 ```math
 D = s_a s_b \widehat D + C 
 ```
+
 ```math
 D = s_a s_b \widehat A \widehat B + C
 ```
 
-
 Epilogue parameters:
+
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
 - `bias` is the bias, is always per-channel (row-vector).
 
-### ScaledEpilogueAzp
+### `ScaledEpilogueAzp`
+
 This epilogue computes the asymmetric per-tensor quantization for activations with bias.
 The output of the GEMM is:
 
 ```math
 \widehat D = \widehat A \widehat B - z_a J_a \widehat B
 ```
+
 ```math
 D = s_a s_b \widehat D + C 
 ```
+
 ```math
 D = s_a s_b \left( \widehat A \widehat B - z_a J_a \widehat B \right) + C
 ```
 
-Because $` z_a `$ is a scalar, the zero-point term $` z_a J_a \widehat B `$ has every row equal to $` z_a \mathbf 1 B `$. 
+Because $` z_a `$ is a scalar, the zero-point term $` z_a J_a \widehat B `$ has every row equal to $` z_a \mathbf 1 B `$.
 That is precomputed and stored in `azp_with_adj` as a row-vector.
 
 Epilogue parameters:
+
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
   - Generally this will be per-tensor as the zero-points are per-tensor.
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
@@ -125,13 +142,15 @@ Epilogue parameters:
 
 To use these kernels efficiently, users must precompute the `azp_with_adj` term offline and pass it to the kernel.
 
-### ScaledEpilogueAzpPerToken
+### `ScaledEpilogueAzpPerToken`
+
 This epilogue computes the asymmetric per-token quantization for activations with bias.
 
 The output of the GEMM is the same as above, but the $` z_a `$ is a column-vector.
 That means the zero-point term $` z_a J_a \widehat B `$ becomes an outer product of $` z_a `$ and $` \mathbf 1 \widehat B `$.
 
 Epilogue parameters:
+
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
   - Generally this will be per-token as the zero-points are per-token.
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
@@ -142,6 +161,7 @@ Epilogue parameters:
 To use these kernels efficiently, users must precompute the `azp_adj` term offline and pass it to the kernel.
 
 The epilogue performs the following computation (where `Dq` is the raw quantized output of the GEMM):
-```
+
+```math
 out = scale_a * scale_b * (Dq - azp_adj * azp) + bias
 ```

csrc/quantization/cutlass_w8a8/c3x/cutlass_gemm_caller.cuh

@@ -30,12 +30,18 @@ static inline cute::Shape<int, int, int, int> get_problem_shape(
 }
 
 template <typename GemmKernel>
-void cutlass_gemm_caller(torch::Device device,
-                         cute::Shape<int, int, int, int> prob_shape,
-                         typename GemmKernel::MainloopArguments mainloop_args,
-                         typename GemmKernel::EpilogueArguments epilogue_args) {
+void cutlass_gemm_caller(
+    torch::Device device, cute::Shape<int, int, int, int> prob_shape,
+    typename GemmKernel::MainloopArguments mainloop_args,
+    typename GemmKernel::EpilogueArguments epilogue_args,
+    typename GemmKernel::TileSchedulerArguments scheduler = {}) {
+  cutlass::KernelHardwareInfo hw_info;
   typename GemmKernel::Arguments args{cutlass::gemm::GemmUniversalMode::kGemm,
-                                      prob_shape, mainloop_args, epilogue_args};
+                                      prob_shape,
+                                      mainloop_args,
+                                      epilogue_args,
+                                      hw_info,
+                                      scheduler};
 
   // Launch the CUTLASS GEMM kernel.
   using GemmOp = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;

csrc/quantization/cutlass_w8a8/c3x/scaled_mm.cuh

@@ -53,12 +53,17 @@ struct cutlass_3x_gemm {
 
   using EVTCompute = typename Epilogue::EVTCompute;
 
+  // These are the minimum alignments needed for the kernels to compile
+  static constexpr int AlignmentAB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+  static constexpr int AlignmentCD = 4;
+
   using CollectiveEpilogue =
       typename cutlass::epilogue::collective::CollectiveBuilder<
           cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp, TileShape,
           ClusterShape, cutlass::epilogue::collective::EpilogueTileAuto,
-          ElementAcc, float, ElementC, StrideC, 4, ElementD, StrideD, 4,
-          EpilogueSchedule, EVTCompute>::CollectiveOp;
+          ElementAcc, float, ElementC, StrideC, AlignmentCD, ElementD, StrideD,
+          AlignmentCD, EpilogueSchedule, EVTCompute>::CollectiveOp;
 
   static constexpr size_t CEStorageSize =
       sizeof(typename CollectiveEpilogue::SharedStorage);
@@ -69,8 +74,8 @@ struct cutlass_3x_gemm {
   using CollectiveMainloop =
       typename cutlass::gemm::collective::CollectiveBuilder<
           cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp, 
-          ElementAB, cutlass::layout::RowMajor, 16, 
-          ElementAB, cutlass::layout::ColumnMajor, 16, 
+          ElementAB, cutlass::layout::RowMajor, AlignmentAB, 
+          ElementAB, cutlass::layout::ColumnMajor, AlignmentAB, 
           ElementAcc, TileShape, ClusterShape,
           Stages,
           KernelSchedule>::CollectiveOp;

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8_dispatch.cuh

@@ -22,8 +22,9 @@ namespace vllm {
 
 using namespace cute;
 
-template <typename OutType, int GroupSizeM_, int GroupSizeN_, int GroupSizeK_,
-          int TileSizeM_ = 128, class ClusterShape = Shape<_1, _2, _1>>
+template <typename SchedulerType, typename OutType, int GroupSizeM_,
+          int GroupSizeN_, int GroupSizeK_, int TileSizeM_ = 128,
+          class ClusterShape = Shape<_1, _2, _1>>
 struct cutlass_3x_gemm_fp8_blockwise {
   using GroupSizeM = Int<GroupSizeM_>;
   using GroupSizeN = Int<GroupSizeN_>;
@@ -84,7 +85,7 @@ struct cutlass_3x_gemm_fp8_blockwise {
 
   using KernelType = enable_sm90_or_later<cutlass::gemm::kernel::GemmUniversal<
       Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue,
-      cutlass::gemm::PersistentScheduler>>;
+      SchedulerType>>;
 
   struct GemmKernel : public KernelType {};
 
@@ -150,8 +151,24 @@ void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
   typename GemmKernel::EpilogueArguments epilogue_args{
       {}, c_ptr, c_stride, c_ptr, c_stride};
 
+  typename GemmKernel::TileSchedulerArguments scheduler;
+
+  static constexpr bool UsesStreamKScheduler =
+      cute::is_same_v<typename GemmKernel::TileSchedulerTag,
+                      cutlass::gemm::StreamKScheduler>;
+
+  if constexpr (UsesStreamKScheduler) {
+    using DecompositionMode = typename cutlass::gemm::kernel::detail::
+        PersistentTileSchedulerSm90StreamKParams::DecompositionMode;
+    using ReductionMode = typename cutlass::gemm::kernel::detail::
+        PersistentTileSchedulerSm90StreamKParams::ReductionMode;
+
+    scheduler.decomposition_mode = DecompositionMode::StreamK;
+    scheduler.reduction_mode = ReductionMode::Nondeterministic;
+  }
+
   c3x::cutlass_gemm_caller<GemmKernel>(a.device(), prob_shape, mainloop_args,
-                                       epilogue_args);
+                                       epilogue_args, scheduler);
 }
 
 template <typename OutType>
@@ -160,9 +177,18 @@ void cutlass_gemm_blockwise_sm90_fp8_dispatch(torch::Tensor& out,
                                               torch::Tensor const& b,
                                               torch::Tensor const& a_scales,
                                               torch::Tensor const& b_scales) {
-  cutlass_gemm_caller_blockwise<
-      cutlass_3x_gemm_fp8_blockwise<OutType, 1, 128, 128>>(out, a, b, a_scales,
-                                                           b_scales);
+  auto k = a.size(1);
+  auto n = b.size(1);
+
+  if (k > 3 * n) {
+    cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
+        cutlass::gemm::StreamKScheduler, OutType, 1, 128, 128>>(
+        out, a, b, a_scales, b_scales);
+  } else {
+    cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
+        cutlass::gemm::PersistentScheduler, OutType, 1, 128, 128>>(
+        out, a, b, a_scales, b_scales);
+  }
 }
 
 }  // namespace vllm

csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cuh

@@ -103,14 +103,19 @@ struct cutlass_2x_gemm {
 
   using EVTD = cutlass::epilogue::threadblock::Sm80EVT<D, EVTCompute>;
 
+  // These are the minimum alignments needed for the kernels to compile
+  static constexpr int AlignmentAB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+  static constexpr int AlignmentCD = 4;
+
   // clang-format off
   using RowMajor = typename cutlass::layout::RowMajor;
   using ColumnMajor = typename cutlass::layout::ColumnMajor;
   using KernelType =
     ArchGuard<typename cutlass::gemm::kernel::DefaultGemmWithVisitor<
-      ElementAB, RowMajor, cutlass::ComplexTransform::kNone, 16,
-      ElementAB, ColumnMajor, cutlass::ComplexTransform::kNone, 16,
-      float, cutlass::layout::RowMajor, 4,
+      ElementAB, RowMajor, cutlass::ComplexTransform::kNone, AlignmentAB,
+      ElementAB, ColumnMajor, cutlass::ComplexTransform::kNone, AlignmentAB,
+      float, cutlass::layout::RowMajor, AlignmentCD,
       ElementAcc, float, cutlass::arch::OpClassTensorOp,
       Arch,
       TileShape, WarpShape, InstructionShape,

csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu

@@ -1,7 +1,7 @@
 #include <cudaTypedefs.h>
 #include "c3x/scaled_mm_kernels.hpp"
 
-#include "core/math.hpp"
+#include "cuda_utils.h"
 
 /*
    This file defines quantized GEMM operations using the CUTLASS 3.x API, for
@@ -33,7 +33,8 @@ void cutlass_scaled_mm_sm90(torch::Tensor& c, torch::Tensor const& a,
     auto make_group_shape = [](torch::Tensor const& x,
                                torch::Tensor const& s) -> GroupShape {
       TORCH_CHECK(s.dim() == 2, "cutlass_scaled_mm group scales must be 2D");
-      return {ceil_div(x.size(0), s.size(0)), ceil_div(x.size(1), s.size(1))};
+      return {cuda_utils::ceil_div(x.size(0), s.size(0)),
+              cuda_utils::ceil_div(x.size(1), s.size(1))};
     };
 
     GroupShape a_scale_group_shape = make_group_shape(a, a_scales);

csrc/quantization/fp4/nvfp4_quant_entry.cu

@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2025, NVIDIA CORPORATION.  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.
+ */
+
+#include <torch/all.h>
+
+#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+void scaled_fp4_quant_sm100a(torch::Tensor const& output,
+                             torch::Tensor const& input,
+                             torch::Tensor const& output_sf,
+                             torch::Tensor const& input_sf);
+#endif
+
+void scaled_fp4_quant(torch::Tensor& output, torch::Tensor const& input,
+                      torch::Tensor& output_sf, torch::Tensor const& input_sf) {
+#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+  return scaled_fp4_quant_sm100a(output, input, output_sf, input_sf);
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled nvfp4 quantization");
+}

csrc/quantization/fp4/nvfp4_quant_kernels.cu

@@ -0,0 +1,376 @@
+/*
+ * Copyright (c) 2025, NVIDIA CORPORATION.  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.
+ */
+
+#include <torch/all.h>
+
+#include <cuda_runtime_api.h>
+#include <cuda_runtime.h>
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <cuda_fp8.h>
+
+#include "cuda_utils.h"
+
+// Get type2 from type or vice versa (applied to half and bfloat16)
+template <typename T>
+struct TypeConverter {
+  using Type = half2;
+};  // keep for generality
+
+template <>
+struct TypeConverter<half2> {
+  using Type = half;
+};
+
+template <>
+struct TypeConverter<half> {
+  using Type = half2;
+};
+
+template <>
+struct TypeConverter<__nv_bfloat162> {
+  using Type = __nv_bfloat16;
+};
+
+template <>
+struct TypeConverter<__nv_bfloat16> {
+  using Type = __nv_bfloat162;
+};
+
+#define ELTS_PER_THREAD 8
+
+constexpr int CVT_FP4_ELTS_PER_THREAD = 8;
+constexpr int CVT_FP4_SF_VEC_SIZE = 16;
+
+// Convert 8 float32 values into 8 e2m1 values (represented as one uint32_t).
+inline __device__ uint32_t fp32_vec_to_e2m1(float (&array)[8]) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000)
+  uint32_t val;
+  asm volatile(
+      "{\n"
+      ".reg .b8 byte0;\n"
+      ".reg .b8 byte1;\n"
+      ".reg .b8 byte2;\n"
+      ".reg .b8 byte3;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte0, %2, %1;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte1, %4, %3;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte2, %6, %5;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte3, %8, %7;\n"
+      "mov.b32 %0, {byte0, byte1, byte2, byte3};\n"
+      "}"
+      : "=r"(val)
+      : "f"(array[0]), "f"(array[1]), "f"(array[2]), "f"(array[3]),
+        "f"(array[4]), "f"(array[5]), "f"(array[6]), "f"(array[7]));
+  return val;
+#else
+  return 0;
+#endif
+}
+
+// Convert 4 float2 values into 8 e2m1 values (represented as one uint32_t).
+inline __device__ uint32_t fp32_vec_to_e2m1(float2 (&array)[4]) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000)
+  uint32_t val;
+  asm volatile(
+      "{\n"
+      ".reg .b8 byte0;\n"
+      ".reg .b8 byte1;\n"
+      ".reg .b8 byte2;\n"
+      ".reg .b8 byte3;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte0, %2, %1;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte1, %4, %3;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte2, %6, %5;\n"
+      "cvt.rn.satfinite.e2m1x2.f32   byte3, %8, %7;\n"
+      "mov.b32 %0, {byte0, byte1, byte2, byte3};\n"
+      "}"
+      : "=r"(val)
+      : "f"(array[0].x), "f"(array[0].y), "f"(array[1].x), "f"(array[1].y),
+        "f"(array[2].x), "f"(array[2].y), "f"(array[3].x), "f"(array[3].y));
+  return val;
+#else
+  return 0;
+#endif
+}
+
+// Fast reciprocal.
+inline __device__ float reciprocal_approximate_ftz(float a) {
+  float b;
+  asm volatile("rcp.approx.ftz.f32 %0, %1;\n" : "=f"(b) : "f"(a));
+  return b;
+}
+
+template <class SFType, int CVT_FP4_NUM_THREADS_PER_SF>
+__device__ uint8_t* cvt_quant_to_fp4_get_sf_out_offset(int rowIdx, int colIdx,
+                                                       int numCols,
+                                                       SFType* SFout) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000)
+  static_assert(CVT_FP4_NUM_THREADS_PER_SF == 1 ||
+                CVT_FP4_NUM_THREADS_PER_SF == 2);
+
+  // One pair of threads write one SF to global memory.
+  // TODO: stage through smem for packed STG.32
+  // is it better than STG.8 from 4 threads ?
+  if (threadIdx.x % CVT_FP4_NUM_THREADS_PER_SF == 0) {
+    // SF vector index (16 elements share one SF in the K dimension).
+    int32_t kIdx = colIdx / CVT_FP4_NUM_THREADS_PER_SF;
+    int32_t mIdx = rowIdx;
+
+    // SF layout [numMTiles, numKTiles, 32 (mTile), 4 (mTile), 4(kTile)]
+    // --> index [mTileIdx, kTileIdx, outerMIdx, innerMIdx, innerKIdx]
+
+    int32_t mTileIdx = mIdx / (32 * 4);
+    // SF vector size 16.
+    int factor = CVT_FP4_SF_VEC_SIZE * 4;
+    int32_t numKTiles = (numCols + factor - 1) / factor;
+    int64_t mTileStride = numKTiles * 32 * 4 * 4;
+
+    int32_t kTileIdx = (kIdx / 4);
+    int64_t kTileStride = 32 * 4 * 4;
+
+    // M tile layout [32, 4] is column-major.
+    int32_t outerMIdx = (mIdx % 32);
+    int64_t outerMStride = 4 * 4;
+
+    int32_t innerMIdx = (mIdx % (32 * 4)) / 32;
+    int64_t innerMStride = 4;
+
+    int32_t innerKIdx = (kIdx % 4);
+    int64_t innerKStride = 1;
+
+    // Compute the global offset.
+    int64_t SFOffset = mTileIdx * mTileStride + kTileIdx * kTileStride +
+                       outerMIdx * outerMStride + innerMIdx * innerMStride +
+                       innerKIdx * innerKStride;
+
+    return reinterpret_cast<uint8_t*>(SFout) + SFOffset;
+  }
+#endif
+  return nullptr;
+}
+
+// Define a 16 bytes packed data type.
+template <class Type>
+struct PackedVec {
+  typename TypeConverter<Type>::Type elts[4];
+};
+
+template <>
+struct PackedVec<__nv_fp8_e4m3> {
+  __nv_fp8x2_e4m3 elts[8];
+};
+
+// Quantizes the provided PackedVec into the uint32_t output
+template <class Type, bool UE8M0_SF = false>
+__device__ uint32_t cvt_warp_fp16_to_fp4(PackedVec<Type>& vec, float SFScaleVal,
+                                         uint8_t* SFout) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000)
+  // Get absolute maximum values among the local 8 values.
+  auto localMax = __habs2(vec.elts[0]);
+
+  // Local maximum value.
+  #pragma unroll
+  for (int i = 1; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) {
+    localMax = __hmax2(localMax, __habs2(vec.elts[i]));
+  }
+
+  // Get the absolute maximum among all 16 values (two threads).
+  localMax = __hmax2(__shfl_xor_sync(uint32_t(-1), localMax, 1), localMax);
+  // Get the final absolute maximum values.
+  float vecMax = float(__hmax(localMax.x, localMax.y));
+
+  // Get the SF (max value of the vector / max value of e2m1).
+  // maximum value of e2m1 = 6.0.
+  // TODO: use half as compute data type.
+  float SFValue = SFScaleVal * (vecMax * reciprocal_approximate_ftz(6.0f));
+  // 8 bits representation of the SF.
+  uint8_t fp8SFVal;
+  // Write the SF to global memory (STG.8).
+  if constexpr (UE8M0_SF) {
+    // Extract the 8 exponent bits from float32.
+    // float 32bits = 1 sign bit + 8 exponent bits + 23 mantissa bits.
+    uint32_t tmp = reinterpret_cast<uint32_t&>(SFValue) >> 23;
+    fp8SFVal = tmp & 0xff;
+    // Convert back to fp32.
+    reinterpret_cast<uint32_t&>(SFValue) = tmp << 23;
+  } else {
+    // Here SFValue is always positive, so E4M3 is the same as UE4M3.
+    __nv_fp8_e4m3 tmp = __nv_fp8_e4m3(SFValue);
+    reinterpret_cast<__nv_fp8_e4m3&>(fp8SFVal) = tmp;
+    // Convert back to fp32.
+    SFValue = float(tmp);
+  }
+  // Get the output scale.
+  // Recipe: final_scale = reciprocal(fp32(fp8(SFValue * SFScaleVal))) *
+  //                       reciprocal(SFScaleVal))
+  float outputScale =
+      SFValue != 0 ? reciprocal_approximate_ftz(
+                         SFValue * reciprocal_approximate_ftz(SFScaleVal))
+                   : 0.0f;
+
+  if (SFout) {
+    // Write the SF to global memory (STG.8).
+    *SFout = fp8SFVal;
+  }
+
+  // Convert the input to float.
+  float2 fp2Vals[CVT_FP4_ELTS_PER_THREAD / 2];
+
+  #pragma unroll
+  for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) {
+    if constexpr (std::is_same_v<Type, half>) {
+      fp2Vals[i] = __half22float2(vec.elts[i]);
+    } else {
+      fp2Vals[i] = __bfloat1622float2(vec.elts[i]);
+    }
+    fp2Vals[i].x *= outputScale;
+    fp2Vals[i].y *= outputScale;
+  }
+
+  // Convert to e2m1 values.
+  uint32_t e2m1Vec = fp32_vec_to_e2m1(fp2Vals);
+
+  // Write the e2m1 values to global memory.
+  return e2m1Vec;
+#else
+  return 0;
+#endif
+}
+
+// Use UE4M3 by default.
+template <class Type, bool UE8M0_SF = false>
+__global__ void
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000)
+__launch_bounds__(512, 4) cvt_fp16_to_fp4(
+#else
+cvt_fp16_to_fp4(
+#endif
+    int32_t numRows, int32_t numCols, Type const* in, float const* SFScale,
+    uint32_t* out, uint32_t* SFout) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000)
+  using PackedVec = PackedVec<Type>;
+  static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
+      (CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
+  static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,
+                "Vec size is not matched.");
+
+  // Get the global scaling factor, which will be applied to the SF.
+  // Note SFScale is the same as next GEMM's alpha, which is
+  // (448.f / (Alpha_A / 6.f)).
+  float const SFScaleVal = SFScale == nullptr ? 1.0f : SFScale[0];
+
+  // Input tensor row/col loops.
+  for (int rowIdx = blockIdx.x; rowIdx < numRows; rowIdx += gridDim.x) {
+    for (int colIdx = threadIdx.x; colIdx < numCols / CVT_FP4_ELTS_PER_THREAD;
+         colIdx += blockDim.x) {
+      int64_t inOffset = rowIdx * (numCols / CVT_FP4_ELTS_PER_THREAD) + colIdx;
+      PackedVec in_vec = reinterpret_cast<PackedVec const*>(in)[inOffset];
+      // Get the output tensor offset.
+      // Same as inOffset because 8 elements are packed into one uint32_t.
+      int64_t outOffset = inOffset;
+      auto& out_pos = out[outOffset];
+
+      auto sf_out =
+          cvt_quant_to_fp4_get_sf_out_offset<uint32_t,
+                                             CVT_FP4_NUM_THREADS_PER_SF>(
+              rowIdx, colIdx, numCols, SFout);
+
+      out_pos =
+          cvt_warp_fp16_to_fp4<Type, UE8M0_SF>(in_vec, SFScaleVal, sf_out);
+    }
+  }
+#endif
+}
+
+template <typename T>
+void invokeFP4Quantization(int m, int n, T const* input, float const* SFScale,
+                           int64_t* output, int32_t* SFOuput, bool useUE8M0,
+                           int multiProcessorCount, cudaStream_t stream) {
+  // Grid, Block size.
+  // Each thread converts 8 values.
+  dim3 block(std::min(int(n / ELTS_PER_THREAD), 512));
+  // Get number of blocks per SM (assume we can fully utilize the SM).
+  int const numBlocksPerSM = 2048 / block.x;
+  dim3 grid(std::min(int(m), multiProcessorCount * numBlocksPerSM));
+
+  // Launch the cvt kernel.
+  if (useUE8M0) {
+    cvt_fp16_to_fp4<T, true><<<grid, block, 0, stream>>>(
+        m, n, input, SFScale, reinterpret_cast<uint32_t*>(output),
+        reinterpret_cast<uint32_t*>(SFOuput));
+  } else {
+    cvt_fp16_to_fp4<T, false><<<grid, block, 0, stream>>>(
+        m, n, input, SFScale, reinterpret_cast<uint32_t*>(output),
+        reinterpret_cast<uint32_t*>(SFOuput));
+  }
+}
+
+// Instantiate the function.
+template void invokeFP4Quantization(int m, int n, half const* input,
+                                    float const* SFScale, int64_t* output,
+                                    int32_t* SFOuput, bool useUE8M0,
+                                    int multiProcessorCount,
+                                    cudaStream_t stream);
+
+template void invokeFP4Quantization(int m, int n, __nv_bfloat16 const* input,
+                                    float const* SFScale, int64_t* output,
+                                    int32_t* SFOuput, bool useUE8M0,
+                                    int multiProcessorCount,
+                                    cudaStream_t stream);
+
+void scaled_fp4_quant_sm100a(torch::Tensor const& output,
+                             torch::Tensor const& input,
+                             torch::Tensor const& output_sf,
+                             torch::Tensor const& input_sf) {
+  int32_t m = input.size(0);
+  int32_t n = input.size(1);
+
+  TORCH_CHECK(n % 16 == 0, "The N dimension must be multiple of 16.");
+
+  int multiProcessorCount =
+      get_device_attribute(cudaDevAttrMultiProcessorCount, -1);
+
+  auto input_sf_ptr = static_cast<float const*>(input_sf.data_ptr());
+  auto sf_out = static_cast<int32_t*>(output_sf.data_ptr());
+  auto output_ptr = static_cast<int64_t*>(output.data_ptr());
+  at::cuda::CUDAGuard device_guard{(char)input.get_device()};
+  auto stream = at::cuda::getCurrentCUDAStream(input.get_device());
+
+  // We don't support e8m0 scales at this moment.
+  bool useUE8M0 = false;
+
+  switch (input.scalar_type()) {
+    case torch::kHalf: {
+      auto input_ptr = reinterpret_cast<half const*>(input.data_ptr());
+      invokeFP4Quantization(m, n, input_ptr, input_sf_ptr, output_ptr, sf_out,
+                            useUE8M0, multiProcessorCount, stream);
+      break;
+    }
+    case torch::kBFloat16: {
+      auto input_ptr = reinterpret_cast<__nv_bfloat16 const*>(input.data_ptr());
+      invokeFP4Quantization(m, n, input_ptr, input_sf_ptr, output_ptr, sf_out,
+                            useUE8M0, multiProcessorCount, stream);
+      break;
+    }
+    default: {
+      std::cerr << "Observing: " << input.scalar_type()
+                << " for the input datatype which is invalid";
+      throw std::runtime_error(
+          "Unsupported input data type for quantize_to_fp4.");
+    }
+  }
+}

csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu

@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2025, NVIDIA CORPORATION.  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.
+ */
+
+#include <torch/all.h>
+
+#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+void cutlass_scaled_fp4_mm_sm100a(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);
+#endif
+
+void cutlass_scaled_fp4_mm(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) {
+#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+  return cutlass_scaled_fp4_mm_sm100a(D, A, B, A_sf, B_sf, alpha);
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(false,
+                              "No compiled nvfp4 mm kernel, vLLM should "
+                              "be compiled using CUDA 12.8 and target "
+                              "compute capability 100 or above.");
+}

csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu

@@ -0,0 +1,281 @@
+/*
+ * Copyright (c) 2025, NVIDIA CORPORATION.  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.
+ */
+
+#include <torch/all.h>
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include "cutlass_extensions/common.hpp"
+
+#include "cutlass/cutlass.h"
+
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+
+#include "cutlass/util/packed_stride.hpp"
+
+using namespace cute;
+
+#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+// Kernel Perf config
+template <typename T>
+struct KernelTraits;
+
+template <>
+struct KernelTraits<float> {
+  using MmaTileShape = Shape<_128, _128, _256>;
+  using ClusterShape = Shape<_1, _1, _1>;
+  using PerSmTileShape_MNK = Shape<_128, _128, _256>;
+};
+
+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>;
+  using LayoutATag = cutlass::layout::RowMajor;
+  static constexpr int AlignmentA = 32;
+
+  // B matrix configuration
+  using ElementB = cutlass::nv_float4_t<cutlass::float_e2m1_t>;
+  using LayoutBTag = cutlass::layout::ColumnMajor;
+  static constexpr int AlignmentB = 32;
+
+  // C/D matrix configuration
+  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;
+
+  // 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<
+          ArchTag, OperatorClass, PerSmTileShape_MNK, ClusterShape,
+          cutlass::epilogue::collective::EpilogueTileAuto, ElementAccumulator,
+          ElementAccumulator, ElementC, LayoutCTag, AlignmentC, ElementD,
+          LayoutDTag, AlignmentD,
+          cutlass::epilogue::collective::EpilogueScheduleAuto>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, ElementA, LayoutATag, AlignmentA, ElementB,
+          LayoutBTag, AlignmentB, ElementAccumulator, MmaTileShape,
+          ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          cutlass::gemm::collective::KernelScheduleAuto>::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>;
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  using StrideA = typename Gemm::GemmKernel::StrideA;
+  using LayoutA = decltype(cute::make_layout(make_shape(0, 0, 0), StrideA{}));
+  using LayoutSFA = typename Gemm::GemmKernel::CollectiveMainloop::LayoutSFA;
+  using StrideB = typename Gemm::GemmKernel::StrideB;
+  using LayoutB = decltype(cute::make_layout(make_shape(0, 0, 0), StrideB{}));
+  using LayoutSFB = typename Gemm::GemmKernel::CollectiveMainloop::LayoutSFB;
+  using StrideC = typename Gemm::GemmKernel::StrideC;
+  using LayoutC = decltype(cute::make_layout(make_shape(0, 0, 0), StrideC{}));
+  using StrideD = typename Gemm::GemmKernel::StrideD;
+  using LayoutD = decltype(cute::make_layout(make_shape(0, 0, 0), StrideD{}));
+};
+
+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 T::Gemm::ElementA;
+  using ElementB = typename T::Gemm::ElementB;
+  using ElementSFA = cutlass::float_ue4m3_t;
+  using ElementSFB = cutlass::float_ue4m3_t;
+  using ElementD = typename T::Gemm::ElementD;
+  using ElementCompute = float;
+  using StrideA = typename T::StrideA;
+  using StrideB = typename T::StrideB;
+  using StrideD = typename T::StrideD;
+  using Sm100BlkScaledConfig =
+      typename T::Gemm::GemmKernel::CollectiveMainloop::Sm100BlkScaledConfig;
+
+  int m = static_cast<int>(M);
+  int n = static_cast<int>(N);
+  int k = static_cast<int>(K);
+  auto stride_A = cutlass::make_cute_packed_stride(StrideA{}, {m, k, 1});
+  auto stride_B = cutlass::make_cute_packed_stride(StrideB{}, {n, k, 1});
+  auto stride_D = cutlass::make_cute_packed_stride(StrideD{}, {m, n, 1});
+
+  auto layout_SFA = Sm100BlkScaledConfig::tile_atom_to_shape_SFA(
+      cute::make_shape(m, n, k, 1));
+  auto layout_SFB = Sm100BlkScaledConfig::tile_atom_to_shape_SFB(
+      cute::make_shape(m, n, k, 1));
+
+  typename T::Gemm::Arguments arguments{
+      cutlass::gemm::GemmUniversalMode::kGemm,
+      {m, n, k, 1},
+      {// Mainloop arguments
+       static_cast<ElementA const*>(A.data_ptr()), stride_A,
+       static_cast<ElementB const*>(B.data_ptr()), stride_B,
+       static_cast<ElementSFA const*>(A_sf.data_ptr()), layout_SFA,
+       static_cast<ElementSFB const*>(B_sf.data_ptr()), layout_SFB},
+      {     // Epilogue arguments
+       {},  // epilogue.thread
+       static_cast<ElementD const*>(D.data_ptr()),
+       stride_D,
+       static_cast<ElementD*>(D.data_ptr()),
+       stride_D}};
+  auto& fusion_args = arguments.epilogue.thread;
+  fusion_args.alpha_ptr = static_cast<ElementCompute const*>(alpha.data_ptr());
+  return arguments;
+}
+
+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 Fp4GemmSm100<T>::Gemm gemm;
+
+  auto arguments =
+      args_from_options<Fp4GemmSm100<T>>(D, A, B, A_sf, B_sf, alpha, m, n, k);
+
+  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);
+
+  CUTLASS_CHECK(gemm.can_implement(arguments));
+
+  CUTLASS_CHECK(gemm.initialize(arguments, workspace.data_ptr(), stream));
+
+  CUTLASS_CHECK(gemm.run(arguments, workspace.data_ptr(), stream));
+}
+#else
+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.");
+}
+#endif  // defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+
+#define CHECK_TYPE(x, st, m) \
+  TORCH_CHECK(x.scalar_type() == st, "Inconsistency of Tensor type:", m)
+#define CHECK_TH_CUDA(x, m) TORCH_CHECK(x.is_cuda(), m, "must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x, m) \
+  TORCH_CHECK(x.is_contiguous(), m, "must be contiguous")
+#define CHECK_INPUT(x, st, m) \
+  CHECK_TH_CUDA(x, m);        \
+  CHECK_CONTIGUOUS(x, m);     \
+  CHECK_TYPE(x, st, m)
+
+constexpr auto FLOAT4_E2M1X2 = at::ScalarType::Byte;
+constexpr auto SF_DTYPE = at::ScalarType::Float8_e4m3fn;
+
+void cutlass_scaled_fp4_mm_sm100a(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) {
+  CHECK_INPUT(A, FLOAT4_E2M1X2, "a");
+  CHECK_INPUT(B, FLOAT4_E2M1X2, "b");
+
+  CHECK_INPUT(A_sf, SF_DTYPE, "scale_a");
+  CHECK_INPUT(B_sf, SF_DTYPE, "scale_b");
+
+  CHECK_INPUT(alpha, at::ScalarType::Float, "alpha");
+
+  TORCH_CHECK(A.dim() == 2, "a must be a matrix");
+  TORCH_CHECK(B.dim() == 2, "b must be a matrix");
+  TORCH_CHECK(A.sizes()[1] == B.sizes()[1],
+              "a and b shapes cannot be multiplied (", A.sizes()[0], "x",
+              A.sizes()[1], " and ", B.sizes()[0], "x", B.sizes()[1], ")");
+
+  auto const m = A.sizes()[0];
+  auto const n = B.sizes()[0];
+  auto const k = A.sizes()[1] * 2;
+
+  constexpr int alignment = 32;
+  TORCH_CHECK(k % alignment == 0, "Expected k to be divisible by ", alignment,
+              ", but got a shape: (", A.sizes()[0], "x", A.sizes()[1],
+              "), k: ", k, ".");
+  TORCH_CHECK(n % alignment == 0, "Expected n to be divisible by ", alignment,
+              ", but got b shape: (", B.sizes()[0], "x", B.sizes()[1], ").");
+
+  auto round_up = [](int x, int y) { return (x + y - 1) / y * y; };
+  int rounded_m = round_up(m, 128);
+  int rounded_n = round_up(n, 128);
+  // Since k is divisible by 32 (alignment), k / 16 is guaranteed to be an
+  // integer.
+  int rounded_k = round_up(k / 16, 4);
+
+  TORCH_CHECK(A_sf.dim() == 2, "scale_a must be a matrix");
+  TORCH_CHECK(B_sf.dim() == 2, "scale_b must be a matrix");
+  TORCH_CHECK(A_sf.sizes()[1] == B_sf.sizes()[1],
+              "scale_a and scale_b shapes cannot be multiplied (",
+              A_sf.sizes()[0], "x", A_sf.sizes()[1], " and ", B_sf.sizes()[0],
+              "x", B_sf.sizes()[1], ")");
+  TORCH_CHECK(A_sf.sizes()[0] == rounded_m && A_sf.sizes()[1] == rounded_k,
+              "scale_a must be padded and swizzled to a shape (", rounded_m,
+              "x", rounded_k, "), but got a shape (", A_sf.sizes()[0], "x",
+              A_sf.sizes()[1], ")");
+  TORCH_CHECK(B_sf.sizes()[0] == rounded_n && B_sf.sizes()[1] == rounded_k,
+              "scale_b must be padded and swizzled to a shape (", rounded_n,
+              "x", rounded_k, "), but got a shape (", B_sf.sizes()[0], "x",
+              B_sf.sizes()[1], ")");
+
+  auto out_dtype = D.dtype();
+  at::cuda::CUDAGuard device_guard{(char)A.get_device()};
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream(A.get_device());
+
+  if (out_dtype == at::ScalarType::Half) {
+    runGemm<cutlass::half_t>(D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
+  } else if (out_dtype == at::ScalarType::BFloat16) {
+    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");
+  }
+}

csrc/quantization/fp8/amd/hip_float8.h

@@ -1,137 +0,0 @@
-#pragma once
-
-#ifdef __HIPCC__
-  #include <hip/hip_runtime.h>
-#else
-  #include <type_traits>
-  #include <stdint.h>
-  #include <math.h>
-  #include <iostream>
-#endif
-
-#include "hip_float8_impl.h"
-
-struct alignas(1) hip_fp8 {
-  struct from_bits_t {};
-  HIP_FP8_HOST_DEVICE static constexpr from_bits_t from_bits() {
-    return from_bits_t();
-  }
-  uint8_t data;
-
-  hip_fp8() = default;
-  HIP_FP8_HOST_DEVICE constexpr hip_fp8(const hip_fp8&) = default;
-  HIP_FP8_HOST_DEVICE constexpr hip_fp8(uint8_t v) = delete;
-  explicit HIP_FP8_HOST_DEVICE constexpr hip_fp8(uint8_t v, from_bits_t)
-      : data(v) {}
-
-#ifdef __HIP__MI300__
-  // NOTE: ON-DEVICE... always optimal bias
-  explicit HIP_FP8_DEVICE hip_fp8(float v)
-      : data(hip_fp8_impl::to_fp8_from_fp32(v)) {}
-
-  explicit HIP_FP8_DEVICE hip_fp8(_Float16 v)
-      : hip_fp8(static_cast<float>(v)) {}
-
-  // Host only implementation using s/w simulation
-  explicit HIP_FP8_HOST
-#else   // __HIP__MI300__
-  // both Host and DEVICE for non-MI300 using s/w simulation
-  explicit HIP_FP8_HOST_DEVICE
-#endif  // __HIP__MI300__
-  hip_fp8(float v) {
-    data = hip_fp8_impl::to_float8<4, 3, float, true /*negative_zero_nan*/,
-                                   true /*clip*/>(v);
-  }
-
-  explicit HIP_FP8_HOST_DEVICE hip_fp8(double v)
-      : hip_fp8(static_cast<float>(v)) {}
-
-#ifdef __HIP__MI300__
-  // upcast using device specific intrinsic
-  explicit inline HIP_FP8_DEVICE operator float() const {
-    float fval;
-    uint32_t i32val = static_cast<uint32_t>(data);
-
-    // upcast
-    asm volatile("v_cvt_f32_fp8 %0, %1 src0_sel:BYTE_0"
-                 : "=v"(fval)
-                 : "v"(i32val));
-
-    return fval;
-  }
-
-  explicit inline HIP_FP8_HOST operator float() const
-#else   // __HIP__MI300__
-  explicit inline HIP_FP8_HOST_DEVICE operator float() const
-#endif  // __HIP__MI300__
-  {
-    return hip_fp8_impl::from_float8<4, 3, float, true /*negative_zero_nan*/>(
-        data);
-  }
-};
-
-namespace std {
-inline hip_fp8 sin(hip_fp8 a) { return hip_fp8(sinf(float(a))); }
-inline hip_fp8 cos(hip_fp8 a) { return hip_fp8(cosf(float(a))); }
-HIP_FP8_HOST_DEVICE constexpr hip_fp8 real(const hip_fp8& a) { return a; }
-}  // namespace std
-
-// Special operator overloading
-inline std::ostream& operator<<(std::ostream& os, const hip_fp8& f8) {
-  return os << float(f8);
-}
-
-// all + operator overloading with mixed types
-// mixed types, always converts to f32, does computation in f32, and returns
-// float
-inline HIP_FP8_HOST_DEVICE float operator+(const float fa, hip_fp8 b) {
-  return (fa + float(b));
-}
-
-inline HIP_FP8_HOST_DEVICE float operator+(hip_fp8 a, const float fb) {
-  return (float(a) + fb);
-}
-
-inline HIP_FP8_HOST_DEVICE hip_fp8 operator+(hip_fp8 a, hip_fp8 b) {
-  return hip_fp8(float(a) + float(b));
-}
-
-inline HIP_FP8_HOST_DEVICE hip_fp8& operator+=(hip_fp8& a, hip_fp8 b) {
-  return a = hip_fp8(float(a) + float(b));
-}
-
-// overloading multiplication, always returns float,
-inline HIP_FP8_HOST_DEVICE float operator*(hip_fp8 a, hip_fp8 b) {
-  return float(a) * float(b);
-}
-
-inline HIP_FP8_HOST_DEVICE float operator*(float a, hip_fp8 b) {
-  return (a * float(b));
-}
-
-inline HIP_FP8_HOST_DEVICE float operator*(hip_fp8 a, float b) {
-  return (float(a) * b);
-}
-
-inline HIP_FP8_HOST_DEVICE float operator*(int32_t a, hip_fp8 b) {
-  return ((float)a * float(b));
-}
-
-inline HIP_FP8_HOST_DEVICE float operator*(double a, hip_fp8 b) {
-  return ((float)a * float(b));
-}
-
-// overloading for compare
-inline HIP_FP8_HOST_DEVICE bool operator==(hip_fp8 a, hip_fp8 b) {
-  return (a.data == b.data);
-}
-inline HIP_FP8_HOST_DEVICE bool operator!=(hip_fp8 a, hip_fp8 b) {
-  return (a.data != b.data);
-}
-
-inline HIP_FP8_HOST_DEVICE bool operator>=(hip_fp8 a, hip_fp8 b) {
-  return static_cast<float>(a) >= static_cast<float>(b);
-}
-inline HIP_FP8_HOST_DEVICE bool operator>(hip_fp8 a, hip_fp8 b) {
-  return static_cast<float>(a) > static_cast<float>(b);
-}

csrc/quantization/fp8/amd/hip_float8_impl.h

@@ -1,316 +0,0 @@
-#pragma once
-
-#if defined(__HIPCC__) && \
-    (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
-  #define __HIP__MI300__
-#endif
-
-#ifdef __HIPCC__
-  #define HIP_FP8_HOST_DEVICE __host__ __device__
-  #define HIP_FP8_HOST __host__
-  #define HIP_FP8_DEVICE __device__
-#else
-  #define HIP_FP8_HOST_DEVICE
-  #define HIP_FP8_HOST
-  #define HIP_FP8_DEVICE
-#endif
-
-namespace hip_fp8_impl {
-
-#ifdef __HIP__MI300__
-HIP_FP8_DEVICE uint8_t to_fp8_from_fp32(float v) {
-  uint8_t i8data;
-  union {
-    float fval;
-    uint32_t i32val;
-    uint8_t i8val[4];  // NOTE: not endian independent
-  } val;
-
-  uint32_t ival = 0;
-  val.fval = v;
-
-  if ((val.i32val & 0x7F800000) !=
-      0x7F800000) {  /// propagate NAN/INF, no clipping
-    val.fval = __builtin_amdgcn_fmed3f(val.fval, 240.0, -240.0);
-  }
-
-  ival = __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival,
-                                         false);  // false -> WORD0
-  val.i32val = ival;
-  i8data = val.i8val[0];
-
-  return i8data;
-}
-#endif  // __HIP__MI300__
-
-HIP_FP8_HOST inline int clz(uint32_t x) { return __builtin_clz(x); }
-#if defined(__HIPCC__) || defined(__CUDA_ARCH__)
-HIP_FP8_DEVICE inline int clz(uint32_t x) { return __clz(x); }
-#endif
-
-template <int we, int wm, typename T, bool negative_zero_nan, bool clip>
-HIP_FP8_HOST_DEVICE uint8_t to_float8(T _x, bool stoch = false,
-                                      uint32_t rng = 0) {
-#ifdef __HIPCC__
-  constexpr bool is_half = std::is_same<T, _Float16>::value;
-#else
-  constexpr bool is_half = false;
-#endif
-  constexpr bool is_float = std::is_same<T, float>::value;
-  static_assert(wm + we == 7, "wm+we==7");
-  static_assert(is_half || is_float, "Only half and float can be cast to f8");
-
-  const int mfmt = (sizeof(T) == 4) ? 23 : 10;
-  uint32_t x;
-  if (sizeof(T) == 4) {
-    x = reinterpret_cast<uint32_t&>(_x);
-  } else {
-    x = reinterpret_cast<uint16_t&>(_x);
-  }
-
-  uint32_t head, mantissa;
-  int exponent, bias;
-  uint32_t sign;
-
-  if (sizeof(T) == 4) {
-    head = x & 0xFF800000;
-    mantissa = x & 0x7FFFFF;
-    exponent = (head >> 23) & 0xFF;
-    sign = head >> 31;
-    bias = 127;
-  } else {
-    head = x & 0xFC00;
-    mantissa = x & 0x3FF;
-    exponent = (head >> 10) & 0x1F;
-    sign = head >> 15;
-    bias = 15;
-  }
-
-  uint32_t signed_inf = (sign << 7) + (((1 << we) - 1) << wm);
-
-  // Deal with inf and NaNs
-  if (negative_zero_nan) {
-    if (sizeof(T) == 4) {
-      if ((x & 0x7F800000) == 0x7F800000) {
-        return 0x80;
-      }
-    } else {
-      // if(__hisinf(x) || __hisnan(x))
-      if ((x & 0x7C00) == 0x7C00) {
-        return 0x80;
-      }
-    }
-  } else {
-    if (sizeof(T) == 4) {
-      if ((x & 0x7F800000) == 0x7F800000) {
-        return signed_inf + (mantissa != 0 ? 1 : 0);
-      }
-    } else {
-      if ((x & 0x7C00) == 0x7C00) {
-        return signed_inf + (mantissa != 0 ? 1 : 0);
-      }
-    }
-  }
-  if (x == 0) {
-    return 0;
-  }
-
-  // First need to check if it is normal or denorm as there is a difference of
-  // implicit 1 Then need to adjust the exponent to align with the F8 exponent,
-  // in the meanwhile, shift The mantissa. Then for stochastic rounding, add rng
-  // to mantissa and truncate. And for RNE, no need to add rng. Then probably
-  // need to check whether there is carry and adjust exponent and mantissa again
-
-  // For IEEE bias mode, the bias is 2^(k-1) -1 where k is the width of exponent
-  // bits
-  const int f8_bias = (1 << (we - 1)) - 1 + (negative_zero_nan ? 1 : 0);
-  const int f8_denormal_act_exponent =
-      1 - f8_bias;  // actual exponent of f8 denormal
-  // act_exponent is the actual exponent of fp32/fp16 (after subtracting bias)
-  // f8_exponent is the converted f8 exponent with bias encoding
-  // exponent_diff is the diff between fp32/fp16 exponent and f8 exponent,
-  // the difference needs to be adjusted and mantissa shifted
-  int act_exponent, f8_exponent, exponent_diff;
-
-  if (exponent == 0) {  // fp32/fp16 is in denormal.
-    /* fp32 denormal is below 2^-127 so it is usually not a concern here, we
-mostly concern fp16 here. In this case, f8 is usually in denormal. But there
-could be exceptions. fp16 denormal has exponent bias 15 while bf8 with NANOO has
-exponent bias 16. It means that there are some numbers in fp16 denormal but they
-are bf8 (NANOO) normals - smallest bf8 (NANOO) normal is 2^-15. fp16 numbers
-where exponent==0 (actual exponent -14) and highest bit of mantissa is 1 are bf8
-(NANOO) normal. In this case, the fp16 mantissa should be shift left by 1  */
-    act_exponent = exponent - bias + 1;
-    exponent_diff =
-        f8_denormal_act_exponent -
-        act_exponent;  // actual exponent is exponent-bias+1 as it is denormal
-  } else {             // fp32/fp16 is normal with implicit 1
-    act_exponent = exponent - bias;
-    if (act_exponent <= f8_denormal_act_exponent) {
-      /* This is the case where fp32/fp16 is normal but it is in f8 denormal
-range. For example fp8 nanoo mode, denormal exponent is -7, but if the
-fp32/fp16 actual exponent is -7, it is actually larger due to the implicit 1,
-Therefore it needs to be adjust to -6 and mantissa shift right by 1.
-So for fp32/fp16, exponent -8 is the cut point to convert to fp8 nanoo */
-      exponent_diff = f8_denormal_act_exponent - act_exponent;
-    } else {              // both fp32/fp16 and f8 are in normal range
-      exponent_diff = 0;  // exponent_diff=0 does not mean there is no
-                          // difference for this case, act_exponent could be
-                          // larger. Just that it does not need shift mantissa
-    }
-    mantissa += (1 << mfmt);  // Add the implicit 1 into mantissa
-  }
-
-  bool midpoint = (mantissa & ((1 << (mfmt - wm + exponent_diff)) - 1)) ==
-                  static_cast<uint32_t>(1 << (mfmt - wm + exponent_diff - 1));
-  /* This part is a bit tricky. The judgment of whether it is a tie needs to be
- done before we shift right as shift right could rip off some residual part
- and make something not midpoint look like midpoint. For example, the fp16
- number 0x1002 (0 00100 0000000010), it is larger than midpoint, but after
- shift right by 4 bits, it would look like midpoint.
-*/
-
-  if (exponent_diff > 0) {
-    mantissa >>= exponent_diff;
-  } else if (exponent_diff == -1) {
-    mantissa <<= -exponent_diff;
-  }
-  bool implicit_one = mantissa & (1 << mfmt);
-  // if there is no implicit 1, it  means the f8 is denormal and need to adjust
-  // to denorm exponent
-  f8_exponent = (act_exponent + exponent_diff) /*actual f8 exponent*/ +
-                f8_bias - (implicit_one ? 0 : 1);
-
-  // Now we have the exponent and mantissa adjusted
-  uint32_t drop_mask = (1 << (mfmt - wm)) - 1;
-  bool odd = mantissa & (1 << (mfmt - wm));  // if the least significant bit
-                                             // that is not truncated is 1
-  mantissa +=
-      (stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1) : mantissa)) &
-      drop_mask;
-
-  // Now we deal with overflow
-  if (f8_exponent == 0) {
-    if ((1 << mfmt) & mantissa) {
-      f8_exponent = 1;  // denormal overflow to become normal, promote exponent
-    }
-  } else {
-    if ((1 << (mfmt + 1)) & mantissa) {
-      mantissa >>= 1;
-      f8_exponent++;
-    }
-  }
-
-  mantissa >>= (mfmt - wm);
-
-  // above range: quantize to maximum possible float of the same sign
-  const int max_exp = (1 << we) - (negative_zero_nan ? 1 : 2);
-  if (f8_exponent > max_exp) {
-    if (clip) {
-      mantissa = (1 << wm) - 1;
-      f8_exponent = max_exp;
-    } else {
-      return signed_inf;
-    }
-  }
-
-  if (f8_exponent == 0 && mantissa == 0) {
-    return negative_zero_nan ? 0 : (sign << 7);
-  }
-  mantissa &= (1 << wm) - 1;
-  return (sign << 7) | (f8_exponent << wm) | mantissa;
-}
-
-template <int we, int wm, typename T = float, bool negative_zero_nan = true>
-inline HIP_FP8_HOST_DEVICE T from_float8(uint8_t x) {
-#ifdef __HIPCC__
-  constexpr bool is_half = std::is_same<T, _Float16>::value;
-#else
-  constexpr bool is_half = false;
-#endif
-  constexpr bool is_float = std::is_same<T, float>::value;
-  static_assert(is_half || is_float, "only half and float are supported");
-
-  constexpr int weo = is_half ? 5 : 8;
-  constexpr int wmo = is_half ? 10 : (is_float ? 23 : 7);
-
-  T fInf, fNegInf, fNaN, fNeg0;
-
-#ifdef __HIPCC__
-  if (is_half) {
-    const uint16_t ihInf = 0x7C00;
-    const uint16_t ihNegInf = 0xFC00;
-    const uint16_t ihNaN = 0x7C01;
-    const uint16_t ihNeg0 = 0x8000;
-    fInf = reinterpret_cast<const _Float16&>(ihInf);
-    fNegInf = reinterpret_cast<const _Float16&>(ihNegInf);
-    fNaN = reinterpret_cast<const _Float16&>(ihNaN);
-    fNeg0 = reinterpret_cast<const _Float16&>(ihNeg0);
-  } else
-#endif
-      if (is_float) {
-    const uint32_t ifInf = 0x7F800000;
-    const uint32_t ifNegInf = 0xFF800000;
-    const uint32_t ifNaN = 0x7F800001;
-    const uint32_t ifNeg0 = 0x80000000;
-    fInf = reinterpret_cast<const float&>(ifInf);
-    fNegInf = reinterpret_cast<const float&>(ifNegInf);
-    fNaN = reinterpret_cast<const float&>(ifNaN);
-    fNeg0 = reinterpret_cast<const float&>(ifNeg0);
-  }
-
-  if (x == 0) {
-    return 0;
-  }
-
-  uint32_t sign = x >> 7;
-  uint32_t mantissa = x & ((1 << wm) - 1);
-  int exponent = (x & 0x7F) >> wm;
-  if (negative_zero_nan) {
-    if (x == 0x80) {
-      return fNaN;
-    }
-  } else {
-    if (x == 0x80) {
-      return fNeg0;
-    }
-    if (exponent == ((1 << we) - 1)) {
-      return (mantissa == 0) ? (sign ? fNegInf : fInf) : fNaN;
-    }
-  }
-  typename std::conditional<sizeof(T) == 2, uint16_t, uint32_t>::type retval;
-  if (we == 5 && is_half && !negative_zero_nan) {
-    retval = x << 8;
-    return reinterpret_cast<const T&>(retval);
-  }
-
-  const int exp_low_cutoff =
-      (1 << (weo - 1)) - (1 << (we - 1)) + 1 - (negative_zero_nan ? 1 : 0);
-
-  // subnormal input
-  if (exponent == 0) {
-    // guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
-    int sh = 1 + clz(mantissa) - (32 - wm);
-    mantissa <<= sh;
-    exponent += 1 - sh;
-    mantissa &= ((1 << wm) - 1);
-  }
-  exponent += exp_low_cutoff - 1;
-  mantissa <<= wmo - wm;
-
-  // subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
-  if (exponent <= 0) {
-    mantissa |= 1 << wmo;
-    mantissa >>= 1 - exponent;
-    exponent = 0;
-  }
-
-  if (sizeof(T) == 2) {
-    retval = (sign << 15) | (exponent << 10) | mantissa;
-  } else {
-    retval = (sign << 31) | (exponent << 23) | mantissa;
-  }
-  return reinterpret_cast<const T&>(retval);
-}
-
-}  // namespace hip_fp8_impl

csrc/quantization/fp8/amd/quant_utils.cuh

@@ -1,13 +1,11 @@
 #pragma once
-#include "hip_float8.h"
+#include <hip/hip_fp8.h>
 
 #include <hip/hip_fp16.h>
 #include <hip/hip_bf16.h>
 #include <hip/hip_bfloat16.h>
 
-#include "../../../attention/dtype_fp8.cuh"
-#include "../../../attention/dtype_float32.cuh"
-#include "../../../attention/dtype_bfloat16.cuh"
+#include "../../../attention/attention_dtypes.h"
 
 namespace vllm {
 #ifdef USE_ROCM
@@ -26,40 +24,31 @@ __inline__ __device__ Tout scaled_vec_conversion(const Tin& x,
   return x;
 }
 
+    #if HIP_FP8_TYPE_OCP
+using fp8_type = __hip_fp8_e4m3;
+using fp8x2_type = __hip_fp8x2_e4m3;
+    #else
+using fp8_type = __hip_fp8_e4m3_fnuz;
+using fp8x2_type = __hip_fp8x2_e4m3_fnuz;
+    #endif
+
 // fp8 -> half
 template <>
 __inline__ __device__ uint16_t
 vec_conversion<uint16_t, uint8_t>(const uint8_t& a) {
-  hip_fp8 f8{a, hip_fp8::from_bits()};
-  __half_raw res;
-  res.data = static_cast<float>(f8);
-  return res.x;
+  return __hip_cvt_fp8_to_halfraw(a, fp8_type::__default_interpret).x;
 }
 
 // fp8x2 -> half2
 template <>
 __inline__ __device__ uint32_t
 vec_conversion<uint32_t, uint16_t>(const uint16_t& a) {
-    #if defined(__HIP__MI300__) && \
-        defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
-  const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
   union {
     __half2_raw h2r;
     uint32_t ui32;
   } tmp;
-  tmp.h2r.x.data = f2[0];
-  tmp.h2r.y.data = f2[1];
+  tmp.h2r = __hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret);
   return tmp.ui32;
-    #else
-  union {
-    uint16_t u16[2];
-    uint32_t u32;
-  } tmp;
-
-  tmp.u16[0] = vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a));
-  tmp.u16[1] = vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a >> 8U));
-  return tmp.u32;
-    #endif
 }
 
 // fp8x4 -> half2x2
@@ -92,9 +81,9 @@ using __nv_bfloat16 = __hip_bfloat16;
 template <>
 __inline__ __device__ __nv_bfloat16
 vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a) {
-  hip_fp8 f8{a, hip_fp8::from_bits()};
-  float f{f8};
-  return __float2bfloat16(f);
+  fp8_type f8;
+  f8.__x = a;
+  return __float2bfloat16(static_cast<float>(f8));
 }
 
 using __nv_bfloat162 = __hip_bfloat162;
@@ -136,27 +125,18 @@ __inline__ __device__ bf16_8_t vec_conversion<bf16_8_t, uint2>(const uint2& a) {
 // fp8 -> float
 template <>
 __inline__ __device__ float vec_conversion<float, uint8_t>(const uint8_t& a) {
-  hip_fp8 fp8{a, hip_fp8::from_bits()};
-  return static_cast<float>(fp8);
+  fp8_type f8;
+  f8.__x = a;
+  return static_cast<float>(f8);
 }
 
 // fp8x2 -> float2
 template <>
 __inline__ __device__ float2
 vec_conversion<float2, uint16_t>(const uint16_t& a) {
-    #if defined(__HIP__MI300__) && \
-        defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
-  float2 res;
-  const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
-  res.x = f2[0];
-  res.y = f2[1];
-  return res;
-    #else
-  float2 res;
-  res.x = vec_conversion<float, uint8_t>(static_cast<uint8_t>(a));
-  res.y = vec_conversion<float, uint8_t>(static_cast<uint8_t>(a >> 8U));
-  return res;
-    #endif
+  fp8x2_type f8x2;
+  f8x2.__x = a;
+  return static_cast<float2>(f8x2);
 }
 
 // fp8x4 -> float4
@@ -169,6 +149,15 @@ vec_conversion<Float4_, uint32_t>(const uint32_t& a) {
   return res;
 }
 
+// fp8x4 -> float4
+template <>
+__inline__ __device__ float4
+vec_conversion<float4, uint32_t>(const uint32_t& a) {
+  Float4_ tmp = vec_conversion<Float4_, uint32_t>(a);
+  float4 res = make_float4(tmp.x.x, tmp.x.y, tmp.y.x, tmp.y.y);
+  return res;
+}
+
 // fp8x8 -> float8
 template <>
 __inline__ __device__ Float8_ vec_conversion<Float8_, uint2>(const uint2& a) {
@@ -189,33 +178,36 @@ __inline__ __device__ uint8_t
 vec_conversion<uint8_t, uint16_t>(const uint16_t& a) {
   __half_raw tmp;
   tmp.x = a;
+  return __hip_cvt_halfraw_to_fp8(tmp, fp8_type::__default_saturation,
+                                  fp8_type::__default_interpret);
+}
 
-  hip_fp8 f8{static_cast<float>(tmp.data)};
-  return f8.data;
+template <>
+__inline__ __device__ uint16_t
+vec_conversion<uint16_t, uint32_t>(const uint32_t& a) {
+  union {
+    uint32_t ui32;
+    __half2_raw h2r;
+  } tmp;
+  tmp.ui32 = a;
+  return __hip_cvt_halfraw2_to_fp8x2(tmp.h2r, fp8_type::__default_saturation,
+                                     fp8_type::__default_interpret);
 }
 
 // bf16 -> fp8
 template <>
 __inline__ __device__ uint8_t
 vec_conversion<uint8_t, __nv_bfloat16>(const __nv_bfloat16& a) {
-  hip_fp8 res{__bfloat162float(a)};
-  return res.data;
+  return __hip_cvt_float_to_fp8(__bfloat162float(a),
+                                fp8_type::__default_saturation,
+                                fp8_type::__default_interpret);
 }
 
 // float -> fp8
 template <>
 __inline__ __device__ uint8_t vec_conversion<uint8_t, float>(const float& a) {
-  hip_fp8 f8(a);
-  return f8.data;
-}
-
-// fp8x4 -> float4
-template <>
-__inline__ __device__ float4
-vec_conversion<float4, uint32_t>(const uint32_t& a) {
-  Float4_ tmp = vec_conversion<Float4_, uint32_t>(a);
-  float4 res = make_float4(tmp.x.x, tmp.x.y, tmp.y.x, tmp.y.y);
-  return res;
+  return __hip_cvt_float_to_fp8(a, fp8_type::__default_saturation,
+                                fp8_type::__default_interpret);
 }
 
 // float2 -> half2
@@ -307,90 +299,22 @@ vec_conversion<bf16_8_t, Float8_>(const Float8_& a) {
 
  */
 
-// fp8 -> half
-template <>
-__inline__ __device__ uint16_t
-scaled_vec_conversion<uint16_t, uint8_t>(const uint8_t& a, const float scale) {
-  hip_fp8 f8{a, hip_fp8::from_bits()};
-  __half_raw res;
-  res.data = static_cast<float>(f8) * scale;
-  return res.x;
-}
-
-// fp8x2 -> half2
-template <>
-__inline__ __device__ uint32_t scaled_vec_conversion<uint32_t, uint16_t>(
-    const uint16_t& a, const float scale) {
-    #if defined(__HIP__MI300__) && \
-        defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
-  const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
-  union {
-    __half2_raw h2r;
-    uint32_t ui32;
-  } tmp;
-  tmp.h2r.x.data = f2[0] * scale;
-  tmp.h2r.y.data = f2[1] * scale;
-  return tmp.ui32;
-    #else
-  union {
-    uint16_t u16[2];
-    uint32_t u32;
-  } tmp;
-
-  tmp.u16[0] =
-      scaled_vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a), scale);
-  tmp.u16[1] = scaled_vec_conversion<uint16_t, uint8_t>(
-      static_cast<uint8_t>(a >> 8U), scale);
-  return tmp.u32;
-    #endif
-}
-
-// fp8x4 -> half2x2
-template <>
-__inline__ __device__ uint2
-scaled_vec_conversion<uint2, uint32_t>(const uint32_t& a, const float scale) {
-  union {
-    uint2 u32x2;
-    uint32_t u32[2];
-  } tmp;
-  tmp.u32[0] = scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)a, scale);
-  tmp.u32[1] =
-      scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)(a >> 16U), scale);
-  return tmp.u32x2;
-}
-
-// fp8x8 -> half2x4
-template <>
-__inline__ __device__ uint4
-scaled_vec_conversion<uint4, uint2>(const uint2& a, const float scale) {
-  union {
-    uint4 u64x2;
-    uint2 u64[2];
-  } tmp;
-  tmp.u64[0] = scaled_vec_conversion<uint2, uint32_t>(a.x, scale);
-  tmp.u64[1] = scaled_vec_conversion<uint2, uint32_t>(a.y, scale);
-  return tmp.u64x2;
-}
-
 using __nv_bfloat16 = __hip_bfloat16;
 
 // fp8 -> __nv_bfloat16
 template <>
 __inline__ __device__ __nv_bfloat16
-scaled_vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a,
-                                              const float scale) {
-  hip_fp8 f8{a, hip_fp8::from_bits()};
-  float f{f8};
-  return __float2bfloat16(f * scale);
+scaled_vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a, float scale) {
+  fp8_type f8;
+  f8.__x = a;
+  return __float2bfloat16(static_cast<float>(f8) * scale);
 }
 
-using __nv_bfloat162 = __hip_bfloat162;
-
 // fp8x2 -> __nv_bfloat162
 template <>
 __inline__ __device__ __nv_bfloat162
 scaled_vec_conversion<__nv_bfloat162, uint16_t>(const uint16_t& a,
-                                                const float scale) {
+                                                float scale) {
   __nv_bfloat162 res;
   res.x = scaled_vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)a, scale);
   res.y =
@@ -400,8 +324,8 @@ scaled_vec_conversion<__nv_bfloat162, uint16_t>(const uint16_t& a,
 
 // fp8x4 -> bf16_4_t
 template <>
-__inline__ __device__ bf16_4_t scaled_vec_conversion<bf16_4_t, uint32_t>(
-    const uint32_t& a, const float scale) {
+__inline__ __device__ bf16_4_t
+scaled_vec_conversion<bf16_4_t, uint32_t>(const uint32_t& a, float scale) {
   bf16_4_t res;
   res.x = scaled_vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)a, scale);
   res.y = scaled_vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)(a >> 16U),
@@ -412,7 +336,7 @@ __inline__ __device__ bf16_4_t scaled_vec_conversion<bf16_4_t, uint32_t>(
 // fp8x8 -> bf16_8_t
 template <>
 __inline__ __device__ bf16_8_t
-scaled_vec_conversion<bf16_8_t, uint2>(const uint2& a, const float scale) {
+scaled_vec_conversion<bf16_8_t, uint2>(const uint2& a, float scale) {
   bf16_4_t tmp1, tmp2;
   tmp1 = scaled_vec_conversion<bf16_4_t, uint32_t>(a.x, scale);
   tmp2 = scaled_vec_conversion<bf16_4_t, uint32_t>(a.y, scale);
@@ -427,29 +351,19 @@ scaled_vec_conversion<bf16_8_t, uint2>(const uint2& a, const float scale) {
 // fp8 -> float
 template <>
 __inline__ __device__ float scaled_vec_conversion<float, uint8_t>(
-    const uint8_t& a, const float scale) {
-  hip_fp8 fp8{a, hip_fp8::from_bits()};
-  return static_cast<float>(fp8) * scale;
+    const uint8_t& a, float scale) {
+  fp8_type f8;
+  f8.__x = a;
+  return static_cast<float>(f8) * scale;
 }
 
 // fp8x2 -> float2
 template <>
 __inline__ __device__ float2
-scaled_vec_conversion<float2, uint16_t>(const uint16_t& a, const float scale) {
-    #if defined(__HIP__MI300__) && \
-        defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
-  float2 res;
-  const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
-  res.x = f2[0] * scale;
-  res.y = f2[1] * scale;
-  return res;
-    #else
-  float2 res;
-  res.x = scaled_vec_conversion<float, uint8_t>(static_cast<uint8_t>(a), scale);
-  res.y = scaled_vec_conversion<float, uint8_t>(static_cast<uint8_t>(a >> 8U),
-                                                scale);
-  return res;
-    #endif
+scaled_vec_conversion<float2, uint16_t>(const uint16_t& a, float scale) {
+  fp8x2_type f8x2;
+  f8x2.__x = a;
+  return static_cast<float2>(f8x2) * scale;
 }
 
 // fp8x4 -> float4
@@ -462,10 +376,18 @@ scaled_vec_conversion<Float4_, uint32_t>(const uint32_t& a, const float scale) {
   return res;
 }
 
+// fp8x4 -> float4
+template <>
+__inline__ __device__ float4
+scaled_vec_conversion<float4, uint32_t>(const uint32_t& a, float scale) {
+  Float4_ res = scaled_vec_conversion<Float4_, uint32_t>(a, scale);
+  return {res.x.x, res.x.y, res.y.x, res.y.y};
+}
+
 // fp8x8 -> float8
 template <>
 __inline__ __device__ Float8_
-scaled_vec_conversion<Float8_, uint2>(const uint2& a, const float scale) {
+scaled_vec_conversion<Float8_, uint2>(const uint2& a, float scale) {
   Float4_ tmp1, tmp2;
   tmp1 = scaled_vec_conversion<Float4_, uint32_t>(a.x, scale);
   tmp2 = scaled_vec_conversion<Float4_, uint32_t>(a.y, scale);
@@ -477,44 +399,184 @@ scaled_vec_conversion<Float8_, uint2>(const uint2& a, const float scale) {
   return res;
 }
 
-/* Quantize(HP / scale) => FP8 */
+// fp8 -> half
+template <>
+__inline__ __device__ uint16_t
+scaled_vec_conversion<uint16_t, uint8_t>(const uint8_t& a, float scale) {
+  __half_raw res;
+  res.data = scaled_vec_conversion<float, uint8_t>(a, scale);
+  return res.x;
+}
 
-// TODO(Hai): vectorized to add
+// fp8x2 -> half2
+template <>
+__inline__ __device__ uint32_t
+scaled_vec_conversion<uint32_t, uint16_t>(const uint16_t& a, float scale) {
+  __half2_raw h2r =
+      __hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret);
+  union {
+    __half2_raw h2r;
+    uint32_t ui32;
+  } tmp;
+  tmp.h2r = __hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret);
+  tmp.h2r.x.data *= scale;
+  tmp.h2r.y.data *= scale;
+  return tmp.ui32;
+}
+
+// fp8x4 -> half2x2
+template <>
+__inline__ __device__ uint2
+scaled_vec_conversion<uint2, uint32_t>(const uint32_t& a, float scale) {
+  union {
+    uint2 u32x2;
+    uint32_t u32[2];
+  } tmp;
+  tmp.u32[0] = scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)a, scale);
+  tmp.u32[1] =
+      scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)(a >> 16U), scale);
+  return tmp.u32x2;
+}
+
+// fp8x8 -> half2x4
+template <>
+__inline__ __device__ uint4 scaled_vec_conversion<uint4, uint2>(const uint2& a,
+                                                                float scale) {
+  union {
+    uint4 u64x2;
+    uint2 u64[2];
+  } tmp;
+  tmp.u64[0] = scaled_vec_conversion<uint2, uint32_t>(a.x, scale);
+  tmp.u64[1] = scaled_vec_conversion<uint2, uint32_t>(a.y, scale);
+  return tmp.u64x2;
+}
 
 // half -> fp8
 template <>
 __inline__ __device__ uint8_t
-scaled_vec_conversion<uint8_t, uint16_t>(const uint16_t& a, const float scale) {
+scaled_vec_conversion<uint8_t, uint16_t>(const uint16_t& a, float scale) {
   __half_raw tmp;
   tmp.x = a;
+  tmp.data /= scale;
+  return __hip_cvt_halfraw_to_fp8(tmp, fp8_type::__default_saturation,
+                                  fp8_type::__default_interpret);
+}
 
-  hip_fp8 f8{static_cast<float>(tmp.data) / scale};
-  return f8.data;
+// halfx2 -> fp8x2
+template <>
+__inline__ __device__ uint16_t
+scaled_vec_conversion<uint16_t, uint32_t>(const uint32_t& a, float scale) {
+  union {
+    uint32_t ui32;
+    __half2_raw h2r;
+  } tmp;
+  tmp.ui32 = a;
+  tmp.h2r.x.data /= scale;
+  tmp.h2r.y.data /= scale;
+  return __hip_cvt_halfraw2_to_fp8x2(tmp.h2r, fp8_type::__default_saturation,
+                                     fp8_type::__default_interpret);
+}
+
+// half2x2 -> fp8x4
+template <>
+__inline__ __device__ uint32_t
+scaled_vec_conversion<uint32_t, uint2>(const uint2& a, float scale) {
+  union {
+    uint16_t ui16[2];
+    uint32_t ui32;
+  } tmp;
+  tmp.ui16[0] = scaled_vec_conversion<uint16_t, uint32_t>(a.x, scale);
+  tmp.ui16[1] = scaled_vec_conversion<uint16_t, uint32_t>(a.y, scale);
+  return tmp.ui32;
+}
+
+// half2x4 -> fp8x8
+template <>
+__inline__ __device__ uint2 scaled_vec_conversion<uint2, uint4>(const uint4& a,
+                                                                float scale) {
+  union {
+    uint2 ui2[2];
+    uint4 ui4;
+  } tmp;
+  tmp.ui4 = a;
+  uint2 res;
+  res.x = scaled_vec_conversion<uint32_t, uint2>(tmp.ui2[0], scale);
+  res.y = scaled_vec_conversion<uint32_t, uint2>(tmp.ui2[1], scale);
+  return res;
 }
 
 // bf16 -> fp8
 template <>
 __inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, __nv_bfloat16>(
-    const __nv_bfloat16& a, const float scale) {
-  hip_fp8 res{__bfloat162float(a) / scale};
-  return res.data;
+    const __nv_bfloat16& a, float scale) {
+  return __hip_cvt_float_to_fp8(__bfloat162float(a) / scale,
+                                fp8_type::__default_saturation,
+                                fp8_type::__default_interpret);
+}
+
+// bf16x2 -> fp8x2
+template <>
+__inline__ __device__ uint16_t scaled_vec_conversion<uint16_t, __nv_bfloat162>(
+    const __nv_bfloat162& a, float scale) {
+  union {
+    uint8_t ui8[2];
+    uint16_t ui16;
+  } tmp;
+  tmp.ui8[0] = scaled_vec_conversion<uint8_t, __nv_bfloat16>(a.x, scale);
+  tmp.ui8[1] = scaled_vec_conversion<uint8_t, __nv_bfloat16>(a.y, scale);
+  return tmp.ui16;
+}
+
+// bf16x4 -> fp8x4
+template <>
+__inline__ __device__ uint32_t
+scaled_vec_conversion<uint32_t, bf16_4_t>(const bf16_4_t& a, float scale) {
+  union {
+    uint16_t ui16[2];
+    uint32_t ui32;
+  } tmp;
+  tmp.ui16[0] = scaled_vec_conversion<uint16_t, __nv_bfloat162>(a.x, scale);
+  tmp.ui16[1] = scaled_vec_conversion<uint16_t, __nv_bfloat162>(a.y, scale);
+  return tmp.ui32;
+}
+
+// bf16x8 -> fp8x8
+template <>
+__inline__ __device__ uint2
+scaled_vec_conversion<uint2, bf16_8_t>(const bf16_8_t& a, float scale) {
+  uint2 res;
+  res.x = scaled_vec_conversion<uint32_t, bf16_4_t>({a.x, a.y}, scale);
+  res.y = scaled_vec_conversion<uint32_t, bf16_4_t>({a.z, a.w}, scale);
+  return res;
 }
 
 // float -> fp8
 template <>
 __inline__ __device__ uint8_t
-scaled_vec_conversion<uint8_t, float>(const float& a, const float scale) {
-  hip_fp8 f8(a / scale);
-  return f8.data;
+scaled_vec_conversion<uint8_t, float>(const float& a, float scale) {
+  return __hip_cvt_float_to_fp8(a / scale, fp8_type::__default_saturation,
+                                fp8_type::__default_interpret);
 }
 
-// fp8x4 -> float4
+// floatx2 -> fp8x2
 template <>
-__inline__ __device__ float4
-scaled_vec_conversion<float4, uint32_t>(const uint32_t& a, const float scale) {
-  Float4_ tmp = scaled_vec_conversion<Float4_, uint32_t>(a, scale);
-  float4 res = make_float4(tmp.x.x, tmp.x.y, tmp.y.x, tmp.y.y);
-  return res;
+__inline__ __device__ uint16_t
+scaled_vec_conversion<uint16_t, float2>(const float2& a, float scale) {
+  return __hip_cvt_float2_to_fp8x2(a / scale, fp8_type::__default_saturation,
+                                   fp8_type::__default_interpret);
+}
+
+// floatx4 -> fp8x4
+template <>
+__inline__ __device__ uint32_t
+scaled_vec_conversion<uint32_t, float4>(const float4& a, float scale) {
+  union {
+    uint16_t ui16[2];
+    uint32_t ui32;
+  } tmp;
+  tmp.ui16[0] = scaled_vec_conversion<uint16_t, float2>({a.x, a.y}, scale);
+  tmp.ui16[1] = scaled_vec_conversion<uint16_t, float2>({a.z, a.w}, scale);
+  return tmp.ui32;
 }
   #endif  // ENABLE_FP8
 

csrc/quantization/fp8/common.cuh

@@ -12,7 +12,7 @@ C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX =
     std::numeric_limits<FP8_TYPE>::max();
 #else
   #include <c10/util/Float8_e4m3fnuz.h>
-  #include "amd/hip_float8.h"
+  #include "amd/quant_utils.cuh"
 using FP8_TYPE = c10::Float8_e4m3fnuz;
 // Using the default max value from pytorch (240.0) will cause accuracy
 // issue when running dynamic quantization. Here use 224.0f for rocm.
@@ -47,8 +47,10 @@ __device__ __forceinline__ FP8_TYPE scaled_fp8_conversion(float const val,
   return static_cast<c10::Float8_e4m3fn>(r);
 #else
   // Use hardware cvt instruction for fp8 on rocm
-  return c10::Float8_e4m3fnuz(hip_fp8(r).data,
-                              c10::Float8_e4m3fnuz::from_bits());
+  return c10::Float8_e4m3fnuz(
+      __hip_cvt_float_to_fp8(r, fp8::fp8_type::__default_saturation,
+                             fp8::fp8_type::__default_interpret),
+      c10::Float8_e4m3fnuz::from_bits());
 #endif
 }
 

csrc/quantization/gguf/vecdotq.cuh

@@ -37,6 +37,8 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t *
     return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
 }
 
+// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
+// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4

csrc/quantization/machete/Readme.md

@@ -6,25 +6,25 @@ Machete is a spiritual successor to the Marlin kernel but optimized for Hopper a
 
 Machete effectively performs
 
-```
+```python
 scale_type = w_s.dtype
 compute_type = a.dtype
 out = (w_q.to(scale_type) * w_s - w_z.to(scale_type)) @ a
 ```
 
-Where `w_q` is a quantized weight matrix, `w_s` is the quantization scales, and 
+Where `w_q` is a quantized weight matrix, `w_s` is the quantization scales, and
 `w_z` is the quantization zeropoints.
 
-> **_NOTE:_**  `w_z` is added after the scales so we can 
+> **_NOTE:_**  `w_z` is added after the scales so we can
 use FMA operations, but this means they must have the scales pre-applied if the
-supplied zeropoints assume that they will be subtracted before the scales are 
+supplied zeropoints assume that they will be subtracted before the scales are
 applied.
 
 ## API
 
 The main optimization within Machete is prepacking the weight matrix to more closely match the tensor core layouts, allowing for wider shared memory loads when loading the weight matrix. This means that the weight matrix must be prepacked before calling `machete_gemm`. The flow looks something like:
 
-```
+```python
 from vllm import _custom_ops as ops
 
 ...
@@ -40,6 +40,6 @@ output = ops.machete_gemm(
 
 ## Code Generation
 
-Since Machete is based on Cutlass, we can generate multiple type pairs and different tile shapes using the same kernel template. We generate multiple instantiations of this template using `generate.py`. 
+Since Machete is based on Cutlass, we can generate multiple type pairs and different tile shapes using the same kernel template. We generate multiple instantiations of this template using `generate.py`.
 
-New type pairs (`TypeConfig`s) can be appended to `impl_configs` (in `generate()`), and these will get automatically generated (assuming they can be supported without issues). For each `TypeConfig`, you must also provide an `ImplConfig`, which bundles a `TypeConfig` with a list of `ScheduleConfig`s, `Specialization`s, and a default heuristic. The `ScheduleConfig`s (which contain info on tile shapes, tile scheduler, etc.) can perform differently for different problem shapes, and there is almost never one `ScheduleConfig` that works well for all problem shapes, so it is generally beneficial to generate different `ScheduleConfig`s for different potential problem shapes. This is where the heuristic comes in. For each `TypeConfig`, a default heuristic should be provided. This maps different problem shapes to different `ScheduleConfig`s and is used when the user does not provide the `schedule` parameter to `machete_gemm`. The `Specialization`s define what feature combinations to generate, i.e., `with_zeropoints`, `with_scales`, etc. We can reduce compile times and the final binary size by limiting the set of feature combinations we generate.
+New type pairs (`TypeConfig`s) can be appended to `impl_configs` (in `generate()`), and these will get automatically generated (assuming they can be supported without issues). For each `TypeConfig`, you must also provide an `ImplConfig`, which bundles a `TypeConfig` with a list of `ScheduleConfig`s, `Specialization`s, and a default heuristic. The `ScheduleConfig`s (which contain info on tile shapes, tile scheduler, etc.) can perform differently for different problem shapes, and there is almost never one `ScheduleConfig` that works well for all problem shapes, so it is generally beneficial to generate different `ScheduleConfig`s for different potential problem shapes. This is where the heuristic comes in. For each `TypeConfig`, a default heuristic should be provided. This maps different problem shapes to different `ScheduleConfig`s and is used when the user does not provide the `schedule` parameter to `machete_gemm`. The `Specialization`s define what feature combinations to generate, i.e., `with_zeropoints`, `with_scales`, etc. We can reduce compile times and the final binary size by limiting the set of feature combinations we generate.

csrc/rocm/attention.cu

@@ -24,8 +24,7 @@
 #include "../attention/dtype_fp8.cuh"
 #include "../quantization/fp8/amd/quant_utils.cuh"
 
-#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx940__) || \
-                           defined(__gfx941__) || defined(__gfx942__))
+#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx942__))
   #define __HIP__MI300_MI250__
 #endif
 
@@ -1122,4 +1121,4 @@ void paged_attention(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/sparse/cutlass/sparse_compressor_c3x.cu

@@ -1,165 +0,0 @@
-// clang-format will break include orders
-// clang-format off
-#include <cudaTypedefs.h>
-
-#if defined CUDA_VERSION && CUDA_VERSION >= 12020
-#include "sparse_scaled_mm_c3x.cuh"
-
-#include "cutlass/numeric_conversion.h"
-#include "cutlass/transform/device/transform_universal_adapter.hpp"
-#include "cutlass/transform/kernel/sparse_gemm_compressor.hpp"
-#include "cutlass/epilogue/collective/default_epilogue.hpp"
-
-#include "cutlass/util/host_tensor.h"
-#include "cutlass/util/packed_stride.hpp"
-// clang-format on
-
-using namespace cute;
-using namespace vllm;
-
-/// Make A structured sparse by replacing elements with 0 and compress it
-template <typename ElementA_, typename ElementAcc_>
-bool cutlass_sparse_compress(torch::Tensor& a_nzs, torch::Tensor& a_meta,
-                             torch::Tensor const& a) {
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 || a.dtype() == torch::kFloat8_e4m3fn ||
-              a.dtype() == torch::kFloat16 || a.dtype() == torch::kBFloat16);
-  TORCH_CHECK(a.dim() == 2)
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(0) % 4 == 0)  // Required for semi-structured sparsity
-  TORCH_CHECK(a.stride(1) == 1)
-
-  int m = a.size(0);
-  int k = a.size(1);
-
-  // Sparse kernel setup; this kernel is not used for matmul,
-  // but just for setting up the compressor utility
-  // A matrix configuration
-  using ElementA = ElementA_;
-  using LayoutTagA = cutlass::layout::RowMajor;
-  constexpr int AlignmentA = 128 / cutlass::sizeof_bits<ElementA>::value;
-  // B matrix configuration
-  using ElementB = ElementA;
-  using LayoutTagB = cutlass::layout::ColumnMajor;
-  constexpr int AlignmentB = 128 / cutlass::sizeof_bits<ElementB>::value;
-  // C/D matrix configuration
-  using ElementC = float;
-  using LayoutTagC = cutlass::layout::ColumnMajor;
-  constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementC>::value;
-  // Core kernel configurations
-  using ElementAccumulator = ElementAcc_;
-  using TileShape = Shape<_128, _128, _128>;
-  using TileShapeRef = Shape<_128, _128, _64>;
-  using ClusterShape = Shape<_1, _2, _1>;
-  using KernelSchedule = typename std::conditional<
-      std::is_same_v<ElementA, cutlass::float_e4m3_t>,
-      cutlass::gemm::KernelTmaWarpSpecializedFP8FastAccum,
-      cutlass::gemm::KernelTmaWarpSpecialized>::type;
-
-  using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecialized;
-  using ProblemShape = Shape<int, int, int, int>;
-
-  using CollectiveEpilogue =
-      typename cutlass::epilogue::collective::CollectiveBuilder<
-          cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp, TileShape,
-          ClusterShape, cutlass::epilogue::collective::EpilogueTileAuto,
-          ElementAccumulator, ElementAccumulator, ElementC, LayoutTagC,
-          AlignmentC, ElementC, LayoutTagC, AlignmentC,
-          EpilogueSchedule>::CollectiveOp;
-
-  using CollectiveMainloop =
-      typename cutlass::gemm::collective::CollectiveBuilder<
-          cutlass::arch::Sm90, cutlass::arch::OpClassSparseTensorOp, ElementA,
-          LayoutTagA, AlignmentA, ElementB, LayoutTagB, AlignmentB,
-          ElementAccumulator, TileShape, ClusterShape,
-          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
-              sizeof(typename CollectiveEpilogue::SharedStorage))>,
-          KernelSchedule>::CollectiveOp;
-
-  using GemmKernel =
-      cutlass::gemm::kernel::GemmUniversal<ProblemShape, CollectiveMainloop,
-                                           CollectiveEpilogue>;
-
-  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
-
-  using StrideA = cutlass::gemm::TagToStrideA_t<LayoutTagA>;
-  using StrideE = StrideA;
-
-  using StrideA = Stride<int64_t, Int<1>, int64_t>;
-
-  // The n (=1) dimension does not matter for the compressor
-  typename GemmKernel::ProblemShape prob_shape{m, 1, k, 1};
-
-  using LayoutA = typename GemmKernel::CollectiveMainloop::LayoutA;
-  using LayoutE = typename GemmKernel::CollectiveMainloop::LayoutE;
-
-  using ElementE = typename GemmKernel::CollectiveMainloop::ElementE;
-  using SparseConfig = typename GemmKernel::CollectiveMainloop::SparseConfig;
-
-  // Offline compressor kernel
-  using CompressorUtility =
-      cutlass::transform::kernel::StructuredSparseCompressorUtility<
-          ProblemShape, ElementA, LayoutTagA, SparseConfig>;
-
-  using CompressorKernel =
-      cutlass::transform::kernel::StructuredSparseCompressor<
-          ProblemShape, ElementA, LayoutTagA, SparseConfig,
-          cutlass::arch::Sm90>;
-
-  using Compressor =
-      cutlass::transform::device::TransformUniversalAdapter<CompressorKernel>;
-
-  auto [M, N, K, L] = prob_shape;
-
-  StrideA stride_A;
-  stride_A =
-      cutlass::make_cute_packed_stride(StrideA{}, cute::make_shape(M, K, L));
-
-  CompressorUtility compressor_utility(prob_shape, stride_A);
-
-  int ME = compressor_utility.get_metadata_m_physical();
-  int KE = compressor_utility.get_metadata_k_physical();
-  int KC = compressor_utility.get_tensorA_k_physical();
-
-  auto a_ptr = static_cast<ElementA*>(a.data_ptr());
-
-  auto a_nzs_ptr = static_cast<ElementA*>(a_nzs.data_ptr());
-  auto a_meta_ptr = static_cast<typename Gemm::CollectiveMainloop::ElementE*>(
-      a_meta.data_ptr());
-
-  cutlass::KernelHardwareInfo hw_info;
-  hw_info.device_id = 0;
-  hw_info.sm_count =
-      cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
-          hw_info.device_id);
-  typename Compressor::Arguments arguments{
-      prob_shape, {a_ptr, stride_A, a_nzs_ptr, a_meta_ptr}, {hw_info}};
-
-  Compressor compressor_op;
-  size_t workspace_size = Compressor::get_workspace_size(arguments);
-  cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
-
-  CUTLASS_CHECK(compressor_op.can_implement(arguments));
-  CUTLASS_CHECK(compressor_op.initialize(arguments, workspace.get()));
-  CUTLASS_CHECK(compressor_op.run());
-  CUDA_CHECK(cudaDeviceSynchronize());
-
-  return true;
-}
-
-bool cutlass_sparse_compress_sm90(torch::Tensor& a_nzs, torch::Tensor& a_meta,
-                                  torch::Tensor const& a) {
-  if (a.dtype() == torch::kBFloat16) {
-    return cutlass_sparse_compress<cutlass::bfloat16_t, float>(a_nzs, a_meta,
-                                                               a);
-  } else if (a.dtype() == torch::kFloat16) {
-    return cutlass_sparse_compress<cutlass::half_t, float>(a_nzs, a_meta, a);
-  } else if (a.dtype() == torch::kFloat8_e4m3fn) {
-    return cutlass_sparse_compress<cutlass::float_e4m3_t, float>(a_nzs, a_meta,
-                                                                 a);
-  } else if (a.dtype() == torch::kInt8) {
-    return cutlass_sparse_compress<int8_t, int32_t>(a_nzs, a_meta, a);
-  }
-  return false;
-}
-#endif

csrc/sparse/cutlass/sparse_compressor_c3x.cuh

@@ -0,0 +1,90 @@
+#pragma once
+
+// clang-format will break include orders
+// clang-format off
+#include <cudaTypedefs.h>
+
+#if defined CUDA_VERSION && CUDA_VERSION >= 12020
+#include "sparse_scaled_mm_c3x.cuh"
+
+#include "cutlass/numeric_conversion.h"
+#include "cutlass/transform/device/transform_universal_adapter.hpp"
+#include "cutlass/transform/kernel/sparse_gemm_compressor.hpp"
+#include "cutlass/epilogue/collective/default_epilogue.hpp"
+
+// clang-format on
+
+using namespace cute;
+using namespace vllm;
+
+using CompressorResult = std::tuple<torch::Tensor, torch::Tensor>;
+/// Make A structured sparse by replacing elements with 0 and compress it
+template <typename Gemm>
+CompressorResult cutlass_sparse_compress(torch::Tensor const& a) {
+  // Checks for conformality
+  TORCH_CHECK(a.dtype() == torch::kInt8 || a.dtype() == torch::kFloat8_e4m3fn ||
+              a.dtype() == torch::kFloat16 || a.dtype() == torch::kBFloat16);
+  TORCH_CHECK(a.dim() == 2)
+  // Check for strides and alignment
+  TORCH_CHECK(a.stride(0) % 4 == 0)  // Required for semi-structured sparsity
+  TORCH_CHECK(a.stride(1) == 1)
+
+  using GemmKernel = typename Gemm::KernelType;
+  using ElementA = typename Gemm::ElementAB;
+  using ElementE = typename GemmKernel::CollectiveMainloop::ElementE;
+
+  int m = a.size(0);
+  int k = a.size(1);
+  using ProblemShape = typename GemmKernel::ProblemShape;
+  ProblemShape prob_shape{m, 1, k, 1};
+
+  int64_t lda = a.stride(0);
+  using StrideA = Stride<int64_t, Int<1>, int64_t>;
+  StrideA a_stride{lda, Int<1>{}, 0};
+
+  using CompressorUtility = typename Gemm::CompressorUtility;
+  CompressorUtility compressor_utility(prob_shape, a_stride);
+
+  // Allocate buffers for the metadata E and the compressed matrix A
+  int ME = compressor_utility.get_metadata_m_physical();
+  int KE = compressor_utility.get_metadata_k_physical();
+  int MC = compressor_utility.get_tensorA_m_physical();
+  int KC = compressor_utility.get_tensorA_k_physical();
+
+  auto const a_meta_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(a.device());
+  auto const a_nzs_options =
+      torch::TensorOptions().dtype(a.dtype()).device(a.device());
+
+  auto a_meta = torch::zeros({ME, KE}, a_meta_options);
+  auto a_nzs = torch::zeros({MC, KC}, a_nzs_options);
+
+  auto a_ptr = static_cast<ElementA*>(a.data_ptr());
+  auto a_nzs_ptr = static_cast<ElementA*>(a_nzs.data_ptr());
+  auto a_meta_ptr = static_cast<ElementE*>(a_meta.data_ptr());
+
+  cutlass::KernelHardwareInfo hw_info;
+  hw_info.device_id = a.device().index();
+  hw_info.sm_count =
+      cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
+          hw_info.device_id);
+
+  using Compressor = typename Gemm::Compressor;
+  typename Compressor::Arguments arguments{
+      prob_shape, {a_ptr, a_stride, a_nzs_ptr, a_meta_ptr}, {hw_info}};
+
+  Compressor compressor_op;
+  size_t workspace_size = Compressor::get_workspace_size(arguments);
+  auto const workspace_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(a.device());
+  auto workspace = torch::empty(workspace_size, workspace_options);
+
+  CUTLASS_CHECK(compressor_op.can_implement(arguments));
+  CUTLASS_CHECK(compressor_op.initialize(arguments, workspace.data_ptr()));
+  CUTLASS_CHECK(compressor_op.run());
+  CUDA_CHECK(cudaDeviceSynchronize());
+
+  return {a_meta, a_nzs};
+}
+
+#endif

csrc/sparse/cutlass/sparse_compressor_entry.cu

@@ -1,42 +0,0 @@
-#include <cudaTypedefs.h>
-
-#include <c10/cuda/CUDAGuard.h>
-#include <torch/all.h>
-
-#include "cutlass_extensions/common.hpp"
-
-#if defined ENABLE_SPARSE_SCALED_MM_C3X && ENABLE_SPARSE_SCALED_MM_C3X
-bool cutlass_sparse_compress_sm90(torch::Tensor& a_nzs, torch::Tensor& a_meta,
-                                  torch::Tensor const& a);
-#endif
-
-bool cutlass_sparse_compress_entry(torch::Tensor& a_nzs, torch::Tensor& a_meta,
-                                   torch::Tensor const& a) {
-  // Checks for conformality
-  TORCH_CHECK(a.dim() == 2 && a_meta.dim() == 2 && a_nzs.dim() == 2);
-  TORCH_CHECK(a.size(0) == a_nzs.size(0) && a.size(0) == a_meta.size(0) &&
-              a_nzs.size(1) * 2 == a.size(1) &&
-              a_meta.size(1) * 2 * 4 == a.size(1));
-  // Considering elemsPerMetaElem = 8b / 2b_per_nz = 4
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && a_nzs.stride(1) == 1 &&
-              a_meta.stride(1) == 1);  // Row-major
-  TORCH_CHECK(a.stride(0) % 8 == 0);   // 8 Byte Alignment for Compression
-
-  at::cuda::OptionalCUDAGuard const device_guard(device_of(a));
-  int32_t version_num = get_sm_version_num();
-
-  // Guard against compilation issues for sm90 kernels
-#if defined ENABLE_SPARSE_SCALED_MM_C3X && ENABLE_SPARSE_SCALED_MM_C3X
-  if (version_num >= 90) {
-    return cutlass_sparse_compress_sm90(a_nzs, a_meta, a);
-  }
-#endif
-
-  TORCH_CHECK_NOT_IMPLEMENTED(
-      false,
-      "No compiled cutlass_scaled_sparse_mm for a compute capability less than "
-      "CUDA device capability: ",
-      version_num);
-}

csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu

@@ -9,17 +9,30 @@
 using namespace cute;
 using namespace vllm;
 
+struct GemmCallerTraits {
+  using return_type = void;
+
+  template <typename GemmConfig, typename... Args>
+  static return_type invoke(Args&&... args) {
+    return cutlass_sparse_gemm_caller<GemmConfig>(std::forward<Args>(args)...);
+  }
+};
+
+struct GemmCompressorTraits {
+  using return_type = CompressorResult;
+
+  template <typename GemmConfig, typename... Args>
+  static return_type invoke(Args&&... args) {
+    return cutlass_sparse_compress<GemmConfig>(std::forward<Args>(args)...);
+  }
+};
+
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue,
-          typename... EpilogueArgs>
-void cutlass_gemm_sm90_fp8_dispatch(torch::Tensor& out, torch::Tensor const& a,
-                                    torch::Tensor const& bt_nzs,
-                                    torch::Tensor const& bt_meta,
-                                    EpilogueArgs&&... args) {
-  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
-  TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
-  TORCH_CHECK(bt_meta.dtype() == torch::kUInt8);
-  TORCH_CHECK(bt_nzs.dtype() == torch::kFloat8_e4m3fn);
+          typename DispatchFunc, typename... Args>
+typename DispatchFunc::return_type cutlass_gemm_sm90_fp8_dispatch(
+    uint32_t m, uint32_t n, Args&&... args) {
+  static_assert(std::is_same_v<InType, cutlass::float_e4m3_t>);
 
   using Cutlass3xGemmDefault =
       typename sm90_config_default<InType, OutType, Epilogue>::Cutlass3xGemm;
@@ -49,122 +62,87 @@ void cutlass_gemm_sm90_fp8_dispatch(torch::Tensor& out, torch::Tensor const& a,
   using Cutlass3xGemm8 =
       typename sm90_fp8_config_8<InType, OutType, Epilogue>::Cutlass3xGemm;
 
-  uint32_t const n = bt_nzs.size(0);
-  uint32_t const m = a.size(0);  // Batch size
   uint32_t const mp2 =
       std::max(static_cast<uint32_t>(64), next_pow_2(m));  // next power of 2
 
   if (mp2 <= 64) {
     if (n == 28672) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm2>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm2>(
+          std::forward<Args>(args)...);
     } else if (n == 4096 || n == 6144) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm1>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm1>(
+          std::forward<Args>(args)...);
     }
   } else if (mp2 <= 128) {
     if (n == 4096) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm3>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm3>(
+          std::forward<Args>(args)...);
     } else if (n == 28672) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm5>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm5>(
+          std::forward<Args>(args)...);
     } else if (n == 6144) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm4>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm4>(
+          std::forward<Args>(args)...);
     }
   } else if (mp2 <= 256) {
     if (n == 4096) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm6>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm6>(
+          std::forward<Args>(args)...);
     } else if (n == 28672) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm8>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm8>(
+          std::forward<Args>(args)...);
     } else if (n == 6144) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm7>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm7>(
+          std::forward<Args>(args)...);
     }
   } else {
     if (n == 6144 || n == 28672) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm8>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm8>(
+          std::forward<Args>(args)...);
     } else if (n == 4096) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemm7>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemm7>(
+          std::forward<Args>(args)...);
     }
   }
 
   // Otherwise the default heuristic
   if (mp2 <= 64) {
     // n in [1, 64]
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmM64>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmM64>(
+        std::forward<Args>(args)...);
   } else if (mp2 <= 128) {
     // n in (64, 128]
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmM128>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmM128>(
+        std::forward<Args>(args)...);
   } else if (mp2 <= 256) {
     // n in (128, 256]
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmM256>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmM256>(
+        std::forward<Args>(args)...);
   } else {
     // n in (256, inf)
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmM512>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmM512>(
+        std::forward<Args>(args)...);
   }
 }
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue,
-          typename... EpilogueArgs>
-void cutlass_gemm_sm90_fp16_dispatch(torch::Tensor& out, torch::Tensor const& a,
-                                     torch::Tensor const& bt_nzs,
-                                     torch::Tensor const& bt_meta,
-                                     EpilogueArgs&&... args) {
-  static_assert(std::is_same<InType, cutlass::half_t>());
-  TORCH_CHECK(a.dtype() == torch::kFloat16);
-  TORCH_CHECK(bt_meta.dtype() == torch::kUInt8);
-  TORCH_CHECK(bt_nzs.dtype() == torch::kFloat16);
-
+          typename DispatchFunc, typename... Args>
+typename DispatchFunc::return_type cutlass_gemm_sm90_16bit_dispatch(
+    uint32_t m, uint32_t n, Args&&... args) {
   using Cutlass3xGemmDefault =
       typename sm90_config_default<InType, OutType, Epilogue>::Cutlass3xGemm;
 
-  // m in (128, inf)
-  return cutlass_sparse_gemm_caller<Cutlass3xGemmDefault>(
-      out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+  return DispatchFunc::template invoke<Cutlass3xGemmDefault>(
+      std::forward<Args>(args)...);
 }
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue,
-          typename... EpilogueArgs>
-void cutlass_gemm_sm90_bf16_dispatch(torch::Tensor& out, torch::Tensor const& a,
-                                     torch::Tensor const& bt_nzs,
-                                     torch::Tensor const& bt_meta,
-                                     EpilogueArgs&&... args) {
-  static_assert(std::is_same<InType, cutlass::bfloat16_t>());
-  TORCH_CHECK(a.dtype() == torch::kBFloat16);
-  TORCH_CHECK(bt_meta.dtype() == torch::kUInt8);
-  TORCH_CHECK(bt_nzs.dtype() == torch::kBFloat16);
-
-  using Cutlass3xGemmDefault =
-      typename sm90_config_default<InType, OutType, Epilogue>::Cutlass3xGemm;
-
-  // m in (128, inf)
-  return cutlass_sparse_gemm_caller<Cutlass3xGemmDefault>(
-      out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
-}
-
-template <typename InType, typename OutType,
-          template <typename, typename, typename> typename Epilogue,
-          typename... EpilogueArgs>
-void cutlass_gemm_sm90_int8_dispatch(torch::Tensor& out, torch::Tensor const& a,
-                                     torch::Tensor const& bt_nzs,
-                                     torch::Tensor const& bt_meta,
-                                     EpilogueArgs&&... args) {
-  static_assert(std::is_same<InType, int8_t>());
-  TORCH_CHECK(a.dtype() == torch::kInt8);
-  TORCH_CHECK(bt_meta.dtype() == torch::kUInt8);
-  TORCH_CHECK(bt_nzs.dtype() == torch::kInt8);
+          typename DispatchFunc, typename... Args>
+typename DispatchFunc::return_type cutlass_gemm_sm90_int8_dispatch(
+    uint32_t m, uint32_t n, Args&&... args) {
+  static_assert(std::is_same_v<InType, int8_t>);
 
   using Cutlass3xGemmDefault =
       typename sm90_config_default<InType, OutType, Epilogue>::Cutlass3xGemm;
@@ -179,37 +157,35 @@ void cutlass_gemm_sm90_int8_dispatch(torch::Tensor& out, torch::Tensor const& a,
       typename sm90_int8_config_M32_NSmall<InType, OutType,
                                            Epilogue>::Cutlass3xGemm;
 
-  uint32_t const n = out.size(1);
   bool const is_small_n = n < 8192;
-
-  uint32_t const m = a.size(0);
   uint32_t const mp2 =
       std::max(static_cast<uint32_t>(32), next_pow_2(m));  // next power of 2
 
   if (mp2 <= 32) {
     // m in [1, 32]
     if (is_small_n) {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemmM32NSmall>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemmM32NSmall>(
+          std::forward<Args>(args)...);
     } else {
-      return cutlass_sparse_gemm_caller<Cutlass3xGemmM32NBig>(
-          out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+      return DispatchFunc::template invoke<Cutlass3xGemmM32NBig>(
+          std::forward<Args>(args)...);
     }
   } else if (mp2 <= 64) {
     // m in (32, 64]
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmM64>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmM64>(
+        std::forward<Args>(args)...);
   } else if (mp2 <= 128) {
     // m in (64, 128]
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmM128>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmM128>(
+        std::forward<Args>(args)...);
   } else {
     // m in (128, inf)
-    return cutlass_sparse_gemm_caller<Cutlass3xGemmDefault>(
-        out, a, bt_nzs, bt_meta, std::forward<EpilogueArgs>(args)...);
+    return DispatchFunc::template invoke<Cutlass3xGemmDefault>(
+        std::forward<Args>(args)...);
   }
 }
 
+// Dispatch to GEMM implementations based on element types
 template <template <typename, typename, typename> typename Epilogue,
           typename... EpilogueArgs>
 void cutlass_scaled_sparse_mm_sm90_epilogue(torch::Tensor& out,
@@ -217,19 +193,24 @@ void cutlass_scaled_sparse_mm_sm90_epilogue(torch::Tensor& out,
                                             torch::Tensor const& bt_nzs,
                                             torch::Tensor const& bt_meta,
                                             EpilogueArgs&&... epilogue_args) {
+  uint32_t const m = out.size(0);
+  uint32_t const n = out.size(1);
+
+  // TODO: add dispatch functions to all of these
   TORCH_CHECK(bt_meta.dtype() == torch::kUInt8);
   if (a.dtype() == torch::kInt8) {
     TORCH_CHECK(bt_nzs.dtype() == torch::kInt8);
 
     if (out.dtype() == torch::kBFloat16) {
       return cutlass_gemm_sm90_int8_dispatch<int8_t, cutlass::bfloat16_t,
-                                             Epilogue>(
-          out, a, bt_nzs, bt_meta,
+                                             Epilogue, GemmCallerTraits>(
+          m, n, out, a, bt_nzs, bt_meta,
           std::forward<EpilogueArgs>(epilogue_args)...);
     } else {
       TORCH_CHECK(out.dtype() == torch::kFloat16);
-      return cutlass_gemm_sm90_int8_dispatch<int8_t, cutlass::half_t, Epilogue>(
-          out, a, bt_nzs, bt_meta,
+      return cutlass_gemm_sm90_int8_dispatch<int8_t, cutlass::half_t, Epilogue,
+                                             GemmCallerTraits>(
+          m, n, out, a, bt_nzs, bt_meta,
           std::forward<EpilogueArgs>(epilogue_args)...);
     }
   } else if (a.dtype() == torch::kFloat8_e4m3fn) {
@@ -237,47 +218,34 @@ void cutlass_scaled_sparse_mm_sm90_epilogue(torch::Tensor& out,
 
     if (out.dtype() == torch::kBFloat16) {
       return cutlass_gemm_sm90_fp8_dispatch<cutlass::float_e4m3_t,
-                                            cutlass::bfloat16_t, Epilogue>(
-          out, a, bt_nzs, bt_meta,
+                                            cutlass::bfloat16_t, Epilogue,
+                                            GemmCallerTraits>(
+          m, n, out, a, bt_nzs, bt_meta,
           std::forward<EpilogueArgs>(epilogue_args)...);
     } else {
       TORCH_CHECK(out.dtype() == torch::kFloat16);
-      return cutlass_gemm_sm90_fp8_dispatch<cutlass::float_e4m3_t,
-                                            cutlass::half_t, Epilogue>(
-          out, a, bt_nzs, bt_meta,
+      return cutlass_gemm_sm90_fp8_dispatch<
+          cutlass::float_e4m3_t, cutlass::half_t, Epilogue, GemmCallerTraits>(
+          m, n, out, a, bt_nzs, bt_meta,
           std::forward<EpilogueArgs>(epilogue_args)...);
     }
   } else if (a.dtype() == torch::kFloat16) {
     TORCH_CHECK(bt_nzs.dtype() == torch::kFloat16);
+    TORCH_CHECK(out.dtype() == torch::kFloat16);
 
-    if (out.dtype() == torch::kBFloat16) {
-      return cutlass_gemm_sm90_fp16_dispatch<cutlass::half_t,
-                                             cutlass::bfloat16_t, Epilogue>(
-          out, a, bt_nzs, bt_meta,
-          std::forward<EpilogueArgs>(epilogue_args)...);
-    } else {
-      TORCH_CHECK(out.dtype() == torch::kFloat16);
-      return cutlass_gemm_sm90_fp16_dispatch<cutlass::half_t, cutlass::half_t,
-                                             Epilogue>(
-          out, a, bt_nzs, bt_meta,
-          std::forward<EpilogueArgs>(epilogue_args)...);
-    }
+    return cutlass_gemm_sm90_16bit_dispatch<cutlass::half_t, cutlass::half_t,
+                                            Epilogue, GemmCallerTraits>(
+        m, n, out, a, bt_nzs, bt_meta,
+        std::forward<EpilogueArgs>(epilogue_args)...);
   } else {  // a.dtype() == torch::kBFloat16
     TORCH_CHECK(a.dtype() == torch::kBFloat16);
     TORCH_CHECK(bt_nzs.dtype() == torch::kBFloat16);
+    TORCH_CHECK(out.dtype() == torch::kBFloat16);
 
-    if (out.dtype() == torch::kBFloat16) {
-      return cutlass_gemm_sm90_bf16_dispatch<cutlass::bfloat16_t,
-                                             cutlass::bfloat16_t, Epilogue>(
-          out, a, bt_nzs, bt_meta,
-          std::forward<EpilogueArgs>(epilogue_args)...);
-    } else {
-      TORCH_CHECK(out.dtype() == torch::kFloat16);
-      return cutlass_gemm_sm90_bf16_dispatch<cutlass::bfloat16_t,
-                                             cutlass::half_t, Epilogue>(
-          out, a, bt_nzs, bt_meta,
-          std::forward<EpilogueArgs>(epilogue_args)...);
-    }
+    return cutlass_gemm_sm90_16bit_dispatch<
+        cutlass::bfloat16_t, cutlass::bfloat16_t, Epilogue, GemmCallerTraits>(
+        m, n, out, a, bt_nzs, bt_meta,
+        std::forward<EpilogueArgs>(epilogue_args)...);
   }
 }
 
@@ -287,17 +255,53 @@ void cutlass_scaled_sparse_mm_sm90(torch::Tensor& out, torch::Tensor const& a,
                                    torch::Tensor const& a_scales,
                                    torch::Tensor const& b_scales,
                                    std::optional<torch::Tensor> const& bias) {
+  TORCH_CHECK(bt_meta.dtype() == torch::kUInt8);
   TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
   TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+
   if (bias) {
     TORCH_CHECK(bias->dtype() == out.dtype(),
-                "currently bias dtype must match output dtype ", out.dtype());
-    return cutlass_scaled_sparse_mm_sm90_epilogue<c3x::ScaledEpilogueBias>(
-        out, a, bt_nzs, bt_meta, b_scales, a_scales, *bias);
+                "CUTLASS scaled_mm bias dtype must match output dtype ",
+                out.dtype());
+    return cutlass_scaled_sparse_mm_sm90_epilogue<
+        c3x::ScaledEpilogueColumnBias>(out, a, bt_nzs, bt_meta, b_scales,
+                                       a_scales, *bias);
   } else {
     return cutlass_scaled_sparse_mm_sm90_epilogue<c3x::ScaledEpilogue>(
         out, a, bt_nzs, bt_meta, b_scales, a_scales);
   }
 }
 
+CompressorResult cutlass_sparse_compress_sm90(torch::Tensor const& a) {
+  // These m and n variables are fordispatching to different GEMM algorithms.
+  uint32_t const m = 1;  // Set M to 1 for compression
+  uint32_t const n = a.size(1);
+
+  // Note: For correctess, the compressed format must be invariant in:
+  //  - M, the flattened number of tokens
+  //  - Whether output dtype is fp16 or bf16
+  //  - CUTLASS epilogues
+
+  if (a.dtype() == torch::kInt8) {
+    return cutlass_gemm_sm90_int8_dispatch<int8_t, cutlass::bfloat16_t,
+                                           c3x::TrivialEpilogue,
+                                           GemmCompressorTraits>(m, n, a);
+  } else if (a.dtype() == torch::kFloat8_e4m3fn) {
+    return cutlass_gemm_sm90_fp8_dispatch<
+        cutlass::float_e4m3_t, cutlass::bfloat16_t, c3x::TrivialEpilogue,
+        GemmCompressorTraits>(m, n, a);
+  } else if (a.dtype() == torch::kFloat16) {
+    return cutlass_gemm_sm90_16bit_dispatch<
+        cutlass::bfloat16_t, cutlass::bfloat16_t, c3x::TrivialEpilogue,
+        GemmCompressorTraits>(m, n, a);
+  } else {
+    TORCH_CHECK(a.dtype() == torch::kBFloat16,
+                "cutlass_sparse_compress only supports int8, fp8_e4m3, fp16, "
+                "and bf16 datatypes");
+    return cutlass_gemm_sm90_16bit_dispatch<cutlass::half_t, cutlass::half_t,
+                                            c3x::TrivialEpilogue,
+                                            GemmCompressorTraits>(m, n, a);
+  }
+}
+
 #endif

csrc/sparse/cutlass/sparse_scaled_mm_c3x.cuh

@@ -1,3 +1,5 @@
+#pragma once
+
 // clang-format will break include orders
 // clang-format off
 #include <cudaTypedefs.h>
@@ -12,6 +14,9 @@
 #include "cutlass/epilogue/collective/collective_builder.hpp"
 #include "cutlass/gemm/collective/collective_builder.hpp"
 
+#include "cutlass/transform/device/transform_universal_adapter.hpp"
+#include "cutlass/transform/kernel/sparse_gemm_compressor.hpp"
+
 #include "core/math.hpp"
 #include "cutlass_extensions/cute_utils.cuh"
 #include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
@@ -22,7 +27,7 @@
 using namespace cute;
 
 /*
-   This file defines sparse quantized GEMM operations using the CUTLASS 3.x API,
+   This file defines 2:4 sparse GEMM operations using the CUTLASS 3.x API,
    for NVIDIA GPUs with sm90a (Hopper) or later.
 */
 
@@ -45,17 +50,20 @@ struct enable_sm90_or_later : Kernel {
 
 using GemmUniversalMode = cutlass::gemm::GemmUniversalMode;
 
+/*
+ * cutlass_sparse_3x_gemm defines a 2:4 sparse GEMM kernel via CUTLASS
+ * for SM90 Hopper systems.
+ */
 template <typename ElementAB_, typename ElementD_,
           template <typename, typename, typename> typename Epilogue_,
           typename TileShape, typename ClusterShape, typename KernelSchedule,
-          typename EpilogueSchedule, typename AccType,
-          typename TileSchedule = cutlass::gemm::PersistentScheduler,
-          GemmUniversalMode Mode_ = GemmUniversalMode::kGemm>
+          typename EpilogueSchedule>
 struct cutlass_sparse_3x_gemm {
-  static const GemmUniversalMode Mode = Mode_;
   using ElementAB = ElementAB_;
   using ElementD = ElementD_;
-  using ElementAcc = AccType;
+  using ElementAcc =
+      typename std::conditional<std::is_same_v<ElementAB, int8_t>, int32_t,
+                                float>::type;
 
   using EpilogueDescriptor =
       cutlass::epilogue::collective::detail::EpilogueDescriptor<
@@ -66,30 +74,22 @@ struct cutlass_sparse_3x_gemm {
 
   using ElementC = void;
   using LayoutC = cutlass::layout::RowMajor;
-  using LayoutD = LayoutC;
-  using StrideC = cutlass::detail::TagToStrideA_t<LayoutC>;
-  using StrideD = cutlass::detail::TagToStrideA_t<LayoutD>;
-
   using LayoutC_Transpose =
       typename cutlass::layout::LayoutTranspose<LayoutC>::type;
-  using LayoutD_Transpose =
-      typename cutlass::layout::LayoutTranspose<LayoutD>::type;
 
   using EVTCompute = typename Epilogue::EVTCompute;
 
-  static constexpr int AlignmentA =
+  // These are the minimum alignments needed for the kernels to compile
+  static constexpr int AlignmentAB =
       128 / cutlass::sizeof_bits<ElementAB>::value;
-  static constexpr int AlignmentB =
-      128 / cutlass::sizeof_bits<ElementAB>::value;
-  static constexpr int AlignmentCD =
-      128 / cutlass::sizeof_bits<ElementD>::value;
+  static constexpr int AlignmentCD = 4;
 
   using CollectiveEpilogue =
       typename cutlass::epilogue::collective::CollectiveBuilder<
           cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp, TileShape,
           ClusterShape, cutlass::epilogue::collective::EpilogueTileAuto,
-          ElementAcc, ElementAcc, ElementC, LayoutC_Transpose, AlignmentCD,
-          ElementD, LayoutD_Transpose, AlignmentCD, EpilogueSchedule,
+          ElementAcc, float, ElementC, LayoutC_Transpose, AlignmentCD, ElementD,
+          LayoutC_Transpose, AlignmentCD, EpilogueSchedule,
           EVTCompute>::CollectiveOp;
 
   static constexpr size_t CEStorageSize =
@@ -101,8 +101,8 @@ struct cutlass_sparse_3x_gemm {
   using CollectiveMainloop =
       typename cutlass::gemm::collective::CollectiveBuilder<
           cutlass::arch::Sm90, cutlass::arch::OpClassSparseTensorOp, 
-          ElementAB, cutlass::layout::RowMajor, AlignmentA, 
-          ElementAB, cutlass::layout::ColumnMajor, AlignmentB, 
+          ElementAB, cutlass::layout::RowMajor, AlignmentAB, 
+          ElementAB, cutlass::layout::ColumnMajor, AlignmentAB, 
           ElementAcc, TileShape, ClusterShape,
           Stages,
           KernelSchedule>::CollectiveOp;
@@ -110,11 +110,100 @@ struct cutlass_sparse_3x_gemm {
 
   using KernelType = enable_sm90_or_later<cutlass::gemm::kernel::GemmUniversal<
       cute::Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue,
-      TileSchedule>>;
+      cutlass::gemm::PersistentScheduler>>;
 
   struct GemmKernel : public KernelType {};
+
+  // Sparse compressor definitions
+  using SparseConfig = typename GemmKernel::CollectiveMainloop::SparseConfig;
+  using LayoutTagA = cutlass::layout::RowMajor;
+  using CompressorUtility =
+      cutlass::transform::kernel::StructuredSparseCompressorUtility<
+          typename GemmKernel::ProblemShape, ElementAB, LayoutTagA,
+          SparseConfig>;
+  using CompressorKernel =
+      cutlass::transform::kernel::StructuredSparseCompressor<
+          typename GemmKernel::ProblemShape, ElementAB, LayoutTagA,
+          SparseConfig, cutlass::arch::Sm90>;
+  using Compressor =
+      cutlass::transform::device::TransformUniversalAdapter<CompressorKernel>;
 };
 
+/*
+ * This class defines kernel to compress a 2:4 sparse matrix.
+ * The particular format is defined by the Gemm template parameter,
+ * which is a cutlass_sparse_3x_gemm.
+ */
+using CompressorResult = std::tuple<torch::Tensor, torch::Tensor>;
+/// Make A structured sparse by replacing elements with 0 and compress it
+template <typename Gemm>
+CompressorResult cutlass_sparse_compress(torch::Tensor const& a) {
+  // Checks for conformality
+  TORCH_CHECK(a.dtype() == torch::kInt8 || a.dtype() == torch::kFloat8_e4m3fn ||
+              a.dtype() == torch::kFloat16 || a.dtype() == torch::kBFloat16);
+  TORCH_CHECK(a.dim() == 2)
+  // Check for strides and alignment
+  TORCH_CHECK(a.stride(0) % 4 == 0)  // Required for semi-structured sparsity
+  TORCH_CHECK(a.stride(1) == 1)
+
+  using GemmKernel = typename Gemm::KernelType;
+  using ElementA = typename Gemm::ElementAB;
+  using ElementE = typename GemmKernel::CollectiveMainloop::ElementE;
+
+  int m = a.size(0);
+  int k = a.size(1);
+  using ProblemShape = typename GemmKernel::ProblemShape;
+  ProblemShape prob_shape{m, 1, k, 1};
+
+  int64_t lda = a.stride(0);
+  using StrideA = Stride<int64_t, Int<1>, int64_t>;
+  StrideA a_stride{lda, Int<1>{}, 0};
+
+  using CompressorUtility = typename Gemm::CompressorUtility;
+  CompressorUtility compressor_utility(prob_shape, a_stride);
+
+  // Allocate buffers for the metadata E and the compressed matrix A
+  int ME = compressor_utility.get_metadata_m_physical();
+  int KE = compressor_utility.get_metadata_k_physical();
+  int MC = compressor_utility.get_tensorA_m_physical();
+  int KC = compressor_utility.get_tensorA_k_physical();
+
+  auto const a_meta_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(a.device());
+  auto const a_nzs_options =
+      torch::TensorOptions().dtype(a.dtype()).device(a.device());
+
+  auto a_meta = torch::zeros({ME, KE}, a_meta_options);
+  auto a_nzs = torch::zeros({MC, KC}, a_nzs_options);
+
+  auto a_ptr = static_cast<ElementA*>(a.data_ptr());
+  auto a_nzs_ptr = static_cast<ElementA*>(a_nzs.data_ptr());
+  auto a_meta_ptr = static_cast<ElementE*>(a_meta.data_ptr());
+
+  cutlass::KernelHardwareInfo hw_info;
+  hw_info.device_id = a.device().index();
+  hw_info.sm_count =
+      cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
+          hw_info.device_id);
+
+  using Compressor = typename Gemm::Compressor;
+  typename Compressor::Arguments arguments{
+      prob_shape, {a_ptr, a_stride, a_nzs_ptr, a_meta_ptr}, {hw_info}};
+
+  Compressor compressor_op;
+  size_t workspace_size = Compressor::get_workspace_size(arguments);
+  auto const workspace_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(a.device());
+  auto workspace = torch::empty(workspace_size, workspace_options);
+
+  CUTLASS_CHECK(compressor_op.can_implement(arguments));
+  CUTLASS_CHECK(compressor_op.initialize(arguments, workspace.data_ptr()));
+  CUTLASS_CHECK(compressor_op.run());
+  CUDA_CHECK(cudaDeviceSynchronize());
+
+  return {a_meta, a_nzs};
+}
+
 template <typename Gemm, typename... EpilogueArgs>
 void cutlass_sparse_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
                                 torch::Tensor const& bt_nzs,
@@ -126,27 +215,25 @@ void cutlass_sparse_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
   // Interface stride expected from the argument a (will get transposed)
   // We compute C^T = B^T * A^T, but we assume B is transposed before
   // compression and hence the bt_* naming
-  using LayoutA = cutlass::layout::RowMajor;
   using LayoutB = typename Gemm::GemmKernel::CollectiveMainloop::LayoutA;
   using LayoutE = typename Gemm::GemmKernel::CollectiveMainloop::LayoutE;
-  using LayoutD = cutlass::layout::RowMajor;
 
-  using StrideA = cutlass::detail::TagToStrideA_t<LayoutA>;
-  using StrideD = cutlass::detail::TagToStrideA_t<LayoutD>;
+  // M, N, K after transposition
+  int32_t m = out.size(1);
+  int32_t n = out.size(0);
+  int32_t k = a.size(1);
 
-  auto layout_A = make_cute_layout<StrideA>(a, "A");
-  auto layout_D = make_cute_layout<StrideD>(out, "D");
+  int64_t lda = a.stride(0);
+  int64_t ldc = out.stride(0);
 
-  // Transpose A and D
-  // A doesn't need to be transposed since cutlass expects a NxK matrix
-  // for B (which is At)
-  auto stride_At = layout_A.stride();
-  auto stride_Dt = permute_layout<1, 0, 2>(layout_D).stride();
+  using StrideA = Stride<int64_t, Int<1>, int64_t>;
+  using StrideC = Stride<Int<1>, int64_t, int64_t>;
+
+  StrideA a_stride{lda, Int<1>{}, Int<0>{}};
+  StrideC c_stride{Int<1>{}, ldc, Int<0>{}};
 
   using GemmKernel = typename Gemm::GemmKernel;
-  typename GemmKernel::ProblemShape prob_shape{
-      static_cast<int>(bt_nzs.size(0)), static_cast<int>(size<0>(layout_A)),
-      static_cast<int>(size<1>(layout_A)), 1};
+  typename GemmKernel::ProblemShape prob_shape{m, n, k, 1};
 
   using ElementE = typename GemmKernel::CollectiveMainloop::ElementE;
   using SparseConfig = typename GemmKernel::CollectiveMainloop::SparseConfig;
@@ -158,13 +245,13 @@ void cutlass_sparse_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
   auto b_ptr = static_cast<ElementAB*>(bt_nzs.data_ptr());
   auto e_ptr = static_cast<ElementE*>(bt_meta.data_ptr());
   typename GemmKernel::MainloopArguments mainloop_args{
-      b_ptr, b_layout, a_ptr, stride_At, e_ptr, e_layout};
+      b_ptr, b_layout, a_ptr, a_stride, e_ptr, e_layout};
 
   auto c_ptr = static_cast<ElementD*>(out.data_ptr());
   typename GemmKernel::EpilogueArguments epilogue_args{
       Gemm::Epilogue::prepare_args(
           std::forward<EpilogueArgs>(epilogue_params)...),
-      c_ptr, stride_Dt, c_ptr, stride_Dt};
+      c_ptr, c_stride, c_ptr, c_stride};
 
   typename GemmKernel::Arguments args{cutlass::gemm::GemmUniversalMode::kGemm,
                                       prob_shape, mainloop_args, epilogue_args};
@@ -185,6 +272,10 @@ void cutlass_sparse_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
   CUTLASS_CHECK(status);
 }
 
+//////////////////////////////////////////////////
+// Gemm Configs are defined below
+//////////////////////////////////////////////////
+
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
 struct sm90_config_default {};
@@ -192,28 +283,25 @@ struct sm90_config_default {};
 template <typename OutType,
           template <typename, typename, typename> typename Epilogue>
 struct sm90_config_default<half_t, OutType, Epilogue> {
-  // M in (128, inf)
-  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecialized;
   using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
   using TileShape = Shape<_128, _128, _128>;
-  using ClusterShape = Shape<_2, _1, _1>;
+  using ClusterShape = Shape<_1, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<half_t, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename OutType,
           template <typename, typename, typename> typename Epilogue>
 struct sm90_config_default<cutlass::bfloat16_t, OutType, Epilogue> {
-  // M in (128, inf)
-  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecialized;
   using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
   using TileShape = Shape<_128, _128, _128>;
-  using ClusterShape = Shape<_2, _1, _1>;
+  using ClusterShape = Shape<_1, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<cutlass::bfloat16_t, OutType, Epilogue, TileShape,
-                             ClusterShape, KernelSchedule, EpilogueSchedule,
-                             float>;
+                             ClusterShape, KernelSchedule, EpilogueSchedule>;
 };
 
 //////////////////////// Cherry-Picking Kernels ////////////////////////
@@ -227,7 +315,7 @@ struct sm90_fp8_config_1 {
   using ClusterShape = Shape<_8, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -242,7 +330,7 @@ struct sm90_fp8_config_2 {
   using ClusterShape = Shape<_8, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -255,7 +343,7 @@ struct sm90_fp8_config_3 {
   using ClusterShape = Shape<_1, _2, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -269,7 +357,7 @@ struct sm90_fp8_config_4 {
   using ClusterShape = Shape<_8, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -283,7 +371,7 @@ struct sm90_fp8_config_5 {
   using ClusterShape = Shape<_8, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -296,7 +384,7 @@ struct sm90_fp8_config_6 {
   using ClusterShape = Shape<_1, _2, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -311,7 +399,7 @@ struct sm90_fp8_config_7 {
   using ClusterShape = Shape<_1, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -326,7 +414,7 @@ struct sm90_fp8_config_8 {
   using ClusterShape = Shape<_8, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 ////////////////////////////////////////////////////////////////////////
 
@@ -341,7 +429,7 @@ struct sm90_config_default<cutlass::float_e4m3_t, OutType, Epilogue> {
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<cutlass::float_e4m3_t, OutType, Epilogue,
                              TileShape, ClusterShape, KernelSchedule,
-                             EpilogueSchedule, float>;
+                             EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -355,12 +443,9 @@ struct sm90_fp8_config_M64 {
   using TileShape = Shape<_64, _64, _256>;
   using ClusterShape = Shape<_1, _1, _1>;
 
-  using TileSchedule = cutlass::gemm::PersistentScheduler;
-
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float,
-                             TileSchedule>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -374,12 +459,9 @@ struct sm90_fp8_config_M128 {
   using TileShape = Shape<_64, _128, _256>;
   using ClusterShape = Shape<_1, _1, _1>;
 
-  using TileSchedule = cutlass::gemm::PersistentScheduler;
-
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float,
-                             TileSchedule>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -394,12 +476,9 @@ struct sm90_fp8_config_M256 {
   using TileShape = Shape<_128, _128, _256>;
   using ClusterShape = Shape<_1, _1, _1>;
 
-  using TileSchedule = cutlass::gemm::PersistentScheduler;
-
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float,
-                             TileSchedule>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -414,12 +493,9 @@ struct sm90_fp8_config_M512 {
   using TileShape = Shape<_128, _128, _256>;
   using ClusterShape = Shape<_1, _1, _1>;
 
-  using TileSchedule = cutlass::gemm::PersistentScheduler;
-
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, float,
-                             TileSchedule>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename OutType,
@@ -433,7 +509,7 @@ struct sm90_config_default<int8_t, OutType, Epilogue> {
   using ClusterShape = Shape<_2, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<int8_t, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, int32_t>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -448,7 +524,7 @@ struct sm90_int8_config_M128 {
   using ClusterShape = Shape<_2, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, int32_t>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -462,7 +538,7 @@ struct sm90_int8_config_M64 {
   using ClusterShape = Shape<_1, _1, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, int32_t>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -476,7 +552,7 @@ struct sm90_int8_config_M32_NBig {
   using ClusterShape = Shape<_1, _4, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, int32_t>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
 template <typename InType, typename OutType,
@@ -490,7 +566,7 @@ struct sm90_int8_config_M32_NSmall {
   using ClusterShape = Shape<_1, _8, _1>;
   using Cutlass3xGemm =
       cutlass_sparse_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
-                             KernelSchedule, EpilogueSchedule, int32_t>;
+                             KernelSchedule, EpilogueSchedule>;
 };
 
-}  // namespace
+}  // namespace

csrc/sparse/cutlass/sparse_scaled_mm_entry.cu

@@ -23,6 +23,9 @@ void cutlass_scaled_sparse_mm_sm90(torch::Tensor& c, torch::Tensor const& a,
                                    torch::Tensor const& a_scales,
                                    torch::Tensor const& b_scales,
                                    std::optional<torch::Tensor> const& bias);
+
+using CompressorResult = std::tuple<torch::Tensor, torch::Tensor>;
+CompressorResult cutlass_sparse_compress_sm90(torch::Tensor const& a);
 #endif
 
 void cutlass_scaled_sparse_mm(torch::Tensor& c, torch::Tensor const& a,
@@ -68,3 +71,30 @@ void cutlass_scaled_sparse_mm(torch::Tensor& c, torch::Tensor const& a,
       "CUDA device capability: ",
       version_num);
 }
+
+std::vector<torch::Tensor> cutlass_sparse_compress(torch::Tensor const& a) {
+  // Check for strides and alignment
+  TORCH_CHECK(a.stride(1) == 1);      // Row-major
+  TORCH_CHECK(a.stride(0) % 8 == 0);  // 8 Byte Alignment for Compression
+
+  at::cuda::OptionalCUDAGuard const device_guard(device_of(a));
+  int32_t version_num = get_sm_version_num();
+
+  // Guard against compilation issues for sm90 kernels
+#if defined ENABLE_SPARSE_SCALED_MM_C3X && ENABLE_SPARSE_SCALED_MM_C3X
+  if (version_num >= 90) {
+    std::vector<torch::Tensor> result_tensors;
+
+    auto [a_meta, a_nzs] = cutlass_sparse_compress_sm90(a);
+    result_tensors.push_back(std::move(a_nzs));
+    result_tensors.push_back(std::move(a_meta));
+    return result_tensors;
+  }
+#endif
+
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      false,
+      "No compiled cutlass_sparse_compress for a compute capability less than "
+      "CUDA device capability: ",
+      version_num);
+}

csrc/torch_bindings.cpp

@@ -302,6 +302,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "SymInt size_k) -> Tensor");
   // conditionally compiled so impl registration is in source file
 
+  // CUTLASS nvfp4 block scaled GEMM
+  ops.def(
+      "cutlass_scaled_fp4_mm(Tensor! out, Tensor a, Tensor b,"
+      "                      Tensor block_scale_a, Tensor block_scale_b,"
+      "                      Tensor alpha) -> ()");
+  ops.impl("cutlass_scaled_fp4_mm", torch::kCUDA, &cutlass_scaled_fp4_mm);
+
   // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
   // quantization, as well as bias
   ops.def(
@@ -348,10 +355,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("cutlass_scaled_sparse_mm", torch::kCUDA, &cutlass_scaled_sparse_mm);
 
   // CUTLASS sparse matrix compressor
-  ops.def(
-      "cutlass_sparse_compress_entry(Tensor! a_nzs, Tensor! a_meta,"
-      "                              Tensor a) -> bool");
-  ops.impl("cutlass_sparse_compress_entry", &cutlass_sparse_compress_entry);
+  ops.def("cutlass_sparse_compress(Tensor a) -> Tensor[]");
+  ops.impl("cutlass_sparse_compress", &cutlass_sparse_compress);
 
   // Mamba selective scan kernel
   ops.def(
@@ -387,6 +392,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "bool silu_activation,"
       "int pad_slot_id) -> ()");
   ops.impl("causal_conv1d_fwd", torch::kCUDA, &causal_conv1d_fwd);
+
+  // Compute NVFP4 block quantized tensor.
+  ops.def(
+      "scaled_fp4_quant(Tensor! output, Tensor input,"
+      "                 Tensor! output_scale, Tensor input_scale) -> ()");
+  ops.impl("scaled_fp4_quant", torch::kCUDA, &scaled_fp4_quant);
+
 #endif
 
   // Quantized GEMM for GPTQ.
@@ -488,6 +500,12 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
       "convert_fp8(Tensor! dst_cache, Tensor src_cache, float scale, "
       "str kv_cache_dtype) -> ()");
   cache_ops.impl("convert_fp8", torch::kCUDA, &convert_fp8);
+
+  // Gather cache blocks from src_cache to dst.
+  cache_ops.def(
+      "gather_cache(Tensor src_cache, Tensor! dst, Tensor block_table, "
+      "Tensor cu_seq_lens, int batch_size, Tensor? seq_starts) -> ()");
+  cache_ops.impl("gather_cache", torch::kCUDA, &gather_cache);
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {

docs/seed_parameter_behavior.md

@@ -0,0 +1,51 @@
+# Seed Parameter Behavior in vLLM
+
+## Overview
+
+The `seed` parameter in vLLM is used to control the random states for various random number generators. This parameter can affect the behavior of random operations in user code, especially when working with models in vLLM.
+
+## Default Behavior
+
+By default, the `seed` parameter is set to `None`. When the `seed` parameter is `None`, the global random states for `random`, `np.random`, and `torch.manual_seed` are not set. This means that the random operations will behave as expected, without any fixed random states.
+
+## Specifying a Seed
+
+If a specific seed value is provided, the global random states for `random`, `np.random`, and `torch.manual_seed` will be set accordingly. This can be useful for reproducibility, as it ensures that the random operations produce the same results across multiple runs.
+
+## Example Usage
+
+### Without Specifying a Seed
+
+```python
+import random
+from vllm import LLM
+
+# Initialize a vLLM model without specifying a seed
+model = LLM(model="Qwen/Qwen2.5-0.5B-Instruct")
+
+# Try generating random numbers
+print(random.randint(0, 100))  # Outputs different numbers across runs
+```
+
+### Specifying a Seed
+
+```python
+import random
+from vllm import LLM
+
+# Initialize a vLLM model with a specific seed
+model = LLM(model="Qwen/Qwen2.5-0.5B-Instruct", seed=42)
+
+# Try generating random numbers
+print(random.randint(0, 100))  # Outputs the same number across runs
+```
+
+## Important Notes
+
+- If the `seed` parameter is not specified, the behavior of global random states remains unaffected.
+- If a specific seed value is provided, the global random states for `random`, `np.random`, and `torch.manual_seed` will be set to that value.
+- This behavior can be useful for reproducibility but may lead to non-intuitive behavior if the user is not explicitly aware of it.
+
+## Conclusion
+
+Understanding the behavior of the `seed` parameter in vLLM is crucial for ensuring the expected behavior of random operations in your code. By default, the `seed` parameter is set to `None`, which means that the global random states are not affected. However, specifying a seed value can help achieve reproducibility in your experiments.

docs/source/_static/custom.css

@@ -0,0 +1,8 @@
+.vertical-table-header th.head:not(.stub) {
+    writing-mode: sideways-lr;
+    white-space: nowrap;
+    max-width: 0;
+    p {
+       margin: 0;
+    }
+}

docs/source/conf.py

@@ -12,6 +12,7 @@
 # add these directories to sys.path here. If the directory is relative to the
 # documentation root, use os.path.abspath to make it absolute, like shown here.
 
+import datetime
 import inspect
 import logging
 import os
@@ -27,7 +28,7 @@ sys.path.append(os.path.abspath("../.."))
 # -- Project information -----------------------------------------------------
 
 project = 'vLLM'
-copyright = '2024, vLLM Team'
+copyright = f'{datetime.datetime.now().year}, vLLM Team'
 author = 'the vLLM Team'
 
 # -- General configuration ---------------------------------------------------
@@ -78,8 +79,12 @@ html_theme_options = {
     'use_repository_button': True,
     'use_edit_page_button': True,
 }
+# Add any paths that contain custom static files (such as style sheets) here,
+# relative to this directory. They are copied after the builtin static files,
+# so a file named "default.css" will overwrite the builtin "default.css".
 html_static_path = ["_static"]
 html_js_files = ["custom.js"]
+html_css_files = ["custom.css"]
 
 myst_url_schemes = {
     'http': None,
@@ -121,11 +126,6 @@ if READTHEDOCS_VERSION_TYPE == "tag":
     if os.path.exists(header_file):
         os.remove(header_file)
 
-# Add any paths that contain custom static files (such as style sheets) here,
-# relative to this directory. They are copied after the builtin static files,
-# so a file named "default.css" will overwrite the builtin "default.css".
-# html_static_path = ['_static']
-
 
 # Generate additional rst documentation here.
 def setup(app):

docs/source/contributing/model/basic.md

@@ -74,8 +74,6 @@ def forward(
     self,
     input_ids: torch.Tensor,
     positions: torch.Tensor,
-    kv_caches: List[torch.Tensor],
-    attn_metadata: AttentionMetadata,
 ) -> torch.Tensor:
     ...
 ```

docs/source/contributing/model/multimodal.md

@@ -16,8 +16,6 @@ Further update the model as follows:
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
   +     pixel_values: torch.Tensor,
     ) -> SamplerOutput:
   ```
@@ -262,6 +260,255 @@ def get_mm_max_tokens_per_item(
 Our [actual code](gh-file:vllm/model_executor/models/llava.py) is more abstracted to support vision encoders other than CLIP.
 :::
 
+::::
+
+::::{tab-item} Non-consecutive feature tokens: Fuyu
+:sync: fuyu
+
+Looking at the code of HF's `FuyuForCausalLM`:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/modeling_fuyu.py#L311-L322
+if image_patches is not None and past_key_values is None:
+    patch_embeddings = [
+        self.vision_embed_tokens(patch.to(self.vision_embed_tokens.weight.dtype))
+        .squeeze(0)
+        .to(inputs_embeds.device)
+        for patch in image_patches
+    ]
+    inputs_embeds = self.gather_continuous_embeddings(
+        word_embeddings=inputs_embeds,
+        continuous_embeddings=patch_embeddings,
+        image_patch_input_indices=image_patches_indices,
+    )
+```
+
+The number of placeholder feature tokens for the `i`th item in the batch is `patch_embeddings[i].shape[0]`,
+which is the same as `image_patches[i].shape[0]`, i.e. `num_total_patches`.
+
+Unlike LLaVA, Fuyu does not define the number of patches inside the modeling file. Where can we get more information?
+Considering that the model input comes from the output of `FuyuProcessor`, let's **look at the preprocessing files**.
+
+The image outputs are obtained by calling `FuyuImageProcessor.preprocess` and then
+`FuyuImageProcessor.preprocess_with_tokenizer_info` inside `FuyuProcessor`.
+
+In `FuyuImageProcessor.preprocess`, the images are resized and padded to the target `FuyuImageProcessor.size`,
+returning the dimensions after resizing (but before padding) as metadata.
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/processing_fuyu.py#L541-L544
+image_encoding = self.image_processor.preprocess(images, **output_kwargs["images_kwargs"])
+batch_images = image_encoding["images"]
+image_unpadded_heights = image_encoding["image_unpadded_heights"]
+image_unpadded_widths = image_encoding["image_unpadded_widths"]
+
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L480-L
+if do_resize:
+    batch_images = [
+        [self.resize(image, size=size, input_data_format=input_data_format) for image in images]
+        for images in batch_images
+    ]
+
+image_sizes = [get_image_size(images[0], channel_dim=input_data_format) for images in batch_images]
+image_unpadded_heights = [[image_size[0]] for image_size in image_sizes]
+image_unpadded_widths = [[image_size[1]] for image_size in image_sizes]
+
+if do_pad:
+    batch_images = [
+        [
+            self.pad_image(
+                image,
+                size=size,
+                mode=padding_mode,
+                constant_values=padding_value,
+                input_data_format=input_data_format,
+            )
+            for image in images
+        ]
+        for images in batch_images
+    ]
+```
+
+In `FuyuImageProcessor.preprocess_with_tokenizer_info`, the images are split into patches based on this metadata:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/processing_fuyu.py#L417-L425
+model_image_input = self.image_processor.preprocess_with_tokenizer_info(
+    image_input=tensor_batch_images,
+    image_present=image_present,
+    image_unpadded_h=image_unpadded_heights,
+    image_unpadded_w=image_unpadded_widths,
+    image_placeholder_id=image_placeholder_id,
+    image_newline_id=image_newline_id,
+    variable_sized=True,
+)
+
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L638-L658
+image_height, image_width = image.shape[1], image.shape[2]
+if variable_sized:  # variable_sized=True
+    new_h = min(
+        image_height,
+        math.ceil(image_unpadded_h[batch_index, subseq_index] / patch_height) * patch_height,
+    )
+    new_w = min(
+        image_width,
+        math.ceil(image_unpadded_w[batch_index, subseq_index] / patch_width) * patch_width,
+    )
+    image = image[:, :new_h, :new_w]
+    image_height, image_width = new_h, new_w
+
+num_patches = self.get_num_patches(image_height=image_height, image_width=image_width)
+tensor_of_image_ids = torch.full(
+    [num_patches], image_placeholder_id, dtype=torch.int32, device=image_input.device
+)
+patches = self.patchify_image(image=image.unsqueeze(0)).squeeze(0)
+assert num_patches == patches.shape[0]
+```
+
+The number of patches is in turn defined by `FuyuImageProcessor.get_num_patches`:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L552-L562
+patch_size = patch_size if patch_size is not None else self.patch_size
+patch_height, patch_width = self.patch_size["height"], self.patch_size["width"]
+
+if image_height % patch_height != 0:
+    raise ValueError(f"{image_height=} must be divisible by {patch_height}")
+if image_width % patch_width != 0:
+    raise ValueError(f"{image_width=} must be divisible by {patch_width}")
+
+num_patches_per_dim_h = image_height // patch_height
+num_patches_per_dim_w = image_width // patch_width
+num_patches = num_patches_per_dim_h * num_patches_per_dim_w
+```
+
+We can calculate this in vLLM using this code:
+
+```python
+def get_num_image_patches(
+    self,
+    *,
+    image_width: int,
+    image_height: int,
+) -> int:
+    image_processor = self.get_image_processor()
+    target_width = image_processor.size["width"]
+    target_height = image_processor.size["height"]
+    patch_width = image_processor.patch_size["width"]
+    patch_height = image_processor.patch_size["height"]
+
+    if not (image_width <= target_width and image_height <= target_height):
+        height_scale_factor = target_height / image_height
+        width_scale_factor = target_width / image_width
+        optimal_scale_factor = min(height_scale_factor, width_scale_factor)
+
+        image_height = int(image_height * optimal_scale_factor)
+        image_width = int(image_width * optimal_scale_factor)
+
+    ncols = math.ceil(image_width / patch_width)
+    nrows = math.ceil(image_height / patch_height)
+    return ncols * nrows
+```
+
+These image patches correspond to placeholder tokens (`|SPEAKER|`). However, the processor also
+inserts newline tokens (`|NEWLINE|`) as shown here:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L654-L670
+tensor_of_image_ids = torch.full(
+    [num_patches], image_placeholder_id, dtype=torch.int32, device=image_input.device
+)
+patches = self.patchify_image(image=image.unsqueeze(0)).squeeze(0)
+assert num_patches == patches.shape[0]
+
+if variable_sized:
+    # Now terminate each line with |NEWLINE|.
+    tensor_of_image_ids = tensor_of_image_ids.reshape(-1, image_width // patch_width)
+    newline_ids = torch.full(
+        [tensor_of_image_ids.shape[0], 1],
+        image_newline_id,
+        dtype=torch.int32,
+        device=image_input.device,
+    )
+    tensor_of_image_ids = torch.cat([tensor_of_image_ids, newline_ids], dim=1)
+    tensor_of_image_ids = tensor_of_image_ids.reshape(-1)
+```
+
+So, the layout of tokens for an image is:
+
+```
+|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
+|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
+...
+|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
+```
+
+This makes the placeholder tokens non-consecutive in the prompt.
+Since vLLM requires the feature tokens to be consecutive, **we also treat the newline tokens as feature tokens**.
+
+So overall, the total number of feature tokens is
+
+```python
+def get_num_image_tokens(
+    self,
+    *,
+    image_width: int,
+    image_height: int,
+) -> int:
+    image_processor = self.get_image_processor()
+    target_width = image_processor.size["width"]
+    target_height = image_processor.size["height"]
+    patch_width = image_processor.patch_size["width"]
+    patch_height = image_processor.patch_size["height"]
+
+    if not (image_width <= target_width and image_height <= target_height):
+        height_scale_factor = target_height / image_height
+        width_scale_factor = target_width / image_width
+        optimal_scale_factor = min(height_scale_factor, width_scale_factor)
+
+        image_height = int(image_height * optimal_scale_factor)
+        image_width = int(image_width * optimal_scale_factor)
+
+    ncols = math.ceil(image_width / patch_width)
+    nrows = math.ceil(image_height / patch_height)
+    return (ncols + 1) * nrows
+```
+
+To calculate the maximum number of image tokens, recall that input images are first resized
+to fit within `image_processor.size`. The maximum possible dimensions of the image before
+being converted into patches is therefore equal to `image_processor.size`.
+
+```python
+def get_image_size_with_most_features(self) -> ImageSize:
+    image_processor = self.get_image_processor()
+    return ImageSize(width=image_processor.size["width"],
+                        height=image_processor.size["height"])
+
+def get_max_image_tokens(self) -> int:
+    target_width, target_height = self.get_image_size_with_most_features()
+
+    return self.get_num_image_tokens(
+        image_width=target_width,
+        image_height=target_height,
+    )
+```
+
+And thus, we can override the method as:
+
+```python
+def get_mm_max_tokens_per_item(
+    self,
+    seq_len: int,
+    mm_counts: Mapping[str, int],
+) -> Mapping[str, int]:
+    return {"image": self.get_max_image_tokens()}
+```
+
+:::{note}
+Our [actual code](gh-file:vllm/model_executor/models/fuyu.py) returns `ncols` and `nrows` directly instead of the total token count.
+This is because `ncols` and `nrows` are used to specify the layout of the feature tokens (as shown in Step 4 of this guide).
+:::
+
 ::::
 :::::
 
@@ -282,7 +529,8 @@ on the code for {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max
 ::::{tab-set}
 :::{tab-item} Basic example: LLaVA
 :sync: llava
-Making use of the `get_image_size_with_most_features` method implemented in the previous section:
+
+Making use of the `get_image_size_with_most_features` method implemented in Step 2:
 
 ```python
 def get_dummy_processor_inputs(
@@ -312,6 +560,39 @@ def get_dummy_processor_inputs(
 ```
 
 :::
+
+:::{tab-item} No input placeholders: Fuyu
+:sync: fuyu
+
+Fuyu does not expect image placeholders in the inputs to HF processor, so
+the dummy prompt text is empty regardless of the number of images.
+Otherwise, the logic of this method is very similar to LLaVA:
+
+```python
+def get_dummy_processor_inputs(
+    self,
+    seq_len: int,
+    mm_counts: Mapping[str, int],
+) -> ProcessorInputs:
+    target_width, target_height = \
+        self.info.get_image_size_with_most_features()
+    num_images = mm_counts.get("image", 0)
+
+    mm_data = {
+        "image":
+        self._get_dummy_images(width=target_width,
+                                height=target_height,
+                                num_images=num_images)
+    }
+
+    return ProcessorInputs(
+        prompt_text="",
+        mm_data=mm_data,
+    )
+```
+
+:::
+
 ::::
 
 ## 4. Specify processing details
@@ -325,40 +606,28 @@ to fill in the missing details about HF processing.
 
 ### Multi-modal fields
 
-Override {class}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config` to
+Override {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config` to
 return a schema of the tensors outputted by the HF processor that are related to the input multi-modal items.
 
 :::::{tab-set}
 ::::{tab-item} Basic example: LLaVA
 :sync: llava
 
-Looking at the model's `forward` method:
+The output of `CLIPImageProcessor` is a simple tensor with shape
+`(num_images, num_channels, image_height, image_width)`:
 
 ```python
-# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L387-L404
-def forward(
-    self,
-    input_ids: torch.LongTensor = None,
-    pixel_values: torch.FloatTensor = None,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[List[torch.FloatTensor]] = None,
-    inputs_embeds: Optional[torch.FloatTensor] = None,
-    vision_feature_layer: Optional[int] = None,
-    vision_feature_select_strategy: Optional[str] = None,
-    labels: Optional[torch.LongTensor] = None,
-    use_cache: Optional[bool] = None,
-    output_attentions: Optional[bool] = None,
-    output_hidden_states: Optional[bool] = None,
-    return_dict: Optional[bool] = None,
-    cache_position: Optional[torch.LongTensor] = None,
-    num_logits_to_keep: int = 0,
-) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/image_processing_clip.py#L339-L345
+images = [
+    to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+    for image in all_images
+]
+
+data = {"pixel_values": images}
+return BatchFeature(data=data, tensor_type=return_tensors)
 ```
 
-The only related keyword argument is `pixel_values` which directly corresponds to input images.
-The shape of `pixel_values` is `(N, C, H, W)` where `N` is the number of images.
-So, we override the method as follows:
+So, we override {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config` as follows:
 
 ```python
 def _get_mm_fields_config(
@@ -377,15 +646,87 @@ pre-computed image embeddings, which can be passed to be model via the `image_em
 :::
 
 ::::
+
+::::{tab-item} With postprocessing: Fuyu
+:sync: fuyu
+
+The `image_patches` output of `FuyuImageProcessor.preprocess_with_tokenizer_info` concatenates
+the patches from each image belonging to an item in the batch:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L673-L679
+        image_input_ids.append(tensor_of_image_ids)
+        image_patches.append(patches)
+    else:
+        image_input_ids.append(torch.tensor([], dtype=torch.int32, device=image_input.device))
+
+batch_image_input_ids.append(image_input_ids)
+batch_image_patches.append(image_patches)
+```
+
+The shape of `image_patches` outputted by `FuyuImageProcessor` is therefore
+`(1, num_images, num_patches, patch_width * patch_height * num_channels)`.
+
+In order to support the use of {func}`MultiModalFieldConfig.batched` like in LLaVA,
+we remove the extra batch dimension by overriding {meth}`BaseMultiModalProcessor._call_hf_processor`:
+
+```python
+def _call_hf_processor(
+    self,
+    prompt: str,
+    mm_data: Mapping[str, object],
+    mm_kwargs: Mapping[str, object],
+) -> BatchFeature:
+    processed_outputs = super()._call_hf_processor(
+        prompt=prompt,
+        mm_data=mm_data,
+        mm_kwargs=mm_kwargs,
+    )
+
+    image_patches = processed_outputs.get("image_patches")
+    if image_patches is not None:
+        images = mm_data["images"]
+        assert isinstance(images, list)
+
+        # Original output: (1, num_images, Pn, Px * Py * C)
+        # New output: (num_images, Pn, Px * Py * C)
+        assert (isinstance(image_patches, list)
+                and len(image_patches) == 1)
+        assert (isinstance(image_patches[0], torch.Tensor)
+                and len(image_patches[0]) == len(images))
+
+        processed_outputs["image_patches"] = image_patches[0]
+
+    return processed_outputs
+```
+
+:::{note}
+Our [actual code](gh-file:vllm/model_executor/models/fuyu.py) has special handling
+for text-only inputs to prevent unnecessary warnings from HF processor.
+:::
+
+This lets us override {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config` as follows:
+
+```python
+def _get_mm_fields_config(
+    self,
+    hf_inputs: BatchFeature,
+    hf_processor_mm_kwargs: Mapping[str, object],
+) -> Mapping[str, MultiModalFieldConfig]:
+    return dict(image_patches=MultiModalFieldConfig.batched("image"))
+```
+
+::::
+
 :::::
 
-### Prompt replacements
+### Prompt updates
 
-Override {class}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements` to
-return a list of {class}`~vllm.multimodal.processing.PromptReplacement` instances.
+Override {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates` to
+return a list of {class}`~vllm.multimodal.processing.PromptUpdate` instances.
 
-Each {class}`~vllm.multimodal.processing.PromptReplacement` instance specifies a find-and-replace
-operation performed by the HF processor.
+Each {class}`~vllm.multimodal.processing.PromptUpdate` instance specifies an update operation
+(e.g.: insertion, replacement) performed by the HF processor.
 
 ::::{tab-set}
 :::{tab-item} Basic example: LLaVA
@@ -402,15 +743,15 @@ for sample in text:
 ```
 
 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 the method as follows:
+Based on this, we override {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates` as follows:
 
 ```python
-def _get_prompt_replacements(
+def _get_prompt_updates(
     self,
     mm_items: MultiModalDataItems,
     hf_processor_mm_kwargs: Mapping[str, object],
     out_mm_kwargs: MultiModalKwargs,
-) -> list[PromptReplacement]:
+) -> Sequence[PromptUpdate]:
     hf_config = self.info.get_hf_config()
     image_token_id = hf_config.image_token_index
 
@@ -435,6 +776,159 @@ def _get_prompt_replacements(
 ```
 
 :::
+
+:::{tab-item} Handling additional tokens: Fuyu
+:sync: fuyu
+
+Recall the layout of feature tokens from Step 2:
+
+```
+|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
+|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
+...
+|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
+```
+
+We define a helper function to return `ncols` and `nrows` directly:
+
+```python
+def get_image_feature_grid_size(
+    self,
+    *,
+    image_width: int,
+    image_height: int,
+) -> tuple[int, int]:
+    image_processor = self.get_image_processor()
+    target_width = image_processor.size["width"]
+    target_height = image_processor.size["height"]
+    patch_width = image_processor.patch_size["width"]
+    patch_height = image_processor.patch_size["height"]
+
+    if not (image_width <= target_width and image_height <= target_height):
+        height_scale_factor = target_height / image_height
+        width_scale_factor = target_width / image_width
+        optimal_scale_factor = min(height_scale_factor, width_scale_factor)
+
+        image_height = int(image_height * optimal_scale_factor)
+        image_width = int(image_width * optimal_scale_factor)
+
+    ncols = math.ceil(image_width / patch_width)
+    nrows = math.ceil(image_height / patch_height)
+    return ncols, nrows
+```
+
+Based on this, we can initially define our replacement tokens as:
+
+```python
+def get_replacement(item_idx: int):
+    images = mm_items.get_items("image", ImageProcessorItems)
+    image_size = images.get_image_size(item_idx)
+
+    ncols, nrows = self.info.get_image_feature_grid_size(
+        image_width=image_size.width,
+        image_height=image_size.height,
+    )
+
+    # `_IMAGE_TOKEN_ID` corresponds to `|SPEAKER|`
+    # `_NEWLINE_TOKEN_ID` corresponds to `|NEWLINE|`
+    return ([_IMAGE_TOKEN_ID] * ncols + [_NEWLINE_TOKEN_ID]) * nrows
+```
+
+However, this is not entirely correct. After `FuyuImageProcessor.preprocess_with_tokenizer_info` is called,
+a BOS token (`<s>`) is also added to the promopt:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/processing_fuyu.py#L417-L435
+model_image_input = self.image_processor.preprocess_with_tokenizer_info(
+    image_input=tensor_batch_images,
+    image_present=image_present,
+    image_unpadded_h=image_unpadded_heights,
+    image_unpadded_w=image_unpadded_widths,
+    image_placeholder_id=image_placeholder_id,
+    image_newline_id=image_newline_id,
+    variable_sized=True,
+)
+prompt_tokens, prompts_length = _tokenize_prompts_with_image_and_batch(
+    tokenizer=self.tokenizer,
+    prompts=prompts,
+    scale_factors=scale_factors,
+    max_tokens_to_generate=self.max_tokens_to_generate,
+    max_position_embeddings=self.max_position_embeddings,
+    add_BOS=True,
+    add_beginning_of_answer_token=True,
+)
+```
+
+To accommodate this, instead of a string you can return an instance of `PromptUpdateDetails`
+with different `full` and `feature` attributes:
+
+```python
+hf_config = self.info.get_hf_config()
+bos_token_id = hf_config.bos_token_id  # `<s>`
+assert isinstance(bos_token_id, int)
+
+def get_replacement_fuyu(item_idx: int):
+    images = mm_items.get_items("image", ImageProcessorItems)
+    image_size = images.get_image_size(item_idx)
+
+    ncols, nrows = self.info.get_image_feature_grid_size(
+        image_width=image_size.width,
+        image_height=image_size.height,
+    )
+    image_tokens = ([_IMAGE_TOKEN_ID] * ncols +
+                    [_NEWLINE_TOKEN_ID]) * nrows
+
+    return PromptUpdateDetails(
+        full=image_tokens + [bos_token_id],
+        features=image_tokens,
+    )
+```
+
+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:
+
+```python
+def _get_prompt_updates(
+    self,
+    mm_items: MultiModalDataItems,
+    hf_processor_mm_kwargs: Mapping[str, object],
+    out_mm_kwargs: MultiModalKwargs,
+) -> Sequence[PromptUpdate]:
+    hf_config = self.info.get_hf_config()
+    bos_token_id = hf_config.bos_token_id
+    assert isinstance(bos_token_id, int)
+
+    tokenizer = self.info.get_tokenizer()
+    eot_token_id = tokenizer.bos_token_id
+    assert isinstance(eot_token_id, int)
+
+    def get_replacement_fuyu(item_idx: int):
+        images = mm_items.get_items("image", ImageProcessorItems)
+        image_size = images.get_image_size(item_idx)
+
+        ncols, nrows = self.info.get_image_feature_grid_size(
+            image_width=image_size.width,
+            image_height=image_size.height,
+        )
+        image_tokens = ([_IMAGE_TOKEN_ID] * ncols +
+                        [_NEWLINE_TOKEN_ID]) * nrows
+
+        return PromptUpdateDetails(
+            full=image_tokens + [bos_token_id],
+            features=image_tokens,
+        )
+
+    return [
+        PromptReplacement(
+            modality="image",
+            target=[eot_token_id],
+            replacement=get_replacement_fuyu,
+        )
+    ]
+```
+
+:::
+
 ::::
 
 ## 5. Register processor-related classes

docs/source/contributing/overview.md

@@ -145,6 +145,9 @@ review process:
 - Please respond to all comments within a reasonable time frame. If a comment
   isn't clear or you disagree with a suggestion, feel free to ask for
   clarification or discuss the suggestion.
+- Note that not all CI checks will be executed due to limited computational
+  resources. The reviewer will add `ready` label to the PR when the PR is
+  ready to merge or a full CI run is needed.
 
 ## Thank You
 

docs/source/contributing/profiling/profiling_index.md

@@ -1,15 +1,15 @@
 # Profiling vLLM
 
+:::{warning}
+Profiling is only intended for vLLM developers and maintainers to understand the proportion of time spent in different parts of the codebase. **vLLM end-users should never turn on profiling** as it will significantly slow down the inference.
+:::
+
 We support tracing vLLM workers using the `torch.profiler` module. You can enable tracing by setting the `VLLM_TORCH_PROFILER_DIR` environment variable to the directory where you want to save the traces: `VLLM_TORCH_PROFILER_DIR=/mnt/traces/`
 
 The OpenAI server also needs to be started with the `VLLM_TORCH_PROFILER_DIR` environment variable set.
 
 When using `benchmarks/benchmark_serving.py`, you can enable profiling by passing the `--profile` flag.
 
-:::{warning}
-Only enable profiling in a development environment.
-:::
-
 Traces can be visualized using <https://ui.perfetto.dev/>.
 
 :::{tip}

docs/source/deployment/docker.md

@@ -27,6 +27,36 @@ container to access the host's shared memory. vLLM uses PyTorch, which uses shar
 memory to share data between processes under the hood, particularly for tensor parallel inference.
 :::
 
+:::{note}
+Optional dependencies are not included in order to avoid licensing issues (e.g. <gh-issue:8030>).
+
+If you need to use those dependencies (having accepted the license terms),
+create a custom Dockerfile on top of the base image with an extra layer that installs them:
+
+```Dockerfile
+FROM vllm/vllm-openai:v0.7.3
+
+# e.g. install the `audio` and `video` optional dependencies
+# NOTE: Make sure the version of vLLM matches the base image!
+RUN uv pip install --system vllm[audio,video]==0.7.3
+```
+
+:::
+
+:::{tip}
+Some new models may only be available on the main branch of [HF Transformers](https://github.com/huggingface/transformers).
+
+To use the development version of `transformers`, create a custom Dockerfile on top of the base image
+with an extra layer that installs their code from source:
+
+```Dockerfile
+FROM vllm/vllm-openai:latest
+
+RUN uv pip install --system git+https://github.com/huggingface/transformers.git
+```
+
+:::
+
 (deployment-docker-build-image-from-source)=
 
 ## Building vLLM's Docker Image from Source

docs/source/deployment/integrations/index.md

@@ -6,4 +6,5 @@
 kserve
 kubeai
 llamastack
+llmaz
 :::

docs/source/deployment/integrations/llmaz.md

@@ -0,0 +1,7 @@
+(deployment-llmaz)=
+
+# llmaz
+
+[llmaz](https://github.com/InftyAI/llmaz) is an easy-to-use and advanced inference platform for large language models on Kubernetes, aimed for production use. It uses vLLM as the default model serving backend.
+
+Please refer to the [Quick Start](https://github.com/InftyAI/llmaz?tab=readme-ov-file#quick-start) for more details.

docs/source/design/arch_overview.md

@@ -66,7 +66,7 @@ This server can be started using the `vllm serve` command.
 vllm serve <model>
 ```
 
-The code for the `vllm` CLI can be found in <gh-file:vllm/scripts.py>.
+The code for the `vllm` CLI can be found in <gh-file:vllm/entrypoints/cli/main.py>.
 
 Sometimes you may see the API server entrypoint used directly instead of via the
 `vllm` CLI command. For example:

docs/source/design/mm_processing.md

@@ -6,11 +6,16 @@ To enable various optimizations in vLLM such as [chunked prefill](#chunked-prefi
 
 Here are the main features of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor`:
 
-## Prompt Replacement Detection
+## Prompt Update Detection
 
-One of the main responsibilies of HF processor is to replace input placeholder tokens (e.g. `<image>` for a single image) with feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size). The information about which tokens have been replaced is key to finding the correspondence between placeholder feature tokens and multi-modal inputs.
+One of the main responsibilies of HF processor is to update the prompt with placeholder tokens. For example:
 
-In vLLM, this information is specified using {class}`~vllm.multimodal.processing.PromptReplacement` in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`. Given this specification, we can automatically detect whether HF has replaced the input placeholder tokens by checking whether the feature placeholder tokens exist in the prompt.
+- Insert feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size) at the start of the string.
+- Replace existing input placeholder tokens (e.g. `<image>` for a single image) with feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size).
+
+The information about which tokens have been updated is key to finding the correspondence between placeholder feature tokens and multi-modal inputs.
+
+In vLLM, this information is specified using {class}`~vllm.multimodal.processing.PromptUpdate` in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates`. We can automatically detect whether HF has updated the prompt by checking the existence of the updated tokens.
 
 ## Tokenized Prompt Inputs
 
@@ -22,7 +27,7 @@ Consider that HF processors follow these main steps:
 
 1. Tokenize the text
 2. Process multi-modal inputs
-3. Perform prompt replacement
+3. Perform prompt updates
 
 And we require that:
 
@@ -44,16 +49,16 @@ Moreover, since the tokenized text has not passed through the HF processor, we h
 
 We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data.
 
-(mm-automatic-prompt-replacement)=
+(mm-automatic-prompt-updating)=
 
-### Automatic prompt replacement
+### Automatic prompt updating
 
 We address the second issue by implementing model-agnostic code in
-{meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_prompt_replacements` to automatically replace input placeholder tokens with feature placeholder tokens based on the specification outputted by {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`.
+{meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_prompt_updates` to automatically update the prompt with feature placeholder tokens based on the specification outputted by {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_updates`.
 
 ### Summary
 
-With the help of dummy text and automatic prompt replacement, our multi-modal processor can finally accept both text and token prompts with multi-modal data. The detailed logic is shown in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_main`.
+With the help of dummy text and automatic prompt updating, our multi-modal processor can finally accept both text and token prompts with multi-modal data. The detailed logic is shown in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_main`.
 
 ## Processor Output Caching
 
@@ -61,4 +66,4 @@ Some HF processors, such as the one for Qwen2-VL, are [very slow](gh-issue:9238)
 
 When new data is passed in, we first check which items are in the cache, and which ones are missing. The missing items are passed into the HF processor in a single batch and cached, before being merged with the existing items in the cache.
 
-Since we only process the missing multi-modal data items, the number of input placeholder tokens no longer corresponds to the number of the multi-modal inputs, so they can't be passed alongside the text prompt to HF processor. Therefore, we process the text and multi-modal inputs separately, using [dummy text](#mm-dummy-text) to avoid HF errors. Since this skips HF's prompt replacement code, we apply [automatic prompt replacement](#mm-automatic-prompt-replacement) afterwards to keep the output tokens and multi-modal data consistent with each other.
+Since we only process the missing multi-modal data items, the number of input placeholder tokens no longer corresponds to the number of the multi-modal inputs, so they can't be passed alongside the text prompt to HF processor. Therefore, we process the text and multi-modal inputs separately, using [dummy text](#mm-dummy-text) to avoid HF errors. Since this skips HF's prompt updating code, we apply [automatic prompt updating](#mm-automatic-prompt-updating) afterwards to keep the output tokens and multi-modal data consistent with each other.

docs/source/features/compatibility_matrix.md

@@ -4,8 +4,14 @@
 
 The tables below show mutually exclusive features and the support on some hardware.
 
+The symbols used have the following meanings:
+
+- ✅ = Full compatibility
+- 🟠 = Partial compatibility
+- ❌ = No compatibility
+
 :::{note}
-Check the '✗' with links to see tracking issue for unsupported feature/hardware combination.
+Check the ❌ or 🟠 with links to see tracking issue for unsupported feature/hardware combination.
 :::
 
 ## Feature x Feature
@@ -29,6 +35,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
 :header-rows: 1
 :stub-columns: 1
 :widths: auto
+:class: vertical-table-header
 
 - * Feature
   * [CP](#chunked-prefill)
@@ -48,7 +55,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * beam-search
   * <abbr title="Guided Decoding">guided dec</abbr>
 - * [CP](#chunked-prefill)
-  *
+  * ✅
   *
   *
   *
@@ -66,7 +73,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   *
 - * [APC](#automatic-prefix-caching)
   * ✅
-  *
+  * ✅
   *
   *
   *
@@ -82,9 +89,9 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   *
   *
 - * [LoRA](#lora-adapter)
-  * [✗](gh-pr:9057)
   * ✅
-  *
+  * ✅
+  * ✅
   *
   *
   *
@@ -102,7 +109,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  *
+  * ✅
   *
   *
   *
@@ -118,9 +125,9 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
 - * [SD](#spec_decode)
   * ✅
   * ✅
-  * ✗
+  * ❌
+  * ✅
   * ✅
-  *
   *
   *
   *
@@ -138,7 +145,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  *
+  * ✅
   *
   *
   *
@@ -150,13 +157,13 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   *
   *
 - * <abbr title="Pooling Models">pooling</abbr>
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  *
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ✅
   *
   *
   *
@@ -167,14 +174,14 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   *
   *
 - * <abbr title="Encoder-Decoder Models">enc-dec</abbr>
-  * ✗
-  * [✗](gh-issue:7366)
-  * ✗
-  * ✗
-  * [✗](gh-issue:7366)
+  * ❌
+  * [❌](gh-issue:7366)
+  * ❌
+  * ❌
+  * [❌](gh-issue:7366)
+  * ✅
   * ✅
   * ✅
-  *
   *
   *
   *
@@ -190,9 +197,9 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * ✗
+  * ❌
+  * ✅
   * ✅
-  *
   *
   *
   *
@@ -205,12 +212,12 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * [✗](gh-pr:8199)
-  * ✅
-  * ✗
   * ✅
   * ✅
-  *
+  * ❌
+  * ✅
+  * ✅
+  * ✅
   *
   *
   *
@@ -222,49 +229,49 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * ✗
+  * ❌
+  * ✅
+  * ❌
+  * ❌
   * ✅
-  * ✗
-  * ✗
   * ✅
   * ✅
-  *
   *
   *
   *
   *
   *
 - * multi-step
-  * ✗
+  * ❌
   * ✅
-  * ✗
+  * ❌
+  * ✅
+  * ❌
+  * ✅
+  * ❌
+  * ❌
   * ✅
-  * ✗
   * ✅
-  * ✗
-  * ✗
   * ✅
-  * [✗](gh-issue:8198)
   * ✅
-  *
   *
   *
   *
   *
 - * <abbr title="Multimodal Inputs">mm</abbr>
   * ✅
-  * [✗](gh-pr:8348)
-  * [✗](gh-pr:7199)
-  * ?
-  * ?
+  * [🟠](gh-pr:8348)
+  * [🟠](gh-pr:4194)
+  * ❔
+  * ❔
   * ✅
   * ✅
   * ✅
   * ✅
   * ✅
   * ✅
-  * ?
-  *
+  * ❔
+  * ✅
   *
   *
   *
@@ -273,16 +280,16 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * [✗](gh-issue:6137)
+  * [❌](gh-issue:6137)
   * ✅
-  * ✗
+  * ❌
   * ✅
   * ✅
   * ✅
-  * ?
-  * [✗](gh-issue:7968)
+  * ❔
+  * [❌](gh-issue:7968)
+  * ✅
   * ✅
-  *
   *
   *
 - * beam-search
@@ -290,35 +297,35 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * [✗](gh-issue:6137)
+  * [❌](gh-issue:6137)
   * ✅
-  * ✗
+  * ❌
   * ✅
   * ✅
   * ✅
-  * ?
-  * [✗](gh-issue:7968>)
-  * ?
+  * ❔
+  * [❌](gh-issue:7968)
+  * ❔
+  * ✅
   * ✅
-  *
   *
 - * <abbr title="Guided Decoding">guided dec</abbr>
   * ✅
   * ✅
-  * ?
-  * ?
-  * [✗](gh-issue:11484)
+  * ❔
+  * ❔
+  * [❌](gh-issue:11484)
   * ✅
-  * ✗
-  * ?
+  * ❌
+  * ❔
   * ✅
   * ✅
   * ✅
-  * [✗](gh-issue:9893)
-  * ?
+  * [❌](gh-issue:9893)
+  * ❔
+  * ✅
   * ✅
   * ✅
-  *
 :::
 
 (feature-x-hardware)=
@@ -339,7 +346,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * CPU
   * AMD
 - * [CP](#chunked-prefill)
-  * [✗](gh-issue:2729)
+  * [❌](gh-issue:2729)
   * ✅
   * ✅
   * ✅
@@ -347,7 +354,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
 - * [APC](#automatic-prefix-caching)
-  * [✗](gh-issue:3687)
+  * [❌](gh-issue:3687)
   * ✅
   * ✅
   * ✅
@@ -368,7 +375,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * [✗](gh-issue:8475)
+  * [❌](gh-issue:8475)
   * ✅
 - * [SD](#spec_decode)
   * ✅
@@ -384,7 +391,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * ✗
+  * ❌
   * ✅
 - * <abbr title="Pooling Models">pooling</abbr>
   * ✅
@@ -393,7 +400,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * ?
+  * ❔
 - * <abbr title="Encoder-Decoder Models">enc-dec</abbr>
   * ✅
   * ✅
@@ -401,7 +408,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * ✗
+  * ❌
 - * <abbr title="Multimodal Inputs">mm</abbr>
   * ✅
   * ✅
@@ -432,15 +439,15 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
   * ✅
   * ✅
   * ✅
-  * ✗
-  * ✗
+  * ❌
+  * ❌
 - * multi-step
   * ✅
   * ✅
   * ✅
   * ✅
   * ✅
-  * [✗](gh-issue:8477)
+  * [❌](gh-issue:8477)
   * ✅
 - * best-of
   * ✅

docs/source/features/lora.md

@@ -170,7 +170,7 @@ Now, you can specify a base_model_name alongside the name and path using JSON fo
 
 To provide the backward compatibility support, you can still use the old key-value format (name=path), but the `base_model_name` will remain unspecified in that case.
 
-## Lora model lineage in model card
+## LoRA model lineage in model card
 
 The new format of `--lora-modules` is mainly to support the display of parent model information in the model card. Here's an explanation of how your current response supports this:
 

docs/source/features/quantization/auto_awq.md

@@ -6,13 +6,13 @@ To create a new 4-bit quantized model, you can leverage [AutoAWQ](https://github
 Quantizing reduces the model's precision from FP16 to INT4 which effectively reduces the file size by ~70%.
 The main benefits are lower latency and memory usage.
 
-You can quantize your own models by installing AutoAWQ or picking one of the [400+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq).
+You can quantize your own models by installing AutoAWQ or picking one of the [6500+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq).
 
 ```console
 pip install autoawq
 ```
 
-After installing AutoAWQ, you are ready to quantize a model. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
+After installing AutoAWQ, you are ready to quantize a model. Please refer to the `AutoAWQ documentation <https://casper-hansen.github.io/AutoAWQ/examples/#basic-quantization>`_ for further details. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
 
 ```python
 from awq import AutoAWQForCausalLM

docs/source/features/quantization/gguf.md

@@ -29,6 +29,13 @@ vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlam
 We recommend using the tokenizer from base model instead of GGUF model. Because the tokenizer conversion from GGUF is time-consuming and unstable, especially for some models with large vocab size.
 :::
 
+GGUF assumes that huggingface can convert the metadata to a config file. In case huggingface doesn't support your model you can manually create a config and pass it as hf-confing-path
+
+```console
+# If you model is not supported by huggingface you can manually provide a huggingface compatible config path
+vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0 --hf-config-path Tinyllama/TInyLlama-1.1B-Chat-v1.0
+```
+
 You can also use the GGUF model directly through the LLM entrypoint:
 
 ```python

docs/source/features/quantization/supported_hardware.md

@@ -20,93 +20,93 @@ The table below shows the compatibility of various quantization implementations
   * AWS Inferentia
   * Google TPU
 - * AWQ
-  * ✗
+  * ❌
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
+  * ❌
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
+  * ❌
+  * ❌
 - * GPTQ
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
+  * ❌
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
+  * ❌
+  * ❌
 - * Marlin (GPTQ/AWQ/FP8)
-  * ✗
-  * ✗
+  * ❌
+  * ❌
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ❌
 - * INT8 (W8A8)
-  * ✗
+  * ❌
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
+  * ❌
+  * ❌
   * ✅︎
-  * ✗
-  * ✗
+  * ❌
+  * ❌
 - * FP8 (W8A8)
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
+  * ❌
 - * AQLM
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ❌
 - * bitsandbytes
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ❌
 - * DeepSpeedFP
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
+  * ❌
+  * ❌
 - * GGUF
   * ✅︎
   * ✅︎
@@ -114,16 +114,16 @@ The table below shows the compatibility of various quantization implementations
   * ✅︎
   * ✅︎
   * ✅︎
-  * ✗
-  * ✗
-  * ✗
-  * ✗
+  * ❌
+  * ❌
+  * ❌
+  * ❌
 
 :::
 
 - Volta refers to SM 7.0, Turing to SM 7.5, Ampere to SM 8.0/8.6, Ada to SM 8.9, and Hopper to SM 9.0.
-- "✅︎" indicates that the quantization method is supported on the specified hardware.
-- "✗" indicates that the quantization method is not supported on the specified hardware.
+- ✅︎ indicates that the quantization method is supported on the specified hardware.
+- ❌ indicates that the quantization method is not supported on the specified hardware.
 
 :::{note}
 This compatibility chart is subject to change as vLLM continues to evolve and expand its support for different hardware platforms and quantization methods.