fix

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

.buildkite/nightly-benchmarks/README.md

@@ -28,6 +28,7 @@ See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performanc
 ## Trigger the benchmark
 
 Performance benchmark will be triggered when:
+
 - A PR being merged into vllm.
 - Every commit for those PRs with `perf-benchmarks` label AND `ready` label.
 
@@ -38,6 +39,7 @@ bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
 ```
 
 Runtime environment variables:
+
 - `ON_CPU`: set the value to '1' on Intel® Xeon® Processors. Default value is 0.
 - `SERVING_JSON`: JSON file to use for the serving tests. Default value is empty string (use default file).
 - `LATENCY_JSON`: JSON file to use for the latency tests. Default value is empty string (use default file).
@@ -46,12 +48,14 @@ Runtime environment variables:
 - `REMOTE_PORT`: Port for the remote vLLM service to benchmark. Default value is empty string.
 
 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.
 > NOTE: For Intel® Xeon® Processors, use `tests/latency-tests-cpu.json`, `tests/throughput-tests-cpu.json`, `tests/serving-tests-cpu.json` instead.
+>
 ### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
@@ -100,7 +104,6 @@ We test the throughput by using `vllm bench serve` with request rate = inf to co
             "tensor_parallel_size": 1,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -149,6 +152,7 @@ Here is an example using the script to compare result_a and result_b without det
 
 Here is an example using the script to compare result_a and result_b with detail test name.
 `python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json`
+
 |   | results_a/benchmark_results.json_name | results_a/benchmark_results.json | results_b/benchmark_results.json_name | results_b/benchmark_results.json | perf_ratio        |
 |---|---------------------------------------------|----------------------------------------|---------------------------------------------|----------------------------------------|----------|
 | 0 | serving_llama8B_tp1_sharegpt_qps_1          | 142.633982                             | serving_llama8B_tp1_sharegpt_qps_1          | 156.526018                             | 1.097396 |

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

@@ -1,3 +1,4 @@
+# Nightly benchmark annotation
 
 ## Description
 
@@ -13,15 +14,15 @@ Please download the visualization scripts in the post
 
 - 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:
+    - Download `nightly-benchmarks.zip`.
+    - In the same folder, run the following code:
 
-  ```bash
-  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
-  ```
+    ```bash
+    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

@@ -13,25 +13,25 @@ Latest reproduction guilde: [github issue link](https://github.com/vllm-project/
 ## Setup
 
 - Docker images:
-  - vLLM: `vllm/vllm-openai:v0.6.2`
-  - SGLang: `lmsysorg/sglang:v0.3.2-cu121`
-  - LMDeploy: `openmmlab/lmdeploy:v0.6.1-cu12`
-  - TensorRT-LLM: `nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3`
-    - *NOTE: we uses r24.07 as the current implementation only works for this version. We are going to bump this up.*
-  - Check [nightly-pipeline.yaml](nightly-pipeline.yaml) for the concrete docker images, specs and commands we use for the benchmark.
+    - vLLM: `vllm/vllm-openai:v0.6.2`
+    - SGLang: `lmsysorg/sglang:v0.3.2-cu121`
+    - LMDeploy: `openmmlab/lmdeploy:v0.6.1-cu12`
+    - TensorRT-LLM: `nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3`
+        - *NOTE: we uses r24.07 as the current implementation only works for this version. We are going to bump this up.*
+    - Check [nightly-pipeline.yaml](nightly-pipeline.yaml) for the concrete docker images, specs and commands we use for the benchmark.
 - Hardware
-  - 8x Nvidia A100 GPUs
+    - 8x Nvidia A100 GPUs
 - Workload:
-  - Dataset
-    - ShareGPT dataset
-    - Prefill-heavy dataset (in average 462 input tokens, 16 tokens as output)
-    - Decode-heavy dataset (in average 462 input tokens, 256 output tokens)
-    - Check [nightly-tests.json](tests/nightly-tests.json) for the concrete configuration of datasets we use.
-  - Models: llama-3 8B, llama-3 70B.
-    - We do not use llama 3.1 as it is incompatible with trt-llm r24.07. ([issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105)).
-  - Average QPS (query per second): 2, 4, 8, 16, 32 and inf.
-    - 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).
+    - Dataset
+        - ShareGPT dataset
+        - Prefill-heavy dataset (in average 462 input tokens, 16 tokens as output)
+        - Decode-heavy dataset (in average 462 input tokens, 256 output tokens)
+        - Check [nightly-tests.json](tests/nightly-tests.json) for the concrete configuration of datasets we use.
+    - Models: llama-3 8B, llama-3 70B.
+        - We do not use llama 3.1 as it is incompatible with trt-llm r24.07. ([issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105)).
+    - Average QPS (query per second): 2, 4, 8, 16, 32 and inf.
+        - 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
 

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

@@ -1,3 +1,4 @@
+# Performance benchmarks descriptions
 
 ## Latency tests
 

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

@@ -44,6 +44,7 @@ serving_column_mapping = {
     "test_name": "Test name",
     "gpu_type": "GPU",
     "completed": "# of req.",
+    "max_concurrency": "# of max concurrency.",
     "request_throughput": "Tput (req/s)",
     "total_token_throughput": "Total Token Tput (tok/s)",
     "output_throughput": "Output Tput (tok/s)",

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

@@ -33,7 +33,7 @@ check_gpus() {
 
 check_cpus() {
   # check the number of CPUs and NUMA Node and GPU type.
-  declare -g numa_count=$(python3 -c  "from numa import info;numa_size = info.get_num_configured_nodes(); print(numa_size)")
+  declare -g numa_count=$(lscpu | grep "NUMA node(s):" | awk '{print $3}')
   if [[ $numa_count -gt 0 ]]; then
     echo "NUMA found."
     echo $numa_count

.buildkite/nightly-benchmarks/tests/genai-perf-tests.json

@@ -11,7 +11,6 @@
         },
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,

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

@@ -35,7 +35,6 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,
@@ -90,7 +89,6 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,
@@ -145,7 +143,6 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,
@@ -197,7 +194,6 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,
@@ -251,7 +247,6 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,
@@ -305,7 +300,6 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
             "num_scheduler_steps": 10,
             "max_num_seqs": 512,

.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc2.json

@@ -0,0 +1,203 @@
+[
+    {
+        "test_name": "serving_llama8B_tp1_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp2_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp1_random_128_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 1000,
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp2_random_128_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 1000,
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_random_128_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 1000,
+            "num_prompts": 1000
+        }
+    }
+]

.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc3.json

@@ -0,0 +1,205 @@
+[
+    {
+        "test_name": "serving_llama8B_pp1_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "pipeline_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_pp3_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp2pp6_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_pp1_random_128_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "pipeline_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 1000,
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_pp3_random_128_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL:": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 1000,
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp2pp3_random_128_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 1000,
+            "num_prompts": 1000
+        }
+    }
+]

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

@@ -6,6 +6,7 @@
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "server_parameters": {
@@ -16,8 +17,9 @@
 	    "block_size": 128,
 	    "trust_remote_code": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -36,6 +38,7 @@
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "server_parameters": {
@@ -46,8 +49,9 @@
 	    "block_size": 128,
 	    "trust_remote_code": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -66,6 +70,7 @@
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "server_parameters": {
@@ -76,8 +81,9 @@
 	    "block_size": 128,
 	    "trust_remote_code": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -96,6 +102,7 @@
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "server_parameters": {
@@ -107,8 +114,9 @@
 	    "trust_remote_code": "",
 	    "enable_chunked_prefill": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -129,6 +137,7 @@
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "server_parameters": {
@@ -140,8 +149,9 @@
 	    "trust_remote_code": "",
 	    "enable_chunked_prefill": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {

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

@@ -7,7 +7,6 @@
             "tensor_parallel_size": 1,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -26,7 +25,6 @@
             "tensor_parallel_size": 4,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -45,7 +43,6 @@
             "tensor_parallel_size": 2,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -60,8 +57,7 @@
         "test_name": "serving_llama70B_tp4_sharegpt_specdecode",
         "qps_list": [2],
         "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "disable_log_requests": "", 
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct", 
             "tensor_parallel_size": 4,
             "swap_space": 16,
             "speculative_config": {

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

@@ -16,8 +16,7 @@ DOCKER_BUILDKIT=1 docker build . \
   --build-arg max_jobs=66 \
   --build-arg nvcc_threads=2 \
   --build-arg RUN_WHEEL_CHECK=false \
-  --build-arg torch_cuda_arch_list="9.0+PTX" \
-  --build-arg vllm_fa_cmake_gpu_arches="90-real"
+  --build-arg torch_cuda_arch_list="9.0+PTX"
 
 # Setup cleanup
 remove_docker_container() { docker rm -f gh200-test || true; }

.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh

@@ -4,8 +4,7 @@ set -xu
 
 
 remove_docker_container() { 
-    docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
+    docker rm -f tpu-test || true;
 }
 
 trap remove_docker_container EXIT

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

@@ -5,7 +5,6 @@ set -xu
 
 remove_docker_container() { 
     docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
 }
 
 trap remove_docker_container EXIT

.buildkite/scripts/tpu/config_v6e_1.env

@@ -1,6 +1,6 @@
 # Environment config
 TEST_NAME=llama8b
-CONTAINER_NAME=vllm-tpu
+CONTAINER_NAME=tpu-test
 
 # vllm config
 MODEL=meta-llama/Llama-3.1-8B-Instruct

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -12,8 +12,6 @@ source /etc/environment
 source $ENV_FILE
 
 remove_docker_container() { 
-    docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
     docker rm -f $CONTAINER_NAME || true;
 }
 

.buildkite/scripts/tpu/quantized_v6e_1.env

@@ -1,6 +1,6 @@
 # Environment config
 TEST_NAME=llama8bw8a8
-CONTAINER_NAME=vllm-tpu
+CONTAINER_NAME=tpu-test
 
 # vllm config
 MODEL=RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8

.buildkite/scripts/tpu/run_bm.sh

@@ -44,7 +44,6 @@ echo
 
 VLLM_USE_V1=1 vllm serve $MODEL \
  --seed 42 \
- --disable-log-requests \
  --max-num-seqs $MAX_NUM_SEQS \
  --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
  --tensor-parallel-size $TENSOR_PARALLEL_SIZE \

.buildkite/test-pipeline.yaml

@@ -82,7 +82,7 @@ steps:
   - bash standalone_tests/python_only_compile.sh
 
 - label: Basic Correctness Test # 30min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: true
   torch_nightly: true
   source_file_dependencies:
@@ -99,7 +99,7 @@ steps:
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Chunked Prefill Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
@@ -108,7 +108,7 @@ steps:
   - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
 
 - label: Core Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: true
   source_file_dependencies:
   - vllm/core
@@ -209,7 +209,7 @@ steps:
   - pytest -v -s distributed/test_eplb_execute.py
 
 - label: Metrics, Tracing Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -228,7 +228,7 @@ steps:
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -280,7 +280,7 @@ steps:
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: Examples Test # 25min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
@@ -305,7 +305,7 @@ steps:
     - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/prefix_caching
@@ -314,7 +314,7 @@ steps:
 
 
 - label: Platform Tests (CUDA)
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/cuda
@@ -353,9 +353,10 @@ steps:
     - pytest -v -s compile/test_silu_mul_quant_fusion.py
     - pytest -v -s compile/test_sequence_parallelism.py
     - pytest -v -s compile/test_async_tp.py
+    - pytest -v -s compile/test_fusion_all_reduce.py
 
 - label: PyTorch Fullgraph Smoke Test # 9min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -368,7 +369,7 @@ steps:
   - pytest -v -s compile/piecewise/test_full_cudagraph.py
 
 - label: PyTorch Fullgraph Test # 18min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -377,7 +378,7 @@ steps:
   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Core Operation Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
@@ -416,7 +417,7 @@ steps:
   parallelism: 2
 
 - label: Kernels Mamba Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -424,7 +425,7 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -437,7 +438,7 @@ steps:
     - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
 
 - label: Model Executor Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/model_executor
   - tests/model_executor
@@ -447,7 +448,7 @@ steps:
     - pytest -v -s model_executor
 
 - label: Benchmarks # 9min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/.buildkite"
   source_file_dependencies:
   - benchmarks/
@@ -455,7 +456,7 @@ steps:
   - bash scripts/run-benchmarks.sh
 
 - label: Benchmarks CLI Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/benchmarks/
@@ -494,7 +495,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/
 
 - label: Encoder Decoder tests # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/encoder_decoder
@@ -502,7 +503,7 @@ steps:
     - pytest -v -s encoder_decoder
 
 - label: OpenAI-Compatible Tool Use # 20 min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: false
   source_file_dependencies:
     - vllm/
@@ -580,7 +581,8 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/multimodal/processing
-    - pytest -v -s --ignore models/multimodal/generation/test_whisper.py models/multimodal -m core_model
+    - pytest -v -s --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/test_tensor_schema.py models/multimodal -m core_model
+    - pytest -v -s models/multimodal/test_tensor_schema.py -m core_model  # Needs mp_method="spawn"
     - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
 - label: Multi-Modal Models Test (Extended) 1
@@ -623,7 +625,7 @@ steps:
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -643,11 +645,40 @@ steps:
     - python3 examples/offline_inference/audio_language.py --model-type whisper
     - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
 
+- label: Blackwell Test
+  working_dir: "/vllm-workspace/"
+  gpu: b200
+  # optional: true
+  source_file_dependencies:
+  - csrc/quantization/fp4/
+  - csrc/attention/mla/
+  - csrc/quantization/cutlass_w8a8/moe/
+  - vllm/model_executor/layers/fused_moe/cutlass_moe.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
+  - vllm/v1/attention/backends/flashinfer.py
+  - vllm/compilation/fusion.py
+  commands:
+    - nvidia-smi
+    - python3 examples/offline_inference/basic/chat.py
+    # Attention
+    # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
+    - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
+    - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
+    - pytest -v -s tests/kernels/test_cutlass_mla_decode.py
+    # Quantization
+    - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
+    - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
+    - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
+    - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
+    # Fusion
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
 - label: Distributed Comm Ops Test # 7min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -718,7 +749,6 @@ steps:
   # this test fails consistently.
   # TODO: investigate and fix
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s models/multimodal/generation/test_maverick.py
 
@@ -744,7 +774,7 @@ steps:
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 - label: Multi-step Tests (4 GPUs) # 36min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -765,7 +795,7 @@ steps:
   - pytest -v -s multi_step/test_correctness_llm.py
 
 - label: Pipeline Parallelism Test # 45min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -779,7 +809,7 @@ steps:
   - pytest -v -s distributed/test_pipeline_parallel.py
 
 - label: LoRA TP Test (Distributed)
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   num_gpus: 4
   source_file_dependencies:
   - vllm/lora
@@ -792,6 +822,7 @@ steps:
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
+    - pytest -v -s -x lora/test_multi_loras_with_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

.github/CODEOWNERS

@@ -61,3 +61,15 @@ mkdocs.yaml @hmellor
 /vllm/v1/worker/^xpu @jikunshang
 /vllm/platforms/xpu.py @jikunshang
 /docker/Dockerfile.xpu @jikunshang
+
+# Qwen-specific files
+/vllm/attention/backends/dual_chunk_flash_attn.py @sighingnow
+/vllm/model_executor/models/qwen* @sighingnow
+
+# Mistral-specific files
+/vllm/model_executor/models/mistral*.py @patrickvonplaten
+/vllm/model_executor/models/mixtral*.py @patrickvonplaten
+/vllm/model_executor/models/voxtral*.py @patrickvonplaten
+/vllm/model_executor/models/pixtral*.py @patrickvonplaten
+/vllm/transformers_utils/configs/mistral.py @patrickvonplaten
+/vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,4 +1,5 @@
-## Essential Elements of an Effective PR Description Checklist
+# Essential Elements of an Effective PR Description Checklist
+
 - [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
 - [ ] The test plan, such as providing test command.
 - [ ] The test results, such as pasting the results comparison before and after, or e2e results
@@ -14,5 +15,4 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE B
 
 ## (Optional) Documentation Update
 
-<!--- pyml disable-next-line no-emphasis-as-heading -->
 **BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing>** (anything written below this line will be removed by GitHub Actions)

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

@@ -2,12 +2,16 @@ name: Lint and Deploy Charts
 
 on: pull_request
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 permissions:
   contents: read
 
 jobs:
   lint-and-deploy:
-    runs-on: ubuntu-24.04-arm
+    runs-on: ubuntu-latest
     steps:
       - name: Checkout
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

.github/workflows/matchers/markdownlint.json

@@ -0,0 +1,17 @@
+{
+  "problemMatcher": [
+    {
+      "owner": "markdownlint",
+      "pattern": [
+        {
+          "regexp": "^([^:]*):(\\d+):?(\\d+)?\\s([\\w-\\/]*)\\s(.*)$",
+          "file": 1,
+          "line": 2,
+          "column": 3,
+          "code": 4,
+          "message": 5
+        }
+      ]
+    }
+  ]
+}

.github/workflows/pre-commit.yml

@@ -5,6 +5,10 @@ on:
   push:
     branches: [main]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
 permissions:
   contents: read
 
@@ -17,6 +21,7 @@ jobs:
       with:
         python-version: "3.12"
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
+    - run: echo "::add-matcher::.github/workflows/matchers/markdownlint.json"
     - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"
     - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1
       with:

.github/workflows/scripts/build.sh

@@ -15,7 +15,6 @@ $python_executable -m pip install -r requirements/build.txt -r requirements/cuda
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
 export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
-export VLLM_FA_CMAKE_GPU_ARCHES="80-real;90-real"
 
 bash tools/check_repo.sh
 

.markdownlint.yaml

@@ -0,0 +1,13 @@
+MD007:
+  indent: 4
+MD013: false
+MD024:
+  siblings_only: true
+MD033: false
+MD042: false
+MD045: false
+MD046: false
+MD051: false
+MD052: false
+MD053: false
+MD059: false

.pre-commit-config.yaml

@@ -35,12 +35,12 @@ repos:
     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: v0.9.29
+- repo: https://github.com/igorshubovych/markdownlint-cli
+  rev: v0.45.0
   hooks:
-  - id: pymarkdown
+  - id: markdownlint
     exclude: '.*\.inc\.md'
-    args: [fix]
+    stages: [manual] # Only run in CI
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:

CMakeLists.txt

@@ -529,6 +529,25 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # The nvfp4_scaled_mm_sm120 kernels for Geforce Blackwell SM120 require
+  # CUDA 12.8 or later
+  cuda_archs_loose_intersection(FP4_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
+    set(SRCS
+      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
+      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_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_SM120=1")
+    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
+  else()
+    message(STATUS "Not building NVFP4 as no compatible archs were found.")
+    # clear FP4_ARCHS
+    set(FP4_ARCHS)
+  endif()
+
   # FP4 Archs and flags
   cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
@@ -541,7 +560,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
     message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
   else()

README.md

@@ -1,3 +1,4 @@
+<!-- markdownlint-disable MD001 MD041 -->
 <p align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/assets/logos/vllm-logo-text-dark.png">
@@ -16,6 +17,7 @@ Easy, fast, and cheap LLM serving for everyone
 ---
 
 *Latest News* 🔥
+
 - [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
 - [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
 - [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
@@ -46,6 +48,7 @@ Easy, fast, and cheap LLM serving for everyone
 </details>
 
 ---
+
 ## About
 
 vLLM is a fast and easy-to-use library for LLM inference and serving.
@@ -75,6 +78,7 @@ vLLM is flexible and easy to use with:
 - Multi-LoRA support
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
+
 - Transformer-like LLMs (e.g., Llama)
 - Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
 - Embedding Models (e.g., E5-Mistral)
@@ -91,6 +95,7 @@ pip install vllm
 ```
 
 Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
+
 - [Installation](https://docs.vllm.ai/en/latest/getting_started/installation.html)
 - [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
 - [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
@@ -107,6 +112,7 @@ vLLM is a community project. Our compute resources for development and testing a
 <!-- Note: Please sort them in alphabetical order. -->
 <!-- Note: Please keep these consistent with docs/community/sponsors.md -->
 Cash Donations:
+
 - a16z
 - Dropbox
 - Sequoia Capital
@@ -114,6 +120,7 @@ Cash Donations:
 - ZhenFund
 
 Compute Resources:
+
 - AMD
 - Anyscale
 - AWS

RELEASE.md

@@ -60,9 +60,10 @@ Please note: **No feature work allowed for cherry picks**. All PRs that are cons
 Before each release, we perform end-to-end performance validation to ensure no regressions are introduced. This validation uses the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) on PyTorch CI.
 
 **Current Coverage:**
+
 * Models: Llama3, Llama4, and Mixtral
 * Hardware: NVIDIA H100 and AMD MI300x
-* *Note: Coverage may change based on new model releases and hardware availability*
+* _Note: Coverage may change based on new model releases and hardware availability_
 
 **Performance Validation Process:**
 
@@ -71,11 +72,13 @@ Request write access to the [pytorch/pytorch-integration-testing](https://github
 
 **Step 2: Review Benchmark Setup**
 Familiarize yourself with the benchmark configurations:
+
 * [CUDA setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/cuda)
 * [ROCm setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/rocm)
 
 **Step 3: Run the Benchmark**
 Navigate to the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) and configure:
+
 * **vLLM branch**: Set to the release branch (e.g., `releases/v0.9.2`)
 * **vLLM commit**: Set to the RC commit hash
 

SECURITY.md

@@ -1,13 +1,45 @@
 # Security Policy
 
-## Reporting a Vulnerability
+## Reporting security issues
 
-If you believe you have found a security vulnerability in vLLM, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.
+Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new).
 
-Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).
+## Issue triage
 
----
+Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).
+
+## Threat model
 
 Please see the [Security Guide in the vLLM documentation](https://docs.vllm.ai/en/latest/usage/security.html) for more information on vLLM's security assumptions and recommendations.
 
 Please see [PyTorch's Security Policy](https://github.com/pytorch/pytorch/blob/main/SECURITY.md) for more information and recommendations on how to securely interact with models.
+
+## Issue severity
+
+We will determine the risk of each issue, taking into account our experience dealing with past issues, versions affected, common defaults, and use cases. We use the following severity categories:
+
+### CRITICAL Severity
+
+Vulnerabilities that allow remote attackers to execute arbitrary code, take full control of the system, or significantly compromise confidentiality, integrity, or availability without any interaction or privileges needed, examples include remote code execution via network, deserialization issues that allow exploit chains. Generally those issues which are rated as CVSS  ≥ 9.0.
+
+### HIGH Severity
+
+Serious security flaws that allow elevated impact—like RCE in specific, limited contexts or significant data loss—but require advanced conditions or some trust, examples include RCE in advanced deployment modes (e.g. multi-node), or high impact issues where some sort of privileged network access is required. These issues typically have CVSS scores between 7.0 and 8.9
+
+### MODERATE Severity
+
+Vulnerabilities that cause denial of service or partial disruption, but do not allow arbitrary code execution or data breach and have limited impact. These issues have a CVSS rating between 4.0 and 6.9
+
+### LOW Severity
+
+Minor issues such as informational disclosures, logging errors, non-exploitable flaws, or weaknesses that require local or high-privilege access and offer negligible impact. Examples include side channel attacks or hash collisions. These issues often have CVSS scores less than 4.0
+
+## Prenotification policy
+
+For certain security issues of CRITICAL, HIGH, or MODERATE severity level, we may prenotify certain organizations or vendors that ship vLLM. The purpose of this prenotification is to allow for a coordinated release of fixes for severe issues.
+
+* This prenotification will be in the form of a private email notification. It may also include adding security contacts to the GitHub security advisory, typically a few days before release.
+
+* If you wish to be added to the prenotification group, please send an email copying all the members of the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html). Each vendor contact will be analyzed on a case-by-case basis.
+
+* We may withdraw organizations from receiving future prenotifications if they release fixes or any other information about issues before they are public. Group membership may also change based on policy refinements for who may be included.

benchmarks/README.md

@@ -4,7 +4,7 @@ This README guides you through running benchmark tests with the extensive
 datasets supported on vLLM. It’s a living document, updated as new features and datasets
 become available.
 
-**Dataset Overview**
+## Dataset Overview
 
 <table style="width:100%; border-collapse: collapse;">
   <thead>
@@ -81,16 +81,17 @@ become available.
 
 **Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
 
----
+## 🚀 Example - Online Benchmark
+
 <details>
-<summary><b>🚀 Example - Online Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 First start serving your model
 
 ```bash
-vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
 ```
 
 Then run the benchmarking script
@@ -109,7 +110,7 @@ vllm bench serve \
 
 If successful, you will see the following output
 
-```
+```text
 ============ Serving Benchmark Result ============
 Successful requests:                     10
 Benchmark duration (s):                  5.78
@@ -133,11 +134,11 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
-**Custom Dataset**
+### Custom Dataset
 
 If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
 
-```
+```json
 {"prompt": "What is the capital of India?"}
 {"prompt": "What is the capital of Iran?"}
 {"prompt": "What is the capital of China?"}
@@ -145,7 +146,7 @@ If the dataset you want to benchmark is not supported yet in vLLM, even then you
 
 ```bash
 # start server
-VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests
+VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct
 ```
 
 ```bash
@@ -166,11 +167,11 @@ vllm bench serve --port 9001 --save-result --save-detailed \
 
 You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
 
-**VisionArena Benchmark for Vision Language Models**
+### VisionArena Benchmark for Vision Language Models
 
 ```bash
 # need a model with vision capability here
-vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+vllm serve Qwen/Qwen2-VL-7B-Instruct
 ```
 
 ```bash
@@ -184,7 +185,7 @@ vllm bench serve \
   --num-prompts 1000
 ```
 
-**InstructCoder Benchmark with Speculative Decoding**
+### InstructCoder Benchmark with Speculative Decoding
 
 ``` bash
 VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
@@ -201,13 +202,13 @@ vllm bench serve \
     --num-prompts 2048
 ```
 
-**Other HuggingFaceDataset Examples**
+### Other HuggingFaceDataset Examples
 
 ```bash
-vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+vllm serve Qwen/Qwen2-VL-7B-Instruct
 ```
 
-**`lmms-lab/LLaVA-OneVision-Data`**
+`lmms-lab/LLaVA-OneVision-Data`:
 
 ```bash
 vllm bench serve \
@@ -221,7 +222,7 @@ vllm bench serve \
   --num-prompts 10
 ```
 
-**`Aeala/ShareGPT_Vicuna_unfiltered`**
+`Aeala/ShareGPT_Vicuna_unfiltered`:
 
 ```bash
 vllm bench serve \
@@ -234,7 +235,7 @@ vllm bench serve \
   --num-prompts 10
 ```
 
-**`AI-MO/aimo-validation-aime`**
+`AI-MO/aimo-validation-aime`:
 
 ``` bash
 vllm bench serve \
@@ -245,7 +246,7 @@ vllm bench serve \
     --seed 42
 ```
 
-**`philschmid/mt-bench`**
+`philschmid/mt-bench`:
 
 ``` bash
 vllm bench serve \
@@ -255,7 +256,7 @@ vllm bench serve \
     --num-prompts 80
 ```
 
-**Running With Sampling Parameters**
+### Running With Sampling Parameters
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling
 parameters can be specified. Example client command:
@@ -273,25 +274,29 @@ vllm bench serve \
   --num-prompts 10
 ```
 
-**Running With Ramp-Up Request Rate**
+### Running With Ramp-Up Request Rate
 
 The benchmark tool also supports ramping up the request rate over the
 duration of the benchmark run. This can be useful for stress testing the
 server or finding the maximum throughput that it can handle, given some latency budget.
 
 Two ramp-up strategies are supported:
+
 - `linear`: Increases the request rate linearly from a start value to an end value.
 - `exponential`: Increases the request rate exponentially.
 
 The following arguments can be used to control the ramp-up:
+
 - `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`).
 - `--ramp-up-start-rps`: The request rate at the beginning of the benchmark.
 - `--ramp-up-end-rps`: The request rate at the end of the benchmark.
 
 </details>
 
+## 📈 Example - Offline Throughput Benchmark
+
 <details>
-<summary><b>📈 Example - Offline Throughput Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
@@ -305,15 +310,15 @@ vllm bench throughput \
 
 If successful, you will see the following output
 
-```
+```text
 Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s
 Total num prompt tokens:  5014
 Total num output tokens:  1500
 ```
 
-**VisionArena Benchmark for Vision Language Models**
+### VisionArena Benchmark for Vision Language Models
 
-``` bash
+```bash
 vllm bench throughput \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --backend vllm-chat \
@@ -325,13 +330,13 @@ vllm bench throughput \
 
 The `num prompt tokens` now includes image token counts
 
-```
+```text
 Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s
 Total num prompt tokens:  14527
 Total num output tokens:  1280
 ```
 
-**InstructCoder Benchmark with Speculative Decoding**
+### InstructCoder Benchmark with Speculative Decoding
 
 ``` bash
 VLLM_WORKER_MULTIPROC_METHOD=spawn \
@@ -349,15 +354,15 @@ vllm bench throughput \
     "prompt_lookup_min": 2}'
 ```
 
-```
+```text
 Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
 Total num prompt tokens:  261136
 Total num output tokens:  204800
 ```
 
-**Other HuggingFaceDataset Examples**
+### Other HuggingFaceDataset Examples
 
-**`lmms-lab/LLaVA-OneVision-Data`**
+`lmms-lab/LLaVA-OneVision-Data`:
 
 ```bash
 vllm bench throughput \
@@ -370,7 +375,7 @@ vllm bench throughput \
   --num-prompts 10
 ```
 
-**`Aeala/ShareGPT_Vicuna_unfiltered`**
+`Aeala/ShareGPT_Vicuna_unfiltered`:
 
 ```bash
 vllm bench throughput \
@@ -382,7 +387,7 @@ vllm bench throughput \
   --num-prompts 10
 ```
 
-**`AI-MO/aimo-validation-aime`**
+`AI-MO/aimo-validation-aime`:
 
 ```bash
 vllm bench throughput \
@@ -394,7 +399,7 @@ vllm bench throughput \
   --num-prompts 10
 ```
 
-**Benchmark with LoRA Adapters**
+Benchmark with LoRA adapters:
 
 ``` bash
 # download dataset
@@ -413,20 +418,22 @@ vllm bench throughput \
 
 </details>
 
+## 🛠️ Example - Structured Output Benchmark
+
 <details>
-<summary><b>🛠️ Example - Structured Output Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the performance of structured output generation (JSON, grammar, regex).
 
-**Server Setup**
+### Server Setup
 
 ```bash
-vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
 ```
 
-**JSON Schema Benchmark**
+### JSON Schema Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -438,7 +445,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**Grammar-based Generation Benchmark**
+### Grammar-based Generation Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -450,7 +457,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**Regex-based Generation Benchmark**
+### Regex-based Generation Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -461,7 +468,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**Choice-based Generation Benchmark**
+### Choice-based Generation Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -472,7 +479,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**XGrammar Benchmark Dataset**
+### XGrammar Benchmark Dataset
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -485,14 +492,16 @@ python3 benchmarks/benchmark_serving_structured_output.py \
 
 </details>
 
+## 📚 Example - Long Document QA Benchmark
+
 <details>
-<summary><b>📚 Example - Long Document QA Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the performance of long document question-answering with prefix caching.
 
-**Basic Long Document QA Test**
+### Basic Long Document QA Test
 
 ```bash
 python3 benchmarks/benchmark_long_document_qa_throughput.py \
@@ -504,7 +513,7 @@ python3 benchmarks/benchmark_long_document_qa_throughput.py \
   --repeat-count 5
 ```
 
-**Different Repeat Modes**
+### Different Repeat Modes
 
 ```bash
 # Random mode (default) - shuffle prompts randomly
@@ -537,14 +546,16 @@ python3 benchmarks/benchmark_long_document_qa_throughput.py \
 
 </details>
 
+## 🗂️ Example - Prefix Caching Benchmark
+
 <details>
-<summary><b>🗂️ Example - Prefix Caching Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the efficiency of automatic prefix caching.
 
-**Fixed Prompt with Prefix Caching**
+### Fixed Prompt with Prefix Caching
 
 ```bash
 python3 benchmarks/benchmark_prefix_caching.py \
@@ -555,7 +566,7 @@ python3 benchmarks/benchmark_prefix_caching.py \
   --input-length-range 128:256
 ```
 
-**ShareGPT Dataset with Prefix Caching**
+### ShareGPT Dataset with Prefix Caching
 
 ```bash
 # download dataset
@@ -572,14 +583,16 @@ python3 benchmarks/benchmark_prefix_caching.py \
 
 </details>
 
+## ⚡ Example - Request Prioritization Benchmark
+
 <details>
-<summary><b>⚡ Example - Request Prioritization Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the performance of request prioritization in vLLM.
 
-**Basic Prioritization Test**
+### Basic Prioritization Test
 
 ```bash
 python3 benchmarks/benchmark_prioritization.py \
@@ -590,7 +603,7 @@ python3 benchmarks/benchmark_prioritization.py \
   --scheduling-policy priority
 ```
 
-**Multiple Sequences per Prompt**
+### Multiple Sequences per Prompt
 
 ```bash
 python3 benchmarks/benchmark_prioritization.py \

benchmarks/auto_tune/README.md

@@ -3,6 +3,7 @@
 This script automates the process of finding the optimal server parameter combination (`max-num-seqs` and `max-num-batched-tokens`) to maximize throughput for a vLLM server. It also supports additional constraints such as E2E latency and prefix cache hit rate.
 
 ## Table of Contents
+
 - [Prerequisites](#prerequisites)
 - [Configuration](#configuration)
 - [How to Run](#how-to-run)
@@ -52,7 +53,7 @@ You must set the following variables at the top of the script before execution.
 1. **Configure**: Edit the script and set the variables in the [Configuration](#configuration) section.
 2. **Execute**: Run the script. Since the process can take a long time, it is highly recommended to use a terminal multiplexer like `tmux` or `screen` to prevent the script from stopping if your connection is lost.
 
-```
+```bash
 cd <FOLDER_OF_THIS_SCRIPT>
 bash auto_tune.sh
 ```
@@ -64,6 +65,7 @@ bash auto_tune.sh
 Here are a few examples of how to configure the script for different goals:
 
 ### 1. Maximize Throughput (No Latency Constraint)
+
 - **Goal**: Find the best `max-num-seqs` and `max-num-batched-tokens` to get the highest possible throughput for 1800 input tokens and 20 output tokens.
 - **Configuration**:
 
@@ -76,6 +78,7 @@ MAX_LATENCY_ALLOWED_MS=100000000000 # A very large number
 ```
 
 #### 2. Maximize Throughput with a Latency Requirement
+
 - **Goal**: Find the best server parameters when P99 end-to-end latency must be below 500ms.
 - **Configuration**:
 
@@ -88,6 +91,7 @@ MAX_LATENCY_ALLOWED_MS=500
 ```
 
 #### 3. Maximize Throughput with Prefix Caching and Latency Requirements
+
 - **Goal**: Find the best server parameters assuming a 60% prefix cache hit rate and a latency requirement of 500ms.
 - **Configuration**:
 
@@ -109,7 +113,7 @@ After the script finishes, you will find the results in a new, timestamped direc
 
 - **Final Result Summary**: A file named `result.txt` is created in the log directory. It contains a summary of each tested combination and concludes with the overall best parameters found.
 
-```
+```text
 # Example result.txt content
 hash:a1b2c3d4...
 max_num_seqs: 128, max_num_batched_tokens: 2048, request_rate: 10.0, e2el: 450.5, throughput: 9.8, goodput: 9.8

benchmarks/auto_tune/auto_tune.sh

@@ -49,6 +49,7 @@ best_throughput=0
 best_max_num_seqs=0
 best_num_batched_tokens=0
 best_goodput=0
+best_request_rate=0
 
 start_server() {
     local gpu_memory_utilization=$1
@@ -57,19 +58,35 @@ start_server() {
     local vllm_log=$4
     local profile_dir=$5
 
-    pkill -f vllm
+    pkill -if vllm
 
-    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
-        --disable-log-requests \
-        --port 8004 \
-        --gpu-memory-utilization $gpu_memory_utilization \
-        --max-num-seqs $max_num_seqs \
-        --max-num-batched-tokens $max_num_batched_tokens \
-        --tensor-parallel-size $TP \
-        --enable-prefix-caching \
-        --load-format dummy \
-        --download-dir "$DOWNLOAD_DIR" \
-        --max-model-len $MAX_MODEL_LEN > "$vllm_log" 2>&1 &
+    # Define the common arguments as a bash array.
+    # Each argument and its value are separate elements.
+    local common_args_array=(
+        "$MODEL"
+        "--disable-log-requests"
+        "--port" "8004"
+        "--gpu-memory-utilization" "$gpu_memory_utilization"
+        "--max-num-seqs" "$max_num_seqs"
+        "--max-num-batched-tokens" "$max_num_batched_tokens"
+        "--tensor-parallel-size" "$TP"
+        "--enable-prefix-caching"
+        "--load-format" "dummy"
+        "--download-dir" "$DOWNLOAD_DIR"
+        "--max-model-len" "$MAX_MODEL_LEN"
+    )
+
+    # Use the array expansion "${common_args_array[@]}"
+    # This correctly passes each element as a separate argument.
+    if [[ -n "$profile_dir" ]]; then
+        # Start server with profiling enabled
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
+            vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
+    else
+        # Start server without profiling
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
+            vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
+    fi
 
     # wait for 10 minutes...
     server_started=0
@@ -83,6 +100,7 @@ start_server() {
             sleep 10
         fi
     done
+
     if (( ! server_started )); then
         echo "server did not start within 10 minutes. Please check server log at $vllm_log".
         return 1
@@ -91,37 +109,20 @@ start_server() {
     fi
 }
 
-update_best_profile() {
-    local profile_dir=$1
-    local profile_index=$2
-    sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort))
-    selected_profile_file=
-    if [[ "$SYSTEM" == "TPU" ]]; then
-        selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
-    fi
-    if [[ "$SYSTEM" == "GPU" ]]; then
-        selected_profile_file="${sorted_paths[$profile_index]}"
-    fi
-    rm -f $PROFILE_PATH/*
-    cp $selected_profile_file $PROFILE_PATH
-}
-
 run_benchmark() {
     local max_num_seqs=$1
     local max_num_batched_tokens=$2
     local gpu_memory_utilization=$3
     echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
     local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
-    local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}"
     echo "vllm_log: $vllm_log"
     echo
     rm -f $vllm_log
-    mkdir -p $profile_dir
-    pkill -f vllm
-    local profile_index=0
+    pkill -if vllm
 
     echo "starting server..."
-    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir
+    # Call start_server without a profile_dir to avoid profiling overhead
+    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log ""
     result=$?
     if [[ "$result" -eq 1 ]]; then
         echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
@@ -135,7 +136,8 @@ run_benchmark() {
     # get a basic qps by using request-rate inf
     bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
     prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
-adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    # --profile flag is removed from this call
     vllm bench serve \
         --backend vllm \
         --model $MODEL  \
@@ -149,8 +151,7 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
         --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
         --num-prompts 1000 \
         --random-prefix-len $prefix_len \
-        --port 8004 \
-        --profile &> "$bm_log"
+        --port 8004 &> "$bm_log"
     throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
     e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
     goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
@@ -164,7 +165,6 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
     # start from request-rate as int(throughput) + 1
         request_rate=$((${throughput%.*} + 1))
         while ((request_rate > 0)); do
-            profile_index=$((profile_index+1))
             # clear prefix cache
             curl -X POST http://0.0.0.0:8004/reset_prefix_cache
             sleep 5
@@ -202,12 +202,7 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
             best_max_num_seqs=$max_num_seqs
             best_num_batched_tokens=$max_num_batched_tokens
             best_goodput=$goodput
-            if [[ "$SYSTEM" == "TPU" ]]; then
-                update_best_profile "$profile_dir/plugins/profile" $profile_index
-            fi
-            if [[ "$SYSTEM" == "GPU" ]]; then
-                update_best_profile "$profile_dir" $profile_index
-            fi
+            best_request_rate=$request_rate
         fi
     else
         echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
@@ -216,7 +211,7 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
 
     echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
 
-    pkill vllm
+    pkill -if vllm
     sleep 10
     printf '=%.0s' $(seq 1 20)
     return 0
@@ -229,7 +224,8 @@ read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
 gpu_memory_utilization=0.98
 find_gpu_memory_utilization=0
 while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
-    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
+    # Pass empty string for profile_dir argument
+    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" ""
     result=$?
     if [[ "$result" -eq 0 ]]; then
         find_gpu_memory_utilization=1
@@ -252,5 +248,45 @@ for num_seqs in "${num_seqs_list[@]}"; do
     done
 done
 echo "finish permutations"
+
+# =================================================================================
+# FINAL PROFILING RUN FOR THE BEST CONFIGURATION
+# =================================================================================
+if (( $(echo "$best_throughput > 0" | bc -l) )); then
+    echo
+    echo "Benchmark tuning finished. Now running profiling on the best configuration found..."
+    echo "Best config: max_num_seqs: $best_max_num_seqs, max_num_batched_tokens: $best_num_batched_tokens, throughput: $best_throughput"
+    echo
+
+    vllm_log="$LOG_FOLDER/vllm_log_BEST_PROFILE.txt"
+    bm_log="$LOG_FOLDER/bm_log_BEST_PROFILE.txt"
+
+    # Start server with the best params and profiling ENABLED
+    echo "Starting server for profiling..."
+    start_server $gpu_memory_utilization $best_max_num_seqs $best_num_batched_tokens "$vllm_log" "$PROFILE_PATH"
+
+    # Run benchmark with the best params and the --profile flag
+    echo "Running benchmark with profiling..."
+    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
+    adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    vllm bench serve \
+        --backend vllm \
+        --model $MODEL \
+        --dataset-name random \
+        --random-input-len $adjusted_input_len \
+        --random-output-len $OUTPUT_LEN \
+        --ignore-eos \
+        --disable-tqdm \
+        --request-rate $best_request_rate \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --num-prompts 100 \
+        --random-prefix-len $prefix_len \
+        --port 8004 \
+        --profile &> "$bm_log"
+else
+    echo "No configuration met the latency requirements. Skipping final profiling run."
+fi
+pkill -if vllm
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"

benchmarks/benchmark_serving.py

@@ -5,8 +5,7 @@ r"""Benchmark online serving throughput.
 On the server side, run one of the following commands:
     vLLM OpenAI API server
     vllm serve <your_model> \
-        --swap-space 16 \
-        --disable-log-requests
+        --swap-space 16
 
 On the client side, run:
     python benchmarks/benchmark_serving.py \
@@ -413,6 +412,10 @@ async def benchmark(
 
     print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
     print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
+    if max_concurrency is not None:
+        print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency))
+    if request_rate != float("inf"):
+        print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate))
     print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
     print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
     print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))

benchmarks/benchmark_serving_structured_output.py

@@ -4,7 +4,7 @@ r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
     (vLLM OpenAI API server)
-    vllm serve <your_model> --disable-log-requests
+    vllm serve <your_model>
 
 On the client side, run:
     python benchmarks/benchmark_serving_structured_output.py \
@@ -555,6 +555,10 @@ async def benchmark(
 
     print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
     print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
+    if max_concurrency is not None:
+        print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency))
+    if request_rate != float("inf"):
+        print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate))
     print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
     print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
     print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))

benchmarks/kernels/benchmark_moe.py

@@ -22,6 +22,13 @@ from vllm.utils import FlexibleArgumentParser
 FP8_DTYPE = current_platform.fp8_dtype()
 
 
+def ensure_divisibility(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator."""
+    assert numerator % denominator == 0, (
+        "intermediate_size {} is not divisible by tp {}.".format(numerator, denominator)
+    )
+
+
 class BenchmarkConfig(TypedDict):
     BLOCK_SIZE_M: int
     BLOCK_SIZE_N: int
@@ -603,7 +610,7 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-
+    ensure_divisibility(intermediate_size, args.tp_size)
     hidden_size = config.hidden_size
     dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"

benchmarks/kernels/benchmark_per_token_group_quant.py

@@ -0,0 +1,159 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import math
+from contextlib import contextmanager
+from typing import Callable
+from unittest.mock import patch
+
+import torch
+
+from vllm.model_executor.layers.quantization.utils import fp8_utils, int8_utils
+from vllm.platforms import current_platform
+
+
+@contextmanager
+def _triton_mode():
+    """Temporarily force the Triton fallback path"""
+    with patch("vllm.platforms.current_platform.is_cuda", return_value=False):
+        yield
+
+
+def _time_cuda(
+    fn: Callable[[], tuple[torch.Tensor, torch.Tensor]],
+    warmup_iters: int,
+    bench_iters: int,
+) -> float:
+    # warmup
+    for _ in range(warmup_iters):
+        fn()
+    torch.cuda.synchronize()
+
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+
+    start.record()
+    for _ in range(bench_iters):
+        fn()
+    end.record()
+    torch.cuda.synchronize()
+
+    return start.elapsed_time(end) / bench_iters  # ms/iter
+
+
+def _run_single(
+    shape: tuple[int, int],
+    group_size: int,
+    dtype: str,
+    *,
+    column_major: bool = False,
+    scale_ue8m0: bool = False,
+    warmup_iters: int,
+    bench_iters: int,
+) -> None:
+    num_tokens, hidden_dim = shape
+
+    device = torch.device("cuda")
+    torch.manual_seed(42)
+    x = torch.randn(num_tokens, hidden_dim, device=device, dtype=torch.bfloat16) * 8
+
+    if dtype == "fp8":
+
+        def cuda_impl():
+            return fp8_utils.per_token_group_quant_fp8(
+                x,
+                group_size,
+                column_major_scales=column_major,
+                use_ue8m0=scale_ue8m0,
+            )
+
+        def triton_impl():
+            with _triton_mode():
+                return fp8_utils.per_token_group_quant_fp8(
+                    x,
+                    group_size,
+                    column_major_scales=column_major,
+                    use_ue8m0=scale_ue8m0,
+                )
+    elif dtype == "int8":
+
+        def cuda_impl():
+            return int8_utils.per_token_group_quant_int8(x, group_size)
+
+        def triton_impl():
+            with _triton_mode():
+                return int8_utils.per_token_group_quant_int8(x, group_size)
+    else:
+        raise ValueError("dtype must be 'fp8' or 'int8'")
+
+    cuda_ms = _time_cuda(cuda_impl, warmup_iters, bench_iters)
+    triton_ms = _time_cuda(triton_impl, warmup_iters, bench_iters)
+
+    speedup = triton_ms / cuda_ms if cuda_ms else math.inf
+
+    cfg_desc = (
+        f"shape={shape}  gs={group_size:<3}  col_major={column_major:<5}  "
+        f"ue8m0={scale_ue8m0:<5}  dtype={dtype}"
+    )
+    print(
+        f"{cfg_desc:55} | CUDA {cuda_ms:7.3f} ms  | Triton {triton_ms:7.3f} ms  | "
+        f"speed-up ×{speedup:5.2f}"
+    )
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--warmup-iters", type=int, default=10)
+    parser.add_argument("--bench-iters", type=int, default=100)
+    parser.add_argument("--dtype", choices=["fp8", "int8", "both"], default="both")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    if not current_platform.is_cuda():
+        raise RuntimeError("CUDA device is required to run this benchmark.")
+
+    args = parse_args()
+    warmup_iters, bench_iters = args.warmup_iters, args.bench_iters
+
+    shapes = [(32, 128), (64, 256), (16, 512)]
+    group_sizes = [64, 128]
+
+    dtypes = ["fp8", "int8"] if args.dtype == "both" else [args.dtype]
+
+    header = (
+        "Configuration".ljust(55)
+        + " | "
+        + "CUDA (ms)".center(12)
+        + " | "
+        + "Triton (ms)".center(13)
+        + " | "
+        + "Speed-up"
+    )
+    print(header)
+    print("-" * len(header))
+
+    for dtype in dtypes:
+        for shape in shapes:
+            for gs in group_sizes:
+                if dtype == "fp8":
+                    for col_major in (False, True):
+                        for ue8m0 in (False, True):
+                            _run_single(
+                                shape,
+                                gs,
+                                dtype,
+                                column_major=col_major,
+                                scale_ue8m0=ue8m0,
+                                warmup_iters=warmup_iters,
+                                bench_iters=bench_iters,
+                            )
+                else:  # INT8 has no col-major / ue8m0 switches
+                    _run_single(
+                        shape,
+                        gs,
+                        dtype,
+                        warmup_iters=warmup_iters,
+                        bench_iters=bench_iters,
+                    )

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from __future__ import annotations
+
+import random
+import time
+
+import torch
+from tabulate import tabulate
+
+from vllm import _custom_ops as ops
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.utils import (
+    STR_DTYPE_TO_TORCH_DTYPE,
+    FlexibleArgumentParser,
+    create_kv_caches_with_random_flash,
+)
+
+logger = init_logger(__name__)
+
+
+@torch.inference_mode()
+def run_benchmark(
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    num_blocks: int,
+    dtype: torch.dtype,
+    kv_cache_dtype: str,
+    kv_cache_layout: str,
+    num_iters: int,
+    device: str = "cuda",
+) -> float:
+    """Return latency (seconds) for given num_tokens."""
+
+    if kv_cache_dtype == "fp8" and head_size % 16:
+        raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
+
+    current_platform.seed_everything(42)
+    torch.set_default_device(device)
+
+    # create random key / value tensors [T, H, D].
+    key = torch.randn(num_tokens, num_heads, head_size, dtype=dtype, device=device)
+    value = torch.randn_like(key)
+
+    # prepare the slot mapping.
+    # each token is assigned a unique slot in the KV-cache.
+    num_slots = block_size * num_blocks
+    if num_tokens > num_slots:
+        raise ValueError("num_tokens cannot exceed the total number of cache slots")
+    slot_mapping_lst = random.sample(range(num_slots), num_tokens)
+    slot_mapping = torch.tensor(slot_mapping_lst, dtype=torch.long, device=device)
+
+    key_caches, value_caches = create_kv_caches_with_random_flash(
+        num_blocks,
+        block_size,
+        1,  # num_layers
+        num_heads,
+        head_size,
+        kv_cache_dtype,
+        dtype,
+        device=device,
+        cache_layout=kv_cache_layout,
+    )
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # compute per-kernel scaling factors for fp8 conversion (if used).
+    k_scale = (key.amax() / 64.0).to(torch.float32)
+    v_scale = (value.amax() / 64.0).to(torch.float32)
+
+    def run_cuda_benchmark(n_iters: int) -> float:
+        nonlocal key, value, key_cache, value_cache, slot_mapping
+        torch.cuda.synchronize()
+        start = time.perf_counter()
+        for _ in range(n_iters):
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+        torch.cuda.synchronize()
+        end = time.perf_counter()
+        return (end - start) / n_iters
+
+    # warm-up
+    run_cuda_benchmark(3)
+
+    lat = run_cuda_benchmark(num_iters)
+
+    # free tensors to mitigate OOM when sweeping
+    del key, value, key_cache, value_cache, slot_mapping
+    torch.cuda.empty_cache()
+
+    return lat
+
+
+def main(args):
+    rows = []
+    for layout in ["NHD", "HND"]:
+        for exp in range(1, 17):
+            n_tok = 2**exp
+            lat = run_benchmark(
+                num_tokens=n_tok,
+                num_heads=args.num_heads,
+                head_size=args.head_size,
+                block_size=args.block_size,
+                num_blocks=args.num_blocks,
+                dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+                kv_cache_dtype=args.kv_cache_dtype,
+                kv_cache_layout=layout,
+                num_iters=args.iters,
+                device="cuda",
+            )
+            rows.append([n_tok, layout, f"{lat * 1e6:.3f}"])
+
+    print(tabulate(rows, headers=["num_tokens", "layout", "latency (µs)"]))
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+
+    parser.add_argument("--num-heads", type=int, default=128)
+    parser.add_argument(
+        "--head-size",
+        type=int,
+        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+        default=128,
+    )
+    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
+    parser.add_argument("--num-blocks", type=int, default=128 * 512)
+
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["half", "bfloat16", "float"],
+        default="bfloat16",
+    )
+
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=["auto", "fp8"],
+        default="auto",
+    )
+
+    parser.add_argument("--iters", type=int, default=100)
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -41,7 +41,6 @@ def benchmark_decode(
     device = "cuda"
     torch.manual_seed(0)
 
-    # Currently only HEAD_GRP_SIZE == 8 is supported
     HEAD_GRP_SIZE = 8
     MAX_SEQ_LEN = max_seq_len
 

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -0,0 +1,250 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import csv
+import os
+import random
+from datetime import datetime
+
+import flashinfer
+import torch
+
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+
+# KV Cache Layout for TRT-LLM
+# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
+
+
+def to_float8(x, dtype=torch.float8_e4m3fn):
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax * 0.1
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype), scale.float().reciprocal()
+
+
+@torch.no_grad()
+def benchmark_prefill(
+    num_seqs,
+    max_seq_len,
+    page_size=16,
+    dtype=torch.bfloat16,
+    kv_layout="HND",
+    num_kv_heads=8,
+    kv_cache_dtype="auto",
+    head_dim=128,
+    warmup=10,
+    trials=20,
+):
+    torch.set_default_device("cuda")
+    torch.manual_seed(0)
+
+    HEAD_GRP_SIZE = 8
+    MAX_SEQ_LEN = max_seq_len
+
+    # large number to reduce kv_cache reuse
+    NUM_BLOCKS = int(256000 / page_size)
+
+    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8)
+
+    num_qo_heads = num_kv_heads * HEAD_GRP_SIZE
+    sm_scale = float(1.0 / (head_dim**0.5))
+
+    q_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    q_lens[-1] = MAX_SEQ_LEN
+    max_q_len = max(q_lens)
+    q_indptr = torch.cat(
+        [
+            torch.tensor([0], dtype=torch.int32),
+            torch.cumsum(
+                torch.tensor(q_lens, dtype=torch.int32), dim=0, dtype=torch.int32
+            ),
+        ]
+    )
+    q = torch.randn(sum(q_lens), num_qo_heads, head_dim, dtype=dtype)
+
+    kv_lens = [random.randint(0, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    kv_lens[-1] = MAX_SEQ_LEN
+
+    seq_lens = [q_len + kv_len for q_len, kv_len in zip(q_lens, kv_lens)]
+    max_seq_len = max(seq_lens)
+    seq_lens_tensor = torch.tensor(seq_lens, dtype=torch.int32)
+
+    max_num_blocks_per_seq = (max_seq_len + page_size - 1) // page_size
+    block_tables = torch.randint(
+        0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
+    )
+
+    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
+    kv_cache = torch.randn(size=kv_cache_shape, dtype=dtype)
+    k_scale = v_scale = 1.0
+
+    if kv_cache_dtype.startswith("fp8"):
+        kv_cache, _ = to_float8(kv_cache)
+
+    output_trtllm = torch.empty(q.shape, dtype=dtype)
+
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = seq_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + page_size - 1) // page_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % page_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = page_size
+        kv_last_page_lens.append(kv_last_page_len)
+
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    output_baseline = torch.empty(q.shape, dtype=dtype)
+
+    wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+        workspace_buffer, kv_layout
+    )
+    wrapper.plan(
+        q_indptr,
+        kv_indptr,
+        kv_indices,
+        kv_last_page_lens,
+        num_qo_heads,
+        num_kv_heads,
+        head_dim,
+        page_size,
+        causal=True,
+        sm_scale=sm_scale,
+        q_data_type=dtype,
+        kv_data_type=kv_cache.dtype,
+    )
+
+    def time_fn(fn, warmup=10, trials=20):
+        torch.cuda.synchronize()
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        times = []
+        for i in range(warmup):
+            fn()
+        for i in range(trials):
+            start.record()
+            fn()
+            end.record()
+            torch.cuda.synchronize()
+            times.append(start.elapsed_time(end))  # ms
+        return sum(times) / len(times), torch.std(torch.tensor(times))
+
+    def baseline_prefill():
+        return wrapper.run(
+            q, kv_cache, k_scale=k_scale, v_scale=v_scale, out=output_baseline
+        )
+
+    def trt_prefill():
+        return flashinfer.prefill.trtllm_batch_context_with_kv_cache(
+            query=q,
+            kv_cache=kv_cache,
+            workspace_buffer=workspace_buffer,
+            block_tables=block_tables,
+            seq_lens=seq_lens_tensor,
+            max_q_len=max_q_len,
+            max_kv_len=max_seq_len,
+            bmm1_scale=k_scale * sm_scale,
+            bmm2_scale=v_scale,
+            batch_size=num_seqs,
+            cum_seq_lens_q=q_indptr,
+            cum_seq_lens_kv=kv_indptr,
+            out=output_trtllm,
+        )
+
+    trt_mean, trt_std = time_fn(trt_prefill)
+    baseline_mean, baseline_std = time_fn(baseline_prefill)
+
+    # Calculate percentage speedup (positive means TRT is faster)
+    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
+
+    print(
+        f"\t{num_seqs}\t{max_seq_len}\t{trt_mean:.5f}\t{trt_std.item():.5f}"
+        f"\t{baseline_mean:.5f}\t{baseline_std.item():.5f}\t{speedup_percent:.5f}"
+    )
+
+    # Return results for CSV writing
+    return {
+        "num_seqs": num_seqs,
+        "trt_mean": trt_mean,
+        "trt_std": trt_std.item(),
+        "baseline_mean": baseline_mean,
+        "baseline_std": baseline_std.item(),
+        "speedup_percent": speedup_percent,
+        "q_dtype": str(dtype),
+        "kv_cache_dtype": kv_cache_dtype,
+        "page_size": page_size,
+        "num_kv_heads": num_kv_heads,
+        "head_dim": head_dim,
+        "max_seq_len": max_seq_len,
+    }
+
+
+def write_results_to_csv(results, filename=None):
+    """Write benchmark results to CSV file."""
+    if filename is None:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"
+
+    fieldnames = [
+        "num_seqs",
+        "trt_mean",
+        "trt_std",
+        "baseline_mean",
+        "baseline_std",
+        "speedup_percent",
+        "q_dtype",
+        "kv_cache_dtype",
+        "page_size",
+        "num_kv_heads",
+        "head_dim",
+        "max_seq_len",
+    ]
+
+    file_exists = os.path.exists(filename)
+
+    with open(filename, "a", newline="") as csvfile:
+        writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
+
+        if not file_exists:
+            writer.writeheader()
+
+        for result in results:
+            writer.writerow(result)
+
+    print(f"Results written to {filename}")
+
+
+if __name__ == "__main__":
+    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
+    max_seq_lens = [1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
+    all_results = []
+
+    print(
+        "Running benchmark for q_dtype = bfloat16, kv_cache_dtype: bfloat16, "
+        "output_dtype: bfloat16"
+    )
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\t"
+        "baseline_std\tspeedup_percent"
+    )
+    for max_seq_len in max_seq_lens:
+        for bs in num_seqs:
+            result = benchmark_prefill(
+                bs,
+                max_seq_len,
+                dtype=torch.bfloat16,
+                kv_cache_dtype="auto",
+            )
+            all_results.append(result)
+
+    # Write all results to CSV
+    write_results_to_csv(all_results)

benchmarks/kernels/deepgemm/README.md

@@ -8,7 +8,7 @@ Currently this just includes dense GEMMs and only works on Hopper GPUs.
 
 You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory:
 
-```
+```bash
 git clone --recursive https://github.com/deepseek-ai/DeepGEMM
 cd DeepGEMM
 python setup.py install
@@ -17,7 +17,7 @@ uv pip install -e .
 
 ## Usage
 
-```
+```console
 python benchmark_fp8_block_dense_gemm.py
 INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda.
 ===== STARTING FP8 GEMM BENCHMARK =====

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -4,49 +4,16 @@
 # ruff: noqa: E501
 import time
 
-# Import DeepGEMM functions
-import deep_gemm
 import torch
-from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor
 
-# Import vLLM functions
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    get_col_major_tma_aligned_tensor,
     per_token_group_quant_fp8,
     w8a8_block_fp8_matmul,
 )
 from vllm.triton_utils import triton
-
-
-# Copied from
-# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9
-def per_token_cast_to_fp8(
-        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """Convert tensor to FP8 format with per-token scaling."""
-    assert x.dim() == 2 and x.size(1) % 128 == 0
-    m, n = x.shape
-    x_view = x.view(m, -1, 128)
-    x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4)
-    return (x_view * (448.0 / x_amax.unsqueeze(2))).to(
-        torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1)
-
-
-# Copied from
-# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17
-def per_block_cast_to_fp8(
-        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """Convert tensor to FP8 format with per-block scaling."""
-    assert x.dim() == 2
-    m, n = x.shape
-    x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128),
-                           dtype=x.dtype,
-                           device=x.device)
-    x_padded[:m, :n] = x
-    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
-    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
-    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
-    return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (
-        x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+from vllm.utils.deep_gemm import calc_diff, fp8_gemm_nt, per_block_cast_to_fp8
 
 
 def benchmark_shape(m: int,
@@ -69,14 +36,14 @@ def benchmark_shape(m: int,
 
     # Pre-quantize B for all implementations
     # (weights can be pre-quantized offline)
-    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B)
-    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B)
+    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B, [128, 128], use_ue8m0=True)
+    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B, [128, 128], use_ue8m0=True)
 
     # Block size configuration
     block_size = [128, 128]
 
     # Pre-quantize A for all implementations
-    A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
+    A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8(A, block_size[1])
     A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
     C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
     A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
@@ -85,7 +52,7 @@ def benchmark_shape(m: int,
 
     # === DeepGEMM Implementation ===
     def deepgemm_gemm():
-        deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm),
+        fp8_gemm_nt((A_deepgemm, A_scale_deepgemm),
                                        (B_deepgemm, B_scale_deepgemm),
                                        C_deepgemm)
         return C_deepgemm

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 1c2624e53c078854e0637ee566c72fe2107e75f4
+          GIT_TAG 6dbc6e011a3ebe9349eeb74578940dd7095436ba
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -467,6 +467,12 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   if (GPU_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.
     add_dependencies(${GPU_MOD_NAME} hipify${GPU_MOD_NAME})
+    # Make sure we include the hipified versions of the headers, and avoid conflicts with the ones in the original source folder
+    target_include_directories(${GPU_MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
+      ${GPU_INCLUDE_DIRECTORIES})
+  else()
+    target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
+      ${GPU_INCLUDE_DIRECTORIES})
   endif()
 
   if (GPU_ARCHITECTURES)
@@ -482,8 +488,6 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   target_compile_definitions(${GPU_MOD_NAME} PRIVATE
     "-DTORCH_EXTENSION_NAME=${GPU_MOD_NAME}")
 
-  target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
-    ${GPU_INCLUDE_DIRECTORIES})
 
   target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${GPU_LIBRARIES})
 

csrc/cache_kernels.cu

@@ -5,6 +5,7 @@
 #include "cuda_utils.h"
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
+#include "quantization/vectorization_utils.cuh"
 
 #ifdef USE_ROCM
   #include "quantization/fp8/amd/quant_utils.cuh"
@@ -261,14 +262,26 @@ __global__ void reshape_and_cache_kernel(
   }
 }
 
+// Used by vectorization_utils to copy/convert one element
+template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
+struct CopyWithScaleOp {
+  float scale;
+
+  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      dst = static_cast<OutT>(src);
+    } else {
+      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
+    }
+  }
+};
+
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_flash_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
     const scalar_t* __restrict__ value,  // [num_tokens, num_heads, head_size]
-    cache_t* __restrict__ key_cache,     // [num_blocks, block_size, num_heads,
-                                         // head_size]
-    cache_t* __restrict__ value_cache,   // [num_blocks, block_size, num_heads,
-                                         // head_size]
+    cache_t* __restrict__ key_cache,     // NHD or HND, shape see comments below
+    cache_t* __restrict__ value_cache,   // same above
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int64_t block_stride, const int64_t page_stride,
     const int64_t head_stride, const int64_t key_stride,
@@ -282,25 +295,58 @@ __global__ void reshape_and_cache_flash_kernel(
   }
   const int64_t block_idx = slot_idx / block_size;
   const int64_t block_offset = slot_idx % block_size;
-  const int n = num_heads * head_size;
-  for (int i = threadIdx.x; i < n; i += blockDim.x) {
-    const int64_t src_key_idx = token_idx * key_stride + i;
-    const int64_t src_value_idx = token_idx * value_stride + i;
-    const int head_idx = i / head_size;
-    const int head_offset = i % head_size;
-    const int64_t tgt_key_value_idx = block_idx * block_stride +
-                                      block_offset * page_stride +
-                                      head_idx * head_stride + head_offset;
-    scalar_t tgt_key = key[src_key_idx];
-    scalar_t tgt_value = value[src_value_idx];
-    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-      key_cache[tgt_key_value_idx] = tgt_key;
-      value_cache[tgt_key_value_idx] = tgt_value;
-    } else {
-      key_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
-      value_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
+  const int n_elems = num_heads * head_size;
+
+  // pointers to the beginning of the source row for this token.
+  const scalar_t* __restrict__ key_src = key + token_idx * key_stride;
+  const scalar_t* __restrict__ value_src = value + token_idx * value_stride;
+
+  // find the start position inside the kv-cache for this token.
+  cache_t* __restrict__ key_dst =
+      key_cache + block_idx * block_stride + block_offset * page_stride;
+  cache_t* __restrict__ value_dst =
+      value_cache + block_idx * block_stride + block_offset * page_stride;
+
+  // this is true for the NHD layout where `head_stride == head_size`
+  const bool is_contiguous_heads = (head_stride == head_size);
+
+  float k_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *k_scale;
+  float v_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *v_scale;
+  constexpr int VEC_SIZE = (sizeof(scalar_t) == 2) ? 8 : 4;
+  CopyWithScaleOp<cache_t, scalar_t, kv_dt> k_op{k_scale_val};
+  CopyWithScaleOp<cache_t, scalar_t, kv_dt> v_op{v_scale_val};
+  if (is_contiguous_heads) {
+    // NHD layout
+    // kv cache: [num_blocks, block_size, num_heads, head_size]
+    vectorize_with_alignment<VEC_SIZE>(key_src, key_dst, n_elems, threadIdx.x,
+                                       blockDim.x, k_op);
+
+    vectorize_with_alignment<VEC_SIZE>(value_src, value_dst, n_elems,
+                                       threadIdx.x, blockDim.x, v_op);
+
+  } else {
+    // HND layout: heads are strided, but each head_size segment is contiguous
+    // kv cache: [num_blocks, num_heads, block_size, head_size]
+    const int lane = threadIdx.x & 31;     // 0..31 within warp
+    const int warp_id = threadIdx.x >> 5;  // warp index within block
+    const int warps_per_block = blockDim.x >> 5;
+
+    for (int head = warp_id; head < num_heads; head += warps_per_block) {
+      const scalar_t* __restrict__ k_src_h = key_src + head * head_size;
+      const scalar_t* __restrict__ v_src_h = value_src + head * head_size;
+
+      cache_t* __restrict__ k_dst_h =
+          key_dst + static_cast<int64_t>(head) * head_stride;
+      cache_t* __restrict__ v_dst_h =
+          value_dst + static_cast<int64_t>(head) * head_stride;
+
+      // within each head, let the 32 threads of the warp perform the vector
+      // copy
+      vectorize_with_alignment<VEC_SIZE>(k_src_h, k_dst_h, head_size, lane, 32,
+                                         k_op);
+
+      vectorize_with_alignment<VEC_SIZE>(v_src_h, v_dst_h, head_size, lane, 32,
+                                         v_op);
     }
   }
 }

csrc/cpu/quant.cpp

@@ -16,12 +16,14 @@ struct KernelVecType<float> {
   using cvt_vec_type = vec_op::FP32Vec16;
 };
 
+#if !defined(__aarch64__) || defined(ARM_BF16_SUPPORT)
 template <>
 struct KernelVecType<c10::BFloat16> {
   using load_vec_type = vec_op::BF16Vec16;
   using azp_adj_load_vec_type = vec_op::INT32Vec16;
   using cvt_vec_type = vec_op::FP32Vec16;
 };
+#endif
 
 template <>
 struct KernelVecType<c10::Half> {

csrc/moe/topk_softmax_kernels.cu

@@ -24,9 +24,12 @@
 #ifndef USE_ROCM
     #include <cub/util_type.cuh>
     #include <cub/cub.cuh>
+    #include <cuda/std/functional>
+    using AddOp = cuda::std::plus<float>;
 #else
     #include <hipcub/util_type.hpp>
     #include <hipcub/hipcub.hpp>
+    using AddOp = cub::Sum; 
 #endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
@@ -62,7 +65,6 @@ __launch_bounds__(TPB) __global__
 
     const int thread_row_offset = blockIdx.x * num_cols;
 
-    cub::Sum sum;
     float threadData(-FLT_MAX);
 
     // Don't touch finished rows.
@@ -92,7 +94,7 @@ __launch_bounds__(TPB) __global__
         threadData += exp((static_cast<float>(input[idx]) - float_max));
     }
 
-    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, sum);
+    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, AddOp());
 
     if (threadIdx.x == 0)
     {

csrc/quantization/cutlass_w8a8/Epilogues.md

@@ -86,6 +86,7 @@ D = s_a s_b \widehat A \widehat B
 ```
 
 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).
 
@@ -135,7 +136,7 @@ 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.
+    - 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).
 - `azp_with_adj` is the precomputed zero-point term ($` z_a J_a \widehat B `$), is per-channel (row-vector).
 - `bias` is the bias, is always per-channel (row-vector).
@@ -152,7 +153,7 @@ That means the zero-point term $` z_a J_a \widehat B `$ becomes an outer product
 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.
+    - 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).
 - `azp_adj` is the precomputed zero-point adjustment term ($` \mathbf 1 \widehat B `$), is per-channel (row-vector).
 - `azp` is the zero-point (`z_a`), is per-token (column-vector).

csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu

@@ -335,7 +335,7 @@ void run_fp4_blockwise_scaled_group_mm(
   TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
 }
 
-#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
 constexpr auto FLOAT4_E2M1X2 = at::ScalarType::Byte;
 constexpr auto SF_DTYPE = at::ScalarType::Float8_e4m3fn;
 #endif
@@ -356,7 +356,7 @@ void cutlass_fp4_group_mm(
     const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
     const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
     const torch::Tensor& expert_offsets, const torch::Tensor& sf_offsets) {
-#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
   // Input validation
   CHECK_INPUT(a, FLOAT4_E2M1X2, "a");
   CHECK_INPUT(b, FLOAT4_E2M1X2, "b");
@@ -398,7 +398,7 @@ void cutlass_fp4_group_mm(
   TORCH_CHECK_NOT_IMPLEMENTED(
       false,
       "No compiled cutlass_fp4_group_mm kernel, vLLM must "
-      "be compiled with ENABLE_NVFP4 for SM100+ and CUDA "
+      "be compiled with ENABLE_NVFP4_SM100 for SM100+ and CUDA "
       "12.8 or above.");
 #endif
 }

csrc/quantization/fp4/nvfp4_quant_entry.cu

@@ -16,14 +16,15 @@
 
 #include <torch/all.h>
 
-#if defined ENABLE_NVFP4 && ENABLE_NVFP4
-void scaled_fp4_quant_sm100a(torch::Tensor const& output,
+#if (defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100) || \
+    (defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120)
+void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
                              torch::Tensor const& input,
                              torch::Tensor const& output_sf,
                              torch::Tensor const& input_sf);
 #endif
 
-#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
 void scaled_fp4_experts_quant_sm100a(
     torch::Tensor& output, torch::Tensor& output_scale,
     torch::Tensor const& input, torch::Tensor const& input_global_scale,
@@ -33,8 +34,9 @@ void scaled_fp4_experts_quant_sm100a(
 
 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);
+#if (defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100) || \
+    (defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120)
+  return scaled_fp4_quant_sm1xxa(output, input, output_sf, input_sf);
 #endif
   TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled nvfp4 quantization kernel");
 }
@@ -44,7 +46,7 @@ void scaled_fp4_experts_quant(
     torch::Tensor const& input, torch::Tensor const& input_global_scale,
     torch::Tensor const& input_offset_by_experts,
     torch::Tensor const& output_scale_offset_by_experts) {
-#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
   return scaled_fp4_experts_quant_sm100a(
       output, output_scale, input, input_global_scale, input_offset_by_experts,
       output_scale_offset_by_experts);

csrc/quantization/fp4/nvfp4_quant_kernels.cu

@@ -332,7 +332,7 @@ template void invokeFP4Quantization(int m, int n, __nv_bfloat16 const* input,
                                     int multiProcessorCount,
                                     cudaStream_t stream);
 
-void scaled_fp4_quant_sm100a(torch::Tensor const& output,
+void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
                              torch::Tensor const& input,
                              torch::Tensor const& output_sf,
                              torch::Tensor const& input_sf) {

csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu

@@ -16,7 +16,7 @@
 
 #include <torch/all.h>
 
-#if defined ENABLE_NVFP4 && ENABLE_NVFP4
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
 void cutlass_scaled_fp4_mm_sm100a(torch::Tensor& D, torch::Tensor const& A,
                                   torch::Tensor const& B,
                                   torch::Tensor const& A_sf,
@@ -24,12 +24,22 @@ void cutlass_scaled_fp4_mm_sm100a(torch::Tensor& D, torch::Tensor const& A,
                                   torch::Tensor const& alpha);
 #endif
 
+#if defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120
+void cutlass_scaled_fp4_mm_sm120a(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
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
   return cutlass_scaled_fp4_mm_sm100a(D, A, B, A_sf, B_sf, alpha);
+#elif defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120
+  return cutlass_scaled_fp4_mm_sm120a(D, A, B, A_sf, B_sf, alpha);
 #endif
   TORCH_CHECK_NOT_IMPLEMENTED(false,
                               "No compiled nvfp4 mm kernel, vLLM should "

csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu

@@ -0,0 +1,285 @@
+/*
+ * 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"
+
+#include "core/math.hpp"
+
+using namespace cute;
+
+#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;
+
+struct sm120_fp4_config_M256 {
+  using ClusterShape = Shape<_1, _1, _1>;
+  using MmaTileShape = Shape<_128, _128, _128>;
+  using PerSmTileShape_MNK = Shape<_128, _128, _128>;
+};
+
+struct sm120_fp4_config_default {
+  using ClusterShape = Shape<_1, _1, _1>;
+  using MmaTileShape = Shape<_256, _128, _128>;
+  using PerSmTileShape_MNK = Shape<_256, _128, _128>;
+};
+
+template <typename Config, typename OutType>
+struct Fp4GemmSm120 {
+  using ElementA = cutlass::nv_float4_t<cutlass::float_e2m1_t>;
+  using LayoutATag = cutlass::layout::RowMajor;
+  static constexpr int AlignmentA = 32;
+
+  using ElementB = cutlass::nv_float4_t<cutlass::float_e2m1_t>;
+  using LayoutBTag = cutlass::layout::ColumnMajor;
+  static constexpr int AlignmentB = 32;
+
+  using ElementD = OutType;
+  using ElementC = OutType;
+  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;
+
+  using ElementAccumulator = float;
+  using ArchTag = cutlass::arch::Sm120;
+  using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp;
+
+  using MmaTileShape = typename Config::MmaTileShape;
+  using ClusterShape = typename Config::ClusterShape;
+  using PerSmTileShape_MNK = typename Config::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>;
+};
+
+template <typename Gemm>
+typename 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,
+                                           torch::Tensor const& alpha, int M,
+                                           int N, int K) {
+  using ElementA = typename Gemm::ElementA;
+  using ElementB = typename Gemm::ElementB;
+  using ElementD = typename Gemm::ElementD;
+  using ElementSFA = cutlass::float_ue4m3_t;
+  using ElementSFB = cutlass::float_ue4m3_t;
+  using ElementCompute = float;
+
+  using StrideA = typename Gemm::GemmKernel::StrideA;
+  using StrideB = typename Gemm::GemmKernel::StrideB;
+  using StrideC = typename Gemm::GemmKernel::StrideC;
+  using StrideD = typename Gemm::GemmKernel::StrideD;
+
+  using Sm1xxBlkScaledConfig =
+      typename Gemm::GemmKernel::CollectiveMainloop::Sm1xxBlkScaledConfig;
+
+  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 = Sm1xxBlkScaledConfig::tile_atom_to_shape_SFA(
+      cute::make_shape(M, N, K, 1));
+  auto layout_SFB = Sm1xxBlkScaledConfig::tile_atom_to_shape_SFB(
+      cute::make_shape(M, N, K, 1));
+
+  typename Gemm::Arguments arguments{
+      cutlass::gemm::GemmUniversalMode::kGemm,
+      {M, N, K, 1},
+      {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},
+      {{},
+       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 Gemm>
+void runGemm(at::Tensor& D, at::Tensor const& A, at::Tensor const& B,
+             at::Tensor const& A_sf, at::Tensor const& B_sf,
+             torch::Tensor const& alpha, int M, int N, int K,
+             cudaStream_t stream) {
+  Gemm gemm;
+
+  auto arguments = args_from_options<Gemm>(D, A, B, A_sf, B_sf, alpha, M, N, K);
+
+  size_t workspace_size = 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));
+}
+
+void cutlass_fp4_bf16_gemm_dispatch(torch::Tensor& D, torch::Tensor const& A,
+                                    torch::Tensor const& B,
+                                    torch::Tensor const& A_sf,
+                                    torch::Tensor const& B_sf,
+                                    torch::Tensor const& alpha, int m, int n,
+                                    int k, cudaStream_t stream) {
+  uint32_t const mp2 = std::max(static_cast<uint32_t>(16), next_pow_2(m));
+  if (mp2 <= 256) {
+    runGemm<Fp4GemmSm120<sm120_fp4_config_M256, cutlass::bfloat16_t>::Gemm>(
+        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
+  } else {
+    runGemm<Fp4GemmSm120<sm120_fp4_config_default, cutlass::bfloat16_t>::Gemm>(
+        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
+  }
+}
+
+void cutlass_fp4_f16_gemm_dispatch(torch::Tensor& D, torch::Tensor const& A,
+                                   torch::Tensor const& B,
+                                   torch::Tensor const& A_sf,
+                                   torch::Tensor const& B_sf,
+                                   torch::Tensor const& alpha, int m, int n,
+                                   int k, cudaStream_t stream) {
+  uint32_t const mp2 = std::max(static_cast<uint32_t>(16), next_pow_2(m));
+  if (mp2 <= 256) {
+    runGemm<Fp4GemmSm120<sm120_fp4_config_M256, cutlass::half_t>::Gemm>(
+        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
+  } else {
+    runGemm<Fp4GemmSm120<sm120_fp4_config_default, cutlass::half_t>::Gemm>(
+        D, A, B, A_sf, B_sf, alpha, m, n, k, stream);
+  }
+}
+
+void cutlass_scaled_fp4_mm_sm120a(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(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
+  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();
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream(A.get_device());
+
+  if (out_dtype == at::ScalarType::BFloat16) {
+    return cutlass_fp4_bf16_gemm_dispatch(D, A, B, A_sf, B_sf, alpha, m, n, k,
+                                          stream);
+  } else if (out_dtype == at::ScalarType::Half) {
+    return cutlass_fp4_f16_gemm_dispatch(D, A, B, A_sf, B_sf, alpha, m, n, k,
+                                         stream);
+  } else {
+    TORCH_CHECK(false, "Unsupported output data type of nvfp4 mm sm120 (",
+                out_dtype, ")");
+  }
+#else
+  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_SM120_SUPPORTED)
+}

csrc/quantization/fp8/common.cu

@@ -1,7 +1,8 @@
 #include "common.cuh"
 #include "dispatch_utils.h"
-
+#include "../vectorization_utils.cuh"
 #include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/Exceptions.h>
 
 #ifndef USE_ROCM
   #include <cub/cub.cuh>
@@ -12,74 +13,127 @@
 namespace vllm {
 
 template <typename scalar_t, typename fp8_type>
-__global__ void scaled_fp8_quant_kernel(fp8_type* __restrict__ out,
-                                        const scalar_t* __restrict__ input,
-                                        const float* __restrict__ scale,
-                                        int64_t num_elems) {
-  int tid = blockDim.x * blockIdx.x + threadIdx.x;
+__global__ void scaled_fp8_quant_kernel_strided(
+    fp8_type* __restrict__ out, const scalar_t* __restrict__ input,
+    const float* __restrict__ scale, int hidden_size, int64_t in_row_stride,
+    int64_t out_row_stride) {
+  const int64_t token_idx = blockIdx.x;  // one token per block
+  const int tid = threadIdx.x;
 
-  // Invert the scale so that we can use multiplications to avoid expensive
-  // division.
-  const float inverted_scale = 1.0f / (*scale);
-  scaled_fp8_conversion_vec<scalar_t, true>(
-      out, input, inverted_scale, num_elems, tid, blockDim.x * gridDim.x);
+  const scalar_t* token_in = input + token_idx * in_row_stride;
+  fp8_type* token_out = out + token_idx * out_row_stride;
+
+  const float inv_scale = 1.0f / (*scale);
+
+  vectorize_with_alignment<16>(
+      token_in, token_out, hidden_size, tid, blockDim.x,
+      [=] __device__(fp8_type & dst, const scalar_t& src) {
+        dst = scaled_fp8_conversion<true, fp8_type>(static_cast<float>(src),
+                                                    inv_scale);
+      });
 }
 
 template <typename scalar_t, typename fp8_type>
-__global__ void dynamic_per_token_scaled_fp8_quant_kernel(
-    fp8_type* __restrict__ out, float* __restrict__ scale,
-    scalar_t const* __restrict__ input, float const* __restrict__ scale_ub,
-    const int hidden_size) {
-  int const tid = threadIdx.x;
-  int const token_idx = blockIdx.x;
+__global__ void segmented_max_reduction_strided(
+    float* __restrict__ scale, const scalar_t* __restrict__ input,
+    int hidden_size, int64_t in_row_stride, int64_t num_tokens) {
+  __shared__ float cache[256];
+  const int tid = threadIdx.x;
+  int64_t token_idx = blockIdx.x;
 
-  // Use int64 to avoid overflowing an int32 when calculating this offset
-  int64_t offset = static_cast<int64_t>(token_idx) * hidden_size;
-  scalar_t const* __restrict__ token_input = &input[offset];
-  fp8_type* __restrict__ token_output = &out[offset];
-
-  // For vectorization, token_input and token_output pointers need to be
-  // aligned at 32-byte and 16-byte addresses respectively.
-  bool const can_vectorize = hidden_size % 16 == 0;
-
-  float absmax_val = 0.0f;
-  if (can_vectorize) {
-    absmax_val = thread_max_vec(token_input, hidden_size, tid, blockDim.x);
-  } else {
-    for (int i = tid; i < hidden_size; i += blockDim.x) {
-      float const x = static_cast<float>(token_input[i]);
-      absmax_val = fmaxf(absmax_val, fabsf(x));
-    }
+  // one block per token. Guard in case gridDim.x > num_tokens.
+  if (token_idx >= num_tokens) {
+    return;
   }
 
+  const scalar_t* row_ptr = input + token_idx * in_row_stride;
+
+  // each thread scans elements of the row in a strided fashion.
+  float thread_max = 0.0f;
+  for (int e = tid; e < hidden_size; e += blockDim.x) {
+    float v = fabsf(static_cast<float>(row_ptr[e]));
+    thread_max = fmaxf(thread_max, v);
+  }
+
+  cache[tid] = thread_max;
+  __syncthreads();
+
+  // parallel reduction to find row max.
+  for (int offset = blockDim.x / 2; offset > 0; offset >>= 1) {
+    if (tid < offset) {
+      cache[tid] = fmaxf(cache[tid], cache[tid + offset]);
+    }
+    __syncthreads();
+  }
+
+  // thread 0 updates global scale (per-tensor) atomically.
+  if (tid == 0) {
+    atomicMaxFloat(scale, cache[0] / quant_type_max_v<fp8_type>);
+  }
+}
+
+template <typename scalar_t, typename fp8_type>
+__global__ void scaled_fp8_quant_kernel_strided_dynamic(
+    fp8_type* __restrict__ out, const scalar_t* __restrict__ input,
+    const float* __restrict__ scale, int hidden_size, int64_t in_row_stride,
+    int64_t out_row_stride) {
+  const int64_t token_idx = blockIdx.x;
+  const int tid = threadIdx.x;
+
+  const scalar_t* token_in = input + token_idx * in_row_stride;
+  fp8_type* token_out = out + token_idx * out_row_stride;
+
+  const float reciprocal_scale = 1.0f / (*scale);
+  vectorize_with_alignment<16>(
+      token_in, token_out, hidden_size, tid, blockDim.x,
+      [=] __device__(fp8_type & dst, const scalar_t& src) {
+        dst = scaled_fp8_conversion<true, fp8_type>(static_cast<float>(src),
+                                                    reciprocal_scale);
+      });
+}
+
+template <typename scalar_t, typename fp8_type>
+__global__ void dynamic_per_token_scaled_fp8_quant_kernel_strided(
+    fp8_type* __restrict__ out, float* __restrict__ scale,
+    const scalar_t* __restrict__ input, const float* __restrict__ scale_ub,
+    int hidden_size, int64_t in_row_stride, int64_t out_row_stride) {
+  const int64_t token_idx = blockIdx.x;
+  const int tid = threadIdx.x;
+
+  // Use int64 to avoid overflowing an int32 when calculating this offset
+  int64_t in_offset = static_cast<int64_t>(token_idx) * in_row_stride;
+  int64_t out_offset = static_cast<int64_t>(token_idx) * out_row_stride;
+  const scalar_t* token_in = input + in_offset;
+  fp8_type* token_out = out + out_offset;
+
+  // 1) per-token absmax
+  float absmax_val = 0.f;
+  vectorize_read_with_alignment<16>(
+      token_in, hidden_size, tid, blockDim.x, [&] __device__(scalar_t v) {
+        absmax_val = fmaxf(absmax_val, fabsf(static_cast<float>(v)));
+      });
+
   using BlockReduce = cub::BlockReduce<float, 256>;
-  __shared__ typename BlockReduce::TempStorage reduceStorage;
-  float const block_absmax_val_maybe =
-      BlockReduce(reduceStorage).Reduce(absmax_val, cub::Max{}, blockDim.x);
+  __shared__ typename BlockReduce::TempStorage tmp;
+  const float block_max =
+      BlockReduce(tmp).Reduce(absmax_val, cub::Max{}, blockDim.x);
+
   __shared__ float token_scale;
   if (tid == 0) {
-    if (scale_ub) {
-      token_scale = fminf(block_absmax_val_maybe, *scale_ub);
-    } else {
-      token_scale = block_absmax_val_maybe;
-    }
-    // token scale computation
+    token_scale = scale_ub ? fminf(block_max, *scale_ub) : block_max;
     token_scale = fmaxf(token_scale / quant_type_max_v<fp8_type>,
                         min_scaling_factor<fp8_type>::val());
     scale[token_idx] = token_scale;
   }
   __syncthreads();
 
-  // Note that we don't use inverted scales so we can match FBGemm impl.
-  if (can_vectorize) {
-    scaled_fp8_conversion_vec<scalar_t, false>(
-        token_output, token_input, token_scale, hidden_size, tid, blockDim.x);
-  } else {
-    for (int i = tid; i < hidden_size; i += blockDim.x) {
-      token_output[i] = scaled_fp8_conversion<false, fp8_type>(
-          static_cast<float>(token_input[i]), token_scale);
-    }
-  }
+  // 2) quantize
+  vectorize_with_alignment<16>(
+      token_in, token_out, hidden_size, tid, blockDim.x,
+      [=] __device__(fp8_type & dst, const scalar_t& src) {
+        dst = scaled_fp8_conversion<false, fp8_type>(static_cast<float>(src),
+                                                     token_scale);
+      });
 }
 
 }  // namespace vllm
@@ -88,23 +142,31 @@ void static_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
                              torch::Tensor const& input,  // [..., d]
                              torch::Tensor const& scale)  // [1]
 {
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
-  int const block_size = 256;
-  int const num_tokens = input.numel() / input.size(-1);
-  int const num_elems = input.numel();
-  dim3 const grid(num_tokens);
-  dim3 const block(block_size);
+  TORCH_CHECK(input.stride(-1) == 1,
+              "last dimension of input must be contiguous");
+  TORCH_CHECK(out.stride(-1) == 1,
+              "last dimension of output must be contiguous");
+
+  const int hidden_size = input.size(-1);
+  const int num_tokens = input.numel() / hidden_size;
+  const int block_size = 256;
+  dim3 grid(num_tokens);
+  dim3 block(block_size);
+
+  const int64_t in_row_stride = input.stride(-2);
+  const int64_t out_row_stride = out.stride(-2);
+
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "scaled_fp8_quant_kernel_scalar_type", [&] {
         VLLM_DISPATCH_FP8_TYPES(
             out.scalar_type(), "scaled_fp8_quant_kernel_fp8_type", [&] {
-              vllm::scaled_fp8_quant_kernel<scalar_t, fp8_t>
+              vllm::scaled_fp8_quant_kernel_strided<scalar_t, fp8_t>
                   <<<grid, block, 0, stream>>>(
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                      scale.data_ptr<float>(), num_elems);
+                      scale.data_ptr<float>(), hidden_size, in_row_stride,
+                      out_row_stride);
             });
       });
 }
@@ -113,27 +175,42 @@ void dynamic_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
                               torch::Tensor const& input,  // [..., d]
                               torch::Tensor& scale)        // [1]
 {
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
-  int const block_size = 256;
-  int const num_tokens = input.numel() / input.size(-1);
-  int const num_elems = input.numel();
-  dim3 const grid(num_tokens);
-  dim3 const block(block_size);
+  TORCH_CHECK(input.stride(-1) == 1,
+              "last dimension of input must be contiguous");
+  TORCH_CHECK(out.stride(-1) == 1,
+              "last dimension of output must be contiguous");
+
+  const int hidden_size = input.size(-1);
+  const int num_tokens = input.numel() / hidden_size;
+  const int block_size = 256;
+  dim3 grid(num_tokens);
+  dim3 block(block_size);
+
+  const int64_t in_row_stride = input.stride(-2);
+  const int64_t out_row_stride = out.stride(-2);
+
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  // scale tensor should be initialised to <=0 before reduction
+  AT_CUDA_CHECK(
+      cudaMemsetAsync(scale.data_ptr<float>(), 0, sizeof(float), stream));
+
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "scaled_fp8_quant_kernel_scalar_type", [&] {
         VLLM_DISPATCH_FP8_TYPES(
             out.scalar_type(), "scaled_fp8_quant_kernel_fp8_type", [&] {
-              vllm::segmented_max_reduction<scalar_t, fp8_t>
-                  <<<grid, block, 0, stream>>>(scale.data_ptr<float>(),
-                                               input.data_ptr<scalar_t>(),
-                                               num_elems);
-              vllm::scaled_fp8_quant_kernel<scalar_t, fp8_t>
+              vllm::segmented_max_reduction_strided<scalar_t, fp8_t>
+                  <<<grid, block, 0, stream>>>(
+                      scale.data_ptr<float>(), input.data_ptr<scalar_t>(),
+                      hidden_size, in_row_stride,
+                      static_cast<int64_t>(num_tokens));
+
+              vllm::scaled_fp8_quant_kernel_strided_dynamic<scalar_t, fp8_t>
                   <<<grid, block, 0, stream>>>(
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                      scale.data_ptr<float>(), num_elems);
+                      scale.data_ptr<float>(), hidden_size, in_row_stride,
+                      out_row_stride);
             });
       });
 }
@@ -142,14 +219,19 @@ void dynamic_per_token_scaled_fp8_quant(
     torch::Tensor& out,          // [..., d]
     torch::Tensor const& input,  // [..., d]
     torch::Tensor& scales, std::optional<at::Tensor> const& scale_ub) {
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(input.stride(-1) == 1,
+              "last dimension of input must be contiguous");
+  TORCH_CHECK(out.stride(-1) == 1,
+              "last dimension of output must be contiguous");
 
-  int const hidden_size = input.size(-1);
-  int const num_tokens = input.numel() / hidden_size;
-  int const block_size = 256;
-  dim3 const grid(num_tokens);
-  dim3 const block(std::min(hidden_size, block_size));
+  const int hidden_size = input.size(-1);
+  const int num_tokens = input.numel() / hidden_size;
+  const int block_size = 256;
+  dim3 grid(num_tokens);
+  dim3 block(std::min(hidden_size, block_size));
+
+  const int64_t in_row_stride = input.stride(-2);
+  const int64_t out_row_stride = out.stride(-2);
 
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
@@ -159,13 +241,12 @@ void dynamic_per_token_scaled_fp8_quant(
         VLLM_DISPATCH_FP8_TYPES(
             out.scalar_type(),
             "dynamic_per_token_scaled_fp8_quant_kernel_fp8_type", [&] {
-              vllm::dynamic_per_token_scaled_fp8_quant_kernel<scalar_t, fp8_t>
-                  <<<grid, block, 0, stream>>>(
-                      out.data_ptr<fp8_t>(), scales.data_ptr<float>(),
-                      input.data_ptr<scalar_t>(),
-                      scale_ub.has_value() ? scale_ub->data_ptr<float>()
-                                           : nullptr,
-                      hidden_size);
+              vllm::dynamic_per_token_scaled_fp8_quant_kernel_strided<
+                  scalar_t, fp8_t><<<grid, block, 0, stream>>>(
+                  out.data_ptr<fp8_t>(), scales.data_ptr<float>(),
+                  input.data_ptr<scalar_t>(),
+                  scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
+                  hidden_size, in_row_stride, out_row_stride);
             });
       });
 }

csrc/quantization/fp8/common.cuh

@@ -55,111 +55,4 @@ __device__ __forceinline__ fp8_type scaled_fp8_conversion(float const val,
 #endif
 }
 
-// Compute the absolute maximum m of the input tensor and store
-// m / float8_e4m3::max() in *scale. Each thread block performs a
-// reduction tree and the memory in scale is atomically updated.
-// So to get the right answer, *scale needs to be initialized to
-// a value <= 0.0 and we need to wait for all thread blocks to
-// finish before consuming *scale.
-template <typename scalar_t, typename fp8_type>
-__global__ void segmented_max_reduction(float* __restrict__ scale,
-                                        const scalar_t* __restrict__ input,
-                                        int64_t num_elems) {
-  __shared__ float cache[256];
-  int64_t i = blockDim.x * blockIdx.x + threadIdx.x;
-
-  // First store maximum for all values processes by
-  // the current thread in cache[threadIdx.x]
-  scalar_t tmp = 0.0;
-  while (i < num_elems) {
-    float x = static_cast<float>(input[i]);
-    tmp = fmaxf(tmp, fabsf(x));
-    i += blockDim.x * gridDim.x;
-  }
-  cache[threadIdx.x] = tmp;
-
-  __syncthreads();
-
-  // Now perform parallel reduction within the thread block
-  int ib = blockDim.x / 2;
-  while (ib != 0) {
-    if (threadIdx.x < ib && cache[threadIdx.x + ib] > cache[threadIdx.x]) {
-      cache[threadIdx.x] = cache[threadIdx.x + ib];
-    }
-    __syncthreads();
-    ib /= 2;
-  }
-  // Finally, since cache[0] contains the maximum for this thread block,
-  // atomically write the max to the target location
-  if (threadIdx.x == 0) {
-    atomicMaxFloat(scale, cache[0] / quant_type_max_v<fp8_type>);
-  }
-}
-
-template <typename scalar_t>
-__device__ float thread_max_vec(scalar_t const* __restrict__ input,
-                                int64_t const num_elems, int const tid,
-                                int const step) {
-  constexpr size_t VEC_SIZE = 16;
-  using scalarxN_t = vec_n_t<scalar_t, VEC_SIZE>;
-  // Vectorized input/output to better utilize memory bandwidth.
-  auto const* vectorized_in = reinterpret_cast<scalarxN_t const*>(input);
-
-  // num_elems / VEC_SIZE (which is 16)
-  int64_t const num_vec_elems = num_elems >> 4;
-  float absmax_val = 0.0f;
-
-#pragma unroll
-  for (int64_t i = tid; i < num_vec_elems; i += step) {
-    scalarxN_t in_vec = vectorized_in[i];
-#pragma unroll
-    for (int j = 0; j < VEC_SIZE; ++j) {
-      absmax_val = fmaxf(absmax_val, fabsf(in_vec.val[j]));
-    }
-  }
-
-  // Handle the remaining elements if num_elems is not divisible by VEC_SIZE
-  for (int64_t i = num_vec_elems * VEC_SIZE + tid; i < num_elems; i += step) {
-    absmax_val = fmaxf(absmax_val, fabsf(input[i]));
-  }
-
-  return absmax_val;
-}
-
-template <typename scalar_t, bool is_scale_inverted, typename fp8_type>
-__device__ void scaled_fp8_conversion_vec(fp8_type* __restrict__ out,
-                                          scalar_t const* __restrict__ input,
-                                          float const scale,
-                                          int64_t const num_elems,
-                                          int const tid, int const step) {
-  constexpr size_t VEC_SIZE = 16;
-  using scalarxN_t = vec_n_t<scalar_t, VEC_SIZE>;
-  using float8xN_t = q8_n_t<fp8_type, VEC_SIZE>;
-  // Vectorized input/output to better utilize memory bandwidth.
-  auto const* vectorized_in = reinterpret_cast<scalarxN_t const*>(input);
-  auto* vectorized_out = reinterpret_cast<float8xN_t*>(out);
-
-  // num_elems / VEC_SIZE (which is 16)
-  int64_t const num_vec_elems = num_elems >> 4;
-
-#pragma unroll
-  for (int64_t i = tid; i < num_vec_elems; i += step) {
-    scalarxN_t in_vec = vectorized_in[i];
-    float8xN_t out_vec;
-
-#pragma unroll
-    for (int j = 0; j < VEC_SIZE; ++j) {
-      out_vec.val[j] = scaled_fp8_conversion<is_scale_inverted, fp8_type>(
-          static_cast<float>(in_vec.val[j]), scale);
-    }
-    vectorized_out[i] = out_vec;
-  }
-
-  // Handle the remaining elements if num_elems is not divisible by VEC_SIZE
-  for (int64_t i = num_vec_elems * VEC_SIZE + tid; i < num_elems; i += step) {
-    out[i] = scaled_fp8_conversion<is_scale_inverted, fp8_type>(
-        static_cast<float>(input[i]), scale);
-  }
-}
-
 }  // namespace vllm

csrc/quantization/fp8/per_token_group_quant.cu

@@ -1,12 +1,10 @@
 #include <ATen/cuda/CUDAContext.h>
-#include <c10/util/Float8_e4m3fn.h>
 
 #include "../per_token_group_quant_8bit.h"
 
 #include <cmath>
 
-#include <cuda_fp16.h>
-#include <cuda_bf16.h>
+#include <cuda_fp8.h>
 
 #include <torch/all.h>
 
@@ -199,7 +197,7 @@ void per_token_group_quant_8bit(const torch::Tensor& input,
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "per_token_group_quant_8bit", ([&] {
         if (dst_type == at::ScalarType::Float8_e4m3fn) {
-          LAUNCH_KERNEL(scalar_t, c10::Float8_e4m3fn);
+          LAUNCH_KERNEL(scalar_t, __nv_fp8_e4m3);
         } else if (dst_type == at::ScalarType::Char) {
           LAUNCH_KERNEL(scalar_t, int8_t);
         }

docker/Dockerfile

@@ -1,4 +1,3 @@
-
 # The vLLM Dockerfile is used to construct vLLM image that can be directly used
 # to run the OpenAI compatible server.
 
@@ -16,6 +15,7 @@ ARG PYTHON_VERSION=3.12
 # Example:
 # docker build --build-arg BUILD_BASE_IMAGE=registry.acme.org/mirror/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
 ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
+# TODO: Restore to base image after FlashInfer AOT wheel fixed
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
 # By parameterizing the Deadsnakes repository URL, we allow third-party to use
@@ -119,6 +119,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Reference: https://github.com/astral-sh/uv/pull/1694
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
 
 # 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
@@ -164,9 +166,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # see https://github.com/pytorch/pytorch/pull/123243
 ARG torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0 10.0 12.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-# Override the arch list for flash-attn to reduce the binary size
-ARG vllm_fa_cmake_gpu_arches='80-real;90-real'
-ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
 #################### BASE BUILD IMAGE ####################
 
 #################### WHEEL BUILD IMAGE ####################
@@ -184,6 +183,8 @@ COPY requirements/build.txt requirements/build.txt
 # Reference: https://github.com/astral-sh/uv/pull/1694
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/build.txt \
@@ -209,7 +210,16 @@ ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
 
 # Flag to control whether to use pre-built vLLM wheels
-ARG VLLM_USE_PRECOMPILED=""
+ARG VLLM_USE_PRECOMPILED
+# TODO: in setup.py VLLM_USE_PRECOMPILED is sensitive to truthiness, it will take =0 as "true", this should be fixed
+ENV VLLM_USE_PRECOMPILED=""
+RUN if [ "${VLLM_USE_PRECOMPILED}" = "1" ]; then \
+        export VLLM_USE_PRECOMPILED=1 && \
+        echo "Using precompiled wheels"; \
+    else \
+        unset VLLM_USE_PRECOMPILED && \
+        echo "Leaving VLLM_USE_PRECOMPILED unset to build wheels from source"; \
+    fi
 
 # if USE_SCCACHE is set, use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -226,8 +236,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
         && export SCCACHE_IDLE_TIMEOUT=0 \
         && export CMAKE_BUILD_TYPE=Release \
-        && export VLLM_USE_PRECOMPILED="${VLLM_USE_PRECOMPILED}" \
-        && export VLLM_DOCKER_BUILD_CONTEXT=1 \
         && sccache --show-stats \
         && python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38 \
         && sccache --show-stats; \
@@ -241,22 +249,9 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         # Clean any existing CMake artifacts
         rm -rf .deps && \
         mkdir -p .deps && \
-        export VLLM_USE_PRECOMPILED="${VLLM_USE_PRECOMPILED}" && \
-        export VLLM_DOCKER_BUILD_CONTEXT=1 && \
         python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
     fi
 
-# When using precompiled wheels, keep only the newest manylinux1 wheel and delete others
-RUN if [ "$VLLM_USE_PRECOMPILED" = "1" ]; then \
-        echo "Cleaning up extra wheels in dist/..." && \
-        # Identify the most recent manylinux1_x86_64 wheel
-        KEEP_WHEEL=$(ls -t dist/*manylinux1_x86_64.whl 2>/dev/null | head -n1) && \
-        if [ -n "$KEEP_WHEEL" ]; then \
-            echo "Keeping wheel: $KEEP_WHEEL"; \
-            find dist/ -type f -name "*.whl" ! -path "${KEEP_WHEEL}" -delete; \
-        fi; \
-    fi
-
 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
 # sync the default value with .buildkite/check-wheel-size.py
@@ -281,6 +276,8 @@ ARG PYTORCH_CUDA_INDEX_BASE_URL
 # Reference: https://github.com/astral-sh/uv/pull/1694
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
 
 COPY requirements/lint.txt requirements/lint.txt
 COPY requirements/test.txt requirements/test.txt
@@ -292,7 +289,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
-# TODO: Restore to base image after FlashInfer AOT wheel fixed
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 ARG CUDA_VERSION
 ARG PYTHON_VERSION
@@ -351,6 +347,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Reference: https://github.com/astral-sh/uv/pull/1694
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -394,7 +392,7 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
 # Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
 # We use `--force-reinstall --no-deps` to avoid issues with the existing FlashInfer wheel.
-ARG FLASHINFER_GIT_REF="v0.2.9rc2"
+ARG FLASHINFER_GIT_REF="v0.2.9"
 RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
   . /etc/environment
     git clone --depth 1 --recursive --shallow-submodules \
@@ -437,6 +435,33 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/build.txt \
         --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
+# Install DeepGEMM from source
+ARG DEEPGEMM_GIT_REPO="https://github.com/deepseek-ai/DeepGEMM.git"
+ARG DEEPGEMM_GIT_REF="187656694f7f69e3e7975617a68bc3387680a7e1"
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
+  . /etc/environment
+    CUDA_MAJOR="${CUDA_VERSION%%.*}"
+    CUDA_MINOR="${CUDA_VERSION#${CUDA_MAJOR}.}"
+    CUDA_MINOR="${CUDA_MINOR%%.*}"
+    if [ "$CUDA_MAJOR" -ge 12 ] && [ "$CUDA_MINOR" -ge 8 ]; then
+        git clone --recursive --shallow-submodules \
+            ${DEEPGEMM_GIT_REPO} deepgemm
+        echo "🏗️  Building DeepGEMM"
+        pushd deepgemm
+            git checkout ${DEEPGEMM_GIT_REF}
+            # Build DeepGEMM
+            # (Based on https://github.com/deepseek-ai/DeepGEMM/blob/main/install.sh)
+            rm -rf build dist
+            rm -rf *.egg-info
+            python3 setup.py bdist_wheel
+            uv pip install --system dist/*.whl
+        popd
+        rm -rf deepgemm
+    else
+        echo "Skipping DeepGEMM installation (requires CUDA 12.8+ but got ${CUDA_VERSION})"
+    fi
+BASH
+
 #################### vLLM installation IMAGE ####################
 
 #################### TEST IMAGE ####################
@@ -455,6 +480,8 @@ ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
 # Reference: https://github.com/astral-sh/uv/pull/1694
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
 
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \

docker/Dockerfile.cpu

@@ -113,7 +113,6 @@ WORKDIR /workspace/vllm
 
 RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
     cp requirements/test.in requirements/cpu-test.in && \
-    sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
     sed -i 's/^torch==.*/torch==2.6.0/g' requirements/cpu-test.in && \
     sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \
     sed -i 's/torchvision.*/torchvision/g' requirements/cpu-test.in && \

docker/Dockerfile.nightly_torch

@@ -114,9 +114,6 @@ RUN cat torch_build_versions.txt
 # explicitly set the list to avoid issues with torch 2.2
 # see https://github.com/pytorch/pytorch/pull/123243
 
-# Override the arch list for flash-attn to reduce the binary size
-ARG vllm_fa_cmake_gpu_arches='80-real;90-real'
-ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
 #################### BASE BUILD IMAGE ####################
 
 #################### WHEEL BUILD IMAGE ####################

docker/Dockerfile.tpu

@@ -1,4 +1,4 @@
-ARG NIGHTLY_DATE="20250724"
+ARG NIGHTLY_DATE="20250730"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.12_tpuvm_$NIGHTLY_DATE"
 
 FROM $BASE_IMAGE

docs/cli/README.md

@@ -6,13 +6,13 @@ toc_depth: 4
 
 The vllm command-line tool is used to run and manage vLLM models. You can start by viewing the help message with:
 
-```
+```bash
 vllm --help
 ```
 
 Available Commands:
 
-```
+```bash
 vllm {chat,complete,serve,bench,collect-env,run-batch}
 ```
 

docs/configuration/tpu.md

@@ -40,13 +40,14 @@ Although the first compilation can take some time, for all subsequent server lau
 Use `VLLM_XLA_CACHE_PATH` environment variable to write to shareable storage for future deployed nodes (like when using autoscaling).
 
 #### Reducing compilation time
+
 This initial compilation time ranges significantly and is impacted by many of the arguments discussed in this optimization doc. Factors that influence the length of time to compile are things like model size and `--max-num-batch-tokens`. Other arguments you can tune are things like `VLLM_TPU_MOST_MODEL_LEN`.
 
 ### Optimize based on your data
 
 #### max model len vs. most model len
 
-![most_model_len](../assets/design/v1/tpu/most_model_len.png)
+![most_model_len](../assets/design/tpu/most_model_len.png)
 
 If most of your requests are shorter than the maximum model length but you still need to accommodate occasional longer requests, setting a high maximum model length can negatively impact performance. In these cases, you can try introducing most model len by specifying the `VLLM_TPU_MOST_MODEL_LEN` environment variable.
 
@@ -71,12 +72,15 @@ The fewer tokens we pad, the less unnecessary computation TPU does, the better p
 
 However, you need to be careful to choose the padding gap. If the gap is too small, it means the number of buckets is large, leading to increased warmup (precompile) time and higher memory to store the compiled graph. Too many compilaed graphs may lead to HBM OOM. Conversely, an overly large gap yields no performance improvement compared to the default exponential padding.
 
-**If possible, use the precision that matches the chip’s hardware acceleration**
+#### Quantization
+
+If possible, use the precision that matches the chip’s hardware acceleration:
 
 - v5e has int4/int8 hardware acceleration in the MXU
 - v6e has int4/int8 hardware acceleration in the MXU
 
-Supported quantized formats and features in vLLM on TPU [Jul '25]
+Supported quantized formats and features in vLLM on TPU [Jul '25]:
+
 - INT8 W8A8
 - INT8 W8A16
 - FP8 KV cache
@@ -84,11 +88,13 @@ Supported quantized formats and features in vLLM on TPU [Jul '25]
 - [WIP] AWQ
 - [WIP] FP4 W4A8
 
-**Don't set TP to be less than the number of chips on a single-host deployment**
+#### Parallelization
+
+Don't set TP to be less than the number of chips on a single-host deployment.
 
 Although it’s common to do this with GPUs, don't try to fragment 2 or 8 different workloads across 8 chips on a single host. If you need 1 or 4 chips, just create an instance with 1 or 4 chips (these are partial-host machine types).
 
-### Tune your workloads!
+### Tune your workloads
 
 Although we try to have great default configs, we strongly recommend you check out the [vLLM auto-tuner](../../benchmarks/auto_tune/README.md) to optimize your workloads for your use case.
 
@@ -99,6 +105,7 @@ Although we try to have great default configs, we strongly recommend you check o
 The auto-tuner provides a profile of optimized configurations as its final step. However, interpreting this profile can be challenging for new users. We plan to expand this section in the future with more detailed guidance. In the meantime, you can learn how to collect a TPU profile using vLLM's native profiling tools [here](../examples/offline_inference/profiling_tpu.md). This profile can provide valuable insights into your workload's performance.
 
 #### SPMD
+
 More details to come.
 
 **Want us to cover something that isn't listed here? Open up an issue please and cite this doc. We'd love to hear your questions or tips.**

docs/contributing/ci/failures.md

@@ -20,19 +20,19 @@ the failure?
 
 - **Use this title format:**
 
-    ```
+    ```text
     [CI Failure]: failing-test-job - regex/matching/failing:test
     ```
 
 - **For the environment field:**
 
-    ```
- Still failing on main as of commit abcdef123
+    ```text
+    Still failing on main as of commit abcdef123
     ```
 
 - **In the description, include failing tests:**
 
-    ```
+    ```text
     FAILED failing/test.py:failing_test1 - Failure description
     FAILED failing/test.py:failing_test2 - Failure description
     https://github.com/orgs/vllm-project/projects/20

docs/contributing/ci/update_pytorch_version.md

@@ -106,6 +106,7 @@ releases (which would take too much time), they can be built from
 source to unblock the update process.
 
 ### FlashInfer
+
 Here is how to build and install it from source with `torch2.7.0+cu128` in vLLM [Dockerfile](https://github.com/vllm-project/vllm/blob/27bebcd89792d5c4b08af7a65095759526f2f9e1/docker/Dockerfile#L259-L271):
 
 ```bash
@@ -121,6 +122,7 @@ public location for immediate installation, such as [this FlashInfer wheel link]
 team if you want to get the package published there.
 
 ### xFormers
+
 Similar to FlashInfer, here is how to build and install xFormers from source:
 
 ```bash
@@ -129,19 +131,6 @@ MAX_JOBS=16 uv pip install --system \
     --no-build-isolation "git+https://github.com/facebookresearch/xformers@v0.0.30"
 ```
 
-### Mamba
-
-```bash
-uv pip install --system \
-    --no-build-isolation "git+https://github.com/state-spaces/mamba@v2.2.5"
-```
-
-### causal-conv1d
-
-```
-uv pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
-```
-
 ## Update all the different vLLM platforms
 
 Rather than attempting to update all vLLM platforms in a single pull request, it's more manageable

docs/contributing/deprecation_policy.md

@@ -31,7 +31,7 @@ Features that fall under this policy include (at a minimum) the following:
 The deprecation process consists of several clearly defined stages that span
 multiple Y releases:
 
-**1. Deprecated (Still On By Default)**
+### 1. Deprecated (Still On By Default)
 
 - **Action**: Feature is marked as deprecated.
 - **Timeline**: A removal version is explicitly stated in the deprecation
@@ -46,7 +46,7 @@ warning (e.g., "This will be removed in v0.10.0").
     - GitHub Issue (RFC) for feedback
     - Documentation and use of the `@typing_extensions.deprecated` decorator for Python APIs
 
-**2.Deprecated (Off By Default)**
+### 2.Deprecated (Off By Default)
 
 - **Action**: Feature is disabled by default, but can still be re-enabled via a
 CLI flag or environment variable. Feature throws an error when used without
@@ -55,7 +55,7 @@ re-enabling.
 while signaling imminent removal. Ensures any remaining usage is clearly
 surfaced and blocks silent breakage before full removal.
 
-**3. Removed**
+### 3. Removed
 
 - **Action**: Feature is completely removed from the codebase.
 - **Note**: Only features that have passed through the previous deprecation

docs/contributing/profiling.md

@@ -5,7 +5,12 @@
 
 ## Profile with PyTorch Profiler
 
-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/`
+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/`. Additionally, you can control the profiling content by specifying the following environment variables:
+
+- `VLLM_TORCH_PROFILER_RECORD_SHAPES=1` to enable recording Tensor Shapes, off by default
+- `VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1` to record memory, off by default
+- `VLLM_TORCH_PROFILER_WITH_STACK=1` to enable recording stack information, on by default
+- `VLLM_TORCH_PROFILER_WITH_FLOPS=1` to enable recording FLOPs, off by default
 
 The OpenAI server also needs to be started with the `VLLM_TORCH_PROFILER_DIR` environment variable set.
 
@@ -112,13 +117,13 @@ vllm bench serve \
 
 In practice, you should set the `--duration` argument to a large value. Whenever you want the server to stop profiling, run:
 
-```
+```bash
 nsys sessions list
 ```
 
 to get the session id in the form of `profile-XXXXX`, then run:
 
-```
+```bash
 nsys stop --session=profile-XXXXX
 ```
 

docs/contributing/vulnerability_management.md

@@ -32,9 +32,9 @@ We prefer to keep all vulnerability-related communication on the security report
 on GitHub. However, if you need to contact the VMT directly for an urgent issue,
 you may contact the following individuals:
 
-- Simon Mo - simon.mo@hey.com
-- Russell Bryant - rbryant@redhat.com
-- Huzaifa Sidhpurwala - huzaifas@redhat.com
+- Simon Mo - <simon.mo@hey.com>
+- Russell Bryant - <rbryant@redhat.com>
+- Huzaifa Sidhpurwala - <huzaifas@redhat.com>
 
 ## Slack Discussion
 

docs/deployment/docker.md

@@ -10,23 +10,23 @@ The image can be used to run OpenAI compatible server and is available on Docker
 ```bash
 docker run --runtime nvidia --gpus all \
     -v ~/.cache/huggingface:/root/.cache/huggingface \
-    --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
+    --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
     -p 8000:8000 \
     --ipc=host \
     vllm/vllm-openai:latest \
-    --model mistralai/Mistral-7B-v0.1
+    --model Qwen/Qwen3-0.6B
 ```
 
 This image can also be used with other container engines such as [Podman](https://podman.io/).
 
 ```bash
-podman run --gpus all \
+podman run --device nvidia.com/gpu=all \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
   -p 8000:8000 \
   --ipc=host \
-  vllm/vllm-openai:latest \
-  --model mistralai/Mistral-7B-v0.1
+  docker.io/vllm/vllm-openai:latest \
+  --model Qwen/Qwen3-0.6B
 ```
 
 You can add any other [engine-args](../configuration/engine_args.md) you need after the image tag (`vllm/vllm-openai:latest`).
@@ -106,8 +106,7 @@ of PyTorch Nightly and should be considered **experimental**. Using the flag `--
     -t vllm/vllm-gh200-openai:latest \
     --build-arg max_jobs=66 \
     --build-arg nvcc_threads=2 \
-    --build-arg torch_cuda_arch_list="9.0 10.0+PTX" \
-    --build-arg vllm_fa_cmake_gpu_arches="90-real"
+    --build-arg torch_cuda_arch_list="9.0 10.0+PTX"
     ```
 
 !!! note

docs/deployment/frameworks/anything-llm.md

@@ -19,9 +19,9 @@ vllm serve Qwen/Qwen1.5-32B-Chat-AWQ --max-model-len 4096
 - Download and install [Anything LLM desktop](https://anythingllm.com/desktop).
 
 - On the bottom left of open settings, AI Prooviders --> LLM:
-  - LLM Provider: Generic OpenAI
-  - Base URL: http://{vllm server host}:{vllm server port}/v1
-  - Chat Model Name: `Qwen/Qwen1.5-32B-Chat-AWQ`
+    - LLM Provider: Generic OpenAI
+    - Base URL: http://{vllm server host}:{vllm server port}/v1
+    - Chat Model Name: `Qwen/Qwen1.5-32B-Chat-AWQ`
 
 ![](../../assets/deployment/anything-llm-provider.png)
 
@@ -30,9 +30,9 @@ vllm serve Qwen/Qwen1.5-32B-Chat-AWQ --max-model-len 4096
 ![](../../assets/deployment/anything-llm-chat-without-doc.png)
 
 - Click the upload button:
-  - upload the doc
-  - select the doc and move to the workspace
-  - save and embed
+    - upload the doc
+    - select the doc and move to the workspace
+    - save and embed
 
 ![](../../assets/deployment/anything-llm-upload-doc.png)
 

docs/deployment/frameworks/chatbox.md

@@ -19,11 +19,11 @@ vllm serve qwen/Qwen1.5-0.5B-Chat
 - Download and install [Chatbox desktop](https://chatboxai.app/en#download).
 
 - On the bottom left of settings, Add Custom Provider
-  - API Mode: `OpenAI API Compatible`
-  - Name: vllm
-  - API Host: `http://{vllm server host}:{vllm server port}/v1`
-  - API Path: `/chat/completions`
-  - Model: `qwen/Qwen1.5-0.5B-Chat`
+    - API Mode: `OpenAI API Compatible`
+    - Name: vllm
+    - API Host: `http://{vllm server host}:{vllm server port}/v1`
+    - API Path: `/chat/completions`
+    - Model: `qwen/Qwen1.5-0.5B-Chat`
 
 ![](../../assets/deployment/chatbox-settings.png)
 

docs/deployment/frameworks/dify.md

@@ -34,11 +34,11 @@ docker compose up -d
 - In the top-right user menu (under the profile icon), go to Settings, then click `Model Provider`, and locate the `vLLM` provider to install it.
 
 - Fill in the model provider details as follows:
-  - **Model Type**: `LLM`
-  - **Model Name**: `Qwen/Qwen1.5-7B-Chat`
-  - **API Endpoint URL**: `http://{vllm_server_host}:{vllm_server_port}/v1`
-  - **Model Name for API Endpoint**: `Qwen/Qwen1.5-7B-Chat`
-  - **Completion Mode**: `Completion`
+    - **Model Type**: `LLM`
+    - **Model Name**: `Qwen/Qwen1.5-7B-Chat`
+    - **API Endpoint URL**: `http://{vllm_server_host}:{vllm_server_port}/v1`
+    - **Model Name for API Endpoint**: `Qwen/Qwen1.5-7B-Chat`
+    - **Completion Mode**: `Completion`
 
 ![](../../assets/deployment/dify-settings.png)
 

docs/deployment/frameworks/haystack.md

@@ -1,7 +1,5 @@
 # Haystack
 
-# Haystack
-
 [Haystack](https://github.com/deepset-ai/haystack) is an end-to-end LLM framework that allows you to build applications powered by LLMs, Transformer models, vector search and more. Whether you want to perform retrieval-augmented generation (RAG), document search, question answering or answer generation, Haystack can orchestrate state-of-the-art embedding models and LLMs into pipelines to build end-to-end NLP applications and solve your use case.
 
 It allows you to deploy a large language model (LLM) server with vLLM as the backend, which exposes OpenAI-compatible endpoints.

docs/deployment/frameworks/retrieval_augmented_generation.md

@@ -3,6 +3,7 @@
 [Retrieval-augmented generation (RAG)](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) is a technique that enables generative artificial intelligence (Gen AI) models to retrieve and incorporate new information. It modifies interactions with a large language model (LLM) so that the model responds to user queries with reference to a specified set of documents, using this information to supplement information from its pre-existing training data. This allows LLMs to use domain-specific and/or updated information. Use cases include providing chatbot access to internal company data or generating responses based on authoritative sources.
 
 Here are the integrations:
+
 - vLLM + [langchain](https://github.com/langchain-ai/langchain) + [milvus](https://github.com/milvus-io/milvus)
 - vLLM + [llamaindex](https://github.com/run-llama/llama_index) + [milvus](https://github.com/milvus-io/milvus)
 

docs/deployment/integrations/production-stack.md

@@ -140,11 +140,12 @@ The core vLLM production stack configuration is managed with YAML. Here is the e
     ```
 
 In this YAML configuration:
+
 * **`modelSpec`** includes:
-  * `name`: A nickname that you prefer to call the model.
-  * `repository`: Docker repository of vLLM.
-  * `tag`: Docker image tag.
-  * `modelURL`: The LLM model that you want to use.
+    * `name`: A nickname that you prefer to call the model.
+    * `repository`: Docker repository of vLLM.
+    * `tag`: Docker image tag.
+    * `modelURL`: The LLM model that you want to use.
 * **`replicaCount`**: Number of replicas.
 * **`requestCPU` and `requestMemory`**: Specifies the CPU and memory resource requests for the pod.
 * **`requestGPU`**: Specifies the number of GPUs required.

docs/deployment/k8s.md

@@ -5,7 +5,7 @@ Deploying vLLM on Kubernetes is a scalable and efficient way to serve machine le
 - [Deployment with CPUs](#deployment-with-cpus)
 - [Deployment with GPUs](#deployment-with-gpus)
 - [Troubleshooting](#troubleshooting)
-  - [Startup Probe or Readiness Probe Failure, container log contains "KeyboardInterrupt: terminated"](#startup-probe-or-readiness-probe-failure-container-log-contains-keyboardinterrupt-terminated)
+    - [Startup Probe or Readiness Probe Failure, container log contains "KeyboardInterrupt: terminated"](#startup-probe-or-readiness-probe-failure-container-log-contains-keyboardinterrupt-terminated)
 - [Conclusion](#conclusion)
 
 Alternatively, you can deploy vLLM to Kubernetes using any of the following:

docs/design/fused_moe_modular_kernel.md

@@ -1,6 +1,7 @@
 # Fused MoE Modular Kernel
 
 ## Introduction
+
 FusedMoEModularKernel is implemented [here](gh-file:/vllm/model_executor/layers/fused_moe/modular_kernel.py)
 
 Based on the format of the input activations, FusedMoE implementations are broadly classified into 2 types.
@@ -31,7 +32,8 @@ As can be seen from the diagrams, there are a lot of operations and there can be
 
 The rest of the document will focus on the Contiguous / Non-Batched case. Extrapolating to the Batched case should be straight-forward.
 
-## ModularKernel Components:
+## ModularKernel Components
+
 FusedMoEModularKernel splits the FusedMoE operation into 3 parts,
 
 1. TopKWeightAndReduce
@@ -39,6 +41,7 @@ FusedMoEModularKernel splits the FusedMoE operation into 3 parts,
 3. FusedMoEPermuteExpertsUnpermute
 
 ### TopKWeightAndReduce
+
 The TopK Weight Application and Reduction components happen right after the Unpermute operation and before the All2All Combine. Note that the `FusedMoEPermuteExpertsUnpermute` is responsible for the Unpermute and `FusedMoEPrepareAndFinalize` is responsible for the All2All Combine. There is value in doing the TopK Weight Application and Reduction in the `FusedMoEPermuteExpertsUnpermute`. But some implementations choose to do it `FusedMoEPrepareAndFinalize`. In order to enable this flexibility, we have a TopKWeightAndReduce abstract class.
 
 Please find the implementations of TopKWeightAndReduce [here](gh-file:vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py).
@@ -50,12 +53,14 @@ The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEPermuteExperts
 * `FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceContiguous` / `TopKWeightAndReduceNaiveBatched` / `TopKWeightAndReduceDelegate` if the `FusedMoEPermuteExpertsUnpermute` implementation needs the `FusedMoEPrepareAndFinalize::finalize()` to do the weight application and reduction.
 
 ### FusedMoEPrepareAndFinalize
+
 The `FusedMoEPrepareAndFinalize` abstract class exposes `prepare` and `finalize` functions.
 The `prepare` function is responsible for input activation Quantization and All2All Dispatch. The `finalize` function is responsible for invoking the All2All Combine. Additionally the `finalize` function may or may not do the TopK weight application and reduction (Please refer to the TopKWeightAndReduce section)
 
 ![](../assets/design/fused_moe_modular_kernel/prepare_and_finalize_blocks.png "FusedMoEPrepareAndFinalize Blocks")
 
 ### FusedMoEPermuteExpertsUnpermute
+
 The `FusedMoEPermuteExpertsUnpermute` class is where the crux of the MoE operations happen. The `FusedMoEPermuteExpertsUnpermute` abstract class exposes a few important functions,
 
 * apply()
@@ -63,6 +68,7 @@ The `FusedMoEPermuteExpertsUnpermute` class is where the crux of the MoE operati
 * finalize_weight_and_reduce_impl()
 
 #### apply()
+
 The `apply` method is where the implementations perform
 
 * Permute
@@ -74,50 +80,56 @@ The `apply` method is where the implementations perform
 * Maybe TopK Weight Application + Reduction
 
 #### workspace_shapes()
+
 The core FusedMoE implementation performs a series of operations. It would be inefficient to create output memory for each of these operations separately. To that effect, implementations are required to declare 2 workspace shapes, the workspace datatype and the FusedMoE output shape as outputs of the workspace_shapes() method. This information is used to allocate the workspace tensors and the output tensor in `FusedMoEModularKernel::forward()` and passed on to the `FusedMoEPermuteExpertsUnpermute::apply()` method. The workspaces could then be used as intermediate buffers in the FusedMoE implementation.
 
 #### finalize_weight_and_reduce_impl()
+
 It is sometimes efficient to perform TopK weight application and Reduction inside the `FusedMoEPermuteExpertsUnpermute::apply()`. Find an example [here](https://github.com/vllm-project/vllm/pull/20228). We have a `TopKWeightAndReduce` abstract class to facilitate such implementations. Please refer to the TopKWeightAndReduce section.
 `FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl()` returns the `TopKWeightAndReduce` object that the implementation wants the `FusedMoEPrepareAndFinalize::finalize()` to use.
 
 ![](../assets/design/fused_moe_modular_kernel/fused_experts_blocks.png "FusedMoEPermuteExpertsUnpermute Blocks")
 
 ### FusedMoEModularKernel
+
 `FusedMoEModularKernel` is composed of the `FusedMoEPrepareAndFinalize` and `FusedMoEPermuteExpertsUnpermute` objects.
 `FusedMoEModularKernel` pseudocode/sketch,
 
-```
-FusedMoEModularKernel::__init__(self,
-            prepare_finalize: FusedMoEPrepareAndFinalize,
-            fused_experts: FusedMoEPermuteExpertsUnpermute):
+```py
+class FusedMoEModularKernel:
+    def __init__(self,
+                 prepare_finalize: FusedMoEPrepareAndFinalize,
+                 fused_experts: FusedMoEPermuteExpertsUnpermute):
 
-    self.prepare_finalize = prepare_finalize
-    self.fused_experts = fused_experts
+        self.prepare_finalize = prepare_finalize
+        self.fused_experts = fused_experts
 
-FusedMoEModularKernel::forward(self, DP_A):
+    def forward(self, DP_A):
 
-    Aq, A_scale, _, _, _ = self.prepare_finalize.prepare(DP_A, ...)
+        Aq, A_scale, _, _, _ = self.prepare_finalize.prepare(DP_A, ...)
 
-    workspace13_shape, workspace2_shape, _, _ = self.fused_experts.workspace_shapes(...)
+        workspace13_shape, workspace2_shape, _, _ = self.fused_experts.workspace_shapes(...)
 
-    # allocate workspaces
-    workspace_13 = torch.empty(workspace13_shape, ...)
-    workspace_2 = torch.empty(workspace2_shape, ...)
+        # allocate workspaces
+        workspace_13 = torch.empty(workspace13_shape, ...)
+        workspace_2 = torch.empty(workspace2_shape, ...)
 
-    # execute fused_experts
-    fe_out = self.fused_experts.apply(Aq, A_scale, workspace13, workspace2, ...)
+        # execute fused_experts
+        fe_out = self.fused_experts.apply(Aq, A_scale, workspace13, workspace2, ...)
 
-    # war_impl is an object of type TopKWeightAndReduceNoOp if the fused_experts implementations performs the TopK Weight Application and Reduction.
-    war_impl = self.fused_experts.finalize_weight_and_reduce_impl()
+        # war_impl is an object of type TopKWeightAndReduceNoOp if the fused_experts implementations
+        # performs the TopK Weight Application and Reduction.
+        war_impl = self.fused_experts.finalize_weight_and_reduce_impl()
 
-    output = self.prepare_finalize.finalize(fe_out, war_impl,...)
-                            
-    return output
+        output = self.prepare_finalize.finalize(fe_out, war_impl,...)
+
+        return output
 ```
 
 ## How-To
 
 ### How To Add a FusedMoEPrepareAndFinalize Type
+
 Typically a FusedMoEPrepareAndFinalize type is backed by an All2All Dispatch & Combine implementation / kernel. For example,
 
 * PplxPrepareAndFinalize type is backed by Pplx All2All kernels,
@@ -125,9 +137,11 @@ Typically a FusedMoEPrepareAndFinalize type is backed by an All2All Dispatch & C
 * DeepEPLLPrepareAndFinalize type is backed by DeepEP Low-Latency All2All kernels.
 
 #### Step 1: Add an All2All manager
+
 The purpose of the All2All Manager is to setup the All2All kernel implementations. The `FusedMoEPrepareAndFinalize` implementations typically fetch a kernel-implementation "handle" from the All2All Manager to invoke the Dispatch and Combine functions. Please look at the All2All Manager implementations [here](gh-file:vllm/distributed/device_communicators/all2all.py).
 
 #### Step 2: Add a FusedMoEPrepareAndFinalize Type
+
 This section describes the significance of the various functions exposed by the `FusedMoEPrepareAndFinalize` abstract class.
 
 `FusedMoEPrepareAndFinalize::prepare()`: The prepare method implements the Quantization and All2All Dispatch. Typically the Dispatch function from the relevant All2All Manager is invoked.
@@ -145,6 +159,7 @@ This section describes the significance of the various functions exposed by the
 We suggest picking an already existing `FusedMoEPrepareAndFinalize` implementation that matches your All2All implementation closely and using it as a reference.
 
 ### How To Add a FusedMoEPermuteExpertsUnpermute Type
+
 FusedMoEPermuteExpertsUnpermute performs the core of the FusedMoE operations. The various functions exposed by the abstract class and their significance is as follows,
 
 `FusedMoEPermuteExpertsUnpermute::activation_formats()`: Return the supported Input and Output activation formats. i.e. Contiguous / Batched format.
@@ -159,12 +174,14 @@ implementations that input `FusedMoEActivationFormat.Standard` support chunking
 `FusedMoEPermuteExpertsUnpermute::apply`: Refer to `FusedMoEPermuteExpertsUnpermute` section above.
 
 ### FusedMoEModularKernel Initialization
+
 `FusedMoEMethodBase` class has 2 methods that are collectively responsible in creating the `FusedMoEModularKernel` object. They are,
 
 * select_gemm_impl, and
 * init_prepare_finalize
 
 #### select_gemm_impl
+
 The `select_gemm_impl` method is undefined in the base class. It is the responsibility of the derived class to implement a method that constructs a valid/appropriate `FusedMoEPermuteExpertsUnpermute` object.
 Please refer to the implementations in,
 
@@ -176,12 +193,14 @@ Please refer to the implementations in,
 dervied classes.
 
 #### init_prepare_finalize
+
 Based on the input and env settings, the `init_prepare_finalize` method creates the appropriate `FusedMoEPrepareAndFinalize` object. The method then queries `select_gemm_impl` for the appropriate `FusedMoEPermuteExpertsUnpermute` object and builds the `FusedMoEModularKernel` object
 
 Please take a look at [init_prepare_finalize](https://github.com/vllm-project/vllm/blob/1cbf951ba272c230823b947631065b826409fa62/vllm/model_executor/layers/fused_moe/layer.py#L188).
 **Important**: The `FusedMoEMethodBase` derived classes use the `FusedMoEMethodBase::fused_experts` object in their `apply` methods. When settings permit the construction of a valid `FusedMoEModularKernel` object, we override `FusedMoEMethodBase::fused_experts` with it. This essentially makes the derived classes agnostic to what FusedMoE implementation is used.
 
 ### How To Unit Test
+
 We have `FusedMoEModularKernel` unit tests at [test_modular_kernel_combinations.py](gh-file:tests/kernels/moe/test_modular_kernel_combinations.py).
 
 The unit test iterates through all combinations of `FusedMoEPrepareAndFinalize` and `FusedMoEPremuteExpertsUnpermute` types and if they are
@@ -196,18 +215,21 @@ If you are adding some `FusedMoEPrepareAndFinalize` / `FusedMoEPermuteExpertsUnp
 Doing this will add the new implementation to the test suite.
 
 ### How To Check `FusedMoEPrepareAndFinalize` & `FusedMoEPermuteExpertsUnpermute` Compatibility
+
 The unit test file [test_modular_kernel_combinations.py](gh-file:tests/kernels/moe/test_modular_kernel_combinations.py) can also be executed as a standalone script.
 Example: `python3 -m tests.kernels.moe.test_modular_kernel_combinations --pf-type PplxPrepareAndFinalize --experts-type BatchedTritonExperts`
 As a side-effect, this script can be used to test `FusedMoEPrepareAndFinalize` & `FusedMoEPermuteExpertsUnpermute` compatibility. When invoked
 with incompatible types, the script will error.
 
 ### How To Profile
+
 Please take a look at [profile_modular_kernel.py](gh-file:tests/kernels/moe/modular_kernel_tools/profile_modular_kernel.py)
 The script can be used to generate Torch traces for a single `FusedMoEModularKernel::forward()` call for any compatible
 `FusedMoEPrepareAndFinalize` and `FusedMoEPermuteExpertsUnpermute` types.
 Example: `python3 -m tests.kernels.moe.modular_kernel_tools.profile_modular_kernel --pf-type PplxPrepareAndFinalize --experts-type BatchedTritonExperts`
 
 ## FusedMoEPrepareAndFinalize Implementations
+
 The following table lists the `FusedMoEPrepareAndFinalize` implementations at the time of writing,
 
 | Implementation | Type | Comments |
@@ -220,6 +242,7 @@ The following table lists the `FusedMoEPrepareAndFinalize` implementations at th
 | BatchedPrepareAndFinalize | Batched | A reference prepare/finalize class that reorganizes the tokens into expert batched format, i.e. E x max_num_tokens x K. (Doesn’t use any all2all kernels. This is primarily used in unit testing) |
 
 ## FusedMoEPermuteExpertsUnpermute
+
 The following table lists the `FusedMoEPermuteExpertsUnpermute` implementations at the time of writing,
 
 | Implementation | Type | Comment |

docs/design/metrics.md

@@ -223,7 +223,7 @@ And the calculated intervals are:
 
 Put another way:
 
-![Interval calculations - common case](../../assets/design/v1/metrics/intervals-1.png)
+![Interval calculations - common case](../assets/design/metrics/intervals-1.png)
 
 We explored the possibility of having the frontend calculate these
 intervals using the timing of events visible by the frontend. However,
@@ -238,13 +238,13 @@ When a preemption occurs during decode, since any already generated
 tokens are reused, we consider the preemption as affecting the
 inter-token, decode, and inference intervals.
 
-![Interval calculations - preempted decode](../../assets/design/v1/metrics/intervals-2.png)
+![Interval calculations - preempted decode](../assets/design/metrics/intervals-2.png)
 
 When a preemption occurs during prefill (assuming such an event
 is possible), we consider the preemption as affecting the
 time-to-first-token and prefill intervals.
 
-![Interval calculations - preempted prefill](../../assets/design/v1/metrics/intervals-3.png)
+![Interval calculations - preempted prefill](../assets/design/metrics/intervals-3.png)
 
 ### Frontend Stats Collection
 
@@ -361,7 +361,7 @@ instances in Prometheus.
 
 We use this concept for the `vllm:cache_config_info` metric:
 
-```
+```text
 # HELP vllm:cache_config_info Information of the LLMEngine CacheConfig
 # TYPE vllm:cache_config_info gauge
 vllm:cache_config_info{block_size="16",cache_dtype="auto",calculate_kv_scales="False",cpu_offload_gb="0",enable_prefix_caching="False",gpu_memory_utilization="0.9",...} 1.0
@@ -686,7 +686,7 @@ documentation for this option states:
 The metrics were added by <gh-pr:7089> and who up in an OpenTelemetry trace
 as:
 
-```
+```text
 -> gen_ai.latency.time_in_scheduler: Double(0.017550230026245117)
 -> gen_ai.latency.time_in_model_forward: Double(3.151565277099609)
 -> gen_ai.latency.time_in_model_execute: Double(3.6468167304992676)

docs/design/p2p_nccl_connector.md

@@ -5,6 +5,7 @@ An implementation of xPyD with dynamic scaling based on point-to-point communica
 ## Detailed Design
 
 ### Overall Process
+
 As shown in Figure 1, the overall process of this **PD disaggregation** solution is described through a request flow:
 
 1. The client sends an HTTP request to the Proxy/Router's `/v1/completions` interface.
@@ -23,7 +24,7 @@ A simple HTTP service acts as the entry point for client requests and starts a b
 
 The Proxy/Router is responsible for selecting 1P1D based on the characteristics of the client request, such as the prompt, and generating a corresponding `request_id`, for example:
 
-```
+```text
 cmpl-___prefill_addr_10.0.1.2:21001___decode_addr_10.0.1.3:22001_93923d63113b4b338973f24d19d4bf11-0
 ```
 
@@ -70,6 +71,7 @@ pip install "vllm>=0.9.2"
 ## Run xPyD
 
 ### Instructions
+
 - The following examples are run on an A800 (80GB) device, using the Meta-Llama-3.1-8B-Instruct model.
 - Pay attention to the setting of the `kv_buffer_size` (in bytes). The empirical value is 10% of the GPU memory size. This is related to the kvcache size. If it is too small, the GPU memory buffer for temporarily storing the received kvcache will overflow, causing the kvcache to be stored in the tensor memory pool, which increases latency. If it is too large, the kvcache available for inference will be reduced, leading to a smaller batch size and decreased throughput.
 - For Prefill instances, when using non-GET mode, the `kv_buffer_size` can be set to 1, as Prefill currently does not need to receive kvcache. However, when using GET mode, a larger `kv_buffer_size` is required because it needs to store the kvcache sent to the D instance.
@@ -107,7 +109,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.9 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_buffer_size":"1e1","kv_port":"21001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20001"}}' > /var/vllm.log 2>&1 &
     ```
@@ -129,7 +130,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.7 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_buffer_size":"8e9","kv_port":"22001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20002"}}' > /var/vllm.log 2>&1 &
     ```
@@ -151,7 +151,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.7 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_buffer_size":"8e9","kv_port":"23001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20003"}}' > /var/vllm.log 2>&1 &
     ```
@@ -173,7 +172,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.7 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_buffer_size":"8e9","kv_port":"24001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20004"}}' > /var/vllm.log 2>&1 &
     ```
@@ -204,7 +202,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.9 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_buffer_size":"1e1","kv_port":"21001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20001"}}' > /var/vllm.log 2>&1 &
     ```
@@ -226,7 +223,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.9 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_buffer_size":"1e1","kv_port":"22001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20002"}}' > /var/vllm.log 2>&1 &
     ```
@@ -248,7 +244,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.9 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_buffer_size":"1e1","kv_port":"23001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20003"}}' > /var/vllm.log 2>&1 &
     ```
@@ -270,7 +265,6 @@ python3 disagg_proxy_p2p_nccl_xpyd.py &
         --max-num-seqs 256 \
         --trust-remote-code \
         --gpu-memory-utilization 0.7 \
-        --disable-log-request \
         --kv-transfer-config \
         '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_buffer_size":"8e9","kv_port":"24001","kv_connector_extra_config":{"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20004"}}' > /var/vllm.log 2>&1 &
     ```

docs/design/paged_attention.md

@@ -1,7 +1,7 @@
-# vLLM Paged Attention
+# Paged Attention
 
 !!! warning
-    This document is being kept in the vLLM documentation for historical purposes.
+    This is a historical document based on the [original paper for vLLM](https://arxiv.org/abs/2309.06180).
     It no longer describes the code used in vLLM today.
 
 Currently, vLLM utilizes its own implementation of a multi-head query
@@ -140,7 +140,7 @@ const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
 ```
 
 <figure markdown="span">
-  ![](../../assets/kernel/query.png){ align="center" alt="query" width="70%" }
+  ![](../assets/design/paged_attention/query.png){ align="center" alt="query" width="70%" }
 </figure>
 
 Each thread defines its own `q_ptr` which points to the assigned
@@ -149,7 +149,7 @@ and `HEAD_SIZE` is 128, the `q_ptr` points to data that contains
 total of 128 elements divided into 128 / 4 = 32 vecs.
 
 <figure markdown="span">
-  ![](../../assets/kernel/q_vecs.png){ align="center" alt="q_vecs" width="70%" }
+  ![](../assets/design/paged_attention/q_vecs.png){ align="center" alt="q_vecs" width="70%" }
 </figure>
 
 ```cpp
@@ -188,7 +188,7 @@ points to key token data based on `k_cache` at assigned block,
 assigned head and assigned token.
 
 <figure markdown="span">
-  ![](../../assets/kernel/key.png){ align="center" alt="key" width="70%" }
+  ![](../assets/design/paged_attention/key.png){ align="center" alt="key" width="70%" }
 </figure>
 
 The diagram above illustrates the memory layout for key data. It
@@ -203,7 +203,7 @@ elements for one token) that will be processed by 2 threads (one
 thread group) separately.
 
 <figure markdown="span">
-  ![](../../assets/kernel/k_vecs.png){ align="center" alt="k_vecs" width="70%" }
+  ![](../assets/design/paged_attention/k_vecs.png){ align="center" alt="k_vecs" width="70%" }
 </figure>
 
 ```cpp
@@ -362,15 +362,15 @@ later steps. Now, it should store the normalized softmax result of
 ## Value
 
 <figure markdown="span">
-  ![](../../assets/kernel/value.png){ align="center" alt="value" width="70%" }
+  ![](../assets/design/paged_attention/value.png){ align="center" alt="value" width="70%" }
 </figure>
 
 <figure markdown="span">
-  ![](../../assets/kernel/logits_vec.png){ align="center" alt="logits_vec" width="50%" }
+  ![](../assets/design/paged_attention/logits_vec.png){ align="center" alt="logits_vec" width="50%" }
 </figure>
 
 <figure markdown="span">
-  ![](../../assets/kernel/v_vec.png){ align="center" alt="v_vec" width="70%" }
+  ![](../assets/design/paged_attention/v_vec.png){ align="center" alt="v_vec" width="70%" }
 </figure>
 
 Now we need to retrieve the value data and perform dot multiplication
@@ -499,3 +499,14 @@ for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
 Finally, we need to iterate over different assigned head positions
 and write out the corresponding accumulated result based on the
 `out_ptr`.
+
+## Citation
+
+```bibtex
+@inproceedings{kwon2023efficient,
+  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
+  author={Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
+  booktitle={Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles},
+  year={2023}
+}
+```