[Misc] Support quantization of MllamaForCausalLM (#8822)

Co-authored-by: simon-mo <xmo@berkeley.edu>
Co-authored-by: Chang Su <chang.s.su@oracle.com>
Co-authored-by: Simon Mo <simon.mo@hey.com>
Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>
Co-authored-by: Roger Wang <ywang@roblox.com>

.buildkite/check-wheel-size.py

@@ -1,36 +1,43 @@
 import os
+import sys
 import zipfile
 
-MAX_SIZE_MB = 200
+# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 250 MB
+VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 250))
 
 
 def print_top_10_largest_files(zip_file):
+    """Print the top 10 largest files in the given zip file."""
     with zipfile.ZipFile(zip_file, 'r') as z:
         file_sizes = [(f, z.getinfo(f).file_size) for f in z.namelist()]
         file_sizes.sort(key=lambda x: x[1], reverse=True)
         for f, size in file_sizes[:10]:
-            print(f"{f}: {size/(1024*1024)} MBs uncompressed.")
+            print(f"{f}: {size / (1024 * 1024):.2f} MBs uncompressed.")
 
 
 def check_wheel_size(directory):
+    """Check the size of .whl files in the given directory."""
     for root, _, files in os.walk(directory):
-        for f in files:
-            if f.endswith(".whl"):
-                wheel_path = os.path.join(root, f)
-                wheel_size = os.path.getsize(wheel_path)
-                wheel_size_mb = wheel_size / (1024 * 1024)
-                if wheel_size_mb > MAX_SIZE_MB:
-                    print(
-                        f"Wheel {wheel_path} is too large ({wheel_size_mb} MB) "
-                        f"compare to the allowed size ({MAX_SIZE_MB} MB).")
+        for file_name in files:
+            if file_name.endswith(".whl"):
+                wheel_path = os.path.join(root, file_name)
+                wheel_size_mb = os.path.getsize(wheel_path) / (1024 * 1024)
+                if wheel_size_mb > VLLM_MAX_SIZE_MB:
+                    print(f"Not allowed: Wheel {wheel_path} is larger "
+                          f"({wheel_size_mb:.2f} MB) than the limit "
+                          f"({VLLM_MAX_SIZE_MB} MB).")
                     print_top_10_largest_files(wheel_path)
                     return 1
                 else:
                     print(f"Wheel {wheel_path} is within the allowed size "
-                          f"({wheel_size_mb} MB).")
+                          f"({wheel_size_mb:.2f} MB).")
     return 0
 
 
 if __name__ == "__main__":
-    import sys
-    sys.exit(check_wheel_size(sys.argv[1]))
+    if len(sys.argv) < 2:
+        print("Usage: python check-wheel-size.py <directory>")
+        sys.exit(1)
+
+    directory = sys.argv[1]
+    sys.exit(check_wheel_size(directory))

.buildkite/lm-eval-harness/configs/DeepSeek-V2-Lite-Chat.yaml

@@ -9,3 +9,4 @@ tasks:
     value: 0.664
 limit: 1000
 num_fewshot: 5
+trust_remote_code: True

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-QQQ.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m HandH1998/QQQ-Llama-3-8b-g128 -b 32 -l 1000 -f 5 -t 1
+model_name: "HandH1998/QQQ-Llama-3-8b-g128"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.419
+  - name: "exact_match,flexible-extract"
+    value: 0.416
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Minitron-4B-Base-FP8.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m mgoin/Minitron-4B-Base-FP8 -b auto -l 1000 -f 5 -t 1
+model_name: "mgoin/Minitron-4B-Base-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.233
+  - name: "exact_match,flexible-extract"
+    value: 0.236
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-FP8W8.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen2-1.5B-Instruct-FP8W8 -b auto -l 1000 -f 5 -t 1
+model_name: "nm-testing/Qwen2-1.5B-Instruct-FP8W8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.578
+  - name: "exact_match,flexible-extract"
+    value: 0.585
+limit: 1000
+num_fewshot: 5

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

@@ -1,7 +1,9 @@
 Meta-Llama-3-8B-Instruct.yaml
-Meta-Llama-3-8B-Instruct-FP8.yaml
 Meta-Llama-3-8B-Instruct-FP8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
+Minitron-4B-Base-FP8.yaml
 Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
+Qwen2-1.5B-Instruct-FP8W8.yaml
+Meta-Llama-3-8B-QQQ.yaml

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

@@ -14,7 +14,7 @@ import lm_eval
 import numpy
 import yaml
 
-RTOL = 0.02
+RTOL = 0.05
 TEST_DATA_FILE = os.environ.get(
     "LM_EVAL_TEST_DATA_FILE",
     ".buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml")
@@ -23,9 +23,12 @@ TP_SIZE = os.environ.get("LM_EVAL_TP_SIZE", 1)
 
 
 def launch_lm_eval(eval_config):
+    trust_remote_code = eval_config.get('trust_remote_code', False)
+
     model_args = f"pretrained={eval_config['model_name']}," \
                  f"tensor_parallel_size={TP_SIZE}," \
-                 f"add_bos_token=true"
+                 f"add_bos_token=true," \
+                 f"trust_remote_code={trust_remote_code}"
 
     results = lm_eval.simple_evaluate(
         model="vllm",

.buildkite/nightly-benchmarks/README.md

@@ -3,30 +3,52 @@
 
 ## Introduction
 
-This directory contains the performance benchmarking CI for vllm.
-The goal is to help developers know the impact of their PRs on the performance of vllm.
+This directory contains two sets of benchmark for vllm.
+- Performance benchmark: benchmark vllm's performance under various workload, for **developers** to gain clarity on whether their PR improves/degrades vllm's performance
+- Nightly benchmark: compare vllm's performance against alternatives (tgi, trt-llm and lmdeploy), for **the public** to know when to choose vllm.
 
-This benchmark will be *triggered* upon:
-- A PR being merged into vllm.
-- Every commit for those PRs with `perf-benchmarks` label.
 
-**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for more GPUs is comming later), with different models.
+See  [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
+
+
+## Performance benchmark quick overview
+
+**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!), with different models.
 
 **Benchmarking Duration**: about 1hr.
 
-**For benchmarking developers**: please try your best to constraint the duration of benchmarking to less than 1.5 hr so that it won't take forever to run.
+**For benchmarking developers**: please try your best to constraint the duration of benchmarking to about 1 hr so that it won't take forever to run.
 
 
-## Configuring the workload
+## Nightly benchmark quick overview
 
-The benchmarking workload contains three parts:
-- Latency tests in `latency-tests.json`.
-- Throughput tests in `throughput-tests.json`.
-- Serving tests in `serving-tests.json`.
+**Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B. 
 
-See [descriptions.md](tests/descriptions.md) for detailed descriptions. 
+**Benchmarking engines**: vllm, TGI, trt-llm and lmdeploy.
 
-### Latency test
+**Benchmarking Duration**: about 3.5hrs.
+
+
+
+## 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.
+
+Nightly benchmark will be triggered when:
+- Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label.
+
+
+
+
+## Performance benchmark details
+
+
+See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
+
+
+#### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
 
@@ -47,19 +69,19 @@ Here is an example of one test inside `latency-tests.json`:
 
 In this example:
 -  The `test_name` attributes is a unique identifier for the test. In `latency-tests.json`, it must start with `latency_`.
--  The `parameters` attribute control the command line arguments to be used for `benchmark_latency.py`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-benchmarks-suite.sh` will convert the underline to dash when feeding the arguments to `benchmark_latency.py`. For example, the corresponding command line arguments for `benchmark_latency.py` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15`
+-  The `parameters` attribute control the command line arguments to be used for `benchmark_latency.py`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-performance-benchmarks.sh` will convert the underline to dash when feeding the arguments to `benchmark_latency.py`. For example, the corresponding command line arguments for `benchmark_latency.py` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15`
 
 Note that the performance numbers are highly sensitive to the value of the parameters. Please make sure the parameters are set correctly.
 
 WARNING: The benchmarking script will save json results by itself, so please do not configure `--output-json` parameter in the json file.
 
 
-### Throughput test
+#### Throughput test
 The tests are specified in `throughput-tests.json`. The syntax is similar to `latency-tests.json`, except for that the parameters will be fed forward to `benchmark_throughput.py`.
 
 The number of this test is also stable -- a slight change on the value of this number might vary the performance numbers by a lot.
 
-### Serving test
+#### Serving test
 We test the throughput by using `benchmark_serving.py` with request rate = inf to cover the online serving overhead. The corresponding parameters are in `serving-tests.json`, and here is an example:
 
 ```
@@ -96,9 +118,36 @@ The number of this test is less stable compared to the delay and latency benchma
 
 WARNING: The benchmarking script will save json results by itself, so please do not configure `--save-results` or other results-saving-related parameters in `serving-tests.json`.
 
-## Visualizing the results
+#### Visualizing the results
 The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results.
 You can find the result presented as a table inside the `buildkite/performance-benchmark` job page.
 If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.
 The raw benchmarking results (in the format of json files) are in the `Artifacts` tab of the benchmarking.
+
+
+
+## Nightly test details
+
+See [nightly-descriptions.md](nightly-descriptions.md) for the detailed description on test workload, models and docker containers of benchmarking other llm engines.
+
+
+#### Workflow
+
+- The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines. 
+- Inside each container, we run [run-nightly-suite.sh](run-nightly-suite.sh), which will probe the serving engine of the current container.
+- The `run-nightly-suite.sh` will redirect the request to `tests/run-[llm serving engine name]-nightly.sh`, which parses the workload described in [nightly-tests.json](tests/nightly-tests.json) and performs the benchmark.
+- At last, we run [scripts/plot-nightly-results.py](scripts/plot-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite.
+
+#### Nightly tests
+
+In [nightly-tests.json](tests/nightly-tests.json), we include the command line arguments for benchmarking commands, together with the benchmarking test cases. The format is highly similar to performance benchmark.
+
+#### Docker containers
+
+The docker containers for benchmarking are specified in `nightly-pipeline.yaml`.
+
+WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `tests/run-[llm serving engine name]-nightly.sh`.
+
+WARNING: populating `trt-llm` to latest version is not easy, as it requires updating several protobuf files in [tensorrt-demo](https://github.com/neuralmagic/tensorrt-demo.git).
+

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

@@ -8,8 +8,7 @@ steps:
           containers:
           - image: badouralix/curl-jq
             command:
-            - sh
-            - .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
+            - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
   - wait
   - label: "A100"
     agents:
@@ -21,7 +20,7 @@ steps:
           containers:
           - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
             command:
-            - bash .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
+            - bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
             resources:
               limits:
                 nvidia.com/gpu: 8
@@ -42,20 +41,20 @@ steps:
           - name: devshm
             emptyDir:
               medium: Memory
-  - label: "H100"
-    agents:
-      queue: H100
-    plugins:
-    - docker#v5.11.0:
-        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-        command:
-        - bash
-        - .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
-        mount-buildkite-agent: true
-        propagate-environment: true
-        ipc: host
-        gpus: all
-        environment:
-        - VLLM_USAGE_SOURCE
-        - HF_TOKEN
+  # - label: "H100"
+  #   agents:
+  #     queue: H100
+  #   plugins:
+  #   - docker#v5.11.0:
+  #       image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+  #       command:
+  #       - bash
+  #       - .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
+  #       mount-buildkite-agent: true
+  #       propagate-environment: true
+  #       ipc: host
+  #       gpus: all
+  #       environment:
+  #       - VLLM_USAGE_SOURCE
+  #       - HF_TOKEN
 

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

@@ -1,47 +1,42 @@
 
 ## Latency tests
 
-This test suite aims to test vllm's end-to-end latency under a controlled setup.
-
 - Input length: 32 tokens.
 - Output length: 128 tokens.
 - Batch size: fixed (8).
-- Models: llama-3 8B, llama-3 70B, mixtral 8x7B.
+- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - Evaluation metrics: end-to-end latency (mean, median, p99).
 
-### Latency benchmarking results
 
 {latency_tests_markdown_table}
 
-## Throughput tests
 
-This test suite aims to test vllm's throughput.
+## Throughput tests
 
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm to achieve maximum throughput.
-- Models: llama-3 8B, llama-3 70B, mixtral 8x7B.
+- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - Evaluation metrics: throughput.
 
-### Throughput benchmarking results
 
 {throughput_tests_markdown_table}
 
-## Serving tests
 
-This test suite aims to test vllm's real serving metrics.
+## Serving tests
 
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm and the arrival pattern of the requests.
 - **Average QPS (query per second)**: 1, 4, 16 and inf. QPS = inf means all requests come at once. For other QPS values, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
-- Models: llama-3 8B, llama-3 70B, mixtral 8x7B.
+- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- We also added a speculative decoding test for llama-3 70B, under QPS 2
 - Evaluation metrics: throughput, TTFT (time to the first token, with mean, median and p99), ITL (inter-token latency, with mean, median and p99).
 
-### Serving benchmarking results
 
 {serving_tests_markdown_table}
 
+
 ## json version of the benchmarking tables
 
 This section contains the data of the markdown tables above in JSON format. 

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

@@ -174,8 +174,8 @@ if __name__ == "__main__":
     # document the result
     with open(results_folder / "benchmark_results.md", "w") as f:
 
-        results = read_markdown(
-            "../.buildkite/nightly-benchmarks/tests/descriptions.md")
+        results = read_markdown("../.buildkite/nightly-benchmarks/" +
+                                "performance-benchmarks-descriptions.md")
         results = results.format(
             latency_tests_markdown_table=latency_md_table,
             throughput_tests_markdown_table=throughput_md_table,

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

@@ -34,6 +34,15 @@ check_hf_token() {
   fi
 }
 
+ensure_sharegpt_downloaded() {
+  local FILE=ShareGPT_V3_unfiltered_cleaned_split.json
+  if [ ! -f "$FILE" ]; then
+    wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/$FILE
+  else
+    echo "$FILE already exists."
+  fi
+}
+
 json2args() {
   # transforms the JSON string to command line args, and '_' is replaced to '-'
   # example:
@@ -59,40 +68,38 @@ wait_for_server() {
     done' && return 0 || return 1
 }
 
-kill_gpu_processes() {
-  # kill all processes on GPU.
-  pids=$(nvidia-smi --query-compute-apps=pid --format=csv,noheader)
-  if [ -z "$pids" ]; then
-      echo "No GPU processes found."
+kill_processes_launched_by_current_bash() {
+  # Kill all python processes launched from current bash script
+  current_shell_pid=$$
+  processes=$(ps -eo pid,ppid,command | awk -v ppid="$current_shell_pid" -v proc="$1" '$2 == ppid && $3 ~ proc {print $1}')
+  if [ -n "$processes" ]; then
+    echo "Killing the following processes matching '$1':"
+    echo "$processes"
+    echo "$processes" | xargs kill -9
   else
-      for pid in $pids; do
-          kill -9 "$pid"
-          echo "Killed process with PID: $pid"
-      done
-
-      echo "All GPU processes have been killed."
+    echo "No processes found matching '$1'."
   fi
+}
 
-  # Sometimes kill with pid doesn't work properly, we can also kill all process running python or python3
-  # since we are in container anyway
-  pkill -9 -f python
-  pkill -9 -f python3
+kill_gpu_processes() {
 
-  # waiting for GPU processes to be fully killed
-  # loop while nvidia-smi returns any processes
-  while [ -n "$(nvidia-smi --query-compute-apps=pid --format=csv,noheader)" ]; do
+  ps -aux
+  lsof -t -i:8000 | xargs -r kill -9
+  pkill -f pt_main_thread
+  # this line doesn't work now
+  # ps aux | grep python | grep openai | awk '{print $2}' | xargs -r kill -9
+  pkill -f python3
+  pkill -f /usr/bin/python3
+
+
+  # wait until GPU memory usage smaller than 1GB
+  while [ $(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1) -ge 1000 ]; do
     sleep 1
-    echo "Waiting for GPU processes to be killed"
   done
 
   # remove vllm config file
   rm -rf ~/.config/vllm
 
-  # Print the GPU memory usage
-  # so that we know if all GPU processes are killed.
-  gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0)
-  # The memory usage should be 0 MB.
-  echo "GPU 0 Memory Usage: $gpu_memory_usage MB"
 }
 
 upload_to_buildkite() {
@@ -110,7 +117,7 @@ upload_to_buildkite() {
   fi
 
   # Use the determined command to annotate and upload artifacts
-  $BUILDKITE_AGENT_COMMAND annotate --style "info" --context "$BUILDKITE_LABEL-benchmark-results" < $RESULTS_FOLDER/benchmark_results.md
+  $BUILDKITE_AGENT_COMMAND annotate --style "info" --context "$BUILDKITE_LABEL-benchmark-results" <$RESULTS_FOLDER/benchmark_results.md
   $BUILDKITE_AGENT_COMMAND artifact upload "$RESULTS_FOLDER/*"
 }
 
@@ -162,7 +169,7 @@ run_latency_tests() {
         latency_command: $latency,
         gpu_type: $gpu
       }')
-    echo "$jq_output" > "$RESULTS_FOLDER/$test_name.commands"
+    echo "$jq_output" >"$RESULTS_FOLDER/$test_name.commands"
 
     # run the benchmark
     eval "$latency_command"
@@ -172,7 +179,6 @@ run_latency_tests() {
   done
 }
 
-
 run_throughput_tests() {
   # run throughput tests using `benchmark_throughput.py`
   # $1: a json file specifying throughput test cases
@@ -220,7 +226,7 @@ run_throughput_tests() {
         throughput_command: $command,
         gpu_type: $gpu
       }')
-    echo "$jq_output" > "$RESULTS_FOLDER/$test_name.commands"
+    echo "$jq_output" >"$RESULTS_FOLDER/$test_name.commands"
 
     # run the benchmark
     eval "$throughput_command"
@@ -252,7 +258,6 @@ run_serving_tests() {
       continue
     fi
 
-
     # get client and server arguments
     server_params=$(echo "$params" | jq -r '.server_parameters')
     client_params=$(echo "$params" | jq -r '.client_parameters')
@@ -330,7 +335,7 @@ run_serving_tests() {
           client_command: $client,
           gpu_type: $gpu
         }')
-      echo "$jq_output" > "$RESULTS_FOLDER/${new_test_name}.commands"
+      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
 
     done
 
@@ -347,6 +352,7 @@ main() {
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
   (which jq) || (apt-get update && apt-get -y install jq)
+  (which lsof) || (apt-get update && apt-get install -y lsof)
 
   # get the current IP address, required by benchmark_serving.py
   export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
@@ -355,7 +361,7 @@ main() {
 
   # prepare for benchmarking
   cd benchmarks || exit 1
-  wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+  ensure_sharegpt_downloaded
   declare -g RESULTS_FOLDER=results/
   mkdir -p $RESULTS_FOLDER
   QUICK_BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
@@ -365,7 +371,6 @@ main() {
   run_latency_tests $QUICK_BENCHMARK_ROOT/tests/latency-tests.json
   run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/throughput-tests.json
 
-
   # postprocess benchmarking results
   pip install tabulate pandas
   python3 $QUICK_BENCHMARK_ROOT/scripts/convert-results-json-to-markdown.py

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

@@ -2,9 +2,11 @@
 TOKEN=$(curl -s -L "https://public.ecr.aws/token?service=public.ecr.aws&scope=repository:q9t5s3a7/vllm-ci-test-repo:pull" | jq -r .token)
 URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-test-repo/manifests/$BUILDKITE_COMMIT"
 
+TIMEOUT_SECONDS=10
+
 retries=0
 while [ $retries -lt 1000 ]; do
-    if [ $(curl -s -L -H "Authorization: Bearer $TOKEN" -o /dev/null -w "%{http_code}" $URL) -eq 200 ]; then
+    if [ $(curl -s --max-time $TIMEOUT_SECONDS -L -H "Authorization: Bearer $TOKEN" -o /dev/null -w "%{http_code}" $URL) -eq 200 ]; then
         exit 0
     fi
 

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

@@ -2,7 +2,7 @@
     {
         "test_name": "latency_llama8B_tp1",
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3-8B",
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
             "load_format": "dummy",
             "num_iters_warmup": 5,
@@ -12,7 +12,7 @@
     {
         "test_name": "latency_llama70B_tp4",
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "tensor_parallel_size": 4,
             "load_format": "dummy",
             "num-iters-warmup": 5,

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

@@ -3,7 +3,7 @@
         "test_name": "serving_llama8B_tp1_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
         "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3-8B",
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
             "swap_space": 16,
             "disable_log_stats": "",
@@ -11,7 +11,7 @@
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3-8B",
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
@@ -22,7 +22,7 @@
         "test_name": "serving_llama70B_tp4_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
         "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "tensor_parallel_size": 4,
             "swap_space": 16,
             "disable_log_stats": "",
@@ -30,7 +30,7 @@
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
@@ -55,5 +55,26 @@
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
             "num_prompts": 200
         }
+    },
+    {
+        "test_name": "serving_llama70B_tp4_sharegpt_specdecode",
+        "qps_list": [2],
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
+            "disable_log_requests": "", 
+            "tensor_parallel_size": 4,
+            "swap_space": 16, 
+            "speculative_model": "turboderp/Qwama-0.5B-Instruct",
+            "num_speculative_tokens": 4,
+            "speculative_draft_tensor_parallel_size": 1,
+            "use_v2_block_manager": ""
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200 
+        }
     }
-]
+]

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

@@ -2,7 +2,7 @@
     {
         "test_name": "throughput_llama8B_tp1",
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3-8B",
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
             "load_format": "dummy",
             "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
@@ -13,7 +13,7 @@
     {
         "test_name": "throughput_llama70B_tp4",
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "tensor_parallel_size": 4,
             "load_format": "dummy",
             "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",

.buildkite/release-pipeline.yaml

@@ -1,9 +1,27 @@
 steps:
-  - label: "Build wheel - CUDA {{matrix.cuda_version}}"
+  - label: "Build wheel - CUDA 12.1"
     agents:
       queue: cpu_queue
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg buildkite_commit=$BUILDKITE_COMMIT --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION={{matrix.cuda_version}} --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg buildkite_commit=$BUILDKITE_COMMIT --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "mkdir artifacts"
+      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+      # rename the files to change linux -> manylinux1
+      - "for f in artifacts/dist/*.whl; do mv -- \"$$f\" \"$${f/linux/manylinux1}\"; done"
+      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
+      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/nightly/"
+    env:
+      DOCKER_BUILDKIT: "1"
+
+  - block: "Build CUDA 11.8 wheel"
+    key: block-build-cu118-wheel
+  
+  - label: "Build wheel - CUDA 11.8"
+    depends_on: block-build-cu118-wheel
+    agents:
+      queue: cpu_queue
+    commands:
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg buildkite_commit=$BUILDKITE_COMMIT --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
       # rename the files to change linux -> manylinux1
@@ -12,8 +30,3 @@ steps:
       - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/nightly/"
     env:
       DOCKER_BUILDKIT: "1"
-    matrix:
-      setup:
-        cuda_version:
-          - "11.8.0"
-          - "12.1.0"

.buildkite/run-amd-test.sh

@@ -1,5 +1,5 @@
 # This script runs test inside the corresponding ROCm docker container.
-set -ex
+set -o pipefail
 
 # Print ROCm version
 echo "--- Confirming Clean Initial State"
@@ -55,7 +55,7 @@ while true; do
 done
 
 echo "--- Pulling container" 
-image_name="rocmshared/vllm-ci:${BUILDKITE_COMMIT}"
+image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 docker pull ${image_name}
 
@@ -70,15 +70,85 @@ HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p ${HF_CACHE}
 HF_MOUNT="/root/.cache/huggingface"
 
-docker run \
+commands=$@
+echo "Commands:$commands"
+#ignore certain kernels tests
+if [[ $commands == *" kernels "* ]]; then
+  commands="${commands} \
+  --ignore=kernels/test_attention.py \
+  --ignore=kernels/test_attention_selector.py \
+  --ignore=kernels/test_blocksparse_attention.py \
+  --ignore=kernels/test_causal_conv1d.py \
+  --ignore=kernels/test_cutlass.py \
+  --ignore=kernels/test_encoder_decoder_attn.py \
+  --ignore=kernels/test_flash_attn.py \
+  --ignore=kernels/test_flashinfer.py \
+  --ignore=kernels/test_gguf.py \
+  --ignore=kernels/test_int8_quant.py \
+  --ignore=kernels/test_machete_gemm.py \
+  --ignore=kernels/test_mamba_ssm.py \
+  --ignore=kernels/test_marlin_gemm.py \
+  --ignore=kernels/test_moe.py \
+  --ignore=kernels/test_prefix_prefill.py \
+  --ignore=kernels/test_rand.py \
+  --ignore=kernels/test_sampler.py"
+fi
+
+#ignore certain Entrypoints tests
+if [[ $commands == *" entrypoints/openai "* ]]; then
+  commands=${commands//" entrypoints/openai "/" entrypoints/openai \
+  --ignore=entrypoints/openai/test_accuracy.py \
+  --ignore=entrypoints/openai/test_audio.py \
+  --ignore=entrypoints/openai/test_encoder_decoder.py \
+  --ignore=entrypoints/openai/test_embedding.py \
+  --ignore=entrypoints/openai/test_oot_registration.py "}
+fi
+
+PARALLEL_JOB_COUNT=8
+# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
+if [[ $commands == *"--shard-id="* ]]; then
+  for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
+    #replace shard arguments
+    commands=${commands//"--shard-id= "/"--shard-id=${GPU} "}
+    commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
+    echo "Shard ${GPU} commands:$commands"
+    docker run \
         --device /dev/kfd --device /dev/dri \
         --network host \
         --shm-size=16gb \
         --rm \
+        -e HIP_VISIBLE_DEVICES=${GPU} \
         -e HF_TOKEN \
         -v ${HF_CACHE}:${HF_MOUNT} \
         -e HF_HOME=${HF_MOUNT} \
-        --name ${container_name} \
+        --name ${container_name}_${GPU}  \
         ${image_name} \
-        /bin/bash -c "${@}"
-
+        /bin/bash -c "${commands}" \
+        |& while read -r line; do echo ">>Shard $GPU: $line"; done &
+    PIDS+=($!)
+  done
+  #wait for all processes to finish and collect exit codes
+  for pid in ${PIDS[@]}; do
+    wait ${pid}
+    STATUS+=($?)
+  done
+  for st in ${STATUS[@]}; do
+    if [[ ${st} -ne 0 ]]; then
+      echo "One of the processes failed with $st"
+      exit ${st}
+    fi
+  done
+else
+  docker run \
+          --device /dev/kfd --device /dev/dri \
+          --network host \
+          --shm-size=16gb \
+          --rm \
+          -e HIP_VISIBLE_DEVICES=0 \
+          -e HF_TOKEN \
+          -v ${HF_CACHE}:${HF_MOUNT} \
+          -e HF_HOME=${HF_MOUNT} \
+          --name ${container_name} \
+          ${image_name} \
+          /bin/bash -c "${commands}"
+fi

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

@@ -0,0 +1,33 @@
+# This script build the CPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Try building the docker image
+docker build -t cpu-test -f Dockerfile.ppc64le .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f cpu-test || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image, setting --shm-size=4g for tensor parallel.
+source /etc/environment
+#docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test cpu-test
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN=$HF_TOKEN --name cpu-test cpu-test
+
+# Run basic model test
+docker exec cpu-test bash -c "
+  pip install pytest matplotlib einops transformers_stream_generator
+  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_oot_registration.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py --ignore=tests/models/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
+
+# online inference
+docker exec cpu-test bash -c "
+  python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m & 
+  timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
+  python3 benchmarks/benchmark_serving.py \
+    --backend vllm \
+    --dataset-name random \
+    --model facebook/opt-125m \
+    --num-prompts 20 \
+    --endpoint /v1/completions \
+    --tokenizer facebook/opt-125m"

.buildkite/run-cpu-test.sh

@@ -3,26 +3,47 @@
 set -ex
 
 # Try building the docker image
-docker build -t cpu-test -f Dockerfile.cpu .
-docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-avx2 -f Dockerfile.cpu .
+numactl -C 48-95 -N 1 docker build -t cpu-test -f Dockerfile.cpu .
+numactl -C 48-95 -N 1 docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-avx2 -f Dockerfile.cpu .
 
 # Setup cleanup
 remove_docker_container() { docker rm -f cpu-test cpu-test-avx2 || true; }
 trap remove_docker_container EXIT
 remove_docker_container
 
-# Run the image
+# Run the image, setting --shm-size=4g for tensor parallel.
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 \
-  --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test cpu-test
+ --cpuset-mems=1 --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test cpu-test
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 \
-  --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test-avx2 cpu-test-avx2
+ --cpuset-mems=1 --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-avx2 cpu-test-avx2
 
 # offline inference
-docker exec cpu-test bash -c "python3 examples/offline_inference.py"
 docker exec cpu-test-avx2 bash -c "python3 examples/offline_inference.py"
 
 # Run basic model test
-docker exec cpu-test bash -c "cd tests;
-  pip install pytest Pillow protobuf
-  cd ../
-  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py" # Mamba on CPU is not supported
+docker exec cpu-test bash -c "
+  pip install pytest matplotlib einops transformers_stream_generator datamodel_code_generator
+  pytest -v -s tests/models/decoder_only/language \
+    --ignore=tests/models/test_fp8.py \
+    --ignore=tests/models/decoder_only/language/test_jamba.py \
+    --ignore=tests/models/decoder_only/language/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
+
+# Run compressed-tensor test
+docker exec cpu-test bash -c "
+  pytest -s -v \
+  tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
+  tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynanmic_per_token"
+
+# online inference
+docker exec cpu-test bash -c "
+  export VLLM_CPU_KVCACHE_SPACE=10 
+  export VLLM_CPU_OMP_THREADS_BIND=48-92 
+  python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m & 
+  timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
+  python3 benchmarks/benchmark_serving.py \
+    --backend vllm \
+    --dataset-name random \
+    --model facebook/opt-125m \
+    --num-prompts 20 \
+    --endpoint /v1/completions \
+    --tokenizer facebook/opt-125m"

.buildkite/run-tpu-test.sh

@@ -12,5 +12,4 @@ remove_docker_container
 # For HF_TOKEN.
 source /etc/environment
 # Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu \
-    python3 /workspace/vllm/examples/offline_inference_tpu.py
+docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git && python3 -m pip install pytest  && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py && python3 /workspace/vllm/tests/tpu/test_compilation.py && python3 /workspace/vllm/examples/offline_inference_tpu.py"

.buildkite/test-pipeline.yaml

@@ -5,285 +5,471 @@
 # https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2 
 # to generate the final pipeline yaml file.
 
+# Documentation
+# label(str): the name of the test. emoji allowed.
+# fast_check(bool): whether to run this on each commit on fastcheck pipeline.
+# fast_check_only(bool): run this test on fastcheck pipeline only
+# command(str): the single command to run for tests. incompatible with commands.
+# commands(list): the list of commands to run for test. incompatbile with command.
+# mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
+# gpu(str): override the GPU selection for the test. default is on L4 GPUs. currently only supports a100
+# num_gpus(int): override the number of GPUs for the test. default to 1 GPU. currently support 2,4.
+# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host, 
+#     in this case, commands must be specified. the first command runs on first host, the second
+#     command runs on the second host.
+# working_dir(str): specify the place where command should execute, default to /vllm-workspace/tests
+# source_file_dependencies(list): the list of prefix to opt-in the test for, if empty, the test will always run.
+
+# When adding a test
+# - If the test belong to an existing group, add it there
+# - If the test is short, add to any existing step
+# - If the test takes more than 10min, then it is okay to create a new step. 
+#   Note that all steps execute in parallel. 
 
 steps:
-- label: Async Engine, Inputs, Utils, Worker Test
-  fast_check: true
-  fast_check_only: true
-  commands:
-  - pytest -v -s async_engine # Async Engine
-  - pytest -v -s test_inputs.py
-  - pytest -v -s multimodal
-  - pytest -v -s test_utils.py # Utils
-  - pytest -v -s worker # Worker
+##### fast check tests  #####
 
-- label: Tensorizer, Metrics, Tracing Test
-  fast_check: true
-  fast_check_only: true
-  commands:
-  - apt-get install -y curl libsodium23 && pytest -v -s tensorizer_loader # Tensorizer
-  - pytest -v -s metrics # Metrics
-  - "pip install \
-      opentelemetry-sdk \
-      opentelemetry-api \
-      opentelemetry-exporter-otlp \
-      opentelemetry-semantic-conventions-ai" # Tracing
-  - pytest -v -s tracing
-
-- label: Regression Test
-  mirror_hardwares: [amd]
-  fast_check: true
-  command: pytest -v -s test_regression.py
-  working_dir: "/vllm-workspace/tests" # optional
-
-- label: AsyncEngine Test
-  #mirror_hardwares: [amd]
-  command: pytest -v -s async_engine
-
-- label: Basic Correctness Test
-  mirror_hardwares: [amd]
-  fast_check: true
-  commands:
-  # This flashinfer installation will fail on AMD ROCm, so it is set as optional.
-  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl || true
-  - pytest -v -s basic_correctness/test_basic_correctness.py
-  - pytest -v -s basic_correctness/test_cpu_offload.py
-  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
-  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
-  - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
-
-- label: Core Test
-  mirror_hardwares: [amd]
-  fast_check: true
-  commands:
-  - pytest -v -s core
-  - pytest -v -s distributed/test_parallel_state.py
-
-- label: Distributed Comm Ops Test
-  #mirror_hardwares: [amd]
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
-  commands:
-  - pytest -v -s distributed/test_comm_ops.py
-  - pytest -v -s distributed/test_shm_broadcast.py
-
-- label: 2 Node Tests (4 GPUs in total)
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
-  num_nodes: 2
-  commands:
-  - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
-    - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
-  - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
-
-- label: Distributed Tests (2 GPUs)
-  mirror_hardwares: [amd]
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
-  commands:
-  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=microsoft/Phi-3-vision-128k-instruct DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=microsoft/Phi-3-vision-128k-instruct DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_multimodal_broadcast.py
-  - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py
-
-- label: Distributed Tests (4 GPUs)
-  #mirror_hardwares: [amd]
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
-  fast_check: true
-  commands:
-  - pytest -v -s distributed/test_pynccl.py
-  # We want to test that models which use 2 GPUs work with 4 GPUs, which is why we duplicate them here.
-  # See https://github.com/vllm-project/vllm/pull/5473#issuecomment-2166601837 for context.
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
-
-- label: Pipeline Parallelism Test
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
-  commands:
-  - pytest -v -s distributed/test_pipeline_parallel.py
-
-- label: Engine Test
-  mirror_hardwares: [amd]
-  commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
-  # OOM in the CI unless we run this separately
-  - pytest -v -s tokenization
-
-- label: Entrypoints Test
-  fast_check: true
-  mirror_hardwares: [amd]
-
-  commands:
-  - pytest -v -s entrypoints/llm
-  - pytest -v -s entrypoints/openai
-
-- label: Examples Test
-  working_dir: "/vllm-workspace/examples"
-  mirror_hardwares: [amd]
-  commands:
-    # install aws cli for llava_example.py
-    # install tensorizer for tensorize_vllm_model.py
-    - pip install awscli tensorizer
-    - python3 offline_inference.py
-    - python3 cpu_offload.py
-    - python3 offline_inference_with_prefix.py
-    - python3 llm_engine_example.py
-    - python3 llava_example.py
-    - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-
-- label: Inputs Test
-  #mirror_hardwares: [amd]
-  commands:
-    - pytest -v -s test_inputs.py
-    - pytest -v -s multimodal
-
-- label: Kernels Test %N
-  #mirror_hardwares: [amd]
-  commands:
-    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
-    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 4
-
-- label: Models Test
-  #mirror_hardwares: [amd]
-  commands:
-    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
-    - pytest -v -s models -m \"not vlm\"
-
-- label: Vision Language Models Test
-  mirror_hardwares: [amd]
-  commands:
-    - pytest -v -s models -m vlm
-
-- label: Prefix Caching Test
-  mirror_hardwares: [amd]
-  commands:
-    - pytest -v -s prefix_caching
-
-- label: Samplers Test
-  #mirror_hardwares: [amd]
-  command: pytest -v -s samplers
-
-- label: LogitsProcessor Test
-  mirror_hardwares: [amd]
-  command: pytest -v -s test_logits_processor.py
-
-- label: Utils Test
-  commands:
-    - pytest -v -s test_utils.py
-    - pytest -v -s test_embedded_commit.py
-
-- label: Worker Test
-  mirror_hardwares: [amd]
-  command: pytest -v -s worker
-
-- label: Speculative decoding tests
-  #mirror_hardwares: [amd]
-  commands:
-    # See https://github.com/vllm-project/vllm/issues/5152
-    - export VLLM_ATTENTION_BACKEND=XFORMERS
-    - pytest -v -s spec_decode
-
-- label: LoRA Test %N
-  #mirror_hardwares: [amd]
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py
-  parallelism: 4
-
-- label: LoRA Long Context (Distributed)
-  #mirror_hardwares: [amd]
-  num_gpus: 4
-  # This test runs llama 13B, so it is required to run on 4 GPUs.
-  commands:
-    # FIXIT: find out which code initialize cuda before running the test
-    # before the fix, we need to use spawn to test it
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -v -s -x lora/test_long_context.py
-
-- label: Tensorizer Test
-  #mirror_hardwares: [amd]
-  commands:
-    - apt-get install -y curl libsodium23
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -v -s tensorizer_loader
-
-- label: Metrics Test
-  mirror_hardwares: [amd]
-  command: pytest -v -s metrics
-
-- label: Quantization Test
-  #mirror_hardwares: [amd]
-  command: pytest -v -s quantization
-
-- label: Tracing Test
-  commands: 
-    - "pip install \
-        opentelemetry-sdk \
-        opentelemetry-api \
-        opentelemetry-exporter-otlp \
-        opentelemetry-semantic-conventions-ai"
-    - pytest -v -s tracing
-
-- label: Benchmarks
-  working_dir: "/vllm-workspace/.buildkite"
-  mirror_hardwares: [amd]
-  commands:
-  - pip install aiohttp
-  - bash run-benchmarks.sh
-
-- label: LM Eval Small Models
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  commands:
-  - pip install lm-eval
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - bash ./run-tests.sh -c configs/models-small.txt -t 1
-
-- label: LM Eval Large Models
-  gpu: a100
-  num_gpus: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  commands:
-  - pip install lm-eval
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - bash ./run-tests.sh -c configs/models-large.txt -t 4
-
-- label: Documentation Build
+- label: Documentation Build # 2min
   working_dir: "/vllm-workspace/test_docs/docs"
   fast_check: true
   no_gpu: True
   commands:
   - pip install -r requirements-docs.txt
   - SPHINXOPTS=\"-W\" make html
+  # Check API reference (if it fails, you may have missing mock imports)
+  - grep \"sig sig-object py\" build/html/dev/sampling_params.html
 
-- label: Distributed Tests (A100)
+- label: Async Engine, Inputs, Utils, Worker Test # 15min
+  fast_check: true
+  source_file_dependencies:
+  - vllm/
+  - tests/mq_llm_engine
+  - tests/async_engine
+  - tests/test_inputs
+  - tests/multimodal
+  - tests/test_utils
+  - tests/worker
+  commands:
+  - pytest -v -s mq_llm_engine # MQLLMEngine
+  - pytest -v -s async_engine # AsyncLLMEngine
+  - NUM_SCHEDULER_STEPS=4 pytest -v -s async_engine/test_async_llm_engine.py
+  - pytest -v -s test_inputs.py
+  - pytest -v -s multimodal
+  - pytest -v -s test_utils.py # Utils
+  - pytest -v -s worker # Worker
+
+- label: Basic Correctness Test # 30min
+  #mirror_hardwares: [amd]
+  fast_check: true
+  source_file_dependencies:
+  - vllm/
+  - tests/basic_correctness
+  commands:
+  - pytest -v -s basic_correctness/test_basic_correctness.py
+  - pytest -v -s basic_correctness/test_cpu_offload.py
+  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
+
+- label: Core Test # 10min
+  mirror_hardwares: [amd]
+  fast_check: true
+  source_file_dependencies:
+  - vllm/core
+  - vllm/distributed
+  - tests/core
+  commands:
+  - pytest -v -s core
+
+- label: Entrypoints Test # 20min
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  commands:
+  - pip install -e ./plugins/vllm_add_dummy_model
+  - pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@a4987bba6e9e9b3f22bd3a6c1ecf0abd04fd5622#egg=lm_eval[api]
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py
+  - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
+  - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
+  - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
+  - pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
+  - pytest -v -s entrypoints/openai
+  - pytest -v -s entrypoints/test_chat_utils.py
+  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+
+- label: Distributed Tests (4 GPUs) # 10min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  fast_check: true
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/core/
+  - tests/distributed
+  - tests/spec_decode/e2e/test_integration_dist_tp4
+  commands:
+  - pytest -v -s distributed/test_pynccl.py
+  - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
+
+- label: Metrics, Tracing Test # 10min
+  num_gpus: 2 
+  fast_check: true
+  source_file_dependencies:
+  - vllm/
+  - tests/metrics
+  - tests/tracing
+  commands:
+  - pytest -v -s metrics 
+  - "pip install \
+      'opentelemetry-sdk>=1.26.0,<1.27.0' \
+      'opentelemetry-api>=1.26.0,<1.27.0' \
+      'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
+      'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'"
+  - pytest -v -s tracing
+
+##### fast check tests  #####
+#####  1 GPU test  #####
+
+- label: Regression Test # 5min
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/test_regression
+  command: pytest -v -s test_regression.py
+  working_dir: "/vllm-workspace/tests" # optional
+
+- label: Engine Test # 10min
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/engine
+  - tests/tokenization
+  commands:
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
+  # OOM in the CI unless we run this separately
+  - pytest -v -s tokenization
+
+- label: Examples Test # 12min
+  working_dir: "/vllm-workspace/examples"
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/entrypoints
+  - examples/
+  commands:
+    - pip install awscli tensorizer # for llava example and tensorizer test
+    - python3 offline_inference.py
+    - python3 cpu_offload.py
+    - python3 offline_inference_chat.py
+    - python3 offline_inference_with_prefix.py
+    - python3 llm_engine_example.py
+    - python3 offline_inference_vision_language.py
+    - python3 offline_inference_vision_language_multi_image.py
+    - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - python3 offline_inference_encoder_decoder.py
+
+- label: Prefix Caching Test # 7min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/prefix_caching
+  commands:
+    - pytest -v -s prefix_caching
+
+- label: Samplers Test # 18min
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/sampling_metadata.py
+  - tests/samplers
+  commands:
+    - pytest -v -s samplers
+    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
+
+- label: LogitsProcessor Test # 5min
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - tests/test_logits_processor
+  command: pytest -v -s test_logits_processor.py
+
+- label: Speculative decoding tests # 22min
+  source_file_dependencies:
+  - vllm/spec_decode
+  - tests/spec_decode
+  commands:
+    # See https://github.com/vllm-project/vllm/issues/5152
+    - export VLLM_ATTENTION_BACKEND=XFORMERS
+    - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
+    - pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py
+
+- label: LoRA Test %N # 30min each
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/lora
+  - tests/lora
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py
+  parallelism: 4
+
+- label: "PyTorch Fullgraph Smoke Test"
+  fast_check: true
+  source_file_dependencies:
+  - vllm/
+  - tests/compile
+  commands:
+  - pytest -v -s compile/test_full_graph_smoke.py
+
+- label: "PyTorch Fullgraph Test"
+  source_file_dependencies:
+  - vllm/
+  - tests/compile
+  commands:
+  - pytest -v -s compile/test_full_graph.py
+
+- label: Kernels Test %N # 30min each
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - csrc/
+  - vllm/attention
+  - tests/kernels
+  commands:
+    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 4
+
+- label: Tensorizer Test # 11min
+  mirror_hardwares: [amd]
+  soft_fail: true
+  source_file_dependencies:
+  - vllm/model_executor/model_loader
+  - tests/tensorizer_loader
+  commands:
+    - apt-get update && apt-get install -y curl libsodium23
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s tensorizer_loader
+
+- label: Benchmarks # 9min
+  working_dir: "/vllm-workspace/.buildkite"
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - benchmarks/
+  commands:
+  - pip install aiohttp
+  - bash run-benchmarks.sh
+
+- label: Quantization Test # 15min
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  - tests/quantization
+  command: pytest -v -s quantization
+
+- label: LM Eval Small Models # 53min
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  commands:
+  - pip install lm-eval
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - bash ./run-tests.sh -c configs/models-small.txt -t 1
+
+- label: Encoder Decoder tests # 5min
+  source_file_dependencies:
+  - vllm/
+  - tests/encoder_decoder
+  commands:
+    - pytest -v -s encoder_decoder
+
+- label: OpenAI-Compatible Tool Use # 20 min
+  fast_check: false
+  mirror_hardwares: [ amd ]
+  source_file_dependencies:
+    - vllm/
+    - tests/tool_use
+  commands:
+    - pytest -v -s tool_use
+
+#####  models test  #####
+
+- label: Basic Models Test # 3min
+  source_file_dependencies:
+  - vllm/
+  - tests/models
+  commands:
+    - pip install -e ./plugins/vllm_add_dummy_model
+    - pytest -v -s models/test_oot_registration.py # it needs a clean process
+    - pytest -v -s models/*.py --ignore=models/test_oot_registration.py
+
+- label: Decoder-only Language Models Test # 1h3min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/models/decoder_only/language
+  commands:
+    - pytest -v -s models/decoder_only/language
+
+- label: Decoder-only Multi-Modal Models Test # 56min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/models/decoder_only/audio_language
+  - tests/models/decoder_only/vision_language
+  commands:
+    - pytest -v -s models/decoder_only/audio_language
+    - pytest -v -s models/decoder_only/vision_language
+
+- label: Other Models Test # 5min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/models/embedding/language
+  - tests/models/encoder_decoder/language
+  commands:
+    - pytest -v -s models/embedding/language
+    - pytest -v -s models/encoder_decoder/language
+
+#####  1 GPU test  #####
+#####  multi gpus test  #####
+
+- label: Distributed Comm Ops Test # 7min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  source_file_dependencies:
+  - vllm/distributed
+  - tests/distributed
+  commands:
+  - pytest -v -s distributed/test_comm_ops.py
+  - pytest -v -s distributed/test_shm_broadcast.py
+
+- label: 2 Node Tests (4 GPUs in total) # 16min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  num_nodes: 2
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/model_executor/models/
+  - tests/distributed/
+  commands:
+  - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
+    - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
+    - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
+  - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
+
+- label: Distributed Tests (2 GPUs) # 28min
+  #mirror_hardwares: [amd]
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/model_executor/models/
+  - tests/distributed/
+  - vllm/compilation
+  commands:
+  - pytest -v -s ./compile/test_full_graph_multi_gpu.py
+  - pytest -v -s ./compile/test_wrapper.py
+  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep -q 'Same node test passed'
+  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m distributed_2_gpus
+  # Avoid importing model tests that cause CUDA reinitialization error
+  - pytest models/encoder_decoder/language/test_bart.py models/decoder_only/vision_language/test_broadcast.py -v -s -m distributed_2_gpus
+  - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
+  - pip install -e ./plugins/vllm_add_dummy_model
+  - pytest -v -s distributed/test_distributed_oot.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py
+
+- label: Multi-step Tests (4 GPUs) # 21min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+  - vllm/model_executor/layers/sampler.py
+  - vllm/sequence.py
+  - vllm/worker/worker_base.py
+  - vllm/worker/worker.py
+  - vllm/worker/multi_step_worker.py
+  - vllm/worker/model_runner_base.py
+  - vllm/worker/model_runner.py
+  - vllm/worker/multi_step_model_runner.py
+  - vllm/engine
+  - tests/multi_step
+  commands:
+  - pytest -v -s multi_step/test_correctness_async_llm.py
+  - pytest -v -s multi_step/test_correctness_llm.py
+
+- label: Pipeline Parallelism Test # 23min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/model_executor/models/
+  - tests/distributed/
+  commands:
+  - pytest -v -s distributed/test_pp_cudagraph.py
+  - pytest -v -s distributed/test_pipeline_parallel.py
+
+- label: LoRA Long Context (Distributed) # 11min
+  # This test runs llama 13B, so it is required to run on 4 GPUs.
+  num_gpus: 4
+  soft_fail: true
+  source_file_dependencies:
+  - vllm/lora
+  - tests/lora/test_long_context
+  commands:
+    # FIXIT: find out which code initialize cuda before running the test
+    # before the fix, we need to use spawn to test it
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s -x lora/test_long_context.py
+
+- label: Weight Loading Multiple GPU Test
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  source_file_dependencies:
+  - vllm/
+  - tests/weight_loading
+  commands:
+    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
+
+- label: Weight Loading Multiple GPU Test - Large Models # optional
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  gpu: a100
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/weight_loading
+  commands:
+    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt 
+
+
+##### multi gpus test #####
+##### A100 test #####
+
+- label: Distributed Tests (A100) # optional
   gpu: a100
   num_gpus: 4
+  source_file_dependencies:
+  - vllm/
   commands: 
   # NOTE: don't test llama model here, it seems hf implementation is buggy
   # see https://github.com/vllm-project/vllm/pull/5689 for details
   - pytest -v -s distributed/test_custom_all_reduce.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
-  - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=meta-llama/Meta-Llama-3-8B DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TARGET_TEST_SUITE=A100 pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s -x lora/test_mixtral.py
+
+- label: LM Eval Large Models # optional
+  gpu: a100
+  num_gpus: 4
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  commands:
+  - pip install lm-eval
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - bash ./run-tests.sh -c configs/models-large.txt -t 4

.dockerignore

@@ -1 +1,4 @@
 vllm/*.so
+/.venv
+/build
+dist

.github/ISSUE_TEMPLATE/100-documentation.yml

@@ -20,3 +20,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/200-installation.yml

@@ -38,3 +38,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/300-usage.yml

@@ -36,3 +36,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/400-bug report.yml

@@ -20,11 +20,25 @@ body:
       ```
       It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
+      <details>
+      <summary>The output of `python collect_env.py`</summary>
+
       ```text
-      The output of `python collect_env.py`
+      Your output of `python collect_env.py` here
       ```
+      
+      </details>
   validations:
     required: true
+- type: textarea
+  attributes:
+    label: Model Input Dumps
+    description: |
+      If you are facing crashing due to illegal memory access or other issues with model execution, vLLM may dump the problematic input of the model. In this case, you will see the message `Error in model execution (input dumped to /tmp/err_xxx.pkl)`. If you see this message, please zip the file (because GitHub doesn't support .pkl file format) and upload it here. This will help us to reproduce the issue and facilitate the debugging process.
+    placeholder: |
+      Upload the dumped input file.
+  validations:
+    required: false
 - type: textarea
   attributes:
     label: 🐛 Describe the bug
@@ -84,3 +98,10 @@ body:
       - If the error only appears in vllm, please provide the detailed script of how you run `transformers` and `vllm`, also highlight the difference and what you expect.
 
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/500-feature request.yml

@@ -29,3 +29,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/600-new model.yml

@@ -31,3 +31,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/700-performance discussion.yml

@@ -50,3 +50,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -47,3 +47,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/ISSUE_TEMPLATE/800-misc discussion.yml

@@ -19,3 +19,10 @@ body:
   attributes:
     value: >
       Thanks for contributing 🎉!
+- type: checkboxes
+  id: askllm
+  attributes:
+    label: Before submitting a new issue...
+    options:
+      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
+        required: true

.github/PULL_REQUEST_TEMPLATE.md

@@ -39,6 +39,16 @@ FIX #xxxx (*link existing issues this PR will resolve*)
     <li>Please add documentation to <code>docs/source/</code> if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.</li>
 </ul>
 
+<h3>Adding or changing kernels</h3>
+<p>Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.</p>
+<ul>
+    <li>Make sure custom ops are registered following PyTorch guidelines: <a href="https://pytorch.org/tutorials/advanced/cpp_custom_ops.html#cpp-custom-ops-tutorial">Custom C++ and CUDA Operators</a> and <a href="https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU">The Custom Operators Manual</a></li>
+    <li>Custom operations that return <code>Tensors</code> require meta-functions. Meta-functions should be implemented and registered in python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions.</li>
+    <li>Use <a href="https://pytorch.org/docs/stable/library.html#torch.library.opcheck"><code>torch.libary.opcheck()</code></a> to test the function registration and meta-function for any registered ops.  See <code>tests/kernels</code> for examples.</li>
+    <li>When changing the C++ signature of an existing op, the schema must be updated to reflect the changes.</li>
+    <li>If a new custom type is needed, see the following document: <a href="https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA">Custom Class Support in PT2</a>.
+</ul>
+
 <h3>Notes for Large Changes</h3>
 <p>Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with <code>rfc-required</code> and might not go through the PR.</p>
 

.github/workflows/add_label_ready_comment.yml

@@ -1,23 +0,0 @@
-name: Add Ready Label on Ready Comment
-
-on:
-  issue_comment:
-    types: [created]
-
-jobs:
-  add-ready-label:
-    runs-on: ubuntu-latest
-    if: github.event.issue.pull_request && contains(github.event.comment.body, '/ready')
-    steps:
-        -   name: Add label
-            uses: actions/github-script@v5
-            with:
-                script: |
-                    github.rest.issues.addLabels({
-                        owner: context.repo.owner,
-                        repo: context.repo.repo,
-                        issue_number: context.issue.number,
-                        labels: ['ready']
-                    })
-            env:
-                GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/clang-format.yml

@@ -30,12 +30,11 @@ jobs:
       run: |
         EXCLUDES=(
             'csrc/moe/topk_softmax_kernels.cu'
-            'csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu'
-            'csrc/punica/bgmv/bgmv_config.h'
-            'csrc/punica/bgmv/bgmv_impl.cuh'
-            'csrc/punica/bgmv/vec_dtypes.cuh'
-            'csrc/punica/punica_ops.cu'
-            'csrc/punica/type_convert.h'
+            'csrc/quantization/gguf/ggml-common.h'
+            'csrc/quantization/gguf/dequantize.cuh'
+            'csrc/quantization/gguf/vecdotq.cuh'
+            'csrc/quantization/gguf/mmq.cuh'
+            'csrc/quantization/gguf/mmvq.cuh'
         )
         find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \
             | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \

.github/workflows/mypy.yaml

@@ -15,7 +15,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11"]
+        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}
@@ -25,29 +25,22 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install mypy==1.9.0
+        pip install mypy==1.11.1
         pip install types-setuptools
         pip install types-PyYAML
         pip install types-requests
         pip install types-setuptools
     - name: Mypy
       run: |
-        mypy tests --config-file pyproject.toml
-        mypy vllm/*.py --config-file pyproject.toml
-        mypy vllm/attention --config-file pyproject.toml
-        mypy vllm/core --config-file pyproject.toml
-        mypy vllm/distributed --config-file pyproject.toml
-        mypy vllm/engine  --config-file pyproject.toml
-        mypy vllm/entrypoints --config-file pyproject.toml
-        mypy vllm/executor --config-file pyproject.toml
-        mypy vllm/inputs --config-file pyproject.toml
-        mypy vllm/logging --config-file pyproject.toml
-        mypy vllm/lora --config-file pyproject.toml
-        mypy vllm/model_executor  --config-file pyproject.toml
-        mypy vllm/multimodal --config-file pyproject.toml
-        mypy vllm/platforms --config-file pyproject.toml
-        mypy vllm/spec_decode --config-file pyproject.toml
-        mypy vllm/transformers_utils --config-file pyproject.toml
-        mypy vllm/usage --config-file pyproject.toml
-        mypy vllm/worker --config-file pyproject.toml
+        mypy
+        mypy tests --follow-imports skip
+        mypy vllm/attention --follow-imports skip
+        mypy vllm/distributed --follow-imports skip
+        mypy vllm/engine  --follow-imports skip
+        mypy vllm/executor --follow-imports skip
+        mypy vllm/lora --follow-imports skip
+        mypy vllm/model_executor  --follow-imports skip
+        mypy vllm/prompt_adapter --follow-imports skip
+        mypy vllm/spec_decode --follow-imports skip
+        mypy vllm/worker --follow-imports skip
 

.github/workflows/publish.yml

@@ -48,8 +48,8 @@ jobs:
       fail-fast: false
       matrix:
           os: ['ubuntu-20.04']
-          python-version: ['3.8', '3.9', '3.10', '3.11']
-          pytorch-version: ['2.3.1']  # Must be the most recent version that meets requirements-cuda.txt.
+          python-version: ['3.8', '3.9', '3.10', '3.11', '3.12']
+          pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
           cuda-version: ['11.8', '12.1']
 
     steps:

.github/workflows/reminder_comment.yml

@@ -15,7 +15,7 @@ jobs:
               owner: context.repo.owner,
               repo: context.repo.repo,
               issue_number: context.issue.number,
-              body: '👋 Hi! Thank you for contributing to the vLLM project.\n Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your `fast-check` build on Buildkite UI. \n\nOnce the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).\n\n To run full CI, you can do one of these:\n- Comment `/ready` on the PR\n- Add `ready` label to the PR\n- Enable auto-merge.\n\n🚀'
+              body: '👋 Hi! Thank you for contributing to the vLLM project.\n Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your `fastcheck` build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping `simon-mo` or `khluu` to add you in our Buildkite org. \n\nOnce the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n To run CI, PR reviewers can do one of these:\n- Add `ready` label to the PR\n- Enable auto-merge.\n\n🚀'
             })
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/ruff.yml

@@ -15,7 +15,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11"]
+        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}
@@ -25,10 +25,10 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install ruff==0.1.5 codespell==2.3.0 tomli==2.0.1 isort==5.13.2
+        pip install -r requirements-lint.txt
     - name: Analysing the code with ruff
       run: |
-        ruff .
+        ruff check .
     - name: Spelling check with codespell
       run: |
         codespell --toml pyproject.toml

.github/workflows/scripts/build.sh

@@ -13,9 +13,8 @@ $python_executable -m pip install -r requirements-cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
-# Make sure punica is built for the release (for LoRA)
-export VLLM_INSTALL_PUNICA_KERNELS=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"
 # Build
 $python_executable setup.py bdist_wheel --dist-dir=dist

.github/workflows/yapf.yml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11"]
+        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}

.gitignore

@@ -1,5 +1,8 @@
-# vllm commit id, generated by setup.py
-vllm/commit_id.py
+# version file generated by setuptools-scm
+/vllm/_version.py
+
+# vllm-flash-attn built from source
+vllm/vllm_flash_attn/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -12,6 +15,8 @@ __pycache__/
 # Distribution / packaging
 .Python
 build/
+cmake-build-*/
+CMakeUserPresets.json
 develop-eggs/
 dist/
 downloads/
@@ -87,6 +92,9 @@ target/
 profile_default/
 ipython_config.py
 
+# generated files
+**/generated/**
+
 # pyenv
 #   For a library or package, you might want to ignore these files since the code is
 #   intended to run in multiple environments; otherwise, check them in:
@@ -189,4 +197,4 @@ _build/
 hip_compat.h
 
 # Benchmark dataset
-*.json
+benchmarks/*.json

.readthedocs.yaml

@@ -10,6 +10,7 @@ build:
 
 sphinx:
    configuration: docs/source/conf.py
+   fail_on_warning: true
 
 # If using Sphinx, optionally build your docs in additional formats such as PDF
 formats:

CMakeLists.txt

@@ -1,5 +1,16 @@
-cmake_minimum_required(VERSION 3.21)
+cmake_minimum_required(VERSION 3.26)
 
+# When building directly using CMake, make sure you run the install step
+# (it places the .so files in the correct location).
+#
+# Example:
+# mkdir build && cd build
+# cmake -G Ninja -DVLLM_PYTHON_EXECUTABLE=`which python3` -DCMAKE_INSTALL_PREFIX=.. ..
+# cmake --build . --target install
+#
+# If you want to only build one target, make sure to install it manually:
+# cmake --build . --target _C
+# cmake --install . --component _C
 project(vllm_extensions LANGUAGES CXX)
 
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
@@ -10,11 +21,17 @@ message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
 
 include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 
+# Suppress potential warnings about unused manually-specified variables
+set(ignoreMe "${VLLM_PYTHON_PATH}")
+
+# Prevent installation of dependencies (cutlass) by default.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11")
+set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11" "3.12")
 
 # Supported NVIDIA architectures.
 set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.9;9.0")
@@ -32,7 +49,7 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx11
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.3.1")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.4.0")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.5.0")
 
 #
@@ -66,6 +83,24 @@ endif()
 #
 find_package(Torch REQUIRED)
 
+#
+message(STATUS "Enabling core extension.")
+
+# Define _core_C extension
+#  built for (almost) every target platform, (excludes TPU and Neuron)
+
+set(VLLM_EXT_SRC
+  "csrc/core/torch_bindings.cpp")
+
+define_gpu_extension_target(
+  _core_C
+  DESTINATION vllm
+  LANGUAGE CXX
+  SOURCES ${VLLM_EXT_SRC}
+  COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
+  USE_SABI 3
+  WITH_SOABI)
+
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -74,7 +109,7 @@ if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda" AND
     if (VLLM_TARGET_DEVICE STREQUAL "cpu")
         include(${CMAKE_CURRENT_LIST_DIR}/cmake/cpu_extension.cmake)
     else()
-        message(FATAL_ERROR "Unsupported vLLM target device: ${VLLM_TARGET_DEVICE}")
+        return()
     endif()
     return()
 endif()
@@ -131,8 +166,10 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
+include(FetchContent)
+
 #
-# Define extension targets
+# Define other extension targets
 #
 
 #
@@ -145,7 +182,6 @@ set(VLLM_EXT_SRC
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
-  "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
@@ -155,26 +191,39 @@ set(VLLM_EXT_SRC
   "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  include(FetchContent)
-  SET(CUTLASS_ENABLE_HEADERS_ONLY=ON)
+  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
+
+  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
+  set(CUTLASS_REVISION "v3.5.1" CACHE STRING "CUTLASS revision to use")
+
   FetchContent_Declare(
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        # CUTLASS 3.5.0
-        GIT_TAG 7d49e6c7e2f8896c47f586706e67e1fb215529dc
+        GIT_TAG v3.5.1
+        GIT_PROGRESS TRUE
+
+        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
+        # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
+        # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
+        GIT_SHALLOW TRUE
   )
   FetchContent_MakeAvailable(cutlass)
 
   list(APPEND VLLM_EXT_SRC
+    "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
+    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
     "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
+    "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
     "csrc/quantization/gptq_marlin/awq_marlin_repack.cu"
+    "csrc/quantization/gguf/gguf_kernel.cu"
     "csrc/quantization/fp8/fp8_marlin.cu"
     "csrc/custom_all_reduce.cu"
+    "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu")
@@ -191,8 +240,55 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
           "-gencode arch=compute_90a,code=sm_90a")
   endif()
 
+
+  #
+  # Machete kernels
+
+  # The machete kernels only work on hopper and require CUDA 12.0 or later.
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0)
+    #
+    # For the Machete kernels we automatically generate sources for various 
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    execute_process(
+      COMMAND ${CMAKE_COMMAND} -E env 
+      PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH 
+        ${Python_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/machete/generate.py
+      RESULT_VARIABLE machete_generation_result
+      OUTPUT_VARIABLE machete_generation_output
+      OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log
+      ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log
+    )
+
+    if (NOT machete_generation_result EQUAL 0)
+      message(FATAL_ERROR "Machete generation failed."
+                          " Result: \"${machete_generation_result}\"" 
+                          "\nCheck the log for details: "
+                          "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
+    else()
+      message(STATUS "Machete generation completed successfully.")
+    endif()
+
+    # Add machete generated sources
+    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
+    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
+    message(STATUS "Machete generated sources: ${MACHETE_GEN_SOURCES}")
+
+    set_source_files_properties(
+          ${MACHETE_GEN_SOURCES}
+          PROPERTIES
+          COMPILE_FLAGS
+          "-gencode arch=compute_90a,code=sm_90a")
+  endif()
+
+  # Add pytorch binding for machete (add on even CUDA < 12.0 so that we can
+  #  raise an error if the user that this was built with an incompatible 
+  #  CUDA version)
+  list(APPEND VLLM_EXT_SRC
+    csrc/quantization/machete/machete_pytorch.cu)
 endif()
 
+message(STATUS "Enabling C extension.")
 define_gpu_extension_target(
   _C
   DESTINATION vllm
@@ -200,10 +296,16 @@ define_gpu_extension_target(
   SOURCES ${VLLM_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
-  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
+# If CUTLASS is compiled on NVCC >= 12.5, it by default uses 
+# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the 
+# driver API. This causes problems when linking with earlier versions of CUDA.
+# Setting this variable sidesteps the issue by calling the driver directly.
+target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+
 #
 # _moe_C extension
 #
@@ -212,6 +314,17 @@ set(VLLM_MOE_EXT_SRC
   "csrc/moe/torch_bindings.cpp"
   "csrc/moe/topk_softmax_kernels.cu")
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_MOE_EXT_SRC
+      "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
+      "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h"
+      "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu"
+      "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h"
+      "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu"
+      "csrc/moe/marlin_moe_ops.cu")
+endif()
+
+message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C
   DESTINATION vllm
@@ -222,90 +335,85 @@ define_gpu_extension_target(
   USE_SABI 3
   WITH_SOABI)
 
-#
-# _punica_C extension
-#
+if(VLLM_GPU_LANG STREQUAL "HIP")
+  #
+  # _rocm_C extension
+  #
+  set(VLLM_ROCM_EXT_SRC
+    "csrc/rocm/torch_bindings.cpp"
+    "csrc/rocm/attention.cu")
 
-set(VLLM_PUNICA_EXT_SRC
-  "csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
-  "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
-  "csrc/punica/punica_ops.cu"
-  "csrc/punica/torch_bindings.cpp")
-
-#
-# Copy GPU compilation flags+update for punica
-#
-set(VLLM_PUNICA_GPU_FLAGS ${VLLM_GPU_FLAGS})
-list(REMOVE_ITEM VLLM_PUNICA_GPU_FLAGS
-  "-D__CUDA_NO_HALF_OPERATORS__"
-  "-D__CUDA_NO_HALF_CONVERSIONS__"
-  "-D__CUDA_NO_BFLOAT16_CONVERSIONS__"
-  "-D__CUDA_NO_HALF2_OPERATORS__")
-
-#
-# Filter out CUDA architectures < 8.0 for punica.
-#
-if (${VLLM_GPU_LANG} STREQUAL "CUDA")
-  set(VLLM_PUNICA_GPU_ARCHES)
-  foreach(ARCH ${VLLM_GPU_ARCHES})
-    string_to_ver(CODE_VER ${ARCH})
-    if (CODE_VER GREATER_EQUAL 8.0)
-      list(APPEND VLLM_PUNICA_GPU_ARCHES ${ARCH})
-    endif()
-  endforeach()
-  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
-elseif(${VLLM_GPU_LANG} STREQUAL "HIP")
-  set(VLLM_PUNICA_GPU_ARCHES ${VLLM_GPU_ARCHES})
-  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
-endif()
-
-if (VLLM_PUNICA_GPU_ARCHES)
   define_gpu_extension_target(
-    _punica_C
+    _rocm_C
     DESTINATION vllm
     LANGUAGE ${VLLM_GPU_LANG}
-    SOURCES ${VLLM_PUNICA_EXT_SRC}
-    COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
-    ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    SOURCES ${VLLM_ROCM_EXT_SRC}
+    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_GPU_ARCHES}
     USE_SABI 3
     WITH_SOABI)
+endif()
+
+# vllm-flash-attn currently only supported on CUDA
+if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda")
+  return()
+endif ()
+
+#
+# Build vLLM flash attention from source
+#
+# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
+# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
+# They should be identical but if they aren't, this is a massive footgun.
+#
+# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
+# To only install vllm-flash-attn, use --component vllm_flash_attn_c.
+# If no component is specified, vllm-flash-attn is still installed.
+
+# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
+# This is to enable local development of vllm-flash-attn within vLLM.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
+  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
+endif()
+
+if(VLLM_FLASH_ATTN_SRC_DIR)
+  FetchContent_Declare(vllm-flash-attn SOURCE_DIR ${VLLM_FLASH_ATTN_SRC_DIR})
 else()
-  message(WARNING "Unable to create _punica_C target because none of the "
-    "requested architectures (${VLLM_GPU_ARCHES}) are supported, i.e. >= 8.0")
+  FetchContent_Declare(
+          vllm-flash-attn
+          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
+          GIT_TAG 013f0c4fc47e6574060879d9734c1df8c5c273bd
+          GIT_PROGRESS TRUE
+  )
 endif()
 
-#
-# Add the `default` target which detects which extensions should be
-# built based on platform/architecture.  This is the same logic that
-# setup.py uses to select which extensions should be built and should
-# be kept in sync.
-#
-# The `default` target makes direct use of cmake easier since knowledge
-# of which extensions are supported has been factored in, e.g.
-#
-# mkdir build && cd build
-# cmake -G Ninja -DVLLM_PYTHON_EXECUTABLE=`which python3` -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=../vllm ..
-# cmake --build . --target default
-#
-add_custom_target(default)
+# Set the parent build flag so that the vllm-flash-attn library does not redo compile flag and arch initialization.
+set(VLLM_PARENT_BUILD ON)
 
-if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
-  message(STATUS "Enabling C extension.")
-  add_dependencies(default _C)
+# Ensure the vllm/vllm_flash_attn directory exists before installation
+install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" COMPONENT vllm_flash_attn_c)
 
-  message(STATUS "Enabling moe extension.")
-  add_dependencies(default _moe_C)
+# Make sure vllm-flash-attn install rules are nested under vllm/
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT vllm_flash_attn_c)
+install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" COMPONENT vllm_flash_attn_c)
 
-  # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
-  # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
-  # there are supported target arches.
-  if (VLLM_PUNICA_GPU_ARCHES AND
-      (ENV{VLLM_INSTALL_PUNICA_KERNELS} OR VLLM_INSTALL_PUNICA_KERNELS))
-    message(STATUS "Enabling punica extension.")
-    add_dependencies(default _punica_C)
-  endif()
-endif()
+# Fetch the vllm-flash-attn library
+FetchContent_MakeAvailable(vllm-flash-attn)
+message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
+
+# Restore the install prefix
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" COMPONENT vllm_flash_attn_c)
+
+# Copy over the vllm-flash-attn python files
+install(
+        DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+        DESTINATION vllm/vllm_flash_attn
+        COMPONENT vllm_flash_attn_c
+        FILES_MATCHING PATTERN "*.py"
+)
+
+# Nothing after vllm-flash-attn, see comment about macros above

CODE_OF_CONDUCT.md

@@ -0,0 +1,128 @@
+
+# vLLM Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socioeconomic status,
+nationality, personal appearance, race, caste, color, religion, or sexual
+identity and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the overall
+  community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or advances of
+  any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email address,
+  without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official email address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline/IRL event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement in the #code-of-conduct
+channel in the [vLLM Discord](https://discord.com/invite/jz7wjKhh6g).
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series of
+actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or permanent
+ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior, harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within the
+community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant](https://www.contributor-covenant.org/),
+version 2.1, available at
+[v2.1](https://www.contributor-covenant.org/version/2/1/code_of_conduct.html).
+
+Community Impact Guidelines were inspired by
+[Mozilla's code of conduct enforcement ladder](https://github.com/mozilla/inclusion).
+
+For answers to common questions about this code of conduct, see the
+[Contributor Covenant FAQ](https://www.contributor-covenant.org/faq). Translations are available at
+[Contributor Covenant translations](https://www.contributor-covenant.org/translations).
+

Dockerfile

@@ -9,28 +9,23 @@ ARG CUDA_VERSION=12.4.1
 #################### BASE BUILD IMAGE ####################
 # prepare basic build environment
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04 AS base
-
 ARG CUDA_VERSION=12.4.1
-ARG PYTHON_VERSION=3.10
-
+ARG PYTHON_VERSION=3.12
 ENV DEBIAN_FRONTEND=noninteractive
 
+# Install Python and other dependencies
 RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
     && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
     && apt-get update -y \
-    && apt-get install -y ccache software-properties-common \
+    && apt-get install -y ccache software-properties-common git curl sudo \
     && add-apt-repository ppa:deadsnakes/ppa \
     && apt-get update -y \
     && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && if [ "${PYTHON_VERSION}" != "3" ]; then update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; fi \
-    && python3 --version
-
-RUN apt-get update -y \
-    && apt-get install -y git curl sudo
-
-# Install pip s.t. it will be compatible with our PYTHON_VERSION
-RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}
-RUN python3 -m pip --version
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -46,9 +41,6 @@ COPY requirements-cuda.txt requirements-cuda.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -r requirements-cuda.txt
 
-COPY requirements-mamba.txt requirements-mamba.txt
-RUN python3 -m pip install packaging
-RUN python3 -m pip install -r requirements-mamba.txt
 
 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
@@ -56,22 +48,20 @@ RUN python3 -m pip install -r requirements-mamba.txt
 # see https://github.com/pytorch/pytorch/pull/123243
 ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
 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 ####################
 FROM base AS build
 
-ARG PYTHON_VERSION=3.10
-
 # install build dependencies
 COPY requirements-build.txt requirements-build.txt
 
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -r requirements-build.txt
 
-# install compiler cache to speed up compilation leveraging local or remote caching
-RUN apt-get update -y && apt-get install -y ccache
-
 # files and directories related to build wheels
 COPY csrc csrc
 COPY setup.py setup.py
@@ -88,27 +78,24 @@ ENV MAX_JOBS=${max_jobs}
 # number of threads used by nvcc
 ARG nvcc_threads=8
 ENV NVCC_THREADS=$nvcc_threads
-# make sure punica kernels are built (for LoRA)
-ENV VLLM_INSTALL_PUNICA_KERNELS=1
-
-ARG buildkite_commit
-ENV BUILDKITE_COMMIT=${buildkite_commit}
 
 ARG USE_SCCACHE
+ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
+ARG SCCACHE_REGION_NAME=us-west-2
+ARG SCCACHE_S3_NO_CREDENTIALS=0
 # if USE_SCCACHE is set, use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
         echo "Installing sccache..." \
         && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
         && tar -xzf sccache.tar.gz \
         && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
         && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
-        && if [ "$CUDA_VERSION" = "11.8.0" ]; then \
-            export SCCACHE_BUCKET=vllm-build-sccache-2; \
-           else \
-            export SCCACHE_BUCKET=vllm-build-sccache; \
-           fi \
-        && export SCCACHE_REGION=us-west-2 \
+        && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
+        && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
+        && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
+        && export SCCACHE_IDLE_TIMEOUT=0 \
         && export CMAKE_BUILD_TYPE=Release \
         && sccache --show-stats \
         && python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38 \
@@ -118,14 +105,22 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,source=.git,target=.git  \
     if [ "$USE_SCCACHE" != "1" ]; then \
         python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
     fi
 
-# check the size of the wheel, we cannot upload wheels larger than 100MB
+# Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
-RUN python3 check-wheel-size.py dist
-
+# Default max size of the wheel is 250MB
+ARG VLLM_MAX_SIZE_MB=250
+ENV VLLM_MAX_SIZE_MB=$VLLM_MAX_SIZE_MB
+ARG RUN_WHEEL_CHECK=true
+RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
+        python3 check-wheel-size.py dist; \
+    else \
+        echo "Skipping wheel size check."; \
+    fi
 #################### EXTENSION Build IMAGE ####################
 
 #################### DEV IMAGE ####################
@@ -138,45 +133,31 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -r requirements-dev.txt
 
 #################### DEV IMAGE ####################
-#################### MAMBA Build IMAGE ####################
-FROM dev as mamba-builder
-# max jobs used for build
-ARG max_jobs=2
-ENV MAX_JOBS=${max_jobs}
-
-WORKDIR /usr/src/mamba
-
-COPY requirements-mamba.txt requirements-mamba.txt
-
-# Download the wheel or build it if a pre-compiled release doesn't exist
-RUN pip --verbose wheel -r requirements-mamba.txt \
-    --no-build-isolation --no-deps --no-cache-dir
-
-#################### MAMBA Build IMAGE ####################
-
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu20.04 AS vllm-base
 ARG CUDA_VERSION=12.4.1
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 WORKDIR /vllm-workspace
+ENV DEBIAN_FRONTEND=noninteractive
 
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
+
+# Install Python and other dependencies
 RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
     && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
     && apt-get update -y \
-    && apt-get install -y ccache software-properties-common \
+    && apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
+    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
     && add-apt-repository ppa:deadsnakes/ppa \
     && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && if [ "${PYTHON_VERSION}" != "3" ]; then update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; fi \
-    && python3 --version
-
-RUN apt-get update -y \
-    && apt-get install -y python3-pip git vim curl libibverbs-dev
-
-# Install pip s.t. it will be compatible with our PYTHON_VERSION
-RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}
-RUN python3 -m pip --version
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv libibverbs-dev \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -189,12 +170,9 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
     --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install dist/*.whl --verbose
 
-RUN --mount=type=bind,from=mamba-builder,src=/usr/src/mamba,target=/usr/src/mamba \
-    --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install /usr/src/mamba/*.whl --no-cache-dir
-
 RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.9/flashinfer-0.0.9+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+    . /etc/environment && \
+    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl
 #################### vLLM installation IMAGE ####################
 
 
@@ -224,7 +202,7 @@ FROM vllm-base AS vllm-openai
 
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer 'modelscope!=1.15.0'
+    pip install accelerate hf_transfer 'modelscope!=1.15.0' bitsandbytes>=0.44.0 timm==0.9.10
 
 ENV VLLM_USAGE_SOURCE production-docker-image
 

Dockerfile.cpu

@@ -2,36 +2,70 @@
 
 FROM ubuntu:22.04 AS cpu-test-1
 
-RUN apt-get update  -y \
-    && apt-get install -y git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 \
+ENV CCACHE_DIR=/root/.cache/ccache
+
+ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
+
+RUN --mount=type=cache,target=/var/cache/apt \
+    apt-get update -y \
+    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
+    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
 
 # https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
 # intel-openmp provides additional performance improvement vs. openmp
 # tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
-RUN pip install intel-openmp
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install intel-openmp
 
-ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so:$LD_PRELOAD"
+ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so"
 
+RUN echo 'ulimit -c 0' >> ~/.bashrc
 
-RUN pip install https://intel-extension-for-pytorch.s3.amazonaws.com/ipex_dev/cpu/intel_extension_for_pytorch-2.3.100%2Bgit0eb3473-cp310-cp310-linux_x86_64.whl
+RUN pip install https://intel-extension-for-pytorch.s3.amazonaws.com/ipex_dev/cpu/intel_extension_for_pytorch-2.4.0%2Bgitfbaa4bc-cp310-cp310-linux_x86_64.whl
 
-RUN pip install --upgrade pip \
-    && pip install wheel packaging ninja "setuptools>=49.4.0" numpy
+WORKDIR /workspace
+
+ENV PIP_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cpu
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
+    pip install --upgrade pip && \
+    pip install -r requirements-build.txt
+
+# install oneDNN
+RUN git clone -b rls-v3.5 https://github.com/oneapi-src/oneDNN.git
+
+RUN --mount=type=cache,target=/root/.cache/ccache \
+    cmake -B ./oneDNN/build -S ./oneDNN -G Ninja -DONEDNN_LIBRARY_TYPE=STATIC \ 
+    -DONEDNN_BUILD_DOC=OFF \ 
+    -DONEDNN_BUILD_EXAMPLES=OFF \ 
+    -DONEDNN_BUILD_TESTS=OFF \ 
+    -DONEDNN_BUILD_GRAPH=OFF \ 
+    -DONEDNN_ENABLE_WORKLOAD=INFERENCE \ 
+    -DONEDNN_ENABLE_PRIMITIVE=MATMUL && \
+    cmake --build ./oneDNN/build --target install --config Release
 
 FROM cpu-test-1 AS build
 
-COPY ./ /workspace/vllm
-
 WORKDIR /workspace/vllm
 
-RUN pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
+    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
+    pip install -v -r requirements-cpu.txt
+
+COPY ./ ./
 
 # Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
 ARG VLLM_CPU_DISABLE_AVX512
 ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
 
-RUN VLLM_TARGET_DEVICE=cpu python3 setup.py install
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
+    pip install dist/*.whl && \
+    rm -rf dist
 
 WORKDIR /workspace/
 

Dockerfile.neuron

@@ -1,12 +1,17 @@
 # default base image
-ARG BASE_IMAGE="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-inference-neuronx:2.1.1-neuronx-py310-sdk2.17.0-ubuntu20.04"
+ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.0-ubuntu20.04"
 
 FROM $BASE_IMAGE
 
 RUN echo "Base image is $BASE_IMAGE"
 
 # Install some basic utilities
-RUN apt-get update && apt-get install python3 python3-pip -y
+RUN apt-get update && \
+    apt-get install -y \
+        git \
+        python3 \
+        python3-pip \
+        ffmpeg libsm6 libxext6 libgl1
 
 ### Mount Point ###
 # When launching the container, mount the code directory to /app
@@ -18,19 +23,19 @@ RUN python3 -m pip install --upgrade pip
 RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
 RUN python3 -m pip install sentencepiece transformers==4.36.2 -U
 RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
-RUN python3 -m pip install --pre neuronx-cc==2.12.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install --pre neuronx-cc==2.15.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
 
-COPY ./vllm /app/vllm/vllm
-COPY ./setup.py /app/vllm/setup.py
-COPY ./requirements-common.txt /app/vllm/requirements-common.txt
-COPY ./requirements-neuron.txt /app/vllm/requirements-neuron.txt
+COPY . /app/vllm
 
 RUN cd /app/vllm \
-    && python3 -m pip install -U -r requirements-neuron.txt
+    && python3 -m pip install -U \
+        cmake>=3.26 ninja packaging setuptools-scm>=8 wheel jinja2 \
+        -r requirements-neuron.txt
 
 ENV VLLM_TARGET_DEVICE neuron
-RUN cd /app/vllm \
-    && pip install -e . \
+RUN --mount=type=bind,source=.git,target=.git \
+    cd /app/vllm \
+    && pip install --no-build-isolation -v -e . \
     && cd ..
 
 CMD ["/bin/bash"]

Dockerfile.openvino

@@ -1,10 +1,12 @@
 # The vLLM Dockerfile is used to construct vLLM image that can be directly used
 # to run the OpenAI compatible server.
 
-FROM ubuntu:20.04 AS dev
+FROM ubuntu:22.04 AS dev
 
 RUN apt-get update -y && \
-    apt-get install -y python3-pip git
+    apt-get install -y \
+        git python3-pip \
+        ffmpeg libsm6 libxext6 libgl1
 WORKDIR /workspace
 
 # copy requirements
@@ -13,12 +15,15 @@ COPY requirements-common.txt /workspace/vllm/
 COPY requirements-openvino.txt /workspace/vllm/
 
 COPY vllm/ /workspace/vllm/vllm
+COPY csrc/core /workspace/vllm/csrc/core
+COPY cmake/utils.cmake /workspace/vllm/cmake/
+COPY CMakeLists.txt /workspace/vllm/
 COPY setup.py /workspace/vllm/
 
 # install build requirements
 RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/vllm/requirements-build.txt
 # build vLLM with OpenVINO backend
-RUN PIP_PRE=1 PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu https://storage.openvinotoolkit.org/simple/wheels/nightly/" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace/vllm/
+RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace/vllm/
 
 COPY examples/ /workspace/vllm/examples
 COPY benchmarks/ /workspace/vllm/benchmarks

Dockerfile.ppc64le

@@ -2,21 +2,32 @@ FROM mambaorg/micromamba
 ARG MAMBA_DOCKERFILE_ACTIVATE=1
 USER root
 
-RUN apt-get update  -y     && apt-get install -y git wget vim numactl gcc-12 g++-12 protobuf-compiler libprotobuf-dev     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
+ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"
+
+RUN apt-get update -y && apt-get install -y git wget curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 
 
 # Some packages in requirements-cpu are installed here
 # IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
 # Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults     python=3.10     pytorch-cpu=2.1.2     torchvision-cpu=0.16.2    &&     micromamba clean --all --yes
+RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 torchvision-cpu=0.16.2 rust && micromamba clean --all --yes
 
 COPY ./ /workspace/vllm
 
 WORKDIR /workspace/vllm
 
 # These packages will be in rocketce eventually
-RUN pip install -v -r requirements-cpu.txt --prefer-binary --extra-index-url https://repo.fury.io/mgiessing
+RUN --mount=type=cache,target=/root/.cache/pip  \
+    pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
+        cmake>=3.26 ninja packaging setuptools-scm>=8 wheel jinja2 \
+        torch==2.3.1 \
+        -r requirements-cpu.txt \
+        xformers uvloop==0.20.0
 
-RUN VLLM_TARGET_DEVICE=cpu python3 setup.py install
+RUN --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py install
 
-WORKDIR /vllm-workspace
-ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]
+WORKDIR /workspace/
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.rocm

@@ -1,5 +1,5 @@
-# Default ROCm 6.1 base image
-ARG BASE_IMAGE="rocm/pytorch:rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging"
+# Default ROCm 6.2 base image
+ARG BASE_IMAGE="rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0"
 
 # Default ROCm ARCHes to build vLLM for.
 ARG PYTORCH_ROCM_ARCH="gfx908;gfx90a;gfx942;gfx1100"
@@ -7,18 +7,12 @@ ARG PYTORCH_ROCM_ARCH="gfx908;gfx90a;gfx942;gfx1100"
 # Whether to install CK-based flash-attention
 # If 0, will not install flash-attention
 ARG BUILD_FA="1"
-# If `TRY_FA_WHEEL=1`, we will try installing flash-attention from `FA_WHEEL_URL`
-# If this succeeds, we use the downloaded wheel and skip building flash-attention.
-# Otherwise, ROCm flash-attention from `FA_BRANCH` will be built for the
-# architectures specified in `FA_GFX_ARCHS`
-ARG TRY_FA_WHEEL="1"
-ARG FA_WHEEL_URL="https://github.com/ROCm/flash-attention/releases/download/v2.5.9post1-cktile-vllm/flash_attn-2.5.9.post1-cp39-cp39-linux_x86_64.whl"
 ARG FA_GFX_ARCHS="gfx90a;gfx942"
-ARG FA_BRANCH="23a2b1c2"
+ARG FA_BRANCH="3cea2fb"
 
 # Whether to build triton on rocm
 ARG BUILD_TRITON="1"
-ARG TRITON_BRANCH="e0fc12c"
+ARG TRITON_BRANCH="e192dba"
 
 ### Base image build stage
 FROM $BASE_IMAGE AS base
@@ -50,14 +44,17 @@ RUN python3 -m pip install --upgrade pip
 # Remove sccache so it doesn't interfere with ccache
 # TODO: implement sccache support across components
 RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)"
-# Install torch == 2.5.0 on ROCm
-RUN case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
-        *"rocm-6.1"*) \
-            python3 -m pip uninstall -y torch torchaudio torchvision \
-            && python3 -m pip install --no-cache-dir --pre \
-                torch==2.5.0.dev20240710 torchaudio==2.4.0.dev20240710 \
-                torchvision==0.20.0.dev20240710 \
-               --index-url https://download.pytorch.org/whl/nightly/rocm6.1;; \
+
+# Install torch == 2.6.0 on ROCm
+RUN --mount=type=cache,target=/root/.cache/pip \
+    case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+        *"rocm-6.2"*) \
+            python3 -m pip uninstall -y torch torchvision \
+            && python3 -m pip install --pre \
+                torch==2.6.0.dev20240918 \
+                setuptools-scm>=8 \
+                torchvision==0.20.0.dev20240918 \
+                --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2;; \
         *) ;; esac
 
 ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer
@@ -79,25 +76,18 @@ RUN cd /opt/rocm/share/amd_smi \
 ### Flash-Attention wheel build stage
 FROM base AS build_fa
 ARG BUILD_FA
-ARG TRY_FA_WHEEL
-ARG FA_WHEEL_URL
 ARG FA_GFX_ARCHS
 ARG FA_BRANCH
 # Build ROCm flash-attention wheel if `BUILD_FA = 1`
 RUN --mount=type=cache,target=${CCACHE_DIR} \
     if [ "$BUILD_FA" = "1" ]; then \
-        if [ "${TRY_FA_WHEEL}" = "1" ] && python3 -m pip install "${FA_WHEEL_URL}"; then \
-            # If a suitable wheel exists, we download it instead of building FA
-            mkdir -p /install && wget -N "${FA_WHEEL_URL}" -P /install; \
-        else \
-            mkdir -p libs \
-            && cd libs \
-            && git clone https://github.com/ROCm/flash-attention.git \
-            && cd flash-attention \
-            && git checkout "${FA_BRANCH}" \
-            && git submodule update --init \
-            && GPU_ARCHS="${FA_GFX_ARCHS}" python3 setup.py bdist_wheel --dist-dir=/install; \
-        fi; \
+        mkdir -p libs \
+        && cd libs \
+        && git clone https://github.com/ROCm/flash-attention.git \
+        && cd flash-attention \
+        && git checkout "${FA_BRANCH}" \
+        && git submodule update --init \
+        && GPU_ARCHS="${FA_GFX_ARCHS}" python3 setup.py bdist_wheel --dist-dir=/install; \
     # Create an empty directory otherwise as later build stages expect one
     else mkdir -p /install; \
     fi
@@ -112,6 +102,7 @@ RUN --mount=type=cache,target=${CCACHE_DIR} \
     if [ "$BUILD_TRITON" = "1" ]; then \
     mkdir -p libs \
     && cd libs \
+    && python3 -m pip install ninja cmake wheel pybind11 \
     && git clone https://github.com/OpenAI/triton.git \
     && cd triton \
     && git checkout "${TRITON_BRANCH}" \
@@ -127,34 +118,20 @@ FROM base AS final
 # Import the vLLM development directory from the build context
 COPY . .
 
-# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
-# Manually remove it so that later steps of numpy upgrade can continue
-RUN case "$(which python3)" in \
-        *"/opt/conda/envs/py_3.9"*) \
-            rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/;; \
-        *) ;; esac
-
 # Package upgrades for useful functionality or to avoid dependency issues
 RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install --upgrade numba scipy huggingface-hub[cli]
+    python3 -m pip install --upgrade numba scipy huggingface-hub[cli] pytest-shard
+
 
-# Make sure punica kernels are built (for LoRA)
-ENV VLLM_INSTALL_PUNICA_KERNELS=1
 # Workaround for ray >= 2.10.0
 ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
 # Silences the HF Tokenizers warning
 ENV TOKENIZERS_PARALLELISM=false
 
 RUN --mount=type=cache,target=${CCACHE_DIR} \
+    --mount=type=bind,source=.git,target=.git \
     --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -Ur requirements-rocm.txt \
-    && case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
-        *"rocm-6.1"*) \
-            # Bring in upgrades to HIP graph earlier than ROCm 6.2 for vLLM
-            wget -N https://github.com/ROCm/vllm/raw/fa78403/rocm_patch/libamdhip64.so.6 -P /opt/rocm/lib \
-            # Prevent interference if torch bundles its own HIP runtime
-            && rm -f "$(python3 -c 'import torch; print(torch.__path__[0])')"/lib/libamdhip64.so* || true;; \
-        *) ;; esac \
     && python3 setup.py clean --all \
     && python3 setup.py develop
 

Dockerfile.tpu

@@ -1,20 +1,29 @@
-ARG NIGHTLY_DATE="20240713"
+ARG NIGHTLY_DATE="20240828"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
 
 FROM $BASE_IMAGE
 WORKDIR /workspace
 
-# Install aiohttp separately to avoid build errors.
-RUN pip install aiohttp
-# Install NumPy 1 instead of NumPy 2.
-RUN pip install "numpy<2"
+# Install some basic utilities
+RUN apt-get update && apt-get install -y \
+    git \
+    ffmpeg libsm6 libxext6 libgl1
+
 # Install the TPU and Pallas dependencies.
-RUN pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
-RUN pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
 
 # Build vLLM.
 COPY . /workspace/vllm
 ENV VLLM_TARGET_DEVICE="tpu"
-RUN cd /workspace/vllm && python setup.py develop
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,source=.git,target=.git \
+     cd /workspace/vllm && \
+    python3 -m pip install \
+        cmake>=3.26 ninja packaging setuptools-scm>=8 wheel jinja2 \
+        -r requirements-tpu.txt
+RUN cd /workspace/vllm && python3 setup.py develop
 
 CMD ["/bin/bash"]

Dockerfile.xpu

@@ -1,22 +1,26 @@
-FROM intel/oneapi-basekit:2024.1.0-devel-ubuntu20.04
+FROM intel/oneapi-basekit:2024.2.1-0-devel-ubuntu22.04
 
 RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
     echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
     chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
-    rm /etc/apt/sources.list.d/intel-graphics.list && \
     wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
     echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
     chmod 644 /usr/share/keyrings/intel-graphics.gpg
 
-RUN apt-get update  -y \
-&& apt-get install -y curl libicu70 lsb-release git wget vim numactl python3 python3-pip
+RUN apt-get update  -y && \
+    apt-get install -y curl libicu70 lsb-release git wget vim numactl python3 python3-pip ffmpeg libsm6 libxext6 libgl1
 
 COPY ./ /workspace/vllm
 
 WORKDIR /workspace/vllm
 
-RUN pip install -v -r requirements-xpu.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -v --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ \
+        cmake>=3.26 ninja packaging setuptools-scm>=8 wheel jinja2 \
+        -r requirements-xpu.txt
 
-RUN VLLM_TARGET_DEVICE=xpu python3 setup.py install
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=xpu python3 setup.py install
 
 CMD ["/bin/bash"]

README.md

@@ -10,22 +10,24 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://discord.gg/jz7wjKhh6g"><b>Discord</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://discord.gg/jz7wjKhh6g"><b>Discord</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> |
 
 </p>
 
+
 ---
 
-**The Fifth vLLM Bay Area Meetup (July 24th 5pm-8pm PT)**
+**vLLM, AMD, Anyscale Meet & Greet at [Ray Summit 2024](http://raysummit.anyscale.com) (Monday, Sept 30th, 5-7pm PT) at Marriott Marquis San Francisco**
 
-We are excited to announce our fifth vLLM Meetup!
-Join us to hear the vLLM's recent updates and the upcoming roadmap.
-Additionally, our collaborators from AWS will be presenting their insights and experiences in deploying vLLM.
-Register now [here](https://lu.ma/lp0gyjqr) and be part of the event!
+We are excited to announce our special vLLM event in collaboration with AMD and Anyscale.
+Join us to learn more about recent advancements of vLLM on MI300X.
+Register [here](https://lu.ma/db5ld9n5) and be a part of the event!
 
 ---
 
 *Latest News* 🔥
+- [2024/09] We hosted [the sixth vLLM meetup](https://lu.ma/87q3nvnh) with NVIDIA! Please find the meetup slides [here](https://docs.google.com/presentation/d/1wrLGwytQfaOTd5wCGSPNhoaW3nq0E-9wqyP7ny93xRs/edit?usp=sharing).
+- [2024/07] We hosted [the fifth vLLM meetup](https://lu.ma/lp0gyjqr) with AWS! Please find the meetup slides [here](https://docs.google.com/presentation/d/1RgUD8aCfcHocghoP3zmXzck9vX3RCI9yfUAB2Bbcl4Y/edit?usp=sharing).
 - [2024/07] In partnership with Meta, vLLM officially supports Llama 3.1 with FP8 quantization and pipeline parallelism! Please check out our blog post [here](https://blog.vllm.ai/2024/07/23/llama31.html).
 - [2024/06] We hosted [the fourth vLLM meetup](https://lu.ma/agivllm) with Cloudflare and BentoML! Please find the meetup slides [here](https://docs.google.com/presentation/d/1iJ8o7V2bQEi0BFEljLTwc5G1S10_Rhv3beed5oB0NJ4/edit?usp=sharing).
 - [2024/04] We hosted [the third vLLM meetup](https://robloxandvllmmeetup2024.splashthat.com/) with Roblox! Please find the meetup slides [here](https://docs.google.com/presentation/d/1A--47JAK4BJ39t954HyTkvtfwn0fkqtsL8NGFuslReM/edit?usp=sharing).
@@ -44,10 +46,12 @@ vLLM is fast with:
 - Efficient management of attention key and value memory with **PagedAttention**
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
-- Quantization: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [SqueezeLLM](https://arxiv.org/abs/2306.07629), FP8 KV Cache
-- Optimized CUDA kernels
+- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), INT4, INT8, and FP8.
+- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer.
+- Speculative decoding
+- Chunked prefill
 
-**Performance benchmark**: We include a [performance benchmark](https://buildkite.com/vllm/performance-benchmark/builds/3924) that compares the performance of vllm against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [text-generation-inference](https://github.com/huggingface/text-generation-inference) and [lmdeploy](https://github.com/InternLM/lmdeploy)).
+**Performance benchmark**: We include a [performance benchmark](https://buildkite.com/vllm/performance-benchmark/builds/4068) that compares the performance of vLLM against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [text-generation-inference](https://github.com/huggingface/text-generation-inference) and [lmdeploy](https://github.com/InternLM/lmdeploy)).
 
 vLLM is flexible and easy to use with:
 
@@ -56,20 +60,21 @@ vLLM is flexible and easy to use with:
 - Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs
-- (Experimental) Prefix caching support
-- (Experimental) Multi-lora support
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron.
+- Prefix caching support
+- 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)
+- Embedding Models (e.g. E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)
 
 Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
 
 ## Getting Started
 
-Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
+Install vLLM with `pip` or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
 
 ```bash
 pip install vllm
@@ -107,6 +112,7 @@ vLLM is a community project. Our compute resources for development and testing a
 - Roblox
 - RunPod
 - Sequoia Capital
+- Skywork AI
 - Trainy
 - UC Berkeley
 - UC San Diego
@@ -125,3 +131,10 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
   year={2023}
 }
 ```
+
+## Contact Us
+
+* For technical questions and feature requests, please use Github issues or discussions.
+* For discussing with fellow users, please use Discord.
+* For security disclosures, please use Github's security advisory feature.
+* For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.

SECURITY.md

@@ -0,0 +1,12 @@
+# Security Policy
+
+## Reporting a Vulnerability
+
+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 using https://github.com/vllm-project/vllm/security/advisories/new
+
+---
+Please see PyTorch Security for more information how to securely interact with models: https://github.com/pytorch/pytorch/blob/main/SECURITY.md
+This document mostly references the recommendation from PyTorch, thank you! 

benchmarks/backend_request_func.py

@@ -24,6 +24,8 @@ class RequestFuncInput:
     model: str
     best_of: int = 1
     use_beam_search: bool = False
+    logprobs: Optional[int] = None
+    multi_modal_content: Optional[dict] = None
 
 
 @dataclass
@@ -225,8 +227,8 @@ async def async_request_openai_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "completions"
-    ), "OpenAI Completions API URL must end with 'completions'."
+        ("completions", "profile")
+    ), "OpenAI Completions API URL must end with 'completions' or 'profile'."
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search
@@ -236,6 +238,7 @@ async def async_request_openai_completions(
             "temperature": 0.0,
             "best_of": request_func_input.best_of,
             "max_tokens": request_func_input.output_len,
+            "logprobs": request_func_input.logprobs,
             "stream": True,
         }
         headers = {
@@ -276,8 +279,9 @@ async def async_request_openai_completions(
                                     output.ttft = ttft
 
                                 # Decoding phase
-                                output.itl.append(timestamp -
-                                                  most_recent_timestamp)
+                                else:
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
                                 generated_text += data["choices"][0]["text"]
@@ -309,12 +313,15 @@ async def async_request_openai_chat_completions(
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search
+        content = [{"type": "text", "text": request_func_input.prompt}]
+        if request_func_input.multi_modal_content:
+            content.append(request_func_input.multi_modal_content)
         payload = {
             "model": request_func_input.model,
             "messages": [
                 {
                     "role": "user",
-                    "content": request_func_input.prompt,
+                    "content": content
                 },
             ],
             "temperature": 0.0,

benchmarks/benchmark_latency.py

@@ -10,7 +10,7 @@ import torch
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
-from vllm.engine.arg_utils import EngineArgs
+from vllm.engine.arg_utils import DEVICE_OPTIONS, EngineArgs
 from vllm.inputs import PromptInputs
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 from vllm.utils import FlexibleArgumentParser
@@ -205,13 +205,11 @@ if __name__ == '__main__':
         default=None,
         help=('path to save the pytorch profiler output. Can be visualized '
               'with ui.perfetto.dev or Tensorboard.'))
-    parser.add_argument(
-        "--device",
-        type=str,
-        default="auto",
-        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
-        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
-        'CPU.')
+    parser.add_argument("--device",
+                        type=str,
+                        default="auto",
+                        choices=DEVICE_OPTIONS,
+                        help='device type for vLLM execution')
     parser.add_argument('--block-size',
                         type=int,
                         default=16,

benchmarks/benchmark_prefix_caching.py

@@ -1,8 +1,45 @@
+"""
+Benchmark the efficiency of prefix caching.
+
+This script allows you to benchmark the performance of
+a model with and without prefix caching using either fixed prompts
+or prompts sampled from the ShareGPT dataset.
+
+Fixed example usage:
+    python benchmark_prefix_caching.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --enable-prefix-caching \
+        --num-prompts 1 \
+        --repeat-count 100
+
+ShareGPT example usage:
+    # This command samples 20 prompts with input lengths
+    # between 128 and 256 tokens from the ShareGPT dataset,
+    # then replicates each prompt 5 times.
+    python benchmark_prefix_caching.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json \
+        --enable-prefix-caching \
+        --num-prompts 20 \
+        --repeat-count 5 \
+        --input-length-range 128:256
+"""
+
+import json
+import random
 import time
+from typing import List, Optional, Tuple
+
+from transformers import PreTrainedTokenizerBase
 
 from vllm import LLM, SamplingParams
 from vllm.utils import FlexibleArgumentParser
 
+try:
+    from vllm.transformers_utils.tokenizer import get_tokenizer
+except ImportError:
+    from backend_request_func import get_tokenizer
+
 PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n"  # noqa: E501
 
 
@@ -15,7 +52,83 @@ def test_prefix(llm=None, sampling_params=None, prompts=None):
     print(f"cost time {end_time - start_time}")
 
 
+def sample_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    input_length_range: Tuple[int, int],
+    fixed_output_len: Optional[int],
+) -> List[Tuple[str, int, int]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # Shuffle the dataset.
+    random.shuffle(dataset)
+
+    min_len, max_len = input_length_range
+
+    # Filter out sequences that are too long or too short
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for i in range(len(dataset)):
+        if len(filtered_dataset) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = dataset[i][1]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if min_len <= prompt_len <= max_len:
+            filtered_dataset.append((prompt, prompt_len, output_len))
+
+    return filtered_dataset
+
+
+def repeat_and_sort_requests(requests: List[Tuple[str, int, int]],
+                             repeat_count: int,
+                             sort: bool = False) -> List[str]:
+    repeated_requests = requests * repeat_count
+    if sort:
+        repeated_requests.sort(key=lambda x: x[1])
+    else:
+        random.shuffle(repeated_requests)
+    return [req[0] for req in repeated_requests]
+
+
 def main(args):
+    tokenizer = get_tokenizer(args.model, trust_remote_code=True)
+    input_length_range = tuple(map(int, args.input_length_range.split(':')))
+
+    if args.dataset_path is not None:
+        print(f"Start to sample {args.num_prompts} prompts"
+              "from {args.dataset_path}")
+        filtered_datasets = sample_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            input_length_range=input_length_range,
+            fixed_output_len=args.output_len,
+        )
+    else:
+        prompt_len = len(tokenizer(PROMPT).input_ids)
+        filtered_datasets = [(PROMPT, prompt_len, args.output_len)
+                             ] * args.num_prompts
+
     llm = LLM(model=args.model,
               tokenizer_mode='auto',
               trust_remote_code=True,
@@ -24,10 +137,13 @@ def main(args):
               tensor_parallel_size=args.tensor_parallel_size,
               enable_prefix_caching=args.enable_prefix_caching)
 
-    num_prompts = 100
-    prompts = [PROMPT] * num_prompts
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
 
+    print("Testing filtered datasets")
+    prompts = repeat_and_sort_requests(filtered_datasets,
+                                       repeat_count=args.repeat_count,
+                                       sort=args.sort)
+
     print("------warm up------")
     test_prefix(
         llm=llm,
@@ -45,11 +161,15 @@ def main(args):
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description='Benchmark the performance with or without automatic '
-        'prefix caching.')
+        description=
+        'Benchmark the performance with or without automatic prefix caching.')
     parser.add_argument('--model',
                         type=str,
                         default='baichuan-inc/Baichuan2-13B-Chat')
+    parser.add_argument("--dataset-path",
+                        type=str,
+                        default=None,
+                        help="Path to the dataset.")
     parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
     parser.add_argument('--output-len', type=int, default=10)
     parser.add_argument('--enable-prefix-caching',
@@ -58,5 +178,21 @@ if __name__ == "__main__":
     parser.add_argument('--use-v2-block-manager',
                         action='store_true',
                         help='Use BlockSpaceMangerV2')
+    parser.add_argument('--num-prompts',
+                        type=int,
+                        default=1,
+                        help="Number of the prompts sampled from dataset")
+    parser.add_argument('--repeat-count',
+                        type=int,
+                        default=100,
+                        help='Number of times to repeat each prompt')
+    parser.add_argument('--sort',
+                        action='store_true',
+                        help='Sort prompts by input length')
+    parser.add_argument('--input-length-range',
+                        type=str,
+                        default='128:256',
+                        help='Range of input lengths for sampling prompts,'
+                        'specified as "min:max" (e.g., "128:256").')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prioritization.py

@@ -0,0 +1,295 @@
+"""Benchmark offline prioritization."""
+import argparse
+import json
+import random
+import time
+from typing import List, Optional, Tuple
+
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
+
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+
+
+def sample_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int],
+) -> List[Tuple[str, int, int]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # Shuffle the dataset.
+    random.shuffle(dataset)
+
+    # Filter out sequences that are too long or too short
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for i in range(len(dataset)):
+        if len(filtered_dataset) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = dataset[i][1]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+
+        #Select a equi-probable random priority
+        priority = 0 if random.random() < 0.5 else 1
+
+        filtered_dataset.append((prompt, prompt_len, output_len, priority))
+
+    return filtered_dataset
+
+
+def run_vllm(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: str,
+    quantization: Optional[str],
+    tensor_parallel_size: int,
+    seed: int,
+    n: int,
+    use_beam_search: bool,
+    trust_remote_code: bool,
+    dtype: str,
+    max_model_len: Optional[int],
+    enforce_eager: bool,
+    kv_cache_dtype: str,
+    quantization_param_path: Optional[str],
+    device: str,
+    enable_prefix_caching: bool,
+    enable_chunked_prefill: bool,
+    max_num_batched_tokens: int,
+    gpu_memory_utilization: float = 0.9,
+    download_dir: Optional[str] = None,
+) -> float:
+    from vllm import LLM, SamplingParams
+    llm = LLM(
+        model=model,
+        tokenizer=tokenizer,
+        quantization=quantization,
+        tensor_parallel_size=tensor_parallel_size,
+        seed=seed,
+        trust_remote_code=trust_remote_code,
+        dtype=dtype,
+        max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_memory_utilization,
+        enforce_eager=enforce_eager,
+        kv_cache_dtype=kv_cache_dtype,
+        quantization_param_path=quantization_param_path,
+        device=device,
+        enable_prefix_caching=enable_prefix_caching,
+        download_dir=download_dir,
+        enable_chunked_prefill=enable_chunked_prefill,
+        max_num_batched_tokens=max_num_batched_tokens,
+        disable_log_stats=False,
+    )
+
+    # Add the requests to the engine.
+    prompts = []
+    sampling_params = []
+    priority = []
+    for prompt, _, output_len, _priority in requests:
+        prompts.append(prompt)
+        priority.append(_priority)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=0.0 if use_beam_search else 1.0,
+                top_p=1.0,
+                use_beam_search=use_beam_search,
+                ignore_eos=True,
+                max_tokens=output_len,
+            ))
+
+    start = time.perf_counter()
+    llm.generate(prompts, sampling_params, priority=priority, use_tqdm=True)
+    end = time.perf_counter()
+    return end - start
+
+
+def main(args: argparse.Namespace):
+    print(args)
+    random.seed(args.seed)
+
+    # Sample the requests.
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.tokenizer, trust_remote_code=args.trust_remote_code)
+    if args.dataset is None:
+        # Synthesize a prompt with the given input length.
+        prompt = "hi" * (args.input_len - 1)
+        requests = [(prompt, args.input_len, args.output_len)
+                    for _ in range(args.num_prompts)]
+    else:
+        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
+                                   args.output_len)
+
+    if args.backend == "vllm":
+        elapsed_time = run_vllm(
+            requests, args.model, args.tokenizer, args.quantization,
+            args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
+            args.trust_remote_code, args.dtype, args.max_model_len,
+            args.enforce_eager, args.kv_cache_dtype,
+            args.quantization_param_path, args.device,
+            args.enable_prefix_caching, args.enable_chunked_prefill,
+            args.max_num_batched_tokens, args.gpu_memory_utilization,
+            args.download_dir)
+    else:
+        raise ValueError(f"Unknown backend: {args.backend}")
+    total_num_tokens = sum(prompt_len + output_len
+                           for _, prompt_len, output_len, priority in requests)
+    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+          f"{total_num_tokens / elapsed_time:.2f} tokens/s")
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "elapsed_time": elapsed_time,
+            "num_requests": len(requests),
+            "total_num_tokens": total_num_tokens,
+            "requests_per_second": len(requests) / elapsed_time,
+            "tokens_per_second": total_num_tokens / elapsed_time,
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Benchmark the throughput.")
+    parser.add_argument("--backend",
+                        type=str,
+                        choices=["vllm", "hf", "mii"],
+                        default="vllm")
+    parser.add_argument("--dataset",
+                        type=str,
+                        default=None,
+                        help="Path to the dataset.")
+    parser.add_argument("--input-len",
+                        type=int,
+                        default=None,
+                        help="Input prompt length for each request")
+    parser.add_argument("--output-len",
+                        type=int,
+                        default=None,
+                        help="Output length for each request. Overrides the "
+                        "output length from the dataset.")
+    parser.add_argument("--model", type=str, default="facebook/opt-125m")
+    parser.add_argument("--tokenizer", type=str, default=None)
+    parser.add_argument('--quantization',
+                        '-q',
+                        choices=[*QUANTIZATION_METHODS, None],
+                        default=None)
+    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
+    parser.add_argument("--n",
+                        type=int,
+                        default=1,
+                        help="Number of generated sequences per prompt.")
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument("--num-prompts",
+                        type=int,
+                        default=200,
+                        help="Number of prompts to process.")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument('--trust-remote-code',
+                        action='store_true',
+                        help='trust remote code from huggingface')
+    parser.add_argument(
+        '--max-model-len',
+        type=int,
+        default=None,
+        help='Maximum length of a sequence (including prompt and output). '
+        'If None, will be derived from the model.')
+    parser.add_argument(
+        '--dtype',
+        type=str,
+        default='auto',
+        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
+        help='data type for model weights and activations. '
+        'The "auto" option will use FP16 precision '
+        'for FP32 and FP16 models, and BF16 precision '
+        'for BF16 models.')
+    parser.add_argument('--gpu-memory-utilization',
+                        type=float,
+                        default=0.9,
+                        help='the fraction of GPU memory to be used for '
+                        'the model executor, which can range from 0 to 1.'
+                        'If unspecified, will use the default value of 0.9.')
+    parser.add_argument("--enforce-eager",
+                        action="store_true",
+                        help="enforce eager execution")
+    parser.add_argument(
+        '--kv-cache-dtype',
+        type=str,
+        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
+        default="auto",
+        help='Data type for kv cache storage. If "auto", will use model '
+        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
+        'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
+    parser.add_argument(
+        '--quantization-param-path',
+        type=str,
+        default=None,
+        help='Path to the JSON file containing the KV cache scaling factors. '
+        'This should generally be supplied, when KV cache dtype is FP8. '
+        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
+        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
+        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
+        'instead supported for common inference criteria.')
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        choices=["cuda", "cpu"],
+        help='device type for vLLM execution, supporting CUDA and CPU.')
+    parser.add_argument(
+        "--enable-prefix-caching",
+        action='store_true',
+        help="enable automatic prefix caching for vLLM backend.")
+    parser.add_argument("--enable-chunked-prefill",
+                        action='store_true',
+                        help="enable chunked prefill for vLLM backend.")
+    parser.add_argument('--max-num-batched-tokens',
+                        type=int,
+                        default=None,
+                        help='maximum number of batched tokens per '
+                        'iteration')
+    parser.add_argument('--download-dir',
+                        type=str,
+                        default=None,
+                        help='directory to download and load the weights, '
+                        'default to the default cache dir of huggingface')
+    parser.add_argument(
+        '--output-json',
+        type=str,
+        default=None,
+        help='Path to save the throughput results in JSON format.')
+
+    args = parser.parse_args()
+    if args.tokenizer is None:
+        args.tokenizer = args.model
+    if args.dataset is None:
+        assert args.input_len is not None
+        assert args.output_len is not None
+    else:
+        assert args.input_len is None
+
+    main(args)

benchmarks/benchmark_serving.py

@@ -24,6 +24,8 @@ On the client side, run:
 """
 import argparse
 import asyncio
+import base64
+import io
 import json
 import os
 import random
@@ -31,11 +33,13 @@ import time
 import warnings
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple
+from typing import Any, AsyncGenerator, Collection, Dict, List, Optional, Tuple
 
 import numpy as np
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
                                   RequestFuncOutput)
+from datasets import load_dataset
+from PIL.Image import Image
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 
@@ -56,20 +60,27 @@ class BenchmarkMetrics:
     total_input: int
     total_output: int
     request_throughput: float
-    input_throughput: float
     output_throughput: float
+    total_token_throughput: float
     mean_ttft_ms: float
     median_ttft_ms: float
     std_ttft_ms: float
-    p99_ttft_ms: float
+    percentiles_ttft_ms: List[Tuple[float, float]]
     mean_tpot_ms: float
     median_tpot_ms: float
     std_tpot_ms: float
-    p99_tpot_ms: float
+    percentiles_tpot_ms: List[Tuple[float, float]]
     mean_itl_ms: float
     median_itl_ms: float
     std_itl_ms: float
-    p99_itl_ms: float
+    percentiles_itl_ms: List[Tuple[float, float]]
+    # E2EL stands for end-to-end latency per request.
+    # It is the time taken on the client side from sending
+    # a request to receiving a complete response.
+    mean_e2el_ms: float
+    median_e2el_ms: float
+    std_e2el_ms: float
+    percentiles_e2el_ms: List[Tuple[float, float]]
 
 
 def sample_sharegpt_requests(
@@ -77,7 +88,7 @@ def sample_sharegpt_requests(
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, int, int]]:
+) -> List[Tuple[str, int, int, None]]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
     # Load the dataset.
@@ -112,7 +123,7 @@ def sample_sharegpt_requests(
         if prompt_len > 1024 or prompt_len + output_len > 2048:
             # Prune too long sequences.
             continue
-        filtered_dataset.append((prompt, prompt_len, output_len))
+        filtered_dataset.append((prompt, prompt_len, output_len, None))
 
     return filtered_dataset
 
@@ -124,7 +135,7 @@ def sample_sonnet_requests(
     output_len: int,
     prefix_len: int,
     tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, str, int, int]]:
+) -> List[Tuple[str, str, int, int, None]]:
     assert (
         input_len > prefix_len
     ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
@@ -182,14 +193,80 @@ def sample_sonnet_requests(
             message, add_generation_prompt=True, tokenize=False)
         prompt_len = len(tokenizer(prompt_formatted).input_ids)
         sampled_requests.append(
-            (prompt, prompt_formatted, prompt_len, output_len))
+            (prompt, prompt_formatted, prompt_len, output_len, None))
+
+    return sampled_requests
+
+
+def sample_hf_requests(
+    dataset_path: str,
+    dataset_subset: str,
+    dataset_split: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int] = None,
+) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
+    dataset = load_dataset(dataset_path,
+                           name=dataset_subset,
+                           split=dataset_split,
+                           streaming=True)
+    assert "conversations" in dataset.features, (
+        "HF Dataset must have 'conversations' column.")
+    filtered_dataset = dataset.shuffle().filter(
+        lambda x: len(x["conversations"]) >= 2)
+    sampled_requests: List[Tuple[str, int, int, Dict[str,
+                                                     Collection[str]]]] = []
+    for data in filtered_dataset:
+        if len(sampled_requests) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = data["conversations"][0]["value"]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = data["conversations"][1]["value"]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+
+        if "image" in data and isinstance(data["image"], Image):
+            image: Image = data["image"]
+            image = image.convert("RGB")
+            image_data = io.BytesIO()
+            image.save(image_data, format='JPEG')
+            image_base64 = base64.b64encode(
+                image_data.getvalue()).decode("utf-8")
+            mm_content = {
+                "type": "image_url",
+                "image_url": {
+                    "url": f"data:image/jpeg;base64,{image_base64}"
+                },
+            }
+        else:
+            mm_content = None
+
+        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
 
     return sampled_requests
 
 
 def sample_random_requests(
-        input_len: int, output_len: int, num_prompts: int, range_ratio: float,
-        tokenizer: PreTrainedTokenizerBase) -> List[Tuple[str, int, int]]:
+    prefix_len: int,
+    input_len: int,
+    output_len: int,
+    num_prompts: int,
+    range_ratio: float,
+    tokenizer: PreTrainedTokenizerBase,
+) -> List[Tuple[str, int, int]]:
+    prefix_token_ids = np.random.randint(0,
+                                         tokenizer.vocab_size,
+                                         size=prefix_len).tolist()
 
     input_lens = np.random.randint(
         int(input_len * range_ratio),
@@ -204,10 +281,12 @@ def sample_random_requests(
     offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)
     input_requests = []
     for i in range(num_prompts):
-        prompt = tokenizer.decode([(offsets[i] + i + j) % tokenizer.vocab_size
+        prompt = tokenizer.decode(prefix_token_ids +
+                                  [(offsets[i] + i + j) % tokenizer.vocab_size
                                    for j in range(input_lens[i])])
-        input_requests.append(
-            (prompt, int(input_lens[i]), int(output_lens[i])))
+
+        input_requests.append((prompt, int(prefix_len + input_lens[i]),
+                               int(output_lens[i]), None))
 
     return input_requests
 
@@ -235,6 +314,8 @@ def calculate_metrics(
     outputs: List[RequestFuncOutput],
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
+    selected_percentile_metrics: List[str],
+    selected_percentiles: List[float],
 ) -> Tuple[BenchmarkMetrics, List[int]]:
     actual_output_lens: List[int] = []
     total_input = 0
@@ -242,6 +323,7 @@ def calculate_metrics(
     itls: List[float] = []
     tpots: List[float] = []
     ttfts: List[float] = []
+    e2els: List[float] = []
     for i in range(len(outputs)):
         if outputs[i].success:
             # We use the tokenizer to count the number of output tokens for all
@@ -258,6 +340,7 @@ def calculate_metrics(
                     (outputs[i].latency - outputs[i].ttft) / (output_len - 1))
             itls += outputs[i].itl
             ttfts.append(outputs[i].ttft)
+            e2els.append(outputs[i].latency)
             completed += 1
         else:
             actual_output_lens.append(0)
@@ -272,21 +355,29 @@ def calculate_metrics(
         total_input=total_input,
         total_output=sum(actual_output_lens),
         request_throughput=completed / dur_s,
-        input_throughput=total_input / dur_s,
         output_throughput=sum(actual_output_lens) / dur_s,
+        total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s,
         mean_ttft_ms=np.mean(ttfts or 0) *
         1000,  # ttfts is empty if streaming is not supported by backend
-        median_ttft_ms=np.median(ttfts or 0) * 1000,
         std_ttft_ms=np.std(ttfts or 0) * 1000,
-        p99_ttft_ms=np.percentile(ttfts or 0, 99) * 1000,
+        median_ttft_ms=np.median(ttfts or 0) * 1000,
+        percentiles_ttft_ms=[(p, np.percentile(ttfts or 0, p) * 1000)
+                             for p in selected_percentiles],
         mean_tpot_ms=np.mean(tpots or 0) * 1000,
-        median_tpot_ms=np.median(tpots or 0) * 1000,
         std_tpot_ms=np.std(tpots or 0) * 1000,
-        p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
+        median_tpot_ms=np.median(tpots or 0) * 1000,
+        percentiles_tpot_ms=[(p, np.percentile(tpots or 0, p) * 1000)
+                             for p in selected_percentiles],
         mean_itl_ms=np.mean(itls or 0) * 1000,
-        median_itl_ms=np.median(itls or 0) * 1000,
         std_itl_ms=np.std(itls or 0) * 1000,
-        p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
+        median_itl_ms=np.median(itls or 0) * 1000,
+        percentiles_itl_ms=[(p, np.percentile(itls or 0, p) * 1000)
+                            for p in selected_percentiles],
+        mean_e2el_ms=np.median(e2els or 0) * 1000,
+        std_e2el_ms=np.std(e2els or 0) * 1000,
+        median_e2el_ms=np.mean(e2els or 0) * 1000,
+        percentiles_e2el_ms=[(p, np.percentile(e2els or 0, p) * 1000)
+                             for p in selected_percentiles],
     )
 
     return metrics, actual_output_lens
@@ -295,13 +386,18 @@ def calculate_metrics(
 async def benchmark(
     backend: str,
     api_url: str,
+    base_url: str,
     model_id: str,
     tokenizer: PreTrainedTokenizerBase,
     input_requests: List[Tuple[str, int, int]],
+    logprobs: Optional[int],
     best_of: int,
     use_beam_search: bool,
     request_rate: float,
     disable_tqdm: bool,
+    profile: bool,
+    selected_percentile_metrics: List[str],
+    selected_percentiles: List[str],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -309,15 +405,22 @@ async def benchmark(
         raise ValueError(f"Unknown backend: {backend}")
 
     print("Starting initial single prompt test run...")
-    test_prompt, test_prompt_len, test_output_len = input_requests[0]
+    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
+        input_requests[0])
+    if backend != "openai-chat" and test_mm_content is not None:
+        # multi-modal benchmark is only available on OpenAI Chat backend.
+        raise ValueError(
+            "Multi-modal content is only supported on 'openai-chat' backend.")
     test_input = RequestFuncInput(
         model=model_id,
         prompt=test_prompt,
         api_url=api_url,
         prompt_len=test_prompt_len,
         output_len=test_output_len,
+        logprobs=logprobs,
         best_of=best_of,
         use_beam_search=use_beam_search,
+        multi_modal_content=test_mm_content,
     )
     test_output = await request_func(request_func_input=test_input)
     if not test_output.success:
@@ -326,6 +429,24 @@ async def benchmark(
             f"are correctly specified. Error: {test_output.error}")
     else:
         print("Initial test run completed. Starting main benchmark run...")
+
+    if profile:
+        print("Starting profiler...")
+        profile_input = RequestFuncInput(
+            model=model_id,
+            prompt=test_prompt,
+            api_url=base_url + "/start_profile",
+            prompt_len=test_prompt_len,
+            output_len=test_output_len,
+            logprobs=logprobs,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+            multi_modal_content=test_mm_content,
+        )
+        profile_output = await request_func(request_func_input=profile_input)
+        if profile_output.success:
+            print("Profiler started")
+
     print(f"Traffic request rate: {request_rate}")
 
     pbar = None if disable_tqdm else tqdm(total=len(input_requests))
@@ -333,15 +454,17 @@ async def benchmark(
     benchmark_start_time = time.perf_counter()
     tasks: List[asyncio.Task] = []
     async for request in get_request(input_requests, request_rate):
-        prompt, prompt_len, output_len = request
+        prompt, prompt_len, output_len, mm_content = request
         request_func_input = RequestFuncInput(
             model=model_id,
             prompt=prompt,
             api_url=api_url,
             prompt_len=prompt_len,
             output_len=output_len,
+            logprobs=logprobs,
             best_of=best_of,
             use_beam_search=use_beam_search,
+            multi_modal_content=mm_content,
         )
         tasks.append(
             asyncio.create_task(
@@ -349,6 +472,22 @@ async def benchmark(
                              pbar=pbar)))
     outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
 
+    if profile:
+        print("Stopping profiler...")
+        profile_input = RequestFuncInput(
+            model=model_id,
+            prompt=test_prompt,
+            api_url=base_url + "/stop_profile",
+            prompt_len=test_prompt_len,
+            output_len=test_output_len,
+            logprobs=logprobs,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+        )
+        profile_output = await request_func(request_func_input=profile_input)
+        if profile_output.success:
+            print("Profiler stopped")
+
     if pbar is not None:
         pbar.close()
 
@@ -359,6 +498,8 @@ async def benchmark(
         outputs=outputs,
         dur_s=benchmark_duration,
         tokenizer=tokenizer,
+        selected_percentile_metrics=selected_percentile_metrics,
+        selected_percentiles=selected_percentiles,
     )
 
     print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -370,27 +511,10 @@ async def benchmark(
                                  metrics.total_output))
     print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                     metrics.request_throughput))
-    print("{:<40} {:<10.2f}".format("Input token throughput (tok/s):",
-                                    metrics.input_throughput))
     print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
                                     metrics.output_throughput))
-    print("{s:{c}^{n}}".format(s='Time to First Token', n=50, c='-'))
-    print("{:<40} {:<10.2f}".format("Mean TTFT (ms):", metrics.mean_ttft_ms))
-    print("{:<40} {:<10.2f}".format("Median TTFT (ms):",
-                                    metrics.median_ttft_ms))
-    print("{:<40} {:<10.2f}".format("P99 TTFT (ms):", metrics.p99_ttft_ms))
-    print("{s:{c}^{n}}".format(s='Time per Output Token (excl. 1st token)',
-                               n=50,
-                               c='-'))
-    print("{:<40} {:<10.2f}".format("Mean TPOT (ms):", metrics.mean_tpot_ms))
-    print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
-                                    metrics.median_tpot_ms))
-    print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
-    print("{s:{c}^{n}}".format(s='Inter-token Latency', n=50, c='-'))
-    print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
-    print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
-    print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
-    print("=" * 50)
+    print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
+                                    metrics.total_token_throughput))
 
     result = {
         "duration": benchmark_duration,
@@ -398,20 +522,8 @@ async def benchmark(
         "total_input_tokens": metrics.total_input,
         "total_output_tokens": metrics.total_output,
         "request_throughput": metrics.request_throughput,
-        "input_throughput": metrics.input_throughput,
         "output_throughput": metrics.output_throughput,
-        "mean_ttft_ms": metrics.mean_ttft_ms,
-        "median_ttft_ms": metrics.median_ttft_ms,
-        "std_ttft_ms": metrics.std_ttft_ms,
-        "p99_ttft_ms": metrics.p99_ttft_ms,
-        "mean_tpot_ms": metrics.mean_tpot_ms,
-        "median_tpot_ms": metrics.median_tpot_ms,
-        "std_tpot_ms": metrics.std_tpot_ms,
-        "p99_tpot_ms": metrics.p99_tpot_ms,
-        "mean_itl_ms": metrics.mean_itl_ms,
-        "median_itl_ms": metrics.median_itl_ms,
-        "std_itl_ms": metrics.std_itl_ms,
-        "p99_itl_ms": metrics.p99_itl_ms,
+        "total_token_throughput": metrics.total_token_throughput,
         "input_lens": [output.prompt_len for output in outputs],
         "output_lens": actual_output_lens,
         "ttfts": [output.ttft for output in outputs],
@@ -419,6 +531,47 @@ async def benchmark(
         "generated_texts": [output.generated_text for output in outputs],
         "errors": [output.error for output in outputs],
     }
+
+    def process_one_metric(
+        # E.g., "ttft"
+        metric_attribute_name: str,
+        # E.g., "TTFT"
+        metric_name: str,
+        # E.g., "Time to First Token"
+        metric_header: str,
+    ):
+        # This function print and add statistics of the specified
+        # metric.
+        if metric_attribute_name not in selected_percentile_metrics:
+            return
+        print("{s:{c}^{n}}".format(s=metric_header, n=50, c='-'))
+        print("{:<40} {:<10.2f}".format(
+            f"Mean {metric_name} (ms):",
+            getattr(metrics, f"mean_{metric_attribute_name}_ms")))
+        print("{:<40} {:<10.2f}".format(
+            f"Median {metric_name} (ms):",
+            getattr(metrics, f"median_{metric_attribute_name}_ms")))
+        result[f"mean_{metric_attribute_name}_ms"] = getattr(
+            metrics, f"mean_{metric_attribute_name}_ms")
+        result[f"median_{metric_attribute_name}_ms"] = getattr(
+            metrics, f"median_{metric_attribute_name}_ms")
+        result[f"std_{metric_attribute_name}_ms"] = getattr(
+            metrics, f"std_{metric_attribute_name}_ms")
+        for p, value in getattr(metrics,
+                                f"percentiles_{metric_attribute_name}_ms"):
+            p_word = str(int(p)) if int(p) == p else str(p)
+            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):",
+                                            value))
+            result[f"p{p_word}_{metric_attribute_name}_ms"] = value
+
+    process_one_metric("ttft", "TTFT", "Time to First Token")
+    process_one_metric("tpot", "TPOT",
+                       "Time per Output Token (excl. 1st token)")
+    process_one_metric("itl", "ITL", "Inter-token Latency")
+    process_one_metric("e2el", "E2EL", "End-to-end Latency")
+
+    print("=" * 50)
+
     return result
 
 
@@ -433,8 +586,10 @@ def main(args: argparse.Namespace):
 
     if args.base_url is not None:
         api_url = f"{args.base_url}{args.endpoint}"
+        base_url = f"{args.base_url}"
     else:
         api_url = f"http://{args.host}:{args.port}{args.endpoint}"
+        base_url = f"http://{args.host}:{args.port}"
 
     tokenizer = get_tokenizer(tokenizer_id,
                               trust_remote_code=args.trust_remote_code)
@@ -471,9 +626,9 @@ def main(args: argparse.Namespace):
                 prefix_len=args.sonnet_prefix_len,
                 tokenizer=tokenizer,
             )
-            input_requests = [(prompt, prompt_len, output_len)
+            input_requests = [(prompt, prompt_len, output_len, None)
                               for prompt, prompt_formatted, prompt_len,
-                              output_len in input_requests]
+                              output_len, _ in input_requests]
         else:
             assert (
                 tokenizer.chat_template or tokenizer.default_chat_template
@@ -486,12 +641,23 @@ def main(args: argparse.Namespace):
                 prefix_len=args.sonnet_prefix_len,
                 tokenizer=tokenizer,
             )
-            input_requests = [(prompt_formatted, prompt_len, output_len)
+            input_requests = [(prompt_formatted, prompt_len, output_len, None)
                               for prompt, prompt_formatted, prompt_len,
-                              output_len in input_requests]
+                              output_len, _ in input_requests]
+
+    elif args.dataset_name == "hf":
+        input_requests = sample_hf_requests(
+            dataset_path=args.dataset_path,
+            dataset_subset=args.hf_subset,
+            dataset_split=args.hf_split,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            fixed_output_len=args.hf_output_len,
+        )
 
     elif args.dataset_name == "random":
         input_requests = sample_random_requests(
+            prefix_len=args.random_prefix_len,
             input_len=args.random_input_len,
             output_len=args.random_output_len,
             num_prompts=args.num_prompts,
@@ -506,13 +672,20 @@ def main(args: argparse.Namespace):
         benchmark(
             backend=backend,
             api_url=api_url,
+            base_url=base_url,
             model_id=model_id,
             tokenizer=tokenizer,
             input_requests=input_requests,
+            logprobs=args.logprobs,
             best_of=args.best_of,
             use_beam_search=args.use_beam_search,
             request_rate=args.request_rate,
             disable_tqdm=args.disable_tqdm,
+            profile=args.profile,
+            selected_percentile_metrics=args.percentile_metrics.split(","),
+            selected_percentiles=[
+                float(p) for p in args.metric_percentiles.split(",")
+            ],
         ))
 
     # Save config and results to json
@@ -592,13 +765,14 @@ if __name__ == "__main__":
         "--dataset-name",
         type=str,
         default="sharegpt",
-        choices=["sharegpt", "sonnet", "random"],
+        choices=["sharegpt", "sonnet", "random", "hf"],
         help="Name of the dataset to benchmark on.",
     )
     parser.add_argument("--dataset-path",
                         type=str,
                         default=None,
-                        help="Path to the dataset.")
+                        help="Path to the sharegpt/sonnet dataset. "
+                        "Or the huggingface dataset ID if using HF dataset.")
     parser.add_argument(
         "--model",
         type=str,
@@ -626,52 +800,14 @@ if __name__ == "__main__":
         help="Number of prompts to process.",
     )
     parser.add_argument(
-        "--sharegpt-output-len",
+        "--logprobs",
         type=int,
         default=None,
-        help="Output length for each request. Overrides the output length "
-        "from the ShareGPT dataset.")
-    parser.add_argument(
-        "--sonnet-input-len",
-        type=int,
-        default=550,
-        help=
-        "Number of input tokens per request, used only for sonnet dataset.",
-    )
-    parser.add_argument(
-        "--sonnet-output-len",
-        type=int,
-        default=150,
-        help=
-        "Number of output tokens per request, used only for sonnet dataset.",
-    )
-    parser.add_argument(
-        "--sonnet-prefix-len",
-        type=int,
-        default=200,
-        help=
-        "Number of prefix tokens per request, used only for sonnet dataset.",
-    )
-    parser.add_argument(
-        "--random-input-len",
-        type=int,
-        default=1024,
-        help=
-        "Number of input tokens per request, used only for random sampling.",
-    )
-    parser.add_argument(
-        "--random-output-len",
-        type=int,
-        default=128,
-        help=
-        "Number of output tokens per request, used only for random sampling.",
-    )
-    parser.add_argument(
-        "--random-range-ratio",
-        type=float,
-        default=1.0,
-        help="Range of sampled ratio of input/output length, "
-        "used only for random sampling.",
+        help=("Number of logprobs-per-token to compute & return as part of "
+              "the request. If unspecified, then either (1) if beam search "
+              "is disabled, no logprobs are computed & a single dummy "
+              "logprob is returned for each token; or (2) if beam search "
+              "is enabled 1 logprob per token is computed"),
     )
     parser.add_argument(
         "--request-rate",
@@ -693,6 +829,12 @@ if __name__ == "__main__":
         action="store_true",
         help="Specify to disable tqdm progress bar.",
     )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="Use Torch Profiler. The endpoint must be launched with "
+        "VLLM_TORCH_PROFILER_DIR to enable profiler.",
+    )
     parser.add_argument(
         "--save-result",
         action="store_true",
@@ -722,6 +864,103 @@ if __name__ == "__main__":
         "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
         " format.",
     )
+    parser.add_argument(
+        "--percentile-metrics",
+        type=str,
+        default="ttft,tpot,itl",
+        help="Comma-seperated list of selected metrics to report percentils. "
+        "This argument specifies the metrics to report percentiles. "
+        "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
+        "Default value is \"ttft,tpot,itl\".")
+    parser.add_argument(
+        "--metric-percentiles",
+        type=str,
+        default="99",
+        help="Comma-seperated list of percentiles for selected metrics. "
+        "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
+        "Default value is \"99\". "
+        "Use \"--percentile-metrics\" to select metrics.",
+    )
+
+    # group for dataset specific arguments
+    sonnet_group = parser.add_argument_group("sonnet dataset options")
+    sonnet_group.add_argument(
+        "--sonnet-input-len",
+        type=int,
+        default=550,
+        help=
+        "Number of input tokens per request, used only for sonnet dataset.",
+    )
+    sonnet_group.add_argument(
+        "--sonnet-output-len",
+        type=int,
+        default=150,
+        help=
+        "Number of output tokens per request, used only for sonnet dataset.",
+    )
+    sonnet_group.add_argument(
+        "--sonnet-prefix-len",
+        type=int,
+        default=200,
+        help=
+        "Number of prefix tokens per request, used only for sonnet dataset.",
+    )
+
+    sharegpt_group = parser.add_argument_group("sharegpt dataset options")
+    sharegpt_group.add_argument(
+        "--sharegpt-output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the output length "
+        "from the ShareGPT dataset.")
+
+    random_group = parser.add_argument_group("random dataset options")
+    random_group.add_argument(
+        "--random-input-len",
+        type=int,
+        default=1024,
+        help=
+        "Number of input tokens per request, used only for random sampling.",
+    )
+    random_group.add_argument(
+        "--random-output-len",
+        type=int,
+        default=128,
+        help=
+        "Number of output tokens per request, used only for random sampling.",
+    )
+    random_group.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=1.0,
+        help="Range of sampled ratio of input/output length, "
+        "used only for random sampling.",
+    )
+    random_group.add_argument(
+        "--random-prefix-len",
+        type=int,
+        default=0,
+        help="Number of fixed prefix tokens before random "
+        " context. The length range of context in a random "
+        " request is [random-prefix-len, "
+        " random-prefix-len + random-prefix-len * random-range-ratio).")
+
+    hf_group = parser.add_argument_group("hf dataset options")
+    hf_group.add_argument("--hf-subset",
+                          type=str,
+                          default=None,
+                          help="Subset of the HF dataset.")
+    hf_group.add_argument("--hf-split",
+                          type=str,
+                          default=None,
+                          help="Split of the HF dataset.")
+    hf_group.add_argument(
+        "--hf-output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the output lengths "
+        "from the sampled HF dataset.",
+    )
 
     args = parser.parse_args()
-    main(args)
+    main(args)

benchmarks/benchmark_throughput.py

@@ -6,13 +6,16 @@ import time
 from typing import List, Optional, Tuple
 
 import torch
+import uvloop
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                           PreTrainedTokenizerBase)
 
-from vllm.engine.arg_utils import EngineArgs
+from vllm.engine.arg_utils import DEVICE_OPTIONS, AsyncEngineArgs, EngineArgs
+from vllm.entrypoints.openai.api_server import (
+    build_async_engine_client_from_engine_args)
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, merge_async_iterators
 
 
 def sample_requests(
@@ -82,8 +85,12 @@ def run_vllm(
     max_num_batched_tokens: int,
     distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
+    num_scheduler_steps: int = 1,
+    use_v2_block_manager: bool = False,
     download_dir: Optional[str] = None,
     load_format: str = EngineArgs.load_format,
+    disable_async_output_proc: bool = False,
+    use_new_beam_search_impl: bool = False,
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(
@@ -106,6 +113,9 @@ def run_vllm(
         max_num_batched_tokens=max_num_batched_tokens,
         distributed_executor_backend=distributed_executor_backend,
         load_format=load_format,
+        num_scheduler_steps=num_scheduler_steps,
+        use_v2_block_manager=use_v2_block_manager,
+        disable_async_output_proc=disable_async_output_proc,
     )
 
     # Add the requests to the engine.
@@ -123,12 +133,112 @@ def run_vllm(
                 max_tokens=output_len,
             ))
 
-    start = time.perf_counter()
-    llm.generate(prompts, sampling_params, use_tqdm=True)
-    end = time.perf_counter()
+    if not use_new_beam_search_impl:
+        start = time.perf_counter()
+        llm.generate(prompts, sampling_params, use_tqdm=True)
+        end = time.perf_counter()
+    else:
+        assert use_beam_search
+        prompts = [prompt for prompt, _, _ in requests]
+        # output_len should be the same for all requests.
+        output_len = requests[0][2]
+        for prompt, input_len, _output_len in requests:
+            assert _output_len == output_len
+        start = time.perf_counter()
+        llm.beam_search(prompts,
+                        beam_width=n,
+                        max_tokens=output_len,
+                        ignore_eos=True)
+        end = time.perf_counter()
     return end - start
 
 
+async def run_vllm_async(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: str,
+    quantization: Optional[str],
+    tensor_parallel_size: int,
+    seed: int,
+    n: int,
+    use_beam_search: bool,
+    trust_remote_code: bool,
+    dtype: str,
+    max_model_len: Optional[int],
+    enforce_eager: bool,
+    kv_cache_dtype: str,
+    quantization_param_path: Optional[str],
+    device: str,
+    enable_prefix_caching: bool,
+    enable_chunked_prefill: bool,
+    max_num_batched_tokens: int,
+    distributed_executor_backend: Optional[str],
+    gpu_memory_utilization: float = 0.9,
+    num_scheduler_steps: int = 1,
+    use_v2_block_manager: bool = False,
+    download_dir: Optional[str] = None,
+    load_format: str = EngineArgs.load_format,
+    disable_async_output_proc: bool = False,
+    disable_frontend_multiprocessing: bool = False,
+) -> float:
+    from vllm import SamplingParams
+    engine_args = AsyncEngineArgs(
+        model=model,
+        tokenizer=tokenizer,
+        quantization=quantization,
+        tensor_parallel_size=tensor_parallel_size,
+        seed=seed,
+        trust_remote_code=trust_remote_code,
+        dtype=dtype,
+        max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_memory_utilization,
+        enforce_eager=enforce_eager,
+        kv_cache_dtype=kv_cache_dtype,
+        quantization_param_path=quantization_param_path,
+        device=device,
+        enable_prefix_caching=enable_prefix_caching,
+        download_dir=download_dir,
+        enable_chunked_prefill=enable_chunked_prefill,
+        max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
+        load_format=load_format,
+        num_scheduler_steps=num_scheduler_steps,
+        use_v2_block_manager=use_v2_block_manager,
+        disable_async_output_proc=disable_async_output_proc,
+        worker_use_ray=False,
+        disable_log_requests=True,
+    )
+
+    async with build_async_engine_client_from_engine_args(
+            engine_args, disable_frontend_multiprocessing) as llm:
+
+        # Add the requests to the engine.
+        prompts: List[str] = []
+        sampling_params: List[SamplingParams] = []
+        for prompt, _, output_len in requests:
+            prompts.append(prompt)
+            sampling_params.append(
+                SamplingParams(
+                    n=n,
+                    temperature=0.0 if use_beam_search else 1.0,
+                    top_p=1.0,
+                    use_beam_search=use_beam_search,
+                    ignore_eos=True,
+                    max_tokens=output_len,
+                ))
+
+        generators = []
+        start = time.perf_counter()
+        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
+            generator = llm.generate(prompt, sp, request_id=f"test{i}")
+            generators.append(generator)
+        all_gens = merge_async_iterators(*generators)
+        async for i, res in all_gens:
+            pass
+        end = time.perf_counter()
+        return end - start
+
+
 def run_hf(
     requests: List[Tuple[str, int, int]],
     model: str,
@@ -224,7 +334,7 @@ def main(args: argparse.Namespace):
                                    args.output_len)
 
     if args.backend == "vllm":
-        elapsed_time = run_vllm(
+        run_args = [
             requests, args.model, args.tokenizer, args.quantization,
             args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
             args.trust_remote_code, args.dtype, args.max_model_len,
@@ -232,7 +342,16 @@ def main(args: argparse.Namespace):
             args.quantization_param_path, args.device,
             args.enable_prefix_caching, args.enable_chunked_prefill,
             args.max_num_batched_tokens, args.distributed_executor_backend,
-            args.gpu_memory_utilization, args.download_dir, args.load_format)
+            args.gpu_memory_utilization, args.num_scheduler_steps,
+            args.use_v2_block_manager, args.download_dir, args.load_format,
+            args.disable_async_output_proc
+        ]
+
+        if args.async_engine:
+            run_args.append(args.disable_frontend_multiprocessing)
+            elapsed_time = uvloop.run(run_vllm_async(*run_args))
+        else:
+            elapsed_time = run_vllm(*run_args, args.use_new_beam_search_impl)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -292,6 +411,7 @@ if __name__ == "__main__":
                         default=1,
                         help="Number of generated sequences per prompt.")
     parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument("--use-new-beam-search-impl", action="store_true")
     parser.add_argument("--num-prompts",
                         type=int,
                         default=1000,
@@ -346,17 +466,23 @@ if __name__ == "__main__":
         'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
         'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
         'instead supported for common inference criteria.')
+    parser.add_argument("--device",
+                        type=str,
+                        default="auto",
+                        choices=DEVICE_OPTIONS,
+                        help='device type for vLLM execution')
     parser.add_argument(
-        "--device",
-        type=str,
-        default="auto",
-        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
-        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
-        'CPU.')
+        "--num-scheduler-steps",
+        type=int,
+        default=1,
+        help="Maximum number of forward steps per scheduler call.")
+    parser.add_argument("--use-v2-block-manager",
+                        action='store_true',
+                        help="Enable block manager v2.")
     parser.add_argument(
         "--enable-prefix-caching",
         action='store_true',
-        help="enable automatic prefix caching for vLLM backend.")
+        help="Enable automatic prefix caching for vLLM backend.")
     parser.add_argument("--enable-chunked-prefill",
                         action='store_true',
                         help="enable chunked prefill for vLLM backend.")
@@ -405,6 +531,19 @@ if __name__ == "__main__":
         'section for more information.\n'
         '* "bitsandbytes" will load the weights using bitsandbytes '
         'quantization.\n')
+    parser.add_argument(
+        "--disable-async-output-proc",
+        action='store_true',
+        default=False,
+        help="Disable async output processor for vLLM backend.")
+    parser.add_argument("--async-engine",
+                        action='store_true',
+                        default=False,
+                        help="Use vLLM async engine rather than LLM class.")
+    parser.add_argument("--disable-frontend-multiprocessing",
+                        action='store_true',
+                        default=False,
+                        help="Disable decoupled async engine frontend.")
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -13,7 +13,7 @@ from weight_shapes import WEIGHT_SHAPES
 from vllm import _custom_ops as ops
 from vllm.utils import FlexibleArgumentParser
 
-DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())[1:]
+DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
 DEFAULT_TP_SIZES = [1]
 
@@ -32,7 +32,6 @@ def to_int8(tensor: torch.Tensor) -> torch.Tensor:
 
 def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
                       k: int) -> Tuple[torch.Tensor, torch.Tensor]:
-
     a = torch.randn((m, k), device='cuda') * 5
     b = torch.randn((n, k), device='cuda').t() * 5
 
@@ -44,59 +43,18 @@ def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
     raise ValueError("unsupported dtype")
 
 
-# impl
-
-
-def pytorch_mm_impl(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
-                    scale_b: torch.Tensor,
-                    out_dtype: torch.dtype) -> torch.Tensor:
-    return torch.mm(a, b)
-
-
-def pytorch_fp8_impl(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
-                     scale_b: torch.Tensor,
-                     out_dtype: torch.dtype) -> torch.Tensor:
-    return torch._scaled_mm(a,
-                            b,
-                            scale_a=scale_a,
-                            scale_b=scale_b,
-                            out_dtype=out_dtype)
-
-
-def pytorch_fp8_impl_fast_accum(a: torch.Tensor, b: torch.Tensor,
-                                scale_a: torch.Tensor, scale_b: torch.Tensor,
-                                out_dtype: torch.dtype) -> torch.Tensor:
-    return torch._scaled_mm(a,
-                            b,
-                            scale_a=scale_a,
-                            scale_b=scale_b,
-                            out_dtype=out_dtype,
-                            use_fast_accum=True)
-
-
-def cutlass_impl(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
-                 scale_b: torch.Tensor,
-                 out_dtype: torch.dtype) -> torch.Tensor:
-    return ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype=out_dtype)
-
-
 # bench
-def bench_fn(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
-             scale_b: torch.Tensor, out_dtype: torch.dtype, label: str,
-             sub_label: str, fn: Callable, description: str) -> TMeasurement:
-
+def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
+             **kwargs) -> TMeasurement:
     min_run_time = 1
 
     globals = {
-        "a": a,
-        "b": b,
-        "scale_a": scale_a,
-        "scale_b": scale_b,
-        "out_dtype": out_dtype,
+        "args": args,
+        "kwargs": kwargs,
         "fn": fn,
     }
     return TBenchmark.Timer(
-        stmt="fn(a, b, scale_a, scale_b, out_dtype)",
+        stmt="fn(*args, **kwargs)",
         globals=globals,
         label=label,
         sub_label=sub_label,
@@ -110,19 +68,58 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
     a, b = make_rand_tensors(torch.int8, m, n, k)
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+    azp = torch.zeros((m, ), device="cuda", dtype=torch.int32)
+    azp_adj = torch.zeros((n, ), device="cuda", dtype=torch.int32)
 
     timers = []
-    # pytorch impl
+    # pytorch impl - bfloat16
     timers.append(
-        bench_fn(a.to(dtype=torch.bfloat16, device="cuda"),
-                 b.to(dtype=torch.bfloat16, device="cuda"), scale_a, scale_b,
-                 torch.bfloat16, label, sub_label, pytorch_mm_impl,
-                 "pytorch_bf16_bf16_bf16_matmul-no-scales"))
+        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+                 torch.mm, a.to(dtype=torch.bfloat16),
+                 b.to(dtype=torch.bfloat16)))
+
+    # pytorch impl - float16
+    timers.append(
+        bench_fn(label, sub_label,
+                 "pytorch_fp16_fp16_fp16_matmul-no-scales", torch.mm,
+                 a.to(dtype=torch.float16), b.to(dtype=torch.float16)))
 
     # cutlass impl
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
-                 cutlass_impl, "cutlass_i8_i8_bf16_scaled_mm"))
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
+                 torch.bfloat16))
+
+    # cutlass with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_bias",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
+                 bias))
+
+    # cutlass with azp per-tensor
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj))
+
+    # cutlass with azp per-tensor + bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_bias",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj, None, bias))
+
+    # cutlass with azp per-token
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj, azp))
+
+    # cutlass with azp per-token + bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj, azp, bias))
 
     return timers
 
@@ -133,46 +130,88 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
     a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
 
     timers = []
 
     # pytorch impl w. bf16
     timers.append(
-        bench_fn(a.to(dtype=torch.bfloat16, device="cuda"),
-                 b.to(dtype=torch.bfloat16, device="cuda"), scale_a, scale_b,
-                 torch.bfloat16, label, sub_label, pytorch_mm_impl,
-                 "pytorch_bf16_bf16_bf16_matmul-no-scales"))
+        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
+                 b.to(dtype=torch.bfloat16, device="cuda")))
 
     # pytorch impl: bf16 output, without fp8 fast accum
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
-                 pytorch_fp8_impl, "pytorch_fp8_fp8_bf16_scaled_mm"))
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_bf16_scaled_mm",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.bfloat16))
 
     # pytorch impl: bf16 output, with fp8 fast accum
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
-                 pytorch_fp8_impl_fast_accum,
-                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum"))
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.bfloat16,
+                 use_fast_accum=True))
 
     # pytorch impl: fp16 output, without fp8 fast accum
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
-                 pytorch_fp8_impl, "pytorch_fp8_fp8_fp16_scaled_mm"))
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_fp16_scaled_mm",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.float16))
 
     # pytorch impl: fp16 output, with fp8 fast accum
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
-                 pytorch_fp8_impl_fast_accum,
-                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum"))
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.float16,
+                 use_fast_accum=True))
 
     # cutlass impl: bf16 output
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
-                 cutlass_impl, "cutlass_fp8_fp8_bf16_scaled_mm"))
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
+                 torch.bfloat16))
     # cutlass impl: fp16 output
     timers.append(
-        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
-                 cutlass_impl, "cutlass_fp8_fp8_fp16_scaled_mm"))
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16))
+
+    # cutlass impl: bf16 output, with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm_bias",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
+                 bias))
+
+    # cutlass impl: fp16 output, with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm_bias",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16,
+                 bias.to(dtype=torch.float16)))
+
     return timers
 
 
@@ -193,7 +232,6 @@ def print_timers(timers: Iterable[TMeasurement]):
 
 def run(dtype: torch.dtype,
         MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
-
     results = []
     for m, k, n in MKNs:
         timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
@@ -209,7 +247,6 @@ def make_output(data: Iterable[TMeasurement],
                 MKNs: Iterable[Tuple[int, int, int]],
                 base_description: str,
                 timestamp=None):
-
     print(f"== All Results {base_description} ====")
     print_timers(data)
 
@@ -244,7 +281,6 @@ def run_range_bench(args):
 
 
 def run_model_bench(args):
-
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")

benchmarks/kernels/benchmark_layernorm.py

@@ -0,0 +1,86 @@
+import time
+
+import torch
+
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
+                        seed_everything)
+
+
+@torch.inference_mode()
+def main(num_tokens: int,
+         hidden_size: int,
+         add_residual: bool,
+         dtype: torch.dtype,
+         seed: int = 0,
+         do_profile: bool = False,
+         num_warmup_iters: int = 5,
+         num_iters: int = 100) -> None:
+    seed_everything(seed)
+    torch.set_default_device("cuda")
+
+    layer = RMSNorm(hidden_size).to(dtype=dtype)
+    layer.weight.data.normal_(mean=1.0, std=0.1)
+    scale = 1 / (2 * hidden_size)
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    x *= scale
+    residual = torch.randn_like(x) * scale if add_residual else None
+
+    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
+        torch.cuda.synchronize()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        start_time = time.perf_counter()
+
+        for _ in range(num_iters):
+            layer(x, residual)
+        torch.cuda.synchronize()
+
+        end_time = time.perf_counter()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        return (end_time - start_time) / num_iters
+
+    # Warmup.
+    print("Warming up...")
+    run_benchmark = run_cuda_benchmark
+    run_benchmark(num_iters=num_warmup_iters, profile=False)
+
+    # Benchmark.
+    if do_profile:
+        latency = run_benchmark(num_iters=1, profile=True)
+    else:
+        latency = run_benchmark(num_iters=num_iters, profile=False)
+    print(f"Kernel running time: {latency * 1000000:.3f} us")
+
+
+if __name__ == '__main__':
+    parser = FlexibleArgumentParser(
+        description="Benchmark the layernorm kernel.")
+    parser.add_argument("--num-tokens", type=int, default=4096)
+    parser.add_argument("--hidden-size", type=int, default=8192)
+    parser.add_argument("--add-residual", action="store_true")
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["half", "bfloat16", "float"],
+                        default="half")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--profile", action="store_true")
+    parser.add_argument("--num-warmup-iters", type=int, default=5)
+    parser.add_argument("--num-iters",
+                        type=int,
+                        default=100,
+                        help="Number of benchmark iterations. "
+                        "If --profile is set, this number is ignored")
+
+    args = parser.parse_args()
+    print(args)
+
+    main(num_tokens=args.num_tokens,
+         hidden_size=args.hidden_size,
+         add_residual=args.add_residual,
+         dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+         seed=args.seed,
+         do_profile=args.profile,
+         num_warmup_iters=args.num_warmup_iters,
+         num_iters=args.num_iters)

benchmarks/kernels/benchmark_machete.py

@@ -0,0 +1,420 @@
+import argparse
+import copy
+import itertools
+import math
+import pickle as pkl
+import time
+from itertools import product
+from typing import Callable, Iterable, List, Optional, Tuple
+
+import pandas as pd
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N, marlin_permute_scales)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
+    MarlinWorkspace)
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    gptq_pack, pack_rows, quantize_weights)
+from vllm.scalar_type import ScalarType, scalar_types
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = ["meta-llama/Llama-3-8b", "meta-llama/Llama-2-70b-hf"]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024]
+DEFAULT_TP_SIZES = [1]
+
+
+def machete_pack_weights(w_q: torch.tensor, wtype: ScalarType) -> torch.tensor:
+    w_q = pack_rows(w_q, wtype.size_bits, *w_q.shape)
+    w_q = w_q.t().contiguous().t()  # make col major
+    return ops.machete_prepack_B(w_q, wtype)
+
+
+def make_bench_tensors(
+    atype: torch.dtype, wtype: ScalarType, group_size: int, m: int, n: int,
+    k: int
+) -> Tuple[torch.tensor, List[Tuple[torch.tensor, torch.tensor, torch.tensor,
+                                    torch.tensor]]]:
+    assert wtype.is_integer(), "TODO: support floating point weights"
+
+    # we want to make sure that weights don't fit into L2 cache between runs so
+    #  we construct enough weights to exceed L2 cache, which is 50mb on a H100
+    #  so we target total weight size > 2*50mb
+    num_weights = math.ceil(2 * 50 * 1024**2 * 8 / (k * n * wtype.size_bits))
+
+    a = torch.randn((m, k), device="cuda", dtype=atype) * 5
+    weights = [
+        torch.randn((k, n), device="cuda", dtype=atype)
+        for _ in range(num_weights)
+    ]
+    quanitized_weights = [
+        quantize_weights(w, wtype, group_size) for w in weights
+    ]
+
+    return a, quanitized_weights
+
+
+# impl
+
+
+# bench
+def bench_fn(label: str, sub_label: str, description: str,
+             fn: Callable) -> TMeasurement:
+
+    min_run_time = 1
+    return TBenchmark.Timer(
+        stmt="fn()",
+        globals={
+            "fn": fn
+        },
+        label=label,
+        sub_label=sub_label,
+        description=description,
+    ).blocked_autorange(min_run_time=min_run_time)
+
+
+def loop_over_weights(
+    a: torch.tensor, weights: List[Tuple[torch.tensor, torch.tensor,
+                                         torch.tensor, torch.tensor]],
+    fn: Callable[[torch.tensor, torch.tensor, torch.tensor, torch.tensor],
+                 None]):
+    for w_ref, w_q, w_s, _ in weights:
+        fn(a, w_ref, w_q, w_s)
+
+
+_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None
+_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None
+
+
+def bench(atype: torch.dtype,
+          wtype: ScalarType,
+          group_size: int,
+          m: int,
+          k: int,
+          n: int,
+          label: str,
+          sub_label: str,
+          benchmark_marlinv1: bool = True,
+          sweep_schedules: bool = True) -> Iterable[TMeasurement]:
+    global _SWEEP_SCHEDULES_RESULTS
+
+    a, weights = make_bench_tensors(atype, wtype, group_size, m, n, k)
+    sub_label += f", L={len(weights)}"
+
+    weights_machete = [(w_ref, machete_pack_weights(w_q, wtype), w_s, w_zp)
+                       for w_ref, w_q, w_s, w_zp in weights]
+
+    timers = []
+    # pytorch impl
+    timers.append(
+        bench_fn(
+            label, sub_label, "torch.matmul", lambda: loop_over_weights(
+                a,
+                weights,
+                lambda a, w_ref, w_q, w_s: torch.matmul(a, w_ref),
+            )))
+
+    if benchmark_marlinv1:
+        w_ref = weights[0][0]
+
+        w_zp_empty = torch.empty(0, dtype=torch.int, device=w_ref.device)
+        sort_indices = torch.empty(0, dtype=torch.int, device=w_ref.device)
+        g_idx = torch.empty(0, dtype=torch.int, device=w_ref.device)
+
+        def marlinv1_pack_weights(w_q: torch.tensor) -> torch.tensor:
+            w_q_gptq = gptq_pack(w_q, wtype.size_bits, *w_ref.shape)
+            return ops.gptq_marlin_repack(w_q_gptq, sort_indices, *w_ref.shape,
+                                          wtype.size_bits)
+
+        def marlinv1_permute_scales(w_s: torch.tensor) -> torch.tensor:
+            return marlin_permute_scales(w_s, *w_ref.shape, group_size)
+
+        weights_marlinv1 = [(w_ref, marlinv1_pack_weights(w_q),
+                             marlinv1_permute_scales(w_s), w_zp)
+                            for w_ref, w_q, w_s, w_zp in weights]
+
+        workspace = MarlinWorkspace(w_ref.shape[1], GPTQ_MARLIN_MIN_THREAD_N,
+                                    GPTQ_MARLIN_MAX_PARALLEL)
+
+        # marlinv1
+        timers.append(
+            bench_fn(
+                label, sub_label, "marlin_orig", lambda: loop_over_weights(
+                    a, weights_marlinv1, lambda a, w_ref, w_q, w_s: ops.
+                    gptq_marlin_gemm(a,
+                                     w_q,
+                                     w_s,
+                                     w_zp_empty,
+                                     g_idx,
+                                     sort_indices,
+                                     workspace.scratch,
+                                     wtype,
+                                     size_m=a.shape[0],
+                                     size_n=w_ref.shape[1],
+                                     size_k=w_ref.shape[0],
+                                     is_k_full=True))))
+
+    # machete
+    timers.append(
+        bench_fn(
+            label, sub_label, "machete_heuristic", lambda: loop_over_weights(
+                a, weights_machete, lambda a, _, w_q, w_s: ops.machete_gemm(
+                    a, w_q, wtype, b_scales=w_s, b_group_size=group_size))))
+
+    if sweep_schedules:
+        print("Finding best schedule for machete")
+        best = None
+        best_schedule = None
+        schedules = ops.machete_supported_schedules(wtype)
+        for schedule in reversed(schedules):
+            schedule_M = int(schedule.split("_")[0].split("x")[1])
+
+            # Prune known bad schedules
+            if schedule_M >= 2 * max(m, 16) or schedule_M < m // 4:
+                continue
+
+            def run(a, _, w_q, w_s, schedule=schedule):
+                ops.machete_gemm(a,
+                                 w_q,
+                                 wtype,
+                                 w_s,
+                                 b_group_size=group_size,
+                                 schedule=schedule)
+
+            res = bench_fn(label, sub_label, "machete_best",
+                           lambda: loop_over_weights(a, weights_machete, run))
+
+            results_row = {
+                "M": m,
+                "K": k,
+                "N": n,
+                "group_size": group_size,
+                "schedule": schedule,
+                "median": res.median,
+            }
+            if _SWEEP_SCHEDULES_RESULTS is None:
+                _SWEEP_SCHEDULES_RESULTS = pd.DataFrame(
+                    columns=results_row.keys())
+            _SWEEP_SCHEDULES_RESULTS.\
+                loc[len(_SWEEP_SCHEDULES_RESULTS)] = results_row
+
+            print(f"  {res.median:5.5} ", schedule)
+            if not best or res.median < best.median:
+                best = res
+                best_schedule = schedule
+        print("Best schedule:", best_schedule)
+        timers.append(best)
+
+    return timers
+
+
+# runner
+def print_timers(timers: Iterable[TMeasurement]):
+    compare = TBenchmark.Compare(timers)
+    compare.print()
+
+
+def run(dtype: torch.dtype, sweep_schedules: bool,
+        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+
+    results = []
+    for m, k, n in MKNs:
+        timers = bench(dtype,
+                       scalar_types.uint4b8,
+                       128,
+                       m,
+                       k,
+                       n,
+                       f"{dtype}-gemm",
+                       f"MKN=({m}x{k}x{n})",
+                       sweep_schedules=sweep_schedules)
+        print_timers(timers)
+        results.extend(timers)
+
+    return results
+
+
+# output makers
+def make_output(
+    data: Iterable[TMeasurement],
+    MKNs: Iterable[Tuple[int, int, int]],
+    base_description: str,
+    timestamp=None,
+):
+
+    print(f"== All Results {base_description} ====")
+    print_timers(data)
+
+    # pickle all the results
+    timestamp = int(time.time()) if timestamp is None else timestamp
+    with open(f"{base_description}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(data, f)
+
+
+# argparse runners
+
+
+def run_square_bench(args):
+    dim_sizes = list(
+        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
+
+    data = run(args.dtype, args.sweep_schedules, MKNs)
+
+    make_output(data, MKNs, f"square_bench-{args.dtype}")
+
+
+def run_range_bench(args):
+    m_start, k_start, n_start = [int(x) for x in args.dim_start.split(",")]
+    m_end, k_end, n_end = [int(x) for x in args.dim_end.split(",")]
+    m_increment, k_increment, n_increment = \
+        [int(x) for x in args.dim_increment.split(",")]
+    Ms = list(range(m_start, m_end + 1, m_increment))
+    Ks = list(range(k_start, k_end + 1, k_increment))
+    Ns = list(range(n_start, n_end + 1, n_increment))
+    MKNs = list(product(Ms, Ks, Ns))
+
+    data = run(args.dtype, args.sweep_schedules, MKNs)
+
+    make_output(data, MKNs, f"range_bench-{args.dtype}")
+
+
+def run_model_bench(args):
+
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+        KNs = []
+        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
+            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
+            KNs.append(KN)
+        return KNs
+
+    model_bench_data = []
+    models_tps = list(itertools.product(args.models, args.tp_sizes))
+    for model, tp_size in models_tps:
+        Ms = args.batch_sizes
+        KNs = model_shapes(model, tp_size)
+        MKNs = []
+        for m in Ms:
+            for k, n in KNs:
+                MKNs.append((m, k, n))
+
+        data = run(args.dtype, args.sweep_schedules, MKNs)
+        model_bench_data.append(data)
+
+    # Print all results
+    for data, model_tp in zip(model_bench_data, models_tps):
+        model, tp_size = model_tp
+        print(f"== Results {args.dtype} {model}-TP{tp_size} ====")
+        print_timers(data)
+
+    timestamp = int(time.time())
+
+    all_data = []
+    for d in model_bench_data:
+        all_data.extend(d)
+    # pickle all data
+    with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(all_data, f)
+
+
+if __name__ == "__main__":
+
+    def to_torch_dtype(dt):
+        if dt == "bfloat16":
+            return torch.bfloat16
+        if dt == "float16":
+            return torch.float16
+        raise ValueError("unsupported dtype")
+
+    parser = FlexibleArgumentParser(
+        description="""
+Benchmark Machete GEMM.
+
+    To run square GEMMs:
+        python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 square_bench --dim-start 128 --dim-end 512 --dim-increment 64
+    
+    To run constant N and K and sweep M:
+        python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384
+    
+    To run dimensions from a model:
+        python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1
+    
+    Output:
+        - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
+            """,  # noqa: E501
+        formatter_class=argparse.RawTextHelpFormatter,
+    )
+
+    parser.add_argument(
+        "--dtype",
+        type=to_torch_dtype,
+        required=True,
+        help="Available options are ['bfloat16', 'float16']",
+    )
+    parser.add_argument(
+        "--sweep-schedules",
+        action="store_true",
+        help="Run a sweep over all supported schedules",
+    )
+    parser.add_argument("--sweep-csv-out",
+                        help="CSV to store sweep results",
+                        default="sch_sweep_results.csv")
+    subparsers = parser.add_subparsers(dest="cmd", required=True)
+
+    square_parser = subparsers.add_parser("square_bench")
+    square_parser.add_argument("--dim-start", type=int, required=True)
+    square_parser.add_argument("--dim-end", type=int, required=True)
+    square_parser.add_argument("--dim-increment", type=int, required=True)
+    square_parser.set_defaults(func=run_square_bench)
+
+    range_parser = subparsers.add_parser("range_bench")
+    range_parser.add_argument(
+        "--dim-start",
+        type=str,
+        required=True,
+        help="Start value for M,K,N as common separated list")
+    range_parser.add_argument(
+        "--dim-end",
+        type=str,
+        required=True,
+        help="End value (inclusive) for M,K,N as common separated list")
+    range_parser.add_argument(
+        "--dim-increment",
+        type=str,
+        required=True,
+        help="Increment value for M,K,N as common separated list")
+    range_parser.set_defaults(func=run_range_bench)
+
+    model_parser = subparsers.add_parser("model_bench")
+    model_parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES.keys(),
+    )
+    model_parser.add_argument("--tp-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_TP_SIZES)
+    model_parser.add_argument("--batch-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_BATCH_SIZES)
+    model_parser.set_defaults(func=run_model_bench)
+
+    args = parser.parse_args()
+
+    _SWEEP_SCHEDULES_RESULTS_CSV = args.sweep_csv_out
+    args.func(args)
+
+    if _SWEEP_SCHEDULES_RESULTS is not None:
+        _SWEEP_SCHEDULES_RESULTS.to_csv(_SWEEP_SCHEDULES_RESULTS_CSV)

benchmarks/kernels/benchmark_marlin.py

@@ -7,16 +7,17 @@ from benchmark_shapes import WEIGHT_SHAPES
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
     GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
-    GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_NUM_BITS)
+    GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES)
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
-    GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS)
+    MARLIN_SUPPORTED_GROUP_SIZES, query_marlin_supported_quant_types)
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
     MarlinWorkspace, marlin_quantize)
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
     marlin_24_quantize)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    gptq_pack, quantize_weights, sort_weights)
+    gptq_pack, gptq_quantize_weights, sort_weights)
+from vllm.scalar_type import ScalarType
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
@@ -27,13 +28,14 @@ K_FULL_OPTS = [False, True]
 
 
 def bench_run(results: List[benchmark.Measurement], model: str,
-              act_order: bool, is_k_full: bool, num_bits: int, group_size: int,
-              size_m: int, size_k: int, size_n: int):
+              act_order: bool, is_k_full: bool, quant_type: ScalarType,
+              group_size: int, size_m: int, size_k: int, size_n: int):
     label = "Quant Matmul"
 
-    sub_label = ("{}, act={} k_full={}, b={}, g={}, "
-                 "MKN=({}x{}x{})".format(model, act_order, is_k_full, num_bits,
-                                         group_size, size_m, size_k, size_n))
+    sub_label = ("{}, act={} k_full={}, q={}, g={}, "
+                 "MKN=({}x{}x{})".format(model, act_order, is_k_full,
+                                         str(quant_type), group_size, size_m,
+                                         size_k, size_n))
 
     print(f"Testing: {sub_label}")
 
@@ -50,16 +52,18 @@ def bench_run(results: List[benchmark.Measurement], model: str,
         marlin_g_idx,
         marlin_sort_indices,
         marlin_rand_perm,
-    ) = marlin_quantize(b, num_bits, group_size, act_order)
+    ) = marlin_quantize(b, quant_type, group_size, act_order)
 
     # Marlin_24 quant
     (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta,
-     marlin_24_s) = marlin_24_quantize(b, num_bits, group_size)
+     marlin_24_s) = marlin_24_quantize(b, quant_type, group_size)
+
+    marlin_zp = torch.empty(0, dtype=torch.int, device=b.device)
 
     # GPTQ quant
     (w_ref, q_w, s, g_idx,
-     rand_perm) = quantize_weights(b, num_bits, group_size, act_order)
-    q_w_gptq = gptq_pack(q_w, num_bits, size_k, size_n)
+     rand_perm) = gptq_quantize_weights(b, quant_type, group_size, act_order)
+    q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
 
     # For act_order, sort the "weights" and "g_idx"
     # so that group ids are increasing
@@ -73,10 +77,11 @@ def bench_run(results: List[benchmark.Measurement], model: str,
 
     marlin_24_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_24_MIN_THREAD_N,
                                           GPTQ_MARLIN_24_MAX_PARALLEL)
+    marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)
 
     globals = {
         # Gen params
-        "num_bits": num_bits,
+        "quant_type": quant_type,
         "group_size": group_size,
         "size_m": size_m,
         "size_n": size_n,
@@ -87,6 +92,7 @@ def bench_run(results: List[benchmark.Measurement], model: str,
         "marlin_w_ref": marlin_w_ref,
         "marlin_q_w": marlin_q_w,
         "marlin_s": marlin_s,
+        "marlin_zp": marlin_zp,
         "marlin_g_idx": marlin_g_idx,
         "marlin_sort_indices": marlin_sort_indices,
         "marlin_rand_perm": marlin_rand_perm,
@@ -125,19 +131,29 @@ def bench_run(results: List[benchmark.Measurement], model: str,
     results.append(
         benchmark.Timer(
             stmt=
-            "output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, num_bits, size_m, size_n, size_k, is_k_full)",  # noqa: E501
+            "output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
-            description="gptq_marlin_gemm",
+            description="gptq_marlin_gemm_fp16",
         ).blocked_autorange(min_run_time=min_run_time))
 
-    if (num_bits in GPTQ_MARLIN_24_SUPPORTED_NUM_BITS
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_gemm_fp32",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    if (quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
             and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES):
         results.append(
             benchmark.Timer(
                 stmt=
-                "output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, num_bits, size_m, size_n, size_k)",  # noqa: E501
+                "output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)",  # noqa: E501
                 globals=globals,
                 label=label,
                 sub_label=sub_label,
@@ -147,7 +163,7 @@ def bench_run(results: List[benchmark.Measurement], model: str,
     results.append(
         benchmark.Timer(
             stmt=
-            "q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, num_bits)",  # noqa: E501
+            "q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -183,12 +199,13 @@ def main(args):
                            ) > 0 and is_k_full not in args.limit_k_full:
                         continue
 
-                    for num_bits in GPTQ_MARLIN_SUPPORTED_NUM_BITS:
-                        if len(args.limit_num_bits
-                               ) > 0 and num_bits not in args.limit_num_bits:
+                    for quant_type in query_marlin_supported_quant_types(
+                            False):
+                        if len(args.limit_num_bits) > 0 and \
+                            quant_type.size_bits not in args.limit_num_bits:
                             continue
 
-                        for group_size in GPTQ_MARLIN_SUPPORTED_GROUP_SIZES:
+                        for group_size in MARLIN_SUPPORTED_GROUP_SIZES:
                             if len(
                                     args.limit_group_size
                             ) > 0 and group_size not in args.limit_group_size:
@@ -202,8 +219,8 @@ def main(args):
 
                             for size_m in args.batch_sizes:
                                 bench_run(results, model, act_order, is_k_full,
-                                          num_bits, group_size, size_m, size_k,
-                                          size_n)
+                                          quant_type, group_size, size_m,
+                                          size_k, size_n)
 
     compare = benchmark.Compare(results)
     compare.print()

benchmarks/kernels/benchmark_moe.py

@@ -10,7 +10,7 @@ from ray.experimental.tqdm_ray import tqdm
 from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, seed_everything
 
 
 class BenchmarkConfig(TypedDict):
@@ -30,19 +30,36 @@ def benchmark_config(
     hidden_size: int,
     topk: int,
     dtype: torch.dtype,
-    use_fp8: bool,
+    use_fp8_w8a8: bool,
+    use_int8_w8a16: bool,
     num_iters: int = 100,
 ) -> float:
-    init_dtype = torch.float16 if use_fp8 else dtype
+    init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
-    w1 = torch.randn(num_experts,
-                     shard_intermediate_size,
-                     hidden_size,
-                     dtype=init_dtype)
-    w2 = torch.randn(num_experts,
-                     hidden_size,
-                     shard_intermediate_size // 2,
-                     dtype=init_dtype)
+    if use_int8_w8a16:
+        w1 = torch.randint(-127,
+                           127, (
+                               num_experts,
+                               shard_intermediate_size,
+                               hidden_size,
+                           ),
+                           dtype=torch.int8)
+        w2 = torch.randint(-127,
+                           127, (
+                               num_experts,
+                               hidden_size,
+                               shard_intermediate_size // 2,
+                           ),
+                           dtype=torch.int8)
+    else:
+        w1 = torch.randn(num_experts,
+                         shard_intermediate_size,
+                         hidden_size,
+                         dtype=init_dtype)
+        w2 = torch.randn(num_experts,
+                         hidden_size,
+                         shard_intermediate_size // 2,
+                         dtype=init_dtype)
     gating_output = torch.randn(num_iters,
                                 num_tokens,
                                 num_experts,
@@ -52,7 +69,11 @@ def benchmark_config(
     w2_scale = None
     a1_scale = None
     a2_scale = None
-    if use_fp8:
+    if use_int8_w8a16:
+        w1_scale = torch.randn((num_experts, 2 * shard_intermediate_size),
+                               dtype=torch.float32)
+        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
+    if use_fp8_w8a8:
         w1_scale = torch.randn(num_experts, dtype=torch.float32)
         w2_scale = torch.randn(num_experts, dtype=torch.float32)
         a1_scale = torch.randn(1, dtype=torch.float32)
@@ -76,7 +97,8 @@ def benchmark_config(
             renormalize=True,
             inplace=True,
             override_config=config,
-            use_fp8=use_fp8,
+            use_fp8_w8a8=use_fp8_w8a8,
+            use_int8_w8a16=use_int8_w8a16,
             w1_scale=w1_scale,
             w2_scale=w2_scale,
             a1_scale=a1_scale,
@@ -144,7 +166,7 @@ class BenchmarkWorker:
 
     def __init__(self, seed: int) -> None:
         torch.set_default_device("cuda")
-        torch.cuda.manual_seed_all(seed)
+        seed_everything(seed)
         self.seed = seed
 
     def benchmark(
@@ -155,11 +177,13 @@ class BenchmarkWorker:
         hidden_size: int,
         topk: int,
         dtype: torch.dtype,
-        use_fp8: bool,
+        use_fp8_w8a8: bool,
+        use_int8_w8a16: bool,
     ) -> Tuple[Dict[str, int], float]:
-        torch.cuda.manual_seed_all(self.seed)
-
-        dtype_str = "float8" if use_fp8 else None
+        seed_everything(self.seed)
+        dtype_str = get_config_dtype_str(dtype,
+                                         use_int8_w8a16=use_int8_w8a16,
+                                         use_fp8_w8a8=use_fp8_w8a8)
         # NOTE(woosuk): The current naming convention uses w2.shape[2], which
         # is the intermediate size after silu_and_mul.
         op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
@@ -173,7 +197,8 @@ class BenchmarkWorker:
                                    key=lambda x: abs(x - num_tokens))]
         kernel_time = benchmark_config(config, num_tokens, num_experts,
                                        shard_intermediate_size, hidden_size,
-                                       topk, dtype, use_fp8)
+                                       topk, dtype, use_fp8_w8a8,
+                                       use_int8_w8a16)
         return config, kernel_time
 
     def tune(
@@ -184,9 +209,10 @@ class BenchmarkWorker:
         hidden_size: int,
         topk: int,
         dtype: torch.dtype,
-        use_fp8: bool,
-        search_space: List[BenchmarkConfig],
-    ) -> BenchmarkConfig:
+        use_fp8_w8a8: bool,
+        use_int8_w8a16: bool,
+        search_space: List[Dict[str, int]],
+    ) -> Dict[str, int]:
         best_config = None
         best_time = float("inf")
         for config in tqdm(search_space):
@@ -198,7 +224,8 @@ class BenchmarkWorker:
                                                hidden_size,
                                                topk,
                                                dtype,
-                                               use_fp8,
+                                               use_fp8_w8a8,
+                                               use_int8_w8a16,
                                                num_iters=10)
             except triton.runtime.autotuner.OutOfResources:
                 # Some configurations may be invalid and fail to compile.
@@ -224,20 +251,19 @@ def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     }
 
 
-def save_configs(
-    configs: Dict[int, BenchmarkConfig],
-    num_experts: int,
-    shard_intermediate_size: int,
-    hidden_size: int,
-    topk: int,
-    dtype: torch.dtype,
-    use_fp8: bool,
-) -> None:
-    dtype_str = "float8" if use_fp8 else None
+def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
+                 shard_intermediate_size: int, hidden_size: int, topk: int,
+                 dtype: torch.dtype, use_fp8_w8a8: bool,
+                 use_int8_w8a16: bool) -> None:
+    dtype_str = get_config_dtype_str(dtype,
+                                     use_int8_w8a16=use_int8_w8a16,
+                                     use_fp8_w8a8=use_fp8_w8a8)
+
     # NOTE(woosuk): The current naming convention uses w2.shape[2], which
     # is the intermediate size after silu_and_mul.
     filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
                                     dtype_str)
+
     print(f"Writing best config to {filename}...")
     with open(filename, "w") as f:
         json.dump(configs, f, indent=4)
@@ -253,6 +279,11 @@ def main(args: argparse.Namespace):
         topk = config.ffn_config.moe_top_k
         intermediate_size = config.ffn_config.ffn_hidden_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    elif config.architectures[0] == "JambaForCausalLM":
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Default: Mixtral.
         E = config.num_local_experts
@@ -262,7 +293,8 @@ def main(args: argparse.Namespace):
 
     hidden_size = config.hidden_size
     dtype = config.torch_dtype
-    use_fp8 = args.dtype == "fp8"
+    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
+    use_int8_w8a16 = args.dtype == "int8_w8a16"
 
     if args.batch_size is None:
         batch_sizes = [
@@ -294,21 +326,21 @@ def main(args: argparse.Namespace):
         start = time.time()
         configs = _distribute(
             "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
-                      topk, dtype, use_fp8, search_space)
+                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space)
                      for batch_size in batch_sizes])
         best_configs = {
             M: sort_config(config)
             for M, config in zip(batch_sizes, configs)
         }
         save_configs(best_configs, E, shard_intermediate_size, hidden_size,
-                     topk, dtype, use_fp8)
+                     topk, dtype, use_fp8_w8a8, use_int8_w8a16)
         end = time.time()
         print(f"Tuning took {end - start:.2f} seconds")
     else:
-        outputs = _distribute("benchmark",
-                              [(batch_size, E, shard_intermediate_size,
-                                hidden_size, topk, dtype, use_fp8)
-                               for batch_size in batch_sizes])
+        outputs = _distribute(
+            "benchmark", [(batch_size, E, shard_intermediate_size, hidden_size,
+                           topk, dtype, use_fp8_w8a8, use_int8_w8a16)
+                          for batch_size in batch_sizes])
 
         for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
             print(f"Batch size: {batch_size}, config: {config}")
@@ -323,7 +355,7 @@ if __name__ == "__main__":
     parser.add_argument("--tp-size", "-tp", type=int, default=2)
     parser.add_argument("--dtype",
                         type=str,
-                        choices=["auto", "fp8"],
+                        choices=["auto", "fp8_w8a8", "int8_w8a16"],
                         default="auto")
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)

benchmarks/kernels/benchmark_paged_attention.py

@@ -6,7 +6,7 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        create_kv_caches_with_random)
+                        create_kv_caches_with_random, seed_everything)
 
 NUM_BLOCKS = 1024
 PARTITION_SIZE = 512
@@ -28,10 +28,7 @@ def main(
     device: str = "cuda",
     kv_cache_dtype: Optional[str] = None,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     scale = float(1.0 / (head_size**0.5))
     query = torch.empty(num_seqs,
@@ -175,7 +172,7 @@ if __name__ == '__main__':
     parser.add_argument("--num-kv-heads", type=int, default=8)
     parser.add_argument("--head-size",
                         type=int,
-                        choices=[64, 80, 96, 112, 128, 192, 256],
+                        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("--use-alibi", action="store_true")

benchmarks/kernels/benchmark_quant.py

@@ -0,0 +1,100 @@
+import time
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
+                        seed_everything)
+
+
+@torch.inference_mode()
+def main(num_tokens: int,
+         hidden_size: int,
+         static_scale: bool,
+         quant_dtype: torch.dtype,
+         dtype: torch.dtype,
+         seed: int = 0,
+         do_profile: bool = False,
+         num_warmup_iters: int = 5,
+         num_iters: int = 100) -> None:
+    seed_everything(seed)
+    torch.set_default_device("cuda")
+
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    scale = torch.randn(1, 1, dtype=torch.float32) if static_scale else None
+
+    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
+        torch.cuda.synchronize()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        start_time = time.perf_counter()
+
+        for _ in range(num_iters):
+            if quant_dtype == torch.int8:
+                ops.scaled_int8_quant(x, scale)
+            else:
+                ops.scaled_fp8_quant(x, scale)
+        torch.cuda.synchronize()
+
+        end_time = time.perf_counter()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        return (end_time - start_time) / num_iters
+
+    # Warmup.
+    print("Warming up...")
+    run_benchmark = run_cuda_benchmark
+    run_benchmark(num_iters=num_warmup_iters, profile=False)
+
+    # Benchmark.
+    if do_profile:
+        latency = run_benchmark(num_iters=1, profile=True)
+    else:
+        latency = run_benchmark(num_iters=num_iters, profile=False)
+    print(f"Kernel running time: {latency * 1000000:.3f} us")
+
+
+if __name__ == '__main__':
+
+    def to_torch_dtype(dt):
+        if dt == "int8":
+            return torch.int8
+        if dt == "fp8":
+            return torch.float8_e4m3fn
+        raise ValueError(f"Unsupported dtype: {dt}")
+
+    parser = FlexibleArgumentParser(
+        description="Benchmark the quantization (fp8 or int8) kernel.")
+    parser.add_argument("--num-tokens", type=int, default=4096)
+    parser.add_argument("--hidden-size", type=int, default=8192)
+    parser.add_argument("--static-scale", action="store_true")
+    parser.add_argument("--quant-dtype",
+                        type=str,
+                        choices=["fp8", "int8"],
+                        default="int8")
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["half", "bfloat16", "float"],
+                        default="half")
+
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--profile", action="store_true")
+    parser.add_argument("--num-warmup-iters", type=int, default=5)
+    parser.add_argument("--num-iters",
+                        type=int,
+                        default=100,
+                        help="Number of benchmark iterations. "
+                        "If --profile is set, this number is ignored")
+
+    args = parser.parse_args()
+    print(args)
+
+    main(num_tokens=args.num_tokens,
+         hidden_size=args.hidden_size,
+         static_scale=args.static_scale,
+         quant_dtype=to_torch_dtype(args.quant_dtype),
+         dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+         seed=args.seed,
+         do_profile=args.profile,
+         num_warmup_iters=args.num_warmup_iters,
+         num_iters=args.num_iters)

benchmarks/kernels/benchmark_rope.py

@@ -6,7 +6,7 @@ import torch
 
 from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
                                                          get_rope)
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, seed_everything
 
 
 def benchmark_rope_kernels_multi_lora(
@@ -22,9 +22,7 @@ def benchmark_rope_kernels_multi_lora(
     max_position: int = 8192,
     base: int = 10000,
 ) -> None:
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     if rotary_dim is None:
         rotary_dim = head_size
@@ -94,7 +92,7 @@ if __name__ == '__main__':
     parser.add_argument("--num-heads", type=int, default=8)
     parser.add_argument("--head-size",
                         type=int,
-                        choices=[64, 80, 96, 112, 128, 192, 256],
+                        choices=[64, 80, 96, 112, 120, 128, 192, 256],
                         default=128)
     parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32)
     parser.add_argument("--dtype",

benchmarks/kernels/graph_machete_bench.py

@@ -0,0 +1,62 @@
+import math
+import pickle
+import re
+from collections import defaultdict
+from typing import List
+
+import matplotlib.pyplot as plt
+import pandas as pd
+import seaborn as sns
+from torch.utils.benchmark import Measurement as TMeasurement
+
+from vllm.utils import FlexibleArgumentParser
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description='Benchmark the latency of processing a single batch of '
+        'requests till completion.')
+    parser.add_argument('filename', type=str)
+
+    args = parser.parse_args()
+
+    with open(args.filename, 'rb') as f:
+        data: List[TMeasurement] = pickle.load(f)
+
+    results = defaultdict(lambda: list())
+    for v in data:
+        result = re.search(r"MKN=\(\d+x(\d+x\d+)\)", v.task_spec.sub_label)
+        if result is not None:
+            KN = result.group(1)
+        else:
+            raise Exception("MKN not found")
+        result = re.search(r"MKN=\((\d+)x\d+x\d+\)", v.task_spec.sub_label)
+        if result is not None:
+            M = result.group(1)
+        else:
+            raise Exception("MKN not found")
+
+        kernel = v.task_spec.description
+        results[KN].append({
+            "kernel": kernel,
+            "batch_size": M,
+            "median": v.median
+        })
+
+    rows = int(math.ceil(len(results) / 2))
+    fig, axs = plt.subplots(rows, 2, figsize=(12, 5 * rows))
+    axs = axs.flatten()
+    for axs_idx, (shape, data) in enumerate(results.items()):
+        plt.sca(axs[axs_idx])
+        df = pd.DataFrame(data)
+        sns.lineplot(data=df,
+                     x="batch_size",
+                     y="median",
+                     hue="kernel",
+                     style="kernel",
+                     markers=True,
+                     dashes=False,
+                     palette="Dark2")
+        plt.title(f"Shape: {shape}")
+        plt.ylabel("time (median, s)")
+    plt.tight_layout()
+    plt.savefig("graph_machete_bench.pdf")

benchmarks/kernels/requirements.txt

@@ -0,0 +1 @@
+pandas

benchmarks/kernels/weight_shapes.py

@@ -0,0 +1,43 @@
+# Weight Shapes are in the format
+# ([K, N], TP_SPLIT_DIM)
+# Example:
+#  A shape of ([14336, 4096], 0) indicates the following GEMM shape,
+#   - TP1 : K = 14336, N = 4096
+#   - TP2 : K = 7168, N = 4096
+#  A shape of ([4096, 6144], 1) indicates the following GEMM shape,
+#   - TP1 : K = 4096, N = 6144
+#   - TP4 : K = 4096, N = 1536
+
+# TP1 shapes
+WEIGHT_SHAPES = {
+    "mistralai/Mistral-7B-v0.1": [
+        ([4096, 6144], 1),
+        ([4096, 4096], 0),
+        ([4096, 28672], 1),
+        ([14336, 4096], 0),
+    ],
+    "meta-llama/Llama-2-7b-hf": [
+        ([4096, 12288], 1),
+        ([4096, 4096], 0),
+        ([4096, 22016], 1),
+        ([11008, 4096], 0),
+    ],
+    "meta-llama/Llama-3-8b": [
+        ([4096, 6144], 1),
+        ([4096, 4096], 0),
+        ([4096, 28672], 1),
+        ([14336, 4096], 0),
+    ],
+    "meta-llama/Llama-2-13b-hf": [
+        ([5120, 15360], 1),
+        ([5120, 5120], 0),
+        ([5120, 27648], 1),
+        ([13824, 5120], 0),
+    ],
+    "meta-llama/Llama-2-70b-hf": [
+        ([8192, 10240], 1),
+        ([8192, 8192], 0),
+        ([8192, 57344], 1),
+        ([28672, 8192], 0),
+    ],
+}

benchmarks/launch_tgi_server.sh

@@ -6,7 +6,7 @@ TOKENS=$2
 
 docker run -e HF_TOKEN=$HF_TOKEN --gpus all --shm-size 1g -p $PORT:80 \
            -v $PWD/data:/data \
-           ghcr.io/huggingface/text-generation-inference:1.4.0 \
+           ghcr.io/huggingface/text-generation-inference:2.2.0 \
            --model-id $MODEL \
            --sharded false  \
            --max-input-length 1024 \

cmake/cpu_extension.cmake

@@ -1,4 +1,5 @@
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+set(CMAKE_CXX_STANDARD 17)
 
 #
 # Define environment variables for special configurations
@@ -83,10 +84,7 @@ endif()
 
 message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
 
-
-#
-# Define extension targets
-#
+list(APPEND LIBS dnnl numa)
 
 #
 # _C extension
@@ -95,20 +93,30 @@ set(VLLM_EXT_SRC
     "csrc/cpu/activation.cpp"
     "csrc/cpu/attention.cpp"
     "csrc/cpu/cache.cpp"
+    "csrc/cpu/utils.cpp"
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/pos_encoding.cpp"
     "csrc/cpu/torch_bindings.cpp")
 
+if (AVX512_FOUND AND NOT AVX512_DISABLED)
+    set(VLLM_EXT_SRC
+        "csrc/cpu/quant.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
+#
+# Define extension targets
+#
+
 define_gpu_extension_target(
     _C
     DESTINATION vllm
     LANGUAGE CXX
     SOURCES ${VLLM_EXT_SRC}
+    LIBRARIES ${LIBS}
     COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
     USE_SABI 3
     WITH_SOABI
 )
 
-add_custom_target(default)
 message(STATUS "Enabling C extension.")
-add_dependencies(default _C)

cmake/utils.cmake

@@ -181,7 +181,7 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
     #
     # The torch cmake setup hardcodes the detected architecture flags in
     # `CMAKE_CUDA_FLAGS`.  Since `CMAKE_CUDA_FLAGS` is a "global" variable, it
-    # can't modified on a per-target basis, e.g. for the `punica` extension.
+    # can't modified on a per-target basis.
     # So, all the `-gencode` flags need to be extracted and removed from
     # `CMAKE_CUDA_FLAGS` for processing so they can be passed by another method.
     # Since it's not possible to use `target_compiler_options` for adding target
@@ -350,17 +350,19 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
     ${GPU_INCLUDE_DIRECTORIES})
 
-  target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${torch_python_LIBRARY}
-    ${GPU_LIBRARIES})
+  target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${GPU_LIBRARIES})
 
   # Don't use `TORCH_LIBRARIES` for CUDA since it pulls in a bunch of
   # dependencies that are not necessary and may not be installed.
   if (GPU_LANGUAGE STREQUAL "CUDA")
+    if ("${CUDA_CUDA_LIB}" STREQUAL "")
+      set(CUDA_CUDA_LIB "${CUDA_CUDA_LIBRARY}")
+    endif()
     target_link_libraries(${GPU_MOD_NAME} PRIVATE ${CUDA_CUDA_LIB}
       ${CUDA_LIBRARIES})
   else()
     target_link_libraries(${GPU_MOD_NAME} PRIVATE ${TORCH_LIBRARIES})
   endif()
 
-  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION})
+  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION} COMPONENT ${GPU_MOD_NAME})
 endfunction()

collect_env.py

@@ -65,6 +65,9 @@ DEFAULT_CONDA_PATTERNS = {
     "optree",
     "nccl",
     "transformers",
+    "zmq",
+    "nvidia",
+    "pynvml",
 }
 
 DEFAULT_PIP_PATTERNS = {
@@ -77,6 +80,9 @@ DEFAULT_PIP_PATTERNS = {
     "onnx",
     "nccl",
     "transformers",
+    "zmq",
+    "nvidia",
+    "pynvml",
 }
 
 
@@ -263,8 +269,9 @@ def get_neuron_sdk_version(run_lambda):
 def get_vllm_version():
     try:
         import vllm
-        return vllm.__version__
-    except ImportError:
+        return vllm.__version__ + "@" + vllm.__commit__
+    except Exception:
+        # old version of vllm does not have __commit__
         return 'N/A'
 
 
@@ -278,9 +285,14 @@ def summarize_vllm_build_flags():
 
 
 def get_gpu_topo(run_lambda):
+    output = None
+
     if get_platform() == 'linux':
-        return run_and_read_all(run_lambda, 'nvidia-smi topo -m')
-    return None
+        output = run_and_read_all(run_lambda, 'nvidia-smi topo -m')
+        if output is None:
+            output = run_and_read_all(run_lambda, 'rocm-smi --showtopo')
+
+    return output
 
 
 # example outputs of CPU infos

csrc/attention/attention_kernels.cu

@@ -706,7 +706,7 @@ void paged_attention_v1_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
-  int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
   assert(head_size % thread_group_size == 0);
 
   // NOTE: alibi_slopes is optional.
@@ -751,6 +751,9 @@ void paged_attention_v1_launcher(
     case 112:
       LAUNCH_PAGED_ATTENTION_V1(112);
       break;
+    case 120:
+      LAUNCH_PAGED_ATTENTION_V1(120);
+      break;
     case 128:
       LAUNCH_PAGED_ATTENTION_V1(128);
       break;
@@ -862,7 +865,7 @@ void paged_attention_v2_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
-  int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
   assert(head_size % thread_group_size == 0);
 
   // NOTE: alibi_slopes is optional.
@@ -912,6 +915,9 @@ void paged_attention_v2_launcher(
     case 112:
       LAUNCH_PAGED_ATTENTION_V2(112);
       break;
+    case 120:
+      LAUNCH_PAGED_ATTENTION_V2(120);
+      break;
     case 128:
       LAUNCH_PAGED_ATTENTION_V2(128);
       break;

csrc/attention/attention_utils.cuh

@@ -34,7 +34,7 @@ inline __device__ float qk_dot_(const Vec (&q)[N], const Vec (&k)[N]) {
   A_vec qk_vec = mul<A_vec, Vec, Vec>(q[0], k[0]);
 #pragma unroll
   for (int ii = 1; ii < N; ++ii) {
-    qk_vec = fma(q[ii], k[ii], qk_vec);
+    qk_vec = vllm::fma(q[ii], k[ii], qk_vec);
   }
 
   // Finalize the reduction across lanes.

csrc/attention/dtype_bfloat16.cuh

@@ -94,6 +94,7 @@ inline __device__ float2 bf1622float2(const __nv_bfloat162 val) {
 #else
   return __bfloat1622float2(val);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 inline __device__ __nv_bfloat162 bf162bf162(const __nv_bfloat16 val) {
@@ -102,6 +103,7 @@ inline __device__ __nv_bfloat162 bf162bf162(const __nv_bfloat16 val) {
 #else
   return __bfloat162bfloat162(val);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 // Vector addition.
@@ -115,6 +117,7 @@ inline __device__ __nv_bfloat16 add(__nv_bfloat16 a, __nv_bfloat16 b) {
   return __hadd(a, b);
   #endif
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 inline __device__ __nv_bfloat162 add(__nv_bfloat162 a, __nv_bfloat162 b) {
@@ -123,6 +126,7 @@ inline __device__ __nv_bfloat162 add(__nv_bfloat162 a, __nv_bfloat162 b) {
 #else
   return __hadd2(a, b);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 inline __device__ bf16_4_t add(bf16_4_t a, bf16_4_t b) {
@@ -170,6 +174,7 @@ inline __device__ __nv_bfloat16 mul(__nv_bfloat16 a, __nv_bfloat16 b) {
 #else
   return __hmul(a, b);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 template <>
@@ -179,6 +184,7 @@ inline __device__ __nv_bfloat162 mul(__nv_bfloat162 a, __nv_bfloat162 b) {
 #else
   return __hmul2(a, b);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 template <>
@@ -289,6 +295,7 @@ inline __device__ __nv_bfloat162 fma(__nv_bfloat162 a, __nv_bfloat162 b,
 #else
   return __hfma2(a, b, c);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 inline __device__ __nv_bfloat162 fma(__nv_bfloat16 a, __nv_bfloat162 b,
@@ -298,6 +305,7 @@ inline __device__ __nv_bfloat162 fma(__nv_bfloat16 a, __nv_bfloat162 b,
 #else
   return __hfma2(bf162bf162(a), b, c);
 #endif
+  __builtin_unreachable();  // Suppress missing return statement warning
 }
 
 inline __device__ bf16_4_t fma(bf16_4_t a, bf16_4_t b, bf16_4_t c) {

csrc/cache.h

@@ -25,7 +25,8 @@ void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
                              torch::Tensor& key_cache,
                              torch::Tensor& value_cache,
                              torch::Tensor& slot_mapping,
-                             const std::string& kv_cache_dtype);
+                             const std::string& kv_cache_dtype,
+                             const double k_scale, const double v_scale);
 
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,

csrc/cache_kernels.cu

@@ -203,17 +203,18 @@ __global__ void reshape_and_cache_kernel(
   }
 }
 
-template <typename scalar_t>
+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]
-    scalar_t* __restrict__ k_cache,      // [num_blocks, block_size, num_heads,
+    cache_t* __restrict__ key_cache,     // [num_blocks, block_size, num_heads,
                                          // head_size]
-    scalar_t* __restrict__ v_cache,      // [num_blocks, block_size, num_heads,
+    cache_t* __restrict__ value_cache,   // [num_blocks, block_size, num_heads,
                                          // head_size]
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int block_stride, const int key_stride, const int value_stride,
-    const int num_heads, const int head_size, const int block_size) {
+    const int num_heads, const int head_size, const int block_size,
+    const float k_scale, const float v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   // NOTE: slot_idx can be -1 if the token is padded
@@ -228,11 +229,20 @@ __global__ void reshape_and_cache_flash_kernel(
     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_value_idx = block_idx * block_stride +
-                                  block_offset * num_heads * head_size +
-                                  head_idx * head_size + head_offset;
-    k_cache[tgt_value_idx] = key[src_key_idx];
-    v_cache[tgt_value_idx] = value[src_value_idx];
+    const int64_t tgt_key_value_idx = block_idx * block_stride +
+                                      block_offset * num_heads * head_size +
+                                      head_idx * head_size + 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);
+    }
   }
 }
 }  // namespace vllm
@@ -278,40 +288,45 @@ void reshape_and_cache(
                              CALL_RESHAPE_AND_CACHE)
 }
 
+// KV_T is the stored data type of kv-cache.
+// CACHE_T is the data type of key and value tensors.
+// KV_DTYPE is the real data type of kv-cache.
+#define CALL_RESHAPE_AND_CACHE_FLASH(KV_T, CACHE_T, KV_DTYPE)         \
+  vllm::reshape_and_cache_flash_kernel<KV_T, CACHE_T, KV_DTYPE>       \
+      <<<grid, block, 0, stream>>>(                                   \
+          reinterpret_cast<KV_T*>(key.data_ptr()),                    \
+          reinterpret_cast<KV_T*>(value.data_ptr()),                  \
+          reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
+          reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
+          slot_mapping.data_ptr<int64_t>(), block_stride, key_stride, \
+          value_stride, num_heads, head_size, block_size, k_scale, v_scale);
+
 void reshape_and_cache_flash(
-    torch::Tensor& key,      // [num_tokens, num_heads, head_size]
-    torch::Tensor& value,    // [num_tokens, num_heads, head_size]
-    torch::Tensor& k_cache,  // [num_blocks, block_size, num_heads, head_size]
-    torch::Tensor& v_cache,  // [num_blocks, block_size, num_heads, head_size]
+    torch::Tensor& key,        // [num_tokens, num_heads, head_size]
+    torch::Tensor& value,      // [num_tokens, num_heads, head_size]
+    torch::Tensor& key_cache,  // [num_blocks, block_size, num_heads, head_size]
+    torch::Tensor&
+        value_cache,  // [num_blocks, block_size, num_heads, head_size]
     torch::Tensor& slot_mapping,  // [num_tokens]
-    const std::string& kv_cache_dtype) {
-  // FIXME: only support auto datatype, does not support fp8
-  if (kv_cache_dtype != "auto") {
-    TORCH_CHECK(false, "Unsupported data type of kv cache: ", kv_cache_dtype);
-  }
+    const std::string& kv_cache_dtype, const double k_scale,
+    const double v_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
-  int block_size = k_cache.size(1);
+  int block_size = key_cache.size(1);
 
   int key_stride = key.stride(0);
   int value_stride = value.stride(0);
-  int block_stride = k_cache.stride(0);
-  TORCH_CHECK(k_cache.stride(0) == v_cache.stride(0));
+  int block_stride = key_cache.stride(0);
+  TORCH_CHECK(key_cache.stride(0) == value_cache.stride(0));
 
   dim3 grid(num_tokens);
   dim3 block(std::min(num_heads * head_size, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  VLLM_DISPATCH_FLOATING_TYPES(
-      key.scalar_type(), "reshape_and_cache_flash", [&] {
-        vllm::reshape_and_cache_flash_kernel<scalar_t>
-            <<<grid, block, 0, stream>>>(
-                key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
-                k_cache.data_ptr<scalar_t>(), v_cache.data_ptr<scalar_t>(),
-                slot_mapping.data_ptr<int64_t>(), block_stride, key_stride,
-                value_stride, num_heads, head_size, block_size);
-      });
+
+  DISPATCH_BY_KV_CACHE_DTYPE(key.dtype(), kv_cache_dtype,
+                             CALL_RESHAPE_AND_CACHE_FLASH);
 }
 
 namespace vllm {

csrc/core/scalar_type.hpp

@@ -0,0 +1,548 @@
+#pragma once
+
+#include <torch/custom_class.h>
+
+namespace vllm {
+
+//
+//  ScalarType can represent a wide range of floating point and integer types,
+//  in particular it can be used to represent sub-byte data types (something
+//  that torch.dtype currently does not support).
+//
+//  ScalarTypeTorch is a subclass of ScalarType that is compatible with
+//  TORCH_LIBRARY, making it accessible from Python as well meaning this class
+//  can be used as a argument for custom operators, helping to simplify these
+//  interfaces.
+//
+//  The type definitions on the Python side can be found in: vllm/_core_ext.pyi
+//  these type definitions should be kept up to date with any Python API changes
+//  here.
+//
+class ScalarType {
+ public:
+  enum NanRepr : uint8_t {
+    NAN_NONE = 0,                // nans are not supported
+    NAN_IEEE_754 = 1,            // nans are: exp all 1s, mantissa not all 0s
+    NAN_EXTD_RANGE_MAX_MIN = 2,  // nans are: exp all 1s, mantissa all 1s
+
+    NAN_REPR_ID_MAX
+  };
+
+  constexpr ScalarType(uint8_t exponent, uint8_t mantissa, bool signed_,
+                       int32_t bias, bool finite_values_only = false,
+                       NanRepr nan_repr = NAN_IEEE_754)
+      : exponent(exponent),
+        mantissa(mantissa),
+        signed_(signed_),
+        bias(bias),
+        finite_values_only(finite_values_only),
+        nan_repr(nan_repr){};
+
+  static constexpr ScalarType int_(uint8_t size_bits, int32_t bias = 0) {
+    return ScalarType(0, size_bits - 1, true, bias);
+  }
+
+  static constexpr ScalarType uint(uint8_t size_bits, int32_t bias = 0) {
+    return ScalarType(0, size_bits, false, bias);
+  }
+
+  // IEEE 754 compliant floating point type
+  static constexpr ScalarType float_IEEE754(uint8_t exponent,
+                                            uint8_t mantissa) {
+    TORCH_CHECK(mantissa > 0 && exponent > 0);
+    return ScalarType(exponent, mantissa, true, 0, false, NAN_IEEE_754);
+  }
+
+  // IEEE 754 non-compliant floating point type
+  static constexpr ScalarType float_(uint8_t exponent, uint8_t mantissa,
+                                     bool finite_values_only,
+                                     NanRepr nan_repr) {
+    TORCH_CHECK(nan_repr < NAN_REPR_ID_MAX, "Invalid NanRepr");
+    TORCH_CHECK(mantissa > 0 && exponent > 0);
+    TORCH_CHECK(nan_repr != NAN_IEEE_754,
+                "use `float_IEEE754` constructor for floating point types that "
+                "follow IEEE 754 conventions");
+    return ScalarType(exponent, mantissa, true, 0, finite_values_only,
+                      nan_repr);
+  }
+
+  uint8_t const exponent;  // size of the exponent field (0 for integer types)
+  uint8_t const mantissa;  // size of the mantissa field (size of the integer
+                           // excluding the sign bit for integer types)
+  bool const signed_;  // flag if the type supports negative numbers (i.e. has a
+                       // sign bit)
+  int32_t const bias;  // stored values equal value + bias,
+                       // used for quantized type
+
+  // Extra Floating point info
+  bool const finite_values_only;  // i.e. no +/-inf if true
+  NanRepr const nan_repr;         // how NaNs are represented
+                                  // (not applicable for integer types)
+
+  using Id = int64_t;
+
+ private:
+  // Field size in id
+  template <typename T_>
+  static constexpr size_t member_id_field_width() {
+    using T = std::decay_t<T_>;
+    return std::is_same_v<T, bool> ? 1 : sizeof(T) * 8;
+  }
+
+  template <typename Fn, typename Init, typename Member, typename... Rest>
+  static constexpr auto reduce_members_helper(Fn f, Init val, Member member,
+                                              Rest... rest) {
+    auto new_val = f(val, member);
+    if constexpr (sizeof...(rest) > 0) {
+      return reduce_members_helper(f, new_val, rest...);
+    } else {
+      return new_val;
+    };
+  }
+
+  template <typename Fn, typename Init>
+  constexpr auto reduce_members(Fn f, Init init) const {
+    // Should be in constructor order for `from_id`
+    return reduce_members_helper(f, init, exponent, mantissa, signed_, bias,
+                                 finite_values_only, nan_repr);
+  };
+
+  template <typename Fn, typename Init>
+  static constexpr auto reduce_member_types(Fn f, Init init) {
+    constexpr auto dummy_type = ScalarType(0, 0, false, 0, false, NAN_NONE);
+    return dummy_type.reduce_members(f, init);
+  };
+
+  static constexpr auto id_size_bits() {
+    return reduce_member_types(
+        [](int acc, auto member) -> int {
+          return acc + member_id_field_width<decltype(member)>();
+        },
+        0);
+  }
+
+ public:
+  // unique id for this scalar type that can be computed at compile time for
+  //  c++17 template specialization this is not needed once we migrate to
+  //  c++20 and can pass literal classes as template parameters
+  constexpr Id id() const {
+    static_assert(id_size_bits() <= sizeof(Id) * 8,
+                  "ScalarType id is too large to be stored");
+
+    auto or_and_advance = [](std::pair<Id, uint32_t> result,
+                             auto member) -> std::pair<Id, uint32_t> {
+      auto [id, bit_offset] = result;
+      auto constexpr bits = member_id_field_width<decltype(member)>();
+      return {id | (int64_t(member) & ((uint64_t(1) << bits) - 1))
+                       << bit_offset,
+              bit_offset + bits};
+    };
+    return reduce_members(or_and_advance, std::pair<Id, uint32_t>{}).first;
+  }
+
+  // create a ScalarType from an id, for c++17 template specialization,
+  //  this is not needed once we migrate to c++20 and can pass literal
+  //  classes as template parameters
+  static constexpr ScalarType from_id(Id id) {
+    auto extract_and_advance = [id](auto result, auto member) {
+      using T = decltype(member);
+      auto [tuple, bit_offset] = result;
+      auto constexpr bits = member_id_field_width<T>();
+      auto extracted_val = static_cast<T>((int64_t(id) >> bit_offset) &
+                                          ((uint64_t(1) << bits) - 1));
+      auto new_tuple = std::tuple_cat(tuple, std::make_tuple(extracted_val));
+      return std::pair<decltype(new_tuple), int>{new_tuple, bit_offset + bits};
+    };
+
+    auto [tuple_args, _] = reduce_member_types(extract_and_advance,
+                                               std::pair<std::tuple<>, int>{});
+    return std::apply([](auto... args) { return ScalarType(args...); },
+                      tuple_args);
+  }
+
+  constexpr int64_t size_bits() const {
+    return mantissa + exponent + is_signed();
+  }
+  constexpr bool is_signed() const { return signed_; }
+  constexpr bool is_integer() const { return exponent == 0; }
+  constexpr bool is_floating_point() const { return exponent > 0; }
+  constexpr bool is_ieee_754() const {
+    return is_floating_point() && finite_values_only == false &&
+           nan_repr == NAN_IEEE_754;
+  }
+  constexpr bool has_nans() const {
+    return is_floating_point() && nan_repr != NAN_NONE;
+  }
+  constexpr bool has_infs() const {
+    return is_floating_point() && finite_values_only == false;
+  }
+  constexpr bool has_bias() const { return bias != 0; }
+
+ private:
+  double _floating_point_max() const {
+    TORCH_CHECK(mantissa <= 52 && exponent <= 11,
+                "Cannot represent max/min as a double for type ", str());
+
+    uint64_t max_mantissa = (uint64_t(1) << mantissa) - 1;
+    if (nan_repr == NAN_EXTD_RANGE_MAX_MIN) {
+      max_mantissa -= 1;
+    }
+
+    uint64_t max_exponent = (uint64_t(1) << exponent) - 2;
+    if (nan_repr == NAN_EXTD_RANGE_MAX_MIN || nan_repr == NAN_NONE) {
+      TORCH_CHECK(exponent < 11,
+                  "Cannot represent max/min as a double for type ", str());
+      max_exponent += 1;
+    }
+
+    // adjust the exponent to match that of a double
+    //  for now we assume the exponent bias is the standard 2^(e-1) -1, (where e
+    //  is the exponent bits), there is some precedent for non-standard biases,
+    //  example `float8_e4m3b11fnuz` here: https://github.com/jax-ml/ml_dtypes
+    //  but to avoid premature over complication we are just assuming the
+    //  standard exponent bias until there is a need to support non-standard
+    //  biases
+    uint64_t exponent_bias = (uint64_t(1) << (exponent - 1)) - 1;
+    uint64_t exponent_bias_double = (uint64_t(1) << 10) - 1;  // double e = 11
+
+    uint64_t max_exponent_double =
+        max_exponent - exponent_bias + exponent_bias_double;
+
+    // shift the mantissa into the position for a double and
+    // the exponent
+    uint64_t double_raw =
+        (max_mantissa << (52 - mantissa)) | (max_exponent_double << 52);
+
+    return *reinterpret_cast<double*>(&double_raw);
+  }
+
+  constexpr std::variant<int64_t, double> _raw_max() const {
+    if (is_floating_point()) {
+      return {_floating_point_max()};
+    } else {
+      TORCH_CHECK(size_bits() < 64 || size_bits() == 64 && is_signed(),
+                  "Cannot represent max as a int64_t");
+      return {(int64_t(1) << mantissa) - 1};
+    }
+  }
+
+  constexpr std::variant<int64_t, double> _raw_min() const {
+    if (is_floating_point()) {
+      TORCH_CHECK(is_signed(),
+                  "We currently assume all floating point types are signed");
+      constexpr uint64_t sign_bit_double = (uint64_t(1) << 63);
+
+      double max = _floating_point_max();
+      uint64_t max_raw = *reinterpret_cast<uint64_t*>(&max);
+      uint64_t min_raw = max_raw | sign_bit_double;
+      return {*reinterpret_cast<double*>(&min_raw)};
+    } else {
+      TORCH_CHECK(!is_signed() || size_bits() <= 64,
+                  "Cannot represent min as a int64_t");
+      if (is_signed()) {
+        // set the top bit to 1 (i.e. INT64_MIN) and the rest to 0
+        // then perform an arithmetic shift right to set all the bits above
+        // (size_bits() - 1) to 1
+        return {INT64_MIN >> (64 - size_bits())};
+      } else {
+        return {int64_t(0)};
+      }
+    }
+  }
+
+ public:
+  // Max representable value for this scalar type.
+  // (accounting for bias if there is one)
+  constexpr std::variant<int64_t, double> max() const {
+    return std::visit(
+        [this](auto x) -> std::variant<int64_t, double> { return {x - bias}; },
+        _raw_max());
+  }
+
+  // Min representable value for this scalar type.
+  // (accounting for bias if there is one)
+  constexpr std::variant<int64_t, double> min() const {
+    return std::visit(
+        [this](auto x) -> std::variant<int64_t, double> { return {x - bias}; },
+        _raw_min());
+  }
+
+  std::string str() const {
+    /* naming generally follows: https://github.com/jax-ml/ml_dtypes
+     * for floating point types (leading f) the scheme is:
+     *  `float<size_bits>_e<exponent_bits>m<mantissa_bits>[flags]`
+     *  flags:
+     *  - no-flags: means it follows IEEE 754 conventions
+     *  - f: means finite values only (no infinities)
+     *  - n: means nans are supported (non-standard encoding)
+     * for integer types the scheme is:
+     *  `[u]int<size_bits>[b<bias>]`
+     *  - if bias is not present it means its zero
+     */
+    if (is_floating_point()) {
+      auto ret = "float" + std::to_string(size_bits()) + "_e" +
+                 std::to_string(exponent) + "m" + std::to_string(mantissa);
+      if (!is_ieee_754()) {
+        if (finite_values_only) {
+          ret += "f";
+        }
+        if (nan_repr != NAN_NONE) {
+          ret += "n";
+        }
+      }
+      return ret;
+    } else {
+      auto ret = ((is_signed()) ? "int" : "uint") + std::to_string(size_bits());
+      if (has_bias()) {
+        ret += "b" + std::to_string(bias);
+      }
+      return ret;
+    }
+  }
+
+  constexpr bool operator==(ScalarType const& other) const {
+    return mantissa == other.mantissa && exponent == other.exponent &&
+           bias == other.bias && signed_ == other.signed_ &&
+           finite_values_only == other.finite_values_only &&
+           nan_repr == other.nan_repr;
+  }
+};
+
+// Create a TORCH_LIBRARY compatible version of ScalarType (i.e. inherit from
+//  torch::CustomClassHolder), we use multiple inheritance here since we cannot
+//  have ScalarType inherit from torch::CustomClassHolder and have a constexpr
+//  constructor at the same time (torch::CustomClassHolder does not have a
+//  constexpr destructor)
+// See also:
+// https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA
+class ScalarTypeTorch : public torch::CustomClassHolder, public ScalarType {
+ public:
+  ScalarTypeTorch(int64_t exponent, int64_t mantissa, int64_t bias,
+                  bool _signed)
+      : ScalarType(exponent, mantissa, bias, _signed){};
+
+  ScalarTypeTorch(ScalarType type) : ScalarType(type){};
+
+  using Base = ScalarType;
+  using Self = ScalarTypeTorch;
+  using SelfPtr = c10::intrusive_ptr<Self>;
+
+  static void check_size_bits(int64_t size_bits, bool signed_) {
+    TORCH_CHECK(
+        size_bits <=
+            std::numeric_limits<decltype(std::declval<Self>().mantissa)>::max(),
+        "size_bits bit width is too large to be represented");
+  }
+
+  static void check_bias(int64_t bias) {
+    using Bias = decltype(std::declval<Self>().bias);
+    TORCH_CHECK(bias <= std::numeric_limits<Bias>::max() &&
+                    bias >= std::numeric_limits<Bias>::min(),
+                "bias too large or small to be represented");
+  }
+
+  static void check_exponent(int64_t exponent) {
+    TORCH_CHECK(
+        exponent <=
+            std::numeric_limits<decltype(std::declval<Self>().exponent)>::max(),
+        "exponent bit width is too large to be represented");
+  }
+
+  static void check_mantissa(int64_t mantissa) {
+    TORCH_CHECK(
+        mantissa <=
+            std::numeric_limits<decltype(std::declval<Self>().mantissa)>::max(),
+        "mantissa bit width is too large to be represented");
+  }
+
+  static SelfPtr int_(int64_t size_bits, c10::optional<int64_t> bias) {
+    check_size_bits(size_bits, true);
+    check_bias(bias.value_or(0));
+    return c10::make_intrusive<Self>(
+        ScalarType::int_(size_bits, bias.value_or(0)));
+  }
+
+  static SelfPtr uint(int64_t size_bits, c10::optional<int64_t> bias) {
+    check_size_bits(size_bits, true);
+    check_bias(bias.value_or(0));
+    return c10::make_intrusive<Self>(
+        ScalarType::uint(size_bits, bias.value_or(0)));
+  }
+
+  static SelfPtr float_IEEE754(int64_t exponent, int64_t mantissa) {
+    check_mantissa(mantissa);
+    check_exponent(exponent);
+    return c10::make_intrusive<Self>(
+        ScalarType::float_IEEE754(exponent, mantissa));
+  }
+
+  static SelfPtr float_(int64_t exponent, int64_t mantissa,
+                        bool finite_values_only, int64_t nan_repr) {
+    check_mantissa(mantissa);
+    check_exponent(exponent);
+    return c10::make_intrusive<Self>(ScalarType::float_(
+        exponent, mantissa, finite_values_only, NanRepr(nan_repr)));
+  }
+
+  // This needs to be implemented and throw a TypeError in order for
+  // PyTorch's opcheck to work on ops that use ScalarTypes.
+  int64_t len() const {
+    throw c10::TypeError({__func__, __FILE__, static_cast<uint32_t>(__LINE__)},
+                         "__len__ not implemented");
+    return 0;
+  }
+
+  // Serialize a ScalarType into a tuple of pairs.  Where each pair
+  // is a (fieldname, value).
+  // For simplicity, we are just going to convert to a ScalarTypeId.
+  std::tuple<std::tuple<std::string, int64_t>> obj_flatten() const {
+    return {{"ScalarType", id()}};
+  }
+
+  // Deserialize a scalar type that has been serialized by obj_flatten,
+  // ostensibly from a tuple of (member name, value) pairs, but in reality
+  // just a ScalarTypeId.
+  static SelfPtr obj_unflatten(
+      std::tuple<std::tuple<std::string, int64_t>> const& flat_type) {
+    return c10::make_intrusive<Self>(
+        from_id(std::get<1>(std::get<0>(flat_type))));
+  }
+
+  template <typename T>
+  static void bind_readonly_property(torch::class_<Self>& cls,
+                                     std::string const& name, T Base::*field) {
+    auto getter_func_helper = [field = std::move(field)](SelfPtr const& self) {
+      if constexpr (std::is_member_function_pointer_v<decltype(field)>) {
+        return (self.get()->*field)();
+      } else {
+        return self.get()->*field;
+      }
+    };
+
+    auto getter_func = [field = std::move(field),
+                        getter_func_helper = std::move(getter_func_helper)](
+                           SelfPtr const& self) {
+      auto val = getter_func_helper(self);
+      // upconvert uint8_t, int32_t etc. to int64_t for python
+      if constexpr (std::is_integral_v<T>) {
+        return static_cast<int64_t>(val);
+      } else {
+        return val;
+      }
+    };
+
+    cls.def_property(name, getter_func);
+  }
+
+  template <typename MemberFunc, typename Cls>
+  static void bind_function(torch::class_<Self>& cls, const std::string& name,
+                            MemberFunc Cls::*member) {
+    cls.def(name, [member = std::move(member)](SelfPtr const& self) {
+      return (self.get()->*member)();
+    });
+  }
+
+  template <typename Func>
+  static void bind_function(torch::class_<Self>& cls, const std::string& name,
+                            Func func) {
+    cls.def(name, func);
+  }
+
+  template <typename Func>
+  static void bind_static_function(torch::class_<Self>& cls,
+                                   const std::string& name, Func func) {
+    cls.def_static(name, func);
+  }
+
+  static void bind_class(torch::Library& lib) {
+    auto cls = lib.class_<ScalarTypeTorch>("ScalarType")
+                   .def(torch::init<int64_t, int64_t, int64_t, bool>());
+
+    // Bind Properties
+    bind_readonly_property(cls, "mantissa", &Base::mantissa);
+    bind_readonly_property(cls, "exponent", &Base::exponent);
+    bind_readonly_property(cls, "bias", &Base::bias);
+    bind_readonly_property(cls, "signed", &Base::is_signed);
+    bind_readonly_property(cls, "size_bits", &Base::size_bits);
+
+    // Bind member functions
+    bind_function(cls, "is_signed", &Base::is_signed);
+    bind_function(cls, "is_integer", &Base::is_integer);
+    bind_function(cls, "is_floating_point", &Base::is_floating_point);
+    bind_function(cls, "is_ieee_754", &Base::is_ieee_754);
+    bind_function(cls, "has_nans", &Base::has_nans);
+    bind_function(cls, "has_infs", &Base::has_infs);
+    bind_function(cls, "has_bias", &Base::has_bias);
+
+    bind_function(cls, "max", [](SelfPtr const& self) {
+      return std::visit([](auto arg) { return c10::IValue(arg); },
+                        self.get()->max());
+    });
+    bind_function(cls, "min", [](SelfPtr const& self) {
+      return std::visit([](auto arg) { return c10::IValue(arg); },
+                        self.get()->min());
+    });
+
+    bind_function(cls, "__len__", &ScalarTypeTorch::len);
+    bind_function(cls, "__str__", &Base::str);
+    bind_function(cls, "__eq__", [](SelfPtr const& self, SelfPtr const& other) {
+      return *self == *other;
+    });
+    bind_function(cls, "__repr__", [](SelfPtr const& self) {
+      return "ScalarType." + self.get()->str();
+    });
+
+    bind_function(cls, "__obj_flatten__", &ScalarTypeTorch::obj_flatten);
+    bind_static_function(cls, "__obj_unflatten__",
+                         &ScalarTypeTorch::obj_unflatten);
+
+    // Bind static functions (convenience constructors)
+    bind_static_function(cls, "int_", &ScalarTypeTorch::int_);
+    bind_static_function(cls, "uint", &ScalarTypeTorch::uint);
+    bind_static_function(cls, "float_IEEE754", &ScalarTypeTorch::float_IEEE754);
+    bind_static_function(cls, "float_", &ScalarTypeTorch::float_);
+  }
+};
+
+using ScalarTypeId = int64_t;
+using ScalarTypeTorchPtr = c10::intrusive_ptr<ScalarTypeTorch>;
+
+// "rust style" names generally following:
+//   https://github.com/pytorch/pytorch/blob/6d9f74f0af54751311f0dd71f7e5c01a93260ab3/torch/csrc/api/include/torch/types.h#L60-L70
+static inline constexpr auto kS4 = ScalarType::int_(4);
+static inline constexpr auto kU4 = ScalarType::uint(4);
+static inline constexpr auto kU4B8 = ScalarType::uint(4, 8);
+static inline constexpr auto kS8 = ScalarType::int_(8);
+static inline constexpr auto kU8 = ScalarType::uint(8);
+static inline constexpr auto kU8B128 = ScalarType::uint(8, 128);
+
+static inline constexpr auto kFE3M2f =
+    ScalarType::float_(3, 2, true, ScalarType::NAN_NONE);
+static inline constexpr auto kFE4M3fn =
+    ScalarType::float_(4, 3, true, ScalarType::NAN_EXTD_RANGE_MAX_MIN);
+static inline constexpr auto kFE5M2 = ScalarType::float_IEEE754(5, 2);
+static inline constexpr auto kFE8M7 = ScalarType::float_IEEE754(8, 7);
+static inline constexpr auto kFE5M10 = ScalarType::float_IEEE754(5, 10);
+
+// Fixed width style names, generally following:
+//  https://github.com/pytorch/pytorch/blob/6d9f74f0af54751311f0dd71f7e5c01a93260ab3/torch/csrc/api/include/torch/types.h#L47-L57
+static inline constexpr auto kInt4 = kS4;
+static inline constexpr auto kUint4 = kU4;
+static inline constexpr auto kUint4b8 = kU4B8;
+static inline constexpr auto kInt8 = kS8;
+static inline constexpr auto kUint8 = kU8;
+static inline constexpr auto kUint8b128 = kU8B128;
+
+static inline constexpr auto kFloat6_e3m2f = kFE3M2f;
+static inline constexpr auto kFloat8_e4m3fn = kFE4M3fn;
+static inline constexpr auto kFloat8_e5m2 = kFE5M2;
+static inline constexpr auto kFloat16_e8m7 = kFE8M7;
+static inline constexpr auto kFloat16_e5m10 = kFE5M10;
+
+// colloquial names
+static inline constexpr auto kHalf = kFE5M10;
+static inline constexpr auto kFloat16 = kHalf;
+static inline constexpr auto kBFloat16 = kFE8M7;
+
+static inline constexpr auto kFloat16Id = kFloat16.id();
+};  // namespace vllm

csrc/core/torch_bindings.cpp

@@ -0,0 +1,16 @@
+#include <torch/library.h>
+
+#include "scalar_type.hpp"
+#include "registration.h"
+
+// Note the CORE exstension will be built for (almost) all hardware targets so
+// new additions must account for this. (currently not built for TPU and Neuron)
+
+TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, lib) {
+  // ScalarType, a custom class for representing data types that supports
+  // quantized types, declared here so it can be used when creating interfaces
+  // for custom ops.
+  vllm::ScalarTypeTorch::bind_class(lib);
+}
+
+REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

csrc/cpu/cpu_types_x86.hpp

@@ -24,8 +24,8 @@ namespace vec_op {
 #define CPU_KERNEL_GUARD_OUT(NAME)
 #else
 #define CPU_KERNEL_GUARD_IN(NAME)                                              \
-  std::cout << #NAME << " invoked." << std::endl;
-#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl;
+  RECORD_FUNCTION(#NAME, c10::ArrayRef<c10::IValue>({}));
+#define CPU_KERNEL_GUARD_OUT(NAME)
 #endif
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
@@ -106,6 +106,12 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
   explicit BF16Vec16(const FP32Vec16 &);
 
   void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; }
+
+  void save(void* ptr, const int elem_num) const {
+    constexpr uint32_t M = 0xFFFFFFFF;
+    __mmask16 mask = _cvtu32_mask16(M >> (32 - elem_num));
+    _mm256_mask_storeu_epi16(ptr, mask, reg);
+  }
 };
 
 #ifdef __AVX512F__
@@ -313,8 +319,28 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return FP32Vec16(_mm512_div_ps(reg, b.reg));
   }
 
+  FP32Vec16 clamp(const FP32Vec16& min, const FP32Vec16& max) const {
+    return FP32Vec16(_mm512_min_ps(max.reg, _mm512_max_ps(min.reg, reg)));
+  }
+
+  FP32Vec16 max(const FP32Vec16& b) const {
+    return FP32Vec16(_mm512_max_ps(reg, b.reg));
+  }
+
+  FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
+    constexpr uint32_t M = 0xFFFFFFFF;
+    __mmask16 mask = _cvtu32_mask16(M >> (32 - elem_num));
+    return FP32Vec16(_mm512_mask_max_ps(reg, mask, reg, b.reg));
+  }
+
+  FP32Vec16 abs() const {
+    return FP32Vec16(_mm512_abs_ps(reg));
+  } 
+
   float reduce_sum() const { return _mm512_reduce_add_ps(reg); }
 
+  float reduce_max() const { return _mm512_reduce_max_ps(reg); }
+
   template <int group_size> float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
     constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size));
@@ -323,6 +349,12 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   }
 
   void save(float *ptr) const { _mm512_storeu_ps(ptr, reg); }
+
+  void save(float* ptr, const int elem_num) const {
+    constexpr uint32_t M = 0xFFFFFFFF;
+    __mmask16 mask = _cvtu32_mask16(M >> (32 - elem_num));
+    _mm512_mask_storeu_ps(ptr, mask, reg);
+  }
 };
 #else
 struct FP32Vec16 : public Vec<FP32Vec16> {
@@ -433,6 +465,32 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 };
 #endif
 
+#ifdef __AVX512F__
+struct INT8Vec16: public Vec<INT8Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    __m128i reg;
+    int8_t values[VEC_ELEM_NUM];
+  };
+
+  __m128i reg;
+  
+  explicit INT8Vec16(const FP32Vec16& vec) : reg(
+    _mm512_cvtepi32_epi8(_mm512_cvt_roundps_epi32(vec.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
+  ) {}
+
+  void save(int8_t* ptr) const {
+    _mm_storeu_epi8(ptr, reg);
+  }
+
+  void save(int8_t* ptr, const int elem_num) const {
+    constexpr uint32_t M = 0xFFFFFFFF;
+    __mmask16 mask = _cvtu32_mask16(M >> (32 - elem_num));
+    _mm_mask_storeu_epi8(ptr, mask, reg);
+  }
+};
+#endif
+
 template <typename T> struct VecType { using vec_type = void; };
 
 template <typename T> using vec_t = typename VecType<T>::vec_type;

csrc/cpu/dnnl_helper.hpp

@@ -0,0 +1,168 @@
+#ifndef DNNL_HELPER_HPP
+#define DNNL_HELPER_HPP
+
+#include <c10/util/BFloat16.h>
+
+#include "oneapi/dnnl/dnnl.hpp"
+
+namespace {
+template <typename T>
+struct DNNLType {
+  static constexpr dnnl::memory::data_type type =
+      dnnl::memory::data_type::undef;
+};
+
+template <>
+struct DNNLType<int8_t> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s8;
+};
+
+template <>
+struct DNNLType<int32_t> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s32;
+};
+
+template <>
+struct DNNLType<float> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f32;
+};
+
+template <>
+struct DNNLType<c10::BFloat16> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::bf16;
+};
+
+template <typename T>
+constexpr inline dnnl::memory::data_type get_dnnl_type() {
+  return DNNLType<std::decay_t<T>>::type;
+}
+};  // namespace
+
+template <bool InputNoScale>
+class DNNLPrimitiveHelper {
+ public:
+  // I8 input GEMM kernel (C = a_scales * A @ (b_scales * B^T) + bias)
+  // A: [M, K], row-major
+  // B: [K, N], column-major
+  // C: [M, N], row-major
+  // bias: [N], row-major, optional
+  // a_scales: [MS]
+  // b_scales: [NS]
+  // Note: Due to the limitation of oneDNN
+  // (https://github.com/oneapi-src/oneDNN/issues/1636), the quantized bias is
+  // not supported.
+  template <typename OutputT, typename BiasT>
+  static void gemm_s8s8_jit(const int8_t* a, const int8_t* b, OutputT* c,
+                            const BiasT* bias, dnnl_dim_t M, dnnl_dim_t N,
+                            dnnl_dim_t K, const float* a_scales,
+                            const float* b_scales, dnnl_dim_t MS,
+                            dnnl_dim_t NS) {
+    auto&& OutputType = get_dnnl_type<OutputT>();
+    auto&& BiasType = get_dnnl_type<BiasT>();
+
+    dnnl::memory::desc a_md({M, K}, dnnl::memory::data_type::s8, {K, 1});
+    dnnl::memory::desc b_md({K, N}, dnnl::memory::data_type::s8, {1, K});
+    dnnl::memory::desc c_md({M, N}, OutputType, {N, 1});
+
+    dnnl::primitive_attr attr;
+    if constexpr (!InputNoScale) {
+      if (MS == 1) {
+        // per-tensor
+        attr.set_scales_mask(DNNL_ARG_SRC, 0);
+      } else {
+        // per-token
+        TORCH_CHECK(false, "per-token quantization is unsupported.");
+      }
+    }
+
+    if (NS == 1) {
+      // per-tensor
+      attr.set_scales_mask(DNNL_ARG_WEIGHTS, 0);
+    } else {
+      // per-channel
+      attr.set_scales_mask(DNNL_ARG_WEIGHTS, 2);
+    }
+
+    dnnl::matmul::primitive_desc matmul_pd;
+    if (bias) {
+      dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
+      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
+                                               bias_md, c_md, attr);
+    } else {
+      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
+                                               c_md, attr);
+    }
+    dnnl::matmul matmul(matmul_pd);
+
+    auto& engine = default_engine();
+
+    dnnl::memory a_m(a_md, engine, (void*)a);
+    dnnl::memory b_m(b_md, engine, (void*)b);
+    dnnl::memory c_m(c_md, engine, (void*)c);
+    dnnl::memory a_scales_m({{MS}, dnnl::memory::data_type::f32, {1}}, engine,
+                            (void*)a_scales);
+    dnnl::memory b_scales_m({{NS}, dnnl::memory::data_type::f32, {1}}, engine,
+                            (void*)b_scales);
+
+    auto& stream = default_stream();
+    if constexpr (InputNoScale) {
+      if (bias) {
+        dnnl::memory::desc bias_md({N}, BiasType, {1});
+        dnnl::memory bias_m(bias_md, engine, (void*)bias);
+        matmul.execute(
+            stream, {
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_BIAS, bias_m},
+                        {DNNL_ARG_DST, c_m},
+                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
+                    });
+      } else {
+        matmul.execute(
+            stream, {
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_DST, c_m},
+                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
+                    });
+      }
+    } else {
+      if (bias) {
+        dnnl::memory::desc bias_md({N}, BiasType, {1});
+        dnnl::memory bias_m(bias_md, engine, (void*)bias);
+        matmul.execute(
+            stream, {
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_BIAS, bias_m},
+                        {DNNL_ARG_DST, c_m},
+                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
+                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
+                    });
+      } else {
+        matmul.execute(
+            stream, {
+                        {DNNL_ARG_SRC, a_m},
+                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_DST, c_m},
+                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
+                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
+                    });
+      }
+    }
+    stream.wait();
+  }
+
+ private:
+  static dnnl::engine& default_engine() {
+    static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
+    return engine;
+  }
+
+  static dnnl::stream& default_stream() {
+    static dnnl::stream stream(default_engine());
+    return stream;
+  }
+};
+
+#endif

csrc/cpu/quant.cpp

@@ -0,0 +1,297 @@
+#include "cpu_types.hpp"
+#include "dnnl_helper.hpp"
+
+namespace {
+template <typename scalar_t>
+struct KernelVecType {
+  using load_vec_type = void;
+  using cvt_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using load_vec_type = vec_op::FP32Vec16;
+  using cvt_vec_type = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using load_vec_type = vec_op::BF16Vec16;
+  using cvt_vec_type = vec_op::FP32Vec16;
+};
+
+#ifdef __AVX512F__
+template <typename scalar_t>
+void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                   const float* scale, const int num_tokens,
+                                   const int hidden_size) {
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  constexpr float i8_min =
+      static_cast<float>(std::numeric_limits<int8_t>::min());
+  constexpr float i8_max =
+      static_cast<float>(std::numeric_limits<int8_t>::max());
+  const cvt_vec_t inv_scale(1.0 / *scale);
+  const cvt_vec_t i8_min_vec(i8_min);
+  const cvt_vec_t i8_max_vec(i8_max);
+
+  #pragma omp parallel for
+  for (int i = 0; i < num_tokens; ++i) {
+    int j = 0;
+    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+      load_vec_t elems(input + i * hidden_size + j);
+      cvt_vec_t elems_fp32(elems);
+      elems_fp32 = (elems_fp32 * inv_scale).clamp(i8_min_vec, i8_max_vec);
+      vec_op::INT8Vec16 elems_int8(elems_fp32);
+      elems_int8.save(output + i * hidden_size + j);
+    }
+
+    load_vec_t elems(input + i * hidden_size + j);
+    cvt_vec_t elems_fp32(elems);
+    elems_fp32 = (elems_fp32 * inv_scale).clamp(i8_min_vec, i8_max_vec);
+    vec_op::INT8Vec16 elems_int8(elems_fp32);
+
+    if (j + vec_elem_num == hidden_size) {
+      elems_int8.save(output + i * hidden_size + j);
+    } else {
+      elems_int8.save(output + i * hidden_size + j, hidden_size - j);
+    }
+  }
+}
+
+template <typename scalar_t>
+void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                    float* scale, const int num_tokens,
+                                    const int hidden_size) {
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  #pragma omp parallel for
+  for (int i = 0; i < num_tokens; ++i) {
+    cvt_vec_t max_abs(0.0);
+    {
+      int j = 0;
+      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+        load_vec_t elems(input + i * hidden_size + j);
+        cvt_vec_t elems_fp32(elems);
+        max_abs = max_abs.max(elems_fp32.abs());
+      }
+
+      load_vec_t elems(input + i * hidden_size + j);
+      cvt_vec_t elems_fp32(elems);
+
+      if (j + vec_elem_num == hidden_size) {
+        max_abs = max_abs.max(elems_fp32.abs());
+      } else {
+        max_abs = max_abs.max(elems_fp32.abs(), hidden_size - j);
+      }
+    }
+
+    float scale_val = max_abs.reduce_max() / 127.0f;
+    scale[i] = scale_val;
+    const cvt_vec_t inv_scale(1.0 / scale_val);
+
+    {
+      int j = 0;
+      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+        load_vec_t elems(input + i * hidden_size + j);
+        cvt_vec_t elems_fp32(elems);
+        elems_fp32 = (elems_fp32 * inv_scale);
+        vec_op::INT8Vec16 elems_int8(elems_fp32);
+        elems_int8.save(output + i * hidden_size + j);
+      }
+
+      load_vec_t elems(input + i * hidden_size + j);
+      cvt_vec_t elems_fp32(elems);
+      elems_fp32 = (elems_fp32 * inv_scale);
+      vec_op::INT8Vec16 elems_int8(elems_fp32);
+
+      if (j + vec_elem_num == hidden_size) {
+        elems_int8.save(output + i * hidden_size + j);
+      } else {
+        elems_int8.save(output + i * hidden_size + j, hidden_size - j);
+      }
+    }
+  }
+}
+
+template <bool Bias, typename scalar_t>
+void dynamic_output_scale_impl(const float* input, scalar_t* output,
+                               const float* scale, const scalar_t* bias,
+                               const int num_tokens, const int hidden_size) {
+  CPU_KERNEL_GUARD_IN(dynamic_output_scale_impl)
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  #pragma omp parallel for
+  for (int i = 0; i < num_tokens; ++i) {
+    int j = 0;
+    cvt_vec_t token_scale_vec(scale[i]);
+    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+      cvt_vec_t elems_fp32(input + i * hidden_size + j);
+      elems_fp32 = elems_fp32 * token_scale_vec;
+
+      if constexpr (Bias) {
+        load_vec_t bias_vec(bias + j);
+        cvt_vec_t bias_vec_fp32(bias_vec);
+        elems_fp32 = elems_fp32 + bias_vec_fp32;
+      }
+
+      load_vec_t elems_out(elems_fp32);
+      elems_out.save(output + i * hidden_size + j);
+    }
+
+    cvt_vec_t elems_fp32(input + i * hidden_size + j);
+    elems_fp32 = elems_fp32 * token_scale_vec;
+
+    if constexpr (Bias) {
+      load_vec_t bias_vec(bias + j);
+      cvt_vec_t bias_vec_fp32(bias_vec);
+      elems_fp32 = elems_fp32 + bias_vec_fp32;
+    }
+
+    load_vec_t elems_out(elems_fp32);
+
+    if (j + vec_elem_num == hidden_size) {
+      elems_out.save(output + i * hidden_size + j);
+    } else {
+      elems_out.save(output + i * hidden_size + j, hidden_size - j);
+    }
+  }
+}
+#else
+template <typename scalar_t>
+void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                   const float* scale, const int num_tokens,
+                                   const int hidden_size) {
+  TORCH_CHECK(false, "static_scaled_int8_quant_impl requires AVX512 support.")
+}
+
+template <typename scalar_t>
+void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                    float* scale, const int num_tokens,
+                                    const int hidden_size) {
+  TORCH_CHECK(false, "dynamic_scaled_int8_quant_impl requires AVX512 support.")
+}
+
+template <typename scalar_t>
+void dynamic_output_scale_impl() {
+  TORCH_CHECK(false, "dynamic_output_scale_impl requires AVX512 support.")
+}
+#endif
+}  // namespace
+
+void int8_scaled_mm(torch::Tensor& c,               // [M, OC], row-major
+                    const torch::Tensor& a,         // [M, IC], row-major
+                    const torch::Tensor& b,         // [IC, OC], column-major
+                    const torch::Tensor& a_scales,  // [1] or [M]
+                    const torch::Tensor& b_scales,  // [1] or [OC]
+                    const c10::optional<torch::Tensor>& bias  // [OC]
+) {
+  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm)
+  // Checks for conformality
+  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
+              "int8_scaled_mm only supports INT8 inputs.")
+  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
+  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
+              b.size(1) == c.size(1));
+  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
+  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
+
+  // Check for strides and alignment
+  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
+  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
+  TORCH_CHECK(c.stride(0) % 16 == 0 &&
+              b.stride(1) % 16 == 0);  // 16 Byte Alignment
+  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
+
+  if (bias) {
+    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous() &&
+                bias->dim() == 1);
+  }
+
+  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "cutlass_scaled_mm", [&] {
+    if (a_scales.numel() != 1) {
+      // per-token
+      // Note: oneDNN doesn't support per-token activation quantization
+      torch::Tensor tmp_fp32_out =
+          torch::empty_like(c, ::at::ScalarType::Float);
+      DNNLPrimitiveHelper<true>::gemm_s8s8_jit(
+          a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
+          tmp_fp32_out.data_ptr<float>(), (void*)(0), a.size(0), b.size(1),
+          a.size(1), (float*)(0), b_scales.data_ptr<float>(), 0,
+          b_scales.numel());
+      if (bias.has_value()) {
+        dynamic_output_scale_impl<true>(
+            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+            a_scales.data_ptr<float>(), bias->data_ptr<scalar_t>(), c.size(0),
+            c.size(1));
+      } else {
+        dynamic_output_scale_impl<false>(
+            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+            a_scales.data_ptr<float>(), (scalar_t*)(0), c.size(0), c.size(1));
+      }
+    } else {
+      // per-tensor
+      if (bias.has_value()) {
+        DNNLPrimitiveHelper<false>::gemm_s8s8_jit(
+            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(), c.data_ptr<scalar_t>(),
+            bias->data_ptr<scalar_t>(), a.size(0), b.size(1), a.size(1),
+            a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
+            a_scales.numel(), b_scales.numel());
+      } else {
+        DNNLPrimitiveHelper<false>::gemm_s8s8_jit(
+            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(), c.data_ptr<scalar_t>(),
+            (void*)(0), a.size(0), b.size(1), a.size(1),
+            a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
+            a_scales.numel(), b_scales.numel());
+      }
+    }
+  });
+}
+
+// static-per-tensor quantization.
+void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
+                              const torch::Tensor& input,  // [..., hidden_size]
+                              const torch::Tensor& scale,
+                              c10::optional<torch::Tensor> const& azp) {
+  CPU_KERNEL_GUARD_IN(static_scaled_int8_quant)
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(scale.numel() == 1);
+  TORCH_CHECK(!azp.has_value(), "Zero point is not supported on CPU.");
+
+  const int hidden_size = input.size(-1);
+  const int num_tokens = input.numel() / hidden_size;
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "static_scaled_int8_quant_impl", [&] {
+        static_scaled_int8_quant_impl(
+            input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+            scale.data_ptr<float>(), num_tokens, hidden_size);
+      });
+}
+
+// dynamic-per-token quantization.
+void dynamic_scaled_int8_quant(
+    torch::Tensor& out,          // [..., hidden_size]
+    const torch::Tensor& input,  // [..., hidden_size]
+    torch::Tensor& scale,        // [..., 1]
+    c10::optional<torch::Tensor> const& azp) {
+  CPU_KERNEL_GUARD_IN(dynamic_scaled_int8_quant)
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(!azp.has_value(), "Zero point is not supported on CPU.");
+
+  int const hidden_size = input.size(-1);
+  int const num_tokens = input.numel() / hidden_size;
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "dynamic_scaled_int8_quant_impl", [&] {
+        dynamic_scaled_int8_quant_impl(
+            input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+            scale.data_ptr<float>(), num_tokens, hidden_size);
+      });
+}

csrc/cpu/torch_bindings.cpp

@@ -1,9 +1,16 @@
 #include "cache.h"
 #include "ops.h"
-#include "registration.h"
+#include "core/registration.h"
 
 #include <torch/library.h>
 
+std::string init_cpu_threads_env(const std::string& cpu_ids);
+
+void int8_scaled_mm(torch::Tensor& c, const torch::Tensor& a,
+                    const torch::Tensor& b, const torch::Tensor& a_scales,
+                    const torch::Tensor& b_scales,
+                    const c10::optional<torch::Tensor>& bias);
+
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
@@ -25,8 +32,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // PagedAttention V2.
   ops.def(
       "paged_attention_v2("
-      "    Tensor! out, Tensor exp_sums, Tensor max_logits,"
-      "    Tensor tmp_out, Tensor query, Tensor key_cache,"
+      "    Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
+      "    Tensor! tmp_out, Tensor query, Tensor key_cache,"
       "    Tensor value_cache, int num_kv_heads, float scale,"
       "    Tensor block_tables, Tensor seq_lens, int block_size,"
       "    int max_seq_len, Tensor? alibi_slopes,"
@@ -82,6 +89,29 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "                 Tensor! key, int head_size,"
       "                 Tensor cos_sin_cache, bool is_neox) -> ()");
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
+
+  // Quantization
+#ifdef __AVX512F__
+  // Compute int8 quantized tensor for given scaling factor.
+  ops.def(
+      "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
+      "Tensor? azp) -> ()");
+  ops.impl("static_scaled_int8_quant", torch::kCPU, &static_scaled_int8_quant);
+
+  // Compute int8 quantized tensor and scaling factor
+  ops.def(
+      "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
+      "Tensor!? azp) -> ()");
+  ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
+           &dynamic_scaled_int8_quant);
+  // W8A8 GEMM, supporting symmetric per-tensor or per-row/column
+  // quantization.
+  ops.def(
+      "cutlass_scaled_mm(Tensor! out, Tensor a,"
+      "                  Tensor b, Tensor a_scales,"
+      "                  Tensor b_scales, Tensor? bias) -> ()");
+  ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm);
+#endif
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
@@ -93,8 +123,8 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
 
   // Copy the cache blocks from src to dst.
   cache_ops.def(
-      "copy_blocks(Tensor[]! key_caches, Tensor[]! value_caches, Tensor "
-      "block_mapping) -> ()");
+      "copy_blocks(Tensor(a!)[] key_caches, Tensor[](b!) value_caches, "
+      "Tensor block_mapping) -> ()");
   cache_ops.impl("copy_blocks", torch::kCPU, &copy_blocks);
 
   // Reshape the key and value tensors and cache them.
@@ -107,4 +137,9 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
   cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache);
 }
 
+TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {
+  // CPU utils
+  utils.def("init_cpu_threads_env(str cpu_ids) -> str", &init_cpu_threads_env);
+}
+
 REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

csrc/cpu/utils.cpp

@@ -0,0 +1,90 @@
+#include <numa.h>
+#include <unistd.h>
+#include <string>
+#include <sched.h>
+
+#include "cpu_types.hpp"
+
+std::string init_cpu_threads_env(const std::string& cpu_ids) {
+  bitmask* omp_cpu_mask = numa_parse_cpustring(cpu_ids.c_str());
+  TORCH_CHECK(omp_cpu_mask->size > 0);
+  std::vector<int> omp_cpu_ids;
+  omp_cpu_ids.reserve(omp_cpu_mask->size);
+
+  constexpr int group_size = 8 * sizeof(*omp_cpu_mask->maskp);
+
+  for (int offset = 0; offset < omp_cpu_mask->size; offset += group_size) {
+    unsigned long group_mask = omp_cpu_mask->maskp[offset / group_size];
+    int i = 0;
+    while (group_mask) {
+      if (group_mask & 1) {
+        omp_cpu_ids.emplace_back(offset + i);
+      }
+      ++i;
+      group_mask >>= 1;
+    }
+  }
+
+  // Memory node binding
+  if (numa_available() != -1) {
+    int mem_node_id = numa_node_of_cpu(omp_cpu_ids.front());
+    bitmask* mask = numa_parse_nodestring(std::to_string(mem_node_id).c_str());
+    bitmask* src_mask = numa_get_membind();
+
+    int pid = getpid();
+
+    // move all existing pages to the specified numa node.
+    *(src_mask->maskp) = *(src_mask->maskp) ^ *(mask->maskp);
+    int page_num = numa_migrate_pages(pid, src_mask, mask);
+    if (page_num == -1) {
+      TORCH_CHECK(false,
+                  "numa_migrate_pages failed. errno: " + std::to_string(errno));
+    }
+
+    // restrict memory allocation node.
+    numa_set_membind(mask);
+    numa_set_strict(1);
+  }
+
+  // OMP threads binding
+  omp_set_num_threads((int)omp_cpu_ids.size());
+  torch::set_num_threads((int)omp_cpu_ids.size());
+  TORCH_CHECK_EQ(omp_cpu_ids.size(), torch::get_num_threads());
+  TORCH_CHECK_EQ(omp_cpu_ids.size(), omp_get_max_threads());
+
+  std::vector<std::pair<int, int>> thread_core_mapping;
+  thread_core_mapping.reserve(omp_cpu_ids.size());
+  omp_lock_t writelock;
+  omp_init_lock(&writelock);
+
+#pragma omp parallel for schedule(static, 1)
+  for (size_t i = 0; i < omp_cpu_ids.size(); ++i) {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    CPU_SET(omp_cpu_ids[i], &mask);
+    int ret = sched_setaffinity(0, sizeof(cpu_set_t), &mask);
+    if (ret == -1) {
+      TORCH_CHECK(false,
+                  "sched_setaffinity failed. errno: " + std::to_string(errno));
+    }
+
+    omp_set_lock(&writelock);
+    thread_core_mapping.emplace_back(gettid(), omp_cpu_ids[i]);
+    omp_unset_lock(&writelock);
+  }
+
+  omp_destroy_lock(&writelock);
+
+  numa_free_nodemask(omp_cpu_mask);
+
+  std::stringstream ss;
+  ss << "OMP threads binding of Process " << getpid() << ":\n";
+  std::sort(thread_core_mapping.begin(), thread_core_mapping.end(),
+            [](auto&& a, auto&& b) { return a.second < b.second; });
+  for (auto&& item : thread_core_mapping) {
+    ss << "\t"
+       << "OMP tid: " << item.first << ", core " << item.second << "\n";
+  }
+
+  return ss.str();
+}

csrc/cuda_utils.h

@@ -1,5 +1,15 @@
 #pragma once
 
+#if defined(__CUDACC__) || defined(_NVHPC_CUDA)
+  #define HOST_DEVICE_INLINE __forceinline__ __host__ __device__
+  #define DEVICE_INLINE __forceinline__ __device__
+  #define HOST_INLINE __forceinline__ __host__
+#else
+  #define HOST_DEVICE_INLINE inline
+  #define DEVICE_INLINE inline
+  #define HOST_INLINE inline
+#endif
+
 int64_t get_device_attribute(int64_t attribute, int64_t device_id);
 
 int64_t get_max_shared_memory_per_block_device_attribute(int64_t device_id);

csrc/custom_all_reduce.cu

@@ -55,18 +55,6 @@ bool _is_weak_contiguous(torch::Tensor& t) {
           t.numel() * t.element_size());
 }
 
-bool should_custom_ar(torch::Tensor& inp, int64_t max_size, int64_t world_size,
-                      bool full_nvlink) {
-  auto inp_size = inp.numel() * inp.element_size();
-  // custom allreduce requires input byte size to be multiples of 16
-  if (inp_size % 16 != 0) return false;
-  if (!_is_weak_contiguous(inp)) return false;
-  if (world_size == 2 || full_nvlink) return inp_size <= max_size;
-  // for 4 or more non NVLink-capable GPUs, custom allreduce provides little
-  // performance improvement over NCCL.
-  return false;
-}
-
 void _all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
                  cudaStream_t stream) {
   auto fa = reinterpret_cast<vllm::CustomAllreduce*>(_fa);

csrc/custom_all_reduce.cuh

@@ -6,6 +6,7 @@
 #include <cuda_runtime.h>
 
 #include <iostream>
+#include <array>
 #include <limits>
 #include <map>
 #include <unordered_map>
@@ -23,17 +24,23 @@
 
 namespace vllm {
 
-constexpr int kMaxBlocks = 64;
-// note: we don't want to use atomics for signals because peer atomics are no
-// supported on PCIe links
+constexpr int kMaxBlocks = 36;
+// Counter may overflow, but it's fine since unsigned int overflow is
+// well-defined behavior.
+using FlagType = uint32_t;
 struct Signal {
-  alignas(128) uint32_t start[kMaxBlocks][8];
-  alignas(128) uint32_t end[kMaxBlocks][8];
+  alignas(128) FlagType self_counter[kMaxBlocks][8];
+  // Two sets of peer counters are needed for two syncs. The reason is that
+  // it's possible for peer GPU block to arrive at the second sync point while
+  // the current GPU block haven't passed the first sync point. Thus, peer GPU
+  // may write counter+1 while current GPU is busy waiting for counter. We use
+  // alternating counter array to avoid this possibility.
+  alignas(128) FlagType peer_counter[2][kMaxBlocks][8];
 };
 
 struct __align__(16) RankData { const void* __restrict__ ptrs[8]; };
 
-struct __align__(16) RankSignals { volatile Signal* signals[8]; };
+struct __align__(16) RankSignals { Signal* signals[8]; };
 
 // like std::array, but aligned
 template <typename T, int sz>
@@ -123,47 +130,71 @@ DINLINE O downcast(array_t<float, O::size> val) {
   }
 }
 
-// This function is meant to be used as the first synchronization in the all
-// reduce kernel. Thus, it doesn't need to make any visibility guarantees for
-// prior memory accesses. Note: volatile writes will not be reordered against
-// other volatile writes.
-template <int ngpus>
-DINLINE void start_sync(const RankSignals& sg, volatile Signal* self_sg,
-                        int rank) {
-  if (threadIdx.x < ngpus) {
-    // reset flag for next time
-    self_sg->end[blockIdx.x][threadIdx.x] = 0;
-    // simultaneously write to the corresponding flag of all ranks.
-    // Latency = 1 p2p write
-    sg.signals[threadIdx.x]->start[blockIdx.x][rank] = 1;
-    // wait until we got true from all ranks
-    while (!self_sg->start[blockIdx.x][threadIdx.x]);
-  }
-  __syncthreads();
+static DINLINE void st_flag_release(FlagType* flag_addr, FlagType flag) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  asm volatile("st.release.sys.global.u32 [%1], %0;" ::"r"(flag),
+               "l"(flag_addr));
+#else
+  asm volatile("membar.sys; st.volatile.global.u32 [%1], %0;" ::"r"(flag),
+               "l"(flag_addr));
+#endif
 }
 
-// This function is meant to be used as the second or the final synchronization
-// barrier in the all reduce kernel. If it's the final synchronization barrier,
-// we don't need to make any visibility guarantees for prior memory accesses.
-template <int ngpus, bool final_sync = false>
-DINLINE void end_sync(const RankSignals& sg, volatile Signal* self_sg,
-                      int rank) {
-  __syncthreads();
-  // eliminate the case that prior writes are not visible after signals become
-  // visible. Note that I did not managed to make this happen through a lot of
-  // testing. Might be the case that hardware provides stronger guarantee than
-  // the memory model.
-  if constexpr (!final_sync) __threadfence_system();
+static DINLINE FlagType ld_flag_acquire(FlagType* flag_addr) {
+  FlagType flag;
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  asm volatile("ld.acquire.sys.global.u32 %0, [%1];"
+               : "=r"(flag)
+               : "l"(flag_addr));
+#else
+  asm volatile("ld.volatile.global.u32 %0, [%1]; membar.gl;"
+               : "=r"(flag)
+               : "l"(flag_addr));
+#endif
+  return flag;
+}
+
+static DINLINE void st_flag_volatile(FlagType* flag_addr, FlagType flag) {
+  asm volatile("st.volatile.global.u32 [%1], %0;" ::"r"(flag), "l"(flag_addr));
+}
+
+static DINLINE FlagType ld_flag_volatile(FlagType* flag_addr) {
+  FlagType flag;
+  asm volatile("ld.volatile.global.u32 %0, [%1];"
+               : "=r"(flag)
+               : "l"(flag_addr));
+  return flag;
+}
+
+// is_start: whether this is the very first synchronization barrier.
+// need_fence: whether a memory fence is needed. If true, a release-acquire
+// semantic is used to enforce memory access order before and after this
+// barrier.
+template <int ngpus, bool is_start, bool need_fence = false>
+DINLINE void multi_gpu_barrier(const RankSignals& sg, Signal* self_sg,
+                               int rank) {
+  if constexpr (!is_start) __syncthreads();
+  static_assert(
+      !(is_start && need_fence));  // Start barrier shouldn't need fence.
   if (threadIdx.x < ngpus) {
-    // reset flag for next time
-    self_sg->start[blockIdx.x][threadIdx.x] = 0;
-    // simultaneously write to the corresponding flag of all ranks.
-    // Latency = 1 p2p write
-    sg.signals[threadIdx.x]->end[blockIdx.x][rank] = 1;
-    // wait until we got true from all ranks
-    while (!self_sg->end[blockIdx.x][threadIdx.x]);
+    // Increment the counter. Technically we only need one counter, but we use
+    // multiple per block to eliminate the need to share the counter via smem.
+    auto val = self_sg->self_counter[blockIdx.x][threadIdx.x] += 1;
+    // Write the expected counter value to peer and wait for correct value from
+    // peer.
+    auto peer_counter_ptr =
+        &sg.signals[threadIdx.x]->peer_counter[val % 2][blockIdx.x][rank];
+    auto self_counter_ptr =
+        &self_sg->peer_counter[val % 2][blockIdx.x][threadIdx.x];
+    if constexpr (need_fence) {
+      st_flag_release(peer_counter_ptr, val);
+      while (ld_flag_acquire(self_counter_ptr) != val);
+    } else {
+      st_flag_volatile(peer_counter_ptr, val);
+      while (ld_flag_volatile(self_counter_ptr) != val);
+    }
   }
-  if constexpr (!final_sync) __syncthreads();
+  if constexpr (is_start || need_fence) __syncthreads();
 }
 
 template <typename P, int ngpus, typename A>
@@ -178,33 +209,31 @@ DINLINE P packed_reduce(const P* ptrs[], int idx) {
 
 template <typename T, int ngpus>
 __global__ void __launch_bounds__(512, 1)
-    cross_device_reduce_1stage(RankData* _dp, RankSignals sg,
-                               volatile Signal* self_sg, T* __restrict__ result,
-                               int rank, int size) {
+    cross_device_reduce_1stage(RankData* _dp, RankSignals sg, Signal* self_sg,
+                               T* __restrict__ result, int rank, int size) {
   using P = typename packed_t<T>::P;
   using A = typename packed_t<T>::A;
   // note: we don't reorder the address so the accumulation order is the same
   // for all ranks, ensuring bitwise identical results
   auto dp = *_dp;
-  start_sync<ngpus>(sg, self_sg, rank);
+  multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
   // do the actual reduction
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
     ((P*)result)[idx] = packed_reduce<P, ngpus, A>((const P**)&dp.ptrs[0], idx);
   }
-  end_sync<ngpus, true>(sg, self_sg, rank);
+  multi_gpu_barrier<ngpus, false>(sg, self_sg, rank);
 }
 
 template <typename P>
-DINLINE P* get_tmp_buf(volatile Signal* sg) {
+DINLINE P* get_tmp_buf(Signal* sg) {
   return (P*)(((Signal*)sg) + 1);
 }
 
 template <typename T, int ngpus>
 __global__ void __launch_bounds__(512, 1)
-    cross_device_reduce_2stage(RankData* _dp, RankSignals sg,
-                               volatile Signal* self_sg, T* __restrict__ result,
-                               int rank, int size) {
+    cross_device_reduce_2stage(RankData* _dp, RankSignals sg, Signal* self_sg,
+                               T* __restrict__ result, int rank, int size) {
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = gridDim.x * blockDim.x;
   using P = typename packed_t<T>::P;
@@ -222,12 +251,12 @@ __global__ void __launch_bounds__(512, 1)
     tmps[i] = get_tmp_buf<P>(sg.signals[target]);
   }
   auto tmp_out = tmps[0];
-  start_sync<ngpus>(sg, self_sg, rank);
+  multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
   // stage 1: reduce scatter
   for (int idx = start + tid; idx < end; idx += stride) {
     tmp_out[idx - start] = packed_reduce<P, ngpus, A>(ptrs, idx);
   }
-  end_sync<ngpus>(sg, self_sg, rank);
+  multi_gpu_barrier<ngpus, false, true>(sg, self_sg, rank);
 
   // stage 2: allgather. Note: it's important to match the tid between
   // the two stages, because visibility across devices is only guaranteed
@@ -437,6 +466,8 @@ class CustomAllreduce {
 #define KL(ngpus, name)                                                       \
   name<T, ngpus><<<blocks, threads, 0, stream>>>(ptrs, sg_, self_sg_, output, \
                                                  rank_, size);
+    // TODO(hanzhi713): Threshold is different for A100 and H100.
+    // Add per device threshold.
 #define REDUCE_CASE(ngpus)                            \
   case ngpus: {                                       \
     if (world_size_ == 2) {                           \

csrc/custom_all_reduce_test.cu

@@ -1,15 +1,15 @@
 /**
  * This is a standalone test for custom allreduce.
  * To compile, make sure you have MPI and NCCL installed in your system.
- * export MPI_HOME=XXX
+ * export MPI_HOME=xxx
  * nvcc -O2 -arch=native -std=c++17 custom_all_reduce_test.cu -o
- * custom_all_reduce_test -lnccl -I${MPI_HOME}/include -lmpi
+ * custom_all_reduce_test -lnccl -I${MPI_HOME} -lmpi
  *
  * Warning: this C++ test is not designed to be very readable and was used
  * during the rapid prototyping process.
  *
  * To run:
- * mpirun -np 8 ./custom_all_reduce_test
+ * mpirun --allow-run-as-root -np 8 ./custom_all_reduce_test
  */
 #include <cuda.h>
 #include <curand_kernel.h>
@@ -44,7 +44,14 @@
   } while (0)
 
 __global__ void dummy_kernel() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
   for (int i = 0; i < 100; i++) __nanosleep(1000000);  // 100ms
+#else
+  for (int i = 0; i < 100; i++) {
+    long long int start = clock64();
+    while (clock64() - start < 150000000);  // approximately 98.4ms on P40
+  }
+#endif
 }
 
 template <typename T>
@@ -302,15 +309,19 @@ int main(int argc, char** argv) {
 
   bool performance_test = true;
   cudaProfilerStart();
-  // for (int threads : {256, 512}) {
+  // Uncomment to scan through different block size configs.
+  // for (int threads : {256, 512, 1024}) {
   //   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
-  //     run<half>(myRank, nRanks, comm, threads, block_limit, 4096 * 1024);
+  //     run<half>(myRank, nRanks, comm, threads, block_limit, 1024 * 1024,
+  //     performance_test);
   //   }
   // }
+  // Scan through different sizes to test performance.
   for (int sz = 512; sz <= (8 << 20); sz *= 2) {
     run<half>(myRank, nRanks, comm, 512, 36, sz + 8 * 47, performance_test);
   }
 
   cudaProfilerStop();
+  MPICHECK(MPI_Finalize());
   return EXIT_SUCCESS;
 }

csrc/cutlass_extensions/cute_utils.cuh

@@ -0,0 +1,68 @@
+#pragma once
+
+#include <cute/tensor.hpp>
+#include <torch/all.h>
+namespace cute {
+
+////////////////////////////////////////////////////////////////////
+// layout utils
+////////////////////////////////////////////////////////////////////
+
+// Permute layout based on indices, example:
+//   permute_layout<1, 0>(layout) will swap the two dimensions
+//   permute_layout<0, 2, 1>(layout) will swap the last two dimensions
+template <size_t... I, typename Layout>
+CUTE_HOST_DEVICE static constexpr auto permute_layout(Layout l) {
+  static_assert(rank(l) == sizeof...(I), "Invalid permutation, rank mismatch");
+  return cute::make_layout(cute::get<I>(l)...);
+}
+
+// is the layout f(x) = x
+template <typename Layout>
+CUTE_HOST_DEVICE static constexpr bool is_identity_layout() {
+  if constexpr (std::is_same_v<Layout, void>)
+    return true;
+  else {
+    constexpr auto coalesced_layout = coalesce(Layout{});
+    if constexpr (rank(coalesced_layout) == 1 &&
+                  stride<0>(coalesced_layout) == 1) {
+      return true;
+    }
+    return false;
+  }
+}
+
+////////////////////////////////////////////////////////////////////
+// Pointer utils
+////////////////////////////////////////////////////////////////////
+
+template <class PointerType>
+static constexpr auto get_logical_ptr(PointerType* ptr) {
+  if constexpr (cute::sizeof_bits_v<PointerType> < 8) {
+    return cute::subbyte_iterator<PointerType>(ptr);
+  } else {
+    return ptr;
+  }
+}
+
+////////////////////////////////////////////////////////////////////
+// Misc utils
+////////////////////////////////////////////////////////////////////
+
+template <typename T, typename Elements>
+CUTE_HOST_DEVICE static constexpr auto create_auto_vectorizing_copy() {
+  constexpr auto bits = sizeof_bits_v<T> * Elements{};
+  if constexpr (bits % 128 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<128>{};
+  } else if constexpr (bits % 64 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<64>{};
+  } else if constexpr (bits % 32 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<32>{};
+  } else if constexpr (bits % 16 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<16>{};
+  } else {
+    return AutoVectorizingCopyWithAssumedAlignment<8>{};
+  }
+}
+
+};  // namespace cute

csrc/cutlass_extensions/torch_utils.hpp

@@ -0,0 +1,160 @@
+#pragma once
+
+#include <torch/all.h>
+
+#include "cute/layout.hpp"
+#include "cutlass/layout/matrix.h"
+#include "cutlass/bfloat16.h"
+#include "cutlass/half.h"
+
+using ColumnMajor = typename cutlass::layout::ColumnMajor;
+using RowMajor = typename cutlass::layout::RowMajor;
+
+namespace cute {
+
+namespace detail {
+
+template <class T, class F, class G, int... I>
+CUTE_HOST_DEVICE constexpr auto tapply_with_idx(T&& t, F&& f, G&& g,
+                                                seq<I...>) {
+  return g(f(cute::get<I>(static_cast<T&&>(t)), I)...);
+}
+
+template <class F, int... I>
+CUTE_HOST_DEVICE constexpr auto make_shape_from_idx(F&& f, seq<I...>) {
+  return make_shape(f(I)...);
+}
+
+};  // namespace detail
+
+template <class T, class F>
+CUTE_HOST_DEVICE constexpr auto transform_with_idx(T const& t, F&& f) {
+  if constexpr (cute::is_tuple<T>::value) {
+    return detail::tapply_with_idx(
+        t, f, [](auto const&... a) { return cute::make_tuple(a...); },
+        tuple_seq<T>{});
+  } else {
+    return f(t);
+  }
+
+  CUTE_GCC_UNREACHABLE;
+}
+
+// calls: make_shape(f(0), f(1), ..., f(N-1))
+template <int N, class F>
+CUTE_HOST_DEVICE constexpr auto make_shape_from_idx(F&& f) {
+  return detail::make_shape_from_idx(f, make_seq<N>{});
+}
+
+};  // namespace cute
+
+// Make a layout from a tensor with `rank(Stride{})`, where the shape is the
+// shape of the passed in tensor and the strides are of type `Stride` and
+// contain the strides of the passed in tensor, checking that any static strides
+// in `Stride{}` match the strides of the passed in tensor.
+// If `tensor.dim() < rank(Stride{})`, the shape is padded with 1s and the extra
+// strides are set to be 0 or 1.
+template <typename Stride>
+static inline auto make_cute_layout(torch::Tensor const& tensor,
+                                    std::string_view name = "tensor") {
+  TORCH_CHECK(tensor.dim() <= rank(Stride{}));
+  auto stride = cute::transform_with_idx(
+      Stride{}, [&](auto const& stride_ele, auto const& idx) {
+        using StrideEle = std::decay_t<decltype(stride_ele)>;
+
+        if (idx < tensor.dim()) {
+          if constexpr (cute::is_static_v<StrideEle>) {
+            TORCH_CHECK(StrideEle::value == tensor.stride(idx), "Expected ",
+                        name, ".stride(", idx, ") to be ", StrideEle::value);
+            return StrideEle{};
+          } else {
+            if (tensor.size(idx) == 1) {
+              // use 0 stride for dim with size 1, this is easier for
+              // cute/cutlass to optimize (helps the TMA code flatten dims)
+              return StrideEle{0};
+            } else {
+              return tensor.stride(idx);
+            }
+          }
+        } else {
+          // Extra strides are assumed to be 0 or 1
+          if constexpr (cute::is_static_v<StrideEle>) {
+            static_assert(StrideEle::value == 0 || StrideEle::value == 1);
+          }
+          return StrideEle{};
+        }
+      });
+
+  auto shape = cute::make_shape_from_idx<rank(Stride{})>([&](auto const& idx) {
+    if (idx < tensor.dim())
+      return tensor.size(idx);
+    else
+      return int64_t(1);
+  });
+
+  return make_layout(shape, stride);
+}
+
+template <typename Stride>
+static inline auto maybe_make_cute_layout(
+    c10::optional<torch::Tensor> const& tensor,
+    std::string_view name = "tensor") {
+  using Layout = decltype(make_cute_layout<Stride>(*tensor));
+
+  if (tensor) {
+    return std::optional<Layout>{make_cute_layout<Stride>(*tensor, name)};
+  } else {
+    return std::optional<Layout>{};
+  }
+}
+
+//
+//  Torch Type to Cutlass Type (equivalent_cutlass_type)
+//
+
+template <typename T>
+struct equivalent_cutlass_type {
+  using type = T;
+};
+
+template <typename T>
+using equivalent_cutlass_type_t = typename equivalent_cutlass_type<T>::type;
+
+template <>
+struct equivalent_cutlass_type<c10::Half> {
+  using type = cutlass::half_t;
+};
+
+template <>
+struct equivalent_cutlass_type<c10::BFloat16> {
+  using type = cutlass::bfloat16_t;
+};
+
+//
+// equivalent_scalar_t (basically inverse of equivalent_cutlass_type)
+//
+
+// Return a `c10::CppTypeToScalarType<T>` compatible type, i.e. get the C++ from
+// c10 that is equivalent to T, e.g.: `cutlass::half_t -> c10::Half`
+template <typename T>
+struct equivalent_scalar_type {
+  using type = T;
+};
+
+template <typename T>
+using equivalent_scalar_type_t = typename equivalent_scalar_type<T>::type;
+
+template <>
+struct equivalent_scalar_type<cutlass::half_t> {
+  using type = c10::Half;
+};
+
+template <>
+struct equivalent_scalar_type<cutlass::bfloat16_t> {
+  using type = c10::BFloat16;
+};
+
+// get equivalent c10::ScalarType tag from compile time type
+template <typename T>
+static inline constexpr c10::ScalarType equivalent_scalar_type_v =
+    c10::CppTypeToScalarType<equivalent_scalar_type_t<T>>::value;

csrc/cutlass_extensions/vllm_collective_builder.cuh

@@ -0,0 +1,43 @@
+#pragma once
+
+#include "cutlass/gemm/collective/collective_builder.hpp"
+
+namespace cutlass::gemm::collective {
+using namespace cute;
+
+//
+// VLLMCollectiveBuilder is a wrapper around CollectiveBuilder that allows for
+// for custom kernel tags, allowing you to build custom collectives. Without
+// touching the cutlass library headers, using `CutlassKernelTag` will mean it
+// will resort to using the standard cutlass collective builder.
+//
+
+// Use the default Cutlass collective builder, i.e. use an unmodified cutless
+// collective
+struct CutlassKernelTag {};
+
+template <class KernelTag, class ArchTag, class OpClass, class ElementA,
+          class GmemLayoutA, int AlignmentA, class ElementB, class GmemLayoutB,
+          int AlignmentB, class ElementAccumulator, class TileShape_MNK,
+          class ClusterShape_MNK, class StageCountType,
+          class KernelScheduleType, class Enable = void>
+struct VLLMCollectiveBuilder {
+  static_assert(sizeof(ElementA) == 0,
+                "Could not build a collective for given parameters.");
+};
+
+template <class ArchTag, class OpClass, class ElementA, class GmemLayoutA,
+          int AlignmentA, class ElementB, class GmemLayoutB, int AlignmentB,
+          class ElementAccumulator, class TileShape_MNK, class ClusterShape_MNK,
+          class StageCountType, class KernelScheduleType>
+struct VLLMCollectiveBuilder<
+    CutlassKernelTag, ArchTag, OpClass, ElementA, GmemLayoutA, AlignmentA,
+    ElementB, GmemLayoutB, AlignmentB, ElementAccumulator, TileShape_MNK,
+    ClusterShape_MNK, StageCountType, KernelScheduleType> {
+  using CollectiveOp = typename CollectiveBuilder<
+      ArchTag, OpClass, ElementA, GmemLayoutA, AlignmentA, ElementB,
+      GmemLayoutB, AlignmentB, ElementAccumulator, TileShape_MNK,
+      ClusterShape_MNK, StageCountType, KernelScheduleType>::CollectiveOp;
+};
+
+};  // namespace cutlass::gemm::collective

csrc/cutlass_extensions/vllm_custom_types.cuh

@@ -0,0 +1,50 @@
+#pragma once
+
+#include "cutlass/integer_subbyte.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int Bits, int Bias, bool Signed = false>
+struct vllm_biased_integer_subbyte : public integer_subbyte<Bits, Signed> {
+  using Base = integer_subbyte<Bits, Signed>;
+
+  using Storage = typename Base::Storage;
+  using xint_t = typename Base::xint_t;
+
+  using Base::bits_mask_;
+  using Base::sign_mask_;
+  using Base::storage;
+
+  //
+  // Methods
+  //
+
+  /// No operation
+  vllm_biased_integer_subbyte() = default;
+
+  /// Conversion from integer type
+  CUTLASS_HOST_DEVICE explicit vllm_biased_integer_subbyte(int value)
+      : Base(value) {}
+  CUTLASS_HOST_DEVICE explicit vllm_biased_integer_subbyte(unsigned value)
+      : Base(value) {}
+  CUTLASS_HOST_DEVICE explicit vllm_biased_integer_subbyte(double value)
+      : Base(value) {}
+};
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+// "GPTQ" types, i.e. symmetric quantization
+using vllm_uint4b8_t = vllm_biased_integer_subbyte<4, 8>;      // u4b8
+using vllm_uint8b128_t = vllm_biased_integer_subbyte<8, 128>;  // u8b128
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int Bits, int Bias, bool Signed>
+struct sizeof_bits<vllm_biased_integer_subbyte<Bits, Bias, Signed>> {
+  static constexpr int value = Bits;
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

csrc/cutlass_extensions/vllm_cutlass_library_extension.py

@@ -0,0 +1,49 @@
+import enum
+from typing import Dict, Union
+
+from cutlass_library import *
+
+#
+#   Extend cutlass library with custom types, and missing values
+#
+
+
+class VLLMDataType(enum.Enum):
+    u4b8 = enum_auto()
+    u8b128 = enum_auto()
+
+
+class MixedInputKernelScheduleType(enum.Enum):
+    TmaWarpSpecializedMixedInput = enum_auto()
+    TmaWarpSpecializedPingpongMixedInput = enum_auto()
+    TmaWarpSpecializedCooperativeMixedInput = enum_auto()
+
+
+VLLMDataTypeNames: Dict[Union[VLLMDataType, DataType], str] = {
+    **DataTypeNames,  # type: ignore
+    **{
+        VLLMDataType.u4b8: "u4b8",
+        VLLMDataType.u8b128: "u8b128",
+    }
+}
+
+VLLMDataTypeTag: Dict[Union[VLLMDataType, DataType], str] = {
+    **DataTypeTag,  # type: ignore
+    **{
+        VLLMDataType.u4b8: "cutlass::vllm_uint4b8_t",
+        VLLMDataType.u8b128: "cutlass::vllm_uint8b128_t",
+    }
+}
+
+VLLMKernelScheduleTag: Dict[Union[
+    MixedInputKernelScheduleType, KernelScheduleType], str] = {
+        **KernelScheduleTag,  # type: ignore
+        **{
+            MixedInputKernelScheduleType.TmaWarpSpecializedMixedInput:
+            "cutlass::gemm::KernelTmaWarpSpecializedMixedInput",
+            MixedInputKernelScheduleType.TmaWarpSpecializedPingpongMixedInput:
+            "cutlass::gemm::KernelTmaWarpSpecializedPingpongMixedInput",
+            MixedInputKernelScheduleType.TmaWarpSpecializedCooperativeMixedInput:
+            "cutlass::gemm::KernelTmaWarpSpecializedCooperativeMixedInput",
+        }
+    }

csrc/cutlass_extensions/vllm_numeric_conversion.cuh

@@ -0,0 +1,795 @@
+#pragma once
+
+#include "cutlass/numeric_conversion.h"
+#include "cutlass_extensions/vllm_custom_types.cuh"
+#include "cutlass_extensions/cute_utils.cuh"
+
+// this file extends:
+//   https://github.com/NVIDIA/cutlass/blob/cutlass-3.5.0/include/cutlass/numeric_conversion.h
+// with vllm specific type conversions, namely: vllm_uint4b8_t, vllm_uint8b128_t
+// as well as adds interleaved numeric array converters for specific types.
+// (interleaved numeric array converters can be more efficient for subbyte
+// types)
+
+namespace cutlass {
+
+// InterleavedNumericArrayConverter is like NumericArrayConverter but also
+// deinterleaves converted elements based on IlvBlkLayout, interleaving can
+// make subbyte converts more efficient by allowing for efficient extraction
+// of subbyte elements from a 32bit register.
+template <typename IlvBlkLayout, typename T, typename S, int N,
+          FloatRoundStyle Round = FloatRoundStyle::round_to_nearest,
+          class Enable = void>
+struct InterleavedNumericArrayConverter {
+  using Converter = NumericArrayConverter<T, S, N, Round>;
+
+  using result_type = typename Converter::result_type;
+  using source_type = typename Converter::source_type;
+
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    CUTE_INVALID_CONTROL_PATH(
+        "InterleavedNumericArrayConverter not implemented\n");
+    return {};
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+template <typename IlvBlkLayout, typename T, typename S, int N,
+          FloatRoundStyle Round>
+struct InterleavedNumericArrayConverter<
+    IlvBlkLayout, T, S, N, Round,
+    std::enable_if_t<is_identity_layout<IlvBlkLayout>()>> {
+  using Converter = NumericArrayConverter<T, S, N, Round>;
+
+  using result_type = typename Converter::result_type;
+  using source_type = typename Converter::source_type;
+
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return Converter::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// TODO (LucasWilkinson): Implement
+// for Array<cutlass::float8_e4m3fn, N> <= Array<vllm_uint4b8_t, N>
+
+// ....
+
+template <typename RegConvert32bit, typename T, typename S, int N>
+struct ArrayConverterPacked32Bit {
+  using result_type = Array<T, N>;
+  using source_type = Array<S, N>;
+
+  using result_packed_8_t = Array<T, 8>;
+  using result_packed_4_t = Array<T, 4>;
+  using result_packed_2_t = Array<T, 2>;
+  using src_packed_8_t = Array<S, 8>;
+  using src_packed_4_t = Array<S, 4>;
+  using src_packed_2_t = Array<S, 2>;
+
+  static_assert(N % 2 == 0, "N must be a multiple of 2");
+  static_assert(cutlass::sizeof_bits_v<S> >= 4);  // TODO: add 16 packed sources
+  static_assert(32 % cutlass::sizeof_bits_v<S> == 0);
+  static constexpr auto src_elems_per_32bit_reg =
+      32 / cutlass::sizeof_bits_v<S>;
+
+  // Maybe not Valid. ScalarConverter will not actually work unless
+  // NumericConverter<T, S, Round> is implemented. However it won't be used
+  // anyways since we assert N % 2 == 0, just here for compliance with
+  // VectorizedConverter.
+  using ScalarConverter = NumericConverter<T, S>;
+
+  template <typename PackedSrc>
+  CUTLASS_DEVICE static uint32_t to_reg(PackedSrc const& source) {
+    if constexpr (sizeof(PackedSrc) == 1) {
+      return static_cast<uint32_t>(reinterpret_cast<const uint8_t&>(source));
+    } else if constexpr (sizeof(PackedSrc) == 2) {
+      return static_cast<uint32_t>(reinterpret_cast<const uint16_t&>(source));
+    } else {
+      static_assert(sizeof(PackedSrc) == 4);
+      return reinterpret_cast<const uint32_t&>(source);
+    }
+  }
+
+  // The core converter uses bit tricks to construct a known FP16 number, then
+  // does a subtraction in FP16 for the final result.
+  template <typename PackedResultType, typename PackedSrcType>
+  CUTLASS_DEVICE static PackedResultType packed_convert(
+      PackedSrcType const& source) {
+    static_assert(PackedSrcType::kElements == PackedResultType::kElements);
+    static_assert(PackedResultType::kElements == 2 ||
+                      PackedResultType::kElements == 4 ||
+                      PackedResultType::kElements == 8,
+                  "Invalid PackedResultType must be 2, 4 or 8.");
+    static_assert(std::is_same_v<typename PackedSrcType::Element, S>);
+    static_assert(std::is_same_v<typename PackedResultType::Element, T>);
+
+    return RegConvert32bit::template convert<PackedResultType>(to_reg(source));
+  }
+
+  friend class detail::VectorizedConverter;
+
+ public:
+  CUTLASS_DEVICE static result_type convert(source_type const& source) {
+    result_type result;
+    using ConverterType =
+        ArrayConverterPacked32Bit<RegConvert32bit,
+                                  typename result_type::Element,
+                                  typename source_type::Element, N>;
+
+    if constexpr (src_elems_per_32bit_reg >= 8) {
+      detail::VectorizedConverter::convert<
+          ConverterType, result_packed_8_t, src_packed_8_t, result_packed_4_t,
+          src_packed_4_t, result_packed_2_t, src_packed_2_t>(result, source);
+    } else if constexpr (src_elems_per_32bit_reg >= 4) {
+      detail::VectorizedConverter::convert<ConverterType, result_packed_4_t,
+                                           src_packed_4_t, result_packed_2_t,
+                                           src_packed_2_t>(result, source);
+    } else {
+      detail::VectorizedConverter::convert<ConverterType, result_packed_2_t,
+                                           src_packed_2_t>(result, source);
+    }
+
+    return result;
+  }
+};
+
+// for Array<cutlass::half_t, N> <= Array<vllm_uint4b8_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::half_t, vllm_uint4b8_t, N, Round> {
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<vllm_uint4b8_t, N>;
+
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      // Below constructs the following temporary:
+      // fp16s_01 = {0x00, i4_01, 0x00, i4_01}
+      // fp16s_23 = {0x00, i4_23, 0x00, i4_23}
+      // fp16s_45 = {0x00, i4_45, 0x00, i4_45}
+      // fp16s_67 = {0x00, i4_67, 0x00, i4_67}
+      // We use inline asm instead of __byte_perm intrinsic since we don't want
+      // the documented (& 0x7) on the index. NVCC might be able to optimize it
+      // out since the index is a constexpr, but we choose to be safe about it
+      // here.
+      uint32_t prmt_indices[4] = {0x4040, 0x4141, 0x4242, 0x4343};
+      static_assert(RegArray::kElements <= 4,
+                    "Too many inputs for F16 -> I4 vector converter");
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  prmt.b32 %0, %1, %2, %3;\n"
+            "}\n"
+            : "=r"(r[ii])
+            : "r"(src), "n"(0), "r"(prmt_indices[ii]));
+      }
+
+      // Since the stored 4bit values are biased by 8 we get stored_val = (x+8)
+      //  we are trying to construct x and a fp16 value
+      // The below XOR does the following:
+      //  1) Sets the exponent bits of the FP16 to the correct value for the
+      //  FP16 magic_num. We will be constructing {1024+16*(x1+8), 1024+(x0+8)},
+      //  where x1 in the high nibble and x0 is the low nibble then using hfma
+      //  to subtract 1032 from that
+      // The AND does the following:
+      //  1) Clear the set bits for the int4 we will ignore.
+      // We use lop3 so that we can use 1 instruction for AND and XOR.
+      static constexpr uint32_t xor_mask = 0x64006400;
+      static constexpr uint32_t and_mask = 0xFFF0FF0F;
+      static constexpr uint32_t immLut = (0xf0 & 0xcc) ^ 0xaa;
+
+      // For each operand, computes:
+      // r[i] = (r[i] & and_mask) ^ xor_mask
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii])
+            : "n"(and_mask), "n"(xor_mask), "n"(immLut));
+      }
+
+      // We will issue 2 hfmas that do the following:
+      // {x1, x0} = {1024+16*(x1+8), 1024+(x0+8)} * {1/16, 1} - {72, 1032}
+      //          = {x1 + 1152, x0 + 1032} * {1/16, 1} - {72, 1032}
+      static constexpr uint32_t hfma_bias_rep = 0xD480E408;   // {72, 1032}
+      static constexpr uint32_t hfma_scale_rep = 0x2C003C00;  // {1 / 16, 1}
+
+      const half2& hfma_bias = reinterpret_cast<const half2&>(hfma_bias_rep);
+      const half2& hfma_scale = reinterpret_cast<const half2&>(hfma_scale_rep);
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii]);
+        fp16x2_val = __hfma2(hfma_scale, fp16x2_val, hfma_bias);
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<vllm_uint4b8_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::half_t, vllm_uint4b8_t, N,
+                                        Round, void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<vllm_uint4b8_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t xor_mask = 0x64006400;
+
+      for (int ii = 0; ii < RegArray::kElements; ii += 2) {
+        auto src_ = src >> (4 * (ii));
+        r[ii + 0] = src_;
+        r[ii + 1] = src_;
+
+        static constexpr uint32_t and_xor_imm_lut = (0xf0 & 0xcc) ^ 0xaa;
+
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+        static constexpr uint32_t high_nib_mask = 0x00F000F0;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 0])
+            : "n"(low_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 1])
+            : "n"(high_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {1024+(x1+8), 1024+(x0+8)} * {1, 1} - {1032, 1032}
+        // For high nibble:
+        //  {x1, x0} = {1024+16*(x1+8), 1024+16*(x0+8)} * {1/16, 1/16}
+        //             - {72, 72}
+        static constexpr uint32_t low_nib_bias = 0x64086408;    // {1032, 1032}
+        static constexpr uint32_t high_nib_scale = 0x2C002C00;  // {1/16, 1/16}
+        static constexpr uint32_t high_nib_bias = 0xD480D480;   // {-72, -72}
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 0]);
+          fp16x2_val =
+              __hsub2(fp16x2_val, reinterpret_cast<const half2&>(low_nib_bias));
+        }
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 1]);
+          fp16x2_val = __hfma2(fp16x2_val,
+                               reinterpret_cast<const half2&>(high_nib_scale),
+                               reinterpret_cast<const half2&>(high_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<uint4_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::half_t, uint4_t, N, Round,
+                                        void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<uint4_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t xor_mask = 0x64006400;
+
+      for (int ii = 0; ii < RegArray::kElements; ii += 2) {
+        auto src_ = src >> (4 * (ii));
+        r[ii + 0] = src_;
+        r[ii + 1] = src_;
+
+        static constexpr uint32_t and_xor_imm_lut = (0xf0 & 0xcc) ^ 0xaa;
+
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+        static constexpr uint32_t high_nib_mask = 0x00F000F0;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 0])
+            : "n"(low_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 1])
+            : "n"(high_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {1024+x1, 1024+x0} - {1024, 1024}
+        // For high nibble:
+        //  {x1, x0} = {1024+16*x1, 1024+16*x0} * {1/16, 1/16} - {64, 64}
+        static constexpr uint32_t low_nib_bias = 0x64006400;    // {1024, 1024}
+        static constexpr uint32_t high_nib_scale = 0x2C002C00;  // {1/16, 1/16}
+        static constexpr uint32_t high_nib_bias = 0xD400D400;   // {-64, -64}
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 0]);
+          fp16x2_val =
+              __hsub2(fp16x2_val, reinterpret_cast<const half2&>(low_nib_bias));
+        }
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 1]);
+          fp16x2_val = __hfma2(fp16x2_val,
+                               reinterpret_cast<const half2&>(high_nib_scale),
+                               reinterpret_cast<const half2&>(high_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<vllm_uint8b128_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::half_t, vllm_uint8b128_t, N, Round> {
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<vllm_uint8b128_t, N>;
+
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      // Hold output FP16s in reg. We need 1 reg for every 2 elements
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      uint32_t const prmt_indices[2] = {0x5150, 0x5352};
+      static constexpr uint32_t start_byte_for_fp16 = 0x64646464;
+
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile("prmt.b32 %0,%1,%2,%3;\n"
+                     : "=r"(r[ii])
+                     : "r"(src), "n"(start_byte_for_fp16),
+                       "r"(prmt_indices[ii]));
+      }
+
+      // -128 is folded into bias subtraction, i.e. the 0x80 in the low bytes
+      static constexpr uint32_t bias_rep = 0x64806480;
+      const half2& bias = reinterpret_cast<const half2&>(bias_rep);
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii]);
+        fp16x2_val = __hsub2(fp16x2_val, bias);
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::float, N> <= Array<vllm_uint8b128_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<float, vllm_uint8b128_t, N, Round> {
+  using result_type = Array<float, N>;
+  using source_type = Array<vllm_uint8b128_t, N>;
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      PackedResultType r;
+
+      // __byte_perm simulates the add.u32 0x4B000000 to every u8 element of
+      // u8x4 source and stores the result in r (without introducing extra
+      // cvt.u32.u8 instruction)
+      uint32_t const prmt_indices[4] = {0x7650, 0x7651, 0x7652, 0x7653};
+      uint32_t* result_as_int = reinterpret_cast<uint32_t*>(&r);
+      for (int ii = 0; ii < PackedResultType::kElements; ++ii) {
+        result_as_int[ii] = __byte_perm(src, 0x4B000000, prmt_indices[ii]);
+        // Subtract the magic number 0x4B000000 from tmp in floating-point
+        // arithmetic to obtain final result
+        r[ii] -= (8388608.f + 128.f);  // fold in -128 bias
+      }
+
+      return r;
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+
+// for Array<cutlass::bfloat16_t, N> <= Array<vllm_uint4b8_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::bfloat16_t, vllm_uint4b8_t, N, Round> {
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<vllm_uint4b8_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src_reg) {
+      // Hold output BF16s in reg. We need 1 reg for every 2 elements
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+      uint32_t src_reg_shifted = src_reg >> 4;
+
+      // Below constructs the following temporary:
+      uint32_t const prmt_indices[4] = {0xF4F0, 0xF5F1, 0xF6F2, 0xF7F3};
+      static_assert(RegArray::kElements <= 4,
+                    "Too many inputs for uint4b8_t -> BF16 vector converter");
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  prmt.b32 %0, %1, %2, %3;\n"
+            "}\n"
+            : "=r"(r[ii])
+            : "r"(src_reg), "r"(src_reg_shifted), "r"(prmt_indices[ii]));
+      }
+
+      // Since the stored 4bit values are biased by 8 we get stored_val = (x+8)
+      //  we are trying to construct x and a BF16 value
+      // The below XOR does the following:
+      //  1) Sets the exponent bits of the BF16 to the correct value for the
+      //  BF16 magic_num. We will be constructing {128 + (x1+8), 128 + (x0+8)}
+      //  and subtracting 136 to get {x1, x0}
+      static constexpr uint32_t xor_mask = 0x43004300;
+      static constexpr uint32_t and_mask = 0x000F000F;
+      static constexpr uint32_t immLut = (0xf0 & 0xcc) ^ 0xaa;
+
+      // For each operand, computes:
+      // r[i] = (r[i] & and_mask) ^ xor_mask
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii])
+            : "n"(and_mask), "n"(xor_mask), "n"(immLut));
+      }
+
+      // We will issue 2 bfmas that do the following:
+      // high BF16:
+      // hi_bf16 - 136, lo_bf16 - 136
+
+      // This is the BF16 {136, 136} represented as an integer.
+      static constexpr uint32_t bias_rep = 0x43084308;
+      const __nv_bfloat162& bias =
+          reinterpret_cast<const __nv_bfloat162&>(bias_rep);
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        __nv_bfloat162& bf16x2_val = reinterpret_cast<__nv_bfloat162&>(r[ii]);
+        bf16x2_val = __hsub2(bf16x2_val, bias);
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    }
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::bfloat16_t, N> <= Array<vllm_uint4b8_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::bfloat16_t, vllm_uint4b8_t, N,
+                                        Round, void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<vllm_uint4b8_t, N>;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t or_mask = 0x43004300;
+
+      // Unlike float16 where the mantissa is large enough to contain 2
+      // nibbles, bfloat16 can only fit one, so we can only convert one
+      // nibble at a time
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        r[ii] = src >> (4 * ii);
+
+        static constexpr uint32_t and_or_imm_lut = (0xf0 & 0xcc) | 0xaa;
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 0])
+            : "n"(low_nib_mask), "n"(or_mask), "n"(and_or_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {128+(x1+8), 128+(x0+8)} * {1, 1} - {136, 136}
+        static constexpr uint32_t low_nib_bias = 0x43084308;  // {136, 136}
+
+        {
+          __nv_bfloat162& fp16x2_val = reinterpret_cast<__nv_bfloat162&>(r[ii]);
+          fp16x2_val =
+              __hsub2(fp16x2_val,
+                      reinterpret_cast<const __nv_bfloat162&>(low_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::bfloat16_t, N> <= Array<uint4_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::bfloat16_t, uint4_t, N, Round,
+                                        void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<uint4_t, N>;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(uint32_t src) {
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t or_mask = 0x43004300;
+
+      // Unlike float16 where the mantissa is large enough to contain 2
+      // nibbles, bfloat16 can only fit one, so we can only convert one
+      // nibble at a time
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        r[ii] = src >> (4 * ii);
+
+        static constexpr uint32_t and_or_imm_lut = (0xf0 & 0xcc) | 0xaa;
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii])
+            : "n"(low_nib_mask), "n"(or_mask), "n"(and_or_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {128 + x1, 128 + x0} * {1, 1} - {128, 128}
+        static constexpr uint32_t low_nib_bias = 0x43004300;  // {128, 128}
+
+        {
+          __nv_bfloat162& fp16x2_val = reinterpret_cast<__nv_bfloat162&>(r[ii]);
+          fp16x2_val =
+              __hsub2(fp16x2_val,
+                      reinterpret_cast<const __nv_bfloat162&>(low_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::bfloat16_t, N> <= Array<vllm_uint8b128_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::bfloat16_t, vllm_uint8b128_t, N, Round> {
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<vllm_uint8b128_t, N>;
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  using result_packed_4_t = Array<cutlass::bfloat16_t, 4>;
+  using result_packed_2_t = Array<cutlass::bfloat16_t, 2>;
+  using src_packed_4_t = Array<vllm_uint8b128_t, 4>;
+  using src_packed_2_t = Array<vllm_uint8b128_t, 2>;
+
+  // Not Valid, not supported, only here to satisfy the interface and to avoid
+  //  a compile error. ScalarConverter will not actually work until
+  //  NumericConverter<cutlass::bfloat16_t, vllm_uint8b128_t, Round> is
+  //  implemented
+  using ScalarConverter =
+      NumericConverter<cutlass::bfloat16_t, vllm_uint8b128_t, Round>;
+
+  template <typename PackedResultType, typename PackedSrcType>
+  CUTLASS_DEVICE static PackedResultType packed_convert(
+      PackedSrcType const& source) {
+    static_assert(
+        (platform::is_same<PackedSrcType, src_packed_2_t>::value &&
+         platform::is_same<PackedResultType, result_packed_2_t>::value) ||
+            (platform::is_same<PackedSrcType, src_packed_4_t>::value &&
+             platform::is_same<PackedResultType, result_packed_4_t>::value),
+        "Invalid PackedSrcType/PackedResultType must be 2 or 4 to use private "
+        "convert dispatch.");
+
+    NumericArrayConverter<float, vllm_uint8b128_t, PackedResultType::kElements,
+                          Round>
+        convert_uint8_to_f32;
+    Array<float, PackedResultType::kElements> tmp =
+        convert_uint8_to_f32(source);
+    NumericArrayConverter<cutlass::bfloat16_t, float,
+                          PackedResultType::kElements, Round>
+        convert_f32_to_bf16_;
+    return convert_f32_to_bf16_(tmp);
+  }
+
+  friend class detail::VectorizedConverter;
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    result_type result;
+    using ConverterType =
+        NumericArrayConverter<typename result_type::Element,
+                              typename source_type::Element, N, Round>;
+    detail::VectorizedConverter::convert<ConverterType, result_packed_4_t,
+                                         src_packed_4_t, result_packed_2_t,
+                                         src_packed_2_t>(result, source);
+
+    return result;
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+#endif
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass
+
+/////////////////////////////////////////////////////////////////////////////////////////////////

csrc/layernorm_kernels.cu

@@ -3,13 +3,16 @@
 #include <c10/cuda/CUDAGuard.h>
 
 #include "dispatch_utils.h"
-#include "reduction_utils.cuh"
 #ifndef USE_ROCM
   #include <cuda_bf16.h>
   #include <cuda_fp16.h>
+  #include <cub/util_type.cuh>
+  #include <cub/cub.cuh>
 #else
   #include <hip/hip_bf16.h>
   #include <hip/hip_fp16.h>
+  #include <hipcub/util_type.hpp>
+  #include <hipcub/hipcub.hpp>
 
 using __nv_bfloat16 = __hip_bfloat16;
 using __nv_bfloat162 = __hip_bfloat162;
@@ -31,7 +34,11 @@ __global__ void rms_norm_kernel(
     const float x = (float)input[blockIdx.x * hidden_size + idx];
     variance += x * x;
   }
-  variance = blockReduceSum<float>(variance);
+
+  using BlockReduce = cub::BlockReduce<float, 1024>;
+  __shared__ typename BlockReduce::TempStorage reduceStore;
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
+
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
   }
@@ -228,12 +235,11 @@ fused_add_rms_norm_kernel(
     variance += temp.sum_squares();
     residual_v[id] = temp;
   }
-  /* Keep the following if-else block in sync with the
-     calculation of max_block_size in fused_add_rms_norm */
-  if (num_tokens < 256) {
-    variance = blockReduceSum<float, 1024>(variance);
-  } else
-    variance = blockReduceSum<float, 256>(variance);
+
+  using BlockReduce = cub::BlockReduce<float, 1024>;
+  __shared__ typename BlockReduce::TempStorage reduceStore;
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
+
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
   }
@@ -268,12 +274,11 @@ fused_add_rms_norm_kernel(
     variance += x * x;
     residual[blockIdx.x * hidden_size + idx] = z;
   }
-  /* Keep the following if-else block in sync with the
-     calculation of max_block_size in fused_add_rms_norm */
-  if (num_tokens < 256) {
-    variance = blockReduceSum<float, 1024>(variance);
-  } else
-    variance = blockReduceSum<float, 256>(variance);
+
+  using BlockReduce = cub::BlockReduce<float, 1024>;
+  __shared__ typename BlockReduce::TempStorage reduceStore;
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
+
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
   }