[Docs] Fix readthedocs for tag build (#6158)

Signed-off-by: Christian Rohmann <christian.rohmann@inovex.de>

.buildkite/download-images.sh

@@ -8,10 +8,6 @@ set -o pipefail
 # aws s3 sync s3://air-example-data-2/vllm_opensource_llava/ images/
 mkdir -p images
 cd images
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign_pixel_values.pt
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign_image_features.pt
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom_pixel_values.pt
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom_image_features.pt
 wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign.jpg
 wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom.jpg
 

.buildkite/lm-eval-harness/configs/Meta-Llama-3-70B-Instruct.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-70B-Instruct -b 32 -l 250 -f 5
+model_name: "meta-llama/Meta-Llama-3-70B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.892
+  - name: "exact_match,flexible-extract"
+    value: 0.892
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-FP8.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m neuralmagic/Meta-Llama-3-8B-Instruct-FP8 -b 32 -l 250 -f 5 -t 1
+model_name: "neuralmagic/Meta-Llama-3-8B-Instruct-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.756
+  - name: "exact_match,flexible-extract"
+    value: 0.752
+limit: 250
+num_fewshot: 5

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

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-8B-Instruct -b 32 -l 250 -f 5 -t 1
+model_name: "meta-llama/Meta-Llama-3-8B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.756
+  - name: "exact_match,flexible-extract"
+    value: 0.752
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Mixtral-8x22B-Instruct-v0.1-FP8-Dynamic.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Mixtral-8x22B-Instruct-v0.1-FP8-dynamic -b "auto" -l 250 -f 5 -t 8
+model_name: "neuralmagic/Mixtral-8x22B-Instruct-v0.1-FP8-dynamic"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.86
+  - name: "exact_match,flexible-extract"
+    value: 0.86
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Mixtral-8x7B-Instruct-v0.1-FP8.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8 -b "auto" -l 250 -f 5 -t 4
+model_name: "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.624
+  - name: "exact_match,flexible-extract"
+    value: 0.624
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Mixtral-8x7B-Instruct-v0.1.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1 -b 32 -l 250 -f 5 -t 4
+model_name: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.616
+  - name: "exact_match,flexible-extract"
+    value: 0.632
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2-57B-A14-Instruct.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m Qwen/Qwen2-57B-A14B-Instruct -b "auto" -l 250 -f 5 -t 4
+model_name: "Qwen/Qwen2-57B-A14B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.792
+  - name: "exact_match,flexible-extract"
+    value: 0.824
+limit: 250
+num_fewshot: 5

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

@@ -0,0 +1,3 @@
+Meta-Llama-3-70B-Instruct.yaml
+Mixtral-8x7B-Instruct-v0.1.yaml
+Qwen2-57B-A14-Instruct.yaml

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

@@ -0,0 +1,2 @@
+Meta-Llama-3-8B-Instruct.yaml
+Meta-Llama-3-8B-Instruct-FP8.yaml

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

@@ -0,0 +1,46 @@
+#!/bin/bash
+# We can use this script to compute baseline accuracy on GSM for transformers.
+#
+# Make sure you have lm-eval-harness installed:
+#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@9516087b81a61d0e220b22cc1b75be76de23bc10
+
+usage() {
+    echo``
+    echo "Runs lm eval harness on GSM8k using huggingface transformers."
+    echo "This pathway is intended to be used to create baselines for "
+    echo "our automated nm-test-accuracy workflow"
+    echo
+    echo "usage: ${0} <options>"
+    echo
+    echo "  -m    - huggingface stub or local directory of the model"
+    echo "  -b    - batch size to run the evaluation at"
+    echo "  -l    - limit number of samples to run"
+    echo "  -f    - number of fewshot samples to use"
+    echo
+}
+
+while getopts "m:b:l:f:" OPT; do
+  case ${OPT} in
+    m ) 
+        MODEL="$OPTARG"
+        ;;
+    b ) 
+        BATCH_SIZE="$OPTARG"
+        ;;
+    l ) 
+        LIMIT="$OPTARG"
+        ;;
+    f ) 
+        FEWSHOT="$OPTARG"
+        ;;
+    \? ) 
+        usage
+        exit 1
+        ;;
+  esac
+done
+
+lm_eval --model hf \
+  --model_args pretrained=$MODEL,parallelize=True \
+  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --batch_size $BATCH_SIZE

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

@@ -0,0 +1,51 @@
+#!/bin/bash
+# We can use this script to compute baseline accuracy on GSM for vllm.
+# We use this for fp8, which HF does not support.
+#
+# Make sure you have lm-eval-harness installed:
+#   pip install lm-eval==0.4.2
+
+usage() {
+    echo``
+    echo "Runs lm eval harness on GSM8k using huggingface transformers."
+    echo "This pathway is intended to be used to create baselines for "
+    echo "our automated nm-test-accuracy workflow"
+    echo
+    echo "usage: ${0} <options>"
+    echo
+    echo "  -m    - huggingface stub or local directory of the model"
+    echo "  -b    - batch size to run the evaluation at"
+    echo "  -l    - limit number of samples to run"
+    echo "  -f    - number of fewshot samples to use"
+    echo "  -t    - tensor parallel size to run at"
+    echo
+}
+
+while getopts "m:b:l:f:t:" OPT; do
+  case ${OPT} in
+    m ) 
+        MODEL="$OPTARG"
+        ;;
+    b ) 
+        BATCH_SIZE="$OPTARG"
+        ;;
+    l ) 
+        LIMIT="$OPTARG"
+        ;;
+    f ) 
+        FEWSHOT="$OPTARG"
+        ;;
+    t )
+        TP_SIZE="$OPTARG"
+        ;;
+    \? ) 
+        usage
+        exit 1
+        ;;
+  esac
+done
+
+lm_eval --model vllm \
+  --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE \
+  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --batch_size $BATCH_SIZE

.buildkite/lm-eval-harness/run-tests.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+
+usage() {
+    echo``
+    echo "Runs lm eval harness on GSM8k using vllm and compares to "
+    echo "precomputed baseline (measured by HF transformers.)"
+    echo
+    echo "usage: ${0} <options>"
+    echo
+    echo "  -c    - path to the test data config (e.g. configs/small-models.txt)"
+    echo "  -t    - tensor parallel size"
+    echo
+}
+
+SUCCESS=0
+
+while getopts "c:t:" OPT; do
+  case ${OPT} in
+    c ) 
+        CONFIG="$OPTARG"
+        ;;
+    t )
+        TP_SIZE="$OPTARG"
+        ;;
+    \? )
+        usage
+        exit 1
+        ;;
+  esac
+done
+
+# Parse list of configs.
+IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < $CONFIG
+
+for MODEL_CONFIG in "${MODEL_CONFIGS[@]}"
+do
+    LOCAL_SUCCESS=0
+    
+    echo "=== RUNNING MODEL: $MODEL_CONFIG WITH TP SIZE: $TP_SIZE==="
+
+    export LM_EVAL_TEST_DATA_FILE=$PWD/configs/${MODEL_CONFIG}
+    export LM_EVAL_TP_SIZE=$TP_SIZE
+    pytest -s test_lm_eval_correctness.py || LOCAL_SUCCESS=$?
+
+    if [[ $LOCAL_SUCCESS == 0 ]]; then
+        echo "=== PASSED MODEL: ${MODEL_CONFIG} ==="
+    else
+        echo "=== FAILED MODEL: ${MODEL_CONFIG} ==="
+    fi
+
+    SUCCESS=$((SUCCESS + LOCAL_SUCCESS))
+
+done
+
+if [ "${SUCCESS}" -eq "0" ]; then
+    exit 0
+else
+    exit 1
+fi

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

@@ -0,0 +1,54 @@
+"""
+LM eval harness on model to compare vs HF baseline computed offline.
+Configs are found in configs/$MODEL.yaml
+
+* export LM_EVAL_TEST_DATA_FILE=configs/Meta-Llama-3-70B-Instruct.yaml
+* export LM_EVAL_TP_SIZE=4 
+* pytest -s test_lm_eval_correctness.py
+"""
+
+import os
+from pathlib import Path
+
+import lm_eval
+import numpy
+import yaml
+
+RTOL = 0.02
+TEST_DATA_FILE = os.environ.get(
+    "LM_EVAL_TEST_DATA_FILE",
+    ".buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml")
+
+TP_SIZE = os.environ.get("LM_EVAL_TP_SIZE", 1)
+
+
+def launch_lm_eval(eval_config):
+    model_args = f"pretrained={eval_config['model_name']}," \
+                 f"tensor_parallel_size={TP_SIZE}"
+
+    results = lm_eval.simple_evaluate(
+        model="vllm",
+        model_args=model_args,
+        tasks=[task["name"] for task in eval_config["tasks"]],
+        num_fewshot=eval_config["num_fewshot"],
+        limit=eval_config["limit"],
+        batch_size="auto")
+
+    return results
+
+
+def test_lm_eval_correctness():
+    eval_config = yaml.safe_load(
+        Path(TEST_DATA_FILE).read_text(encoding="utf-8"))
+
+    # Launch eval requests.
+    results = launch_lm_eval(eval_config)
+
+    # Confirm scores match ground truth.
+    for task in eval_config["tasks"]:
+        for metric in task["metrics"]:
+            ground_truth = metric["value"]
+            measured_value = results["results"][task["name"]][metric["name"]]
+            print(f'{task["name"]} | {metric["name"]}: '
+                  f'ground_truth={ground_truth} | measured={measured_value}')
+            assert numpy.isclose(ground_truth, measured_value, rtol=RTOL)

.buildkite/nightly-benchmarks/README.md

@@ -0,0 +1,103 @@
+# vLLM benchmark suite
+
+## 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 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.
+
+**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.
+
+
+## Configuring the workload
+
+The benchmarking workload contains three parts:
+- Latency tests in `latency-tests.json`.
+- Throughput tests in `throughput-tests.json`.
+- Serving tests in `serving-tests.json`.
+
+See [descriptions.md](tests/descriptions.md) for detailed descriptions. 
+
+### Latency test
+
+Here is an example of one test inside `latency-tests.json`:
+
+```json
+[
+    {
+        "test_name": "latency_llama8B_tp1",
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    },
+]
+```
+
+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`
+
+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
+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
+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:
+
+```
+[
+    {
+        "test_name": "serving_llama8B_tp1_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "tensor_parallel_size": 1,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+]
+```
+
+Inside this example:
+- The `test_name` attribute is also a unique identifier for the test. It must start with `serving_`.
+- The `server-parameters` includes the command line arguments for vLLM server.
+- The `client-parameters` includes the command line arguments for `benchmark_serving.py`.
+- The `qps_list` controls the list of qps for test. It will be used to configure the `--request-rate` parameter in `benchmark_serving.py`
+
+The number of this test is less stable compared to the delay and latency benchmarks (due to randomized sharegpt dataset sampling inside `benchmark_serving.py`), but a large change on this number (e.g. 5% change) still vary the output greatly.
+
+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
+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.

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

@@ -0,0 +1,62 @@
+steps:
+  - label: "Wait for container to be ready"
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          containers:
+          - image: badouralix/curl-jq
+            command:
+            - sh
+            - .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
+  - wait
+  - label: "A100 Benchmark"
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          priorityClassName: perf-benchmark
+          containers:
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+            command:
+            - bash .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+            env:
+            - name: VLLM_USAGE_SOURCE
+              value: ci-test
+            - name: HF_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: hf-token-secret
+                  key: token
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+  # - label: "H100: NVIDIA SMI"
+  #   agents:
+  #     queue: H100
+  #   plugins:
+  #   - docker#v5.11.0:
+  #       image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+  #       command:
+  #       - bash
+  #       - .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
+  #       mount-buildkite-agent: true
+  #       propagate-environment: true
+  #       propagate-uid-gid: false
+  #       ipc: host
+  #       gpus: all
+  #       environment:
+  #       - VLLM_USAGE_SOURCE
+  #       - HF_TOKEN
+

.buildkite/nightly-benchmarks/kickoff-pipeline.sh

@@ -1,5 +1,6 @@
 #!/usr/bin/env bash
 
+# NOTE(simon): this script runs inside a buildkite agent with CPU only access.
 set -euo pipefail
 
 # Install system packages
@@ -23,4 +24,4 @@ if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
 fi
 
 # Upload sample.yaml
-buildkite-agent pipeline upload .buildkite/nightly-benchmarks/sample.yaml
+buildkite-agent pipeline upload .buildkite/nightly-benchmarks/benchmark-pipeline.yaml

.buildkite/nightly-benchmarks/run-benchmarks-suite.sh

@@ -0,0 +1,358 @@
+#!/bin/bash
+
+# This script should be run inside the CI process
+# This script assumes that we are already inside the vllm/ directory
+# Benchmarking results will be available inside vllm/benchmarks/results/
+
+# Do not set -e, as the mixtral 8x22B model tends to crash occasionally
+# and we still want to see other benchmarking results even when mixtral crashes.
+set -o pipefail
+
+check_gpus() {
+  # check the number of GPUs and GPU type.
+  declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  if [[ $gpu_count -gt 0 ]]; then
+    echo "GPU found."
+  else
+    echo "Need at least 1 GPU to run benchmarking."
+    exit 1
+  fi
+  declare -g gpu_type=$(echo $(nvidia-smi --query-gpu=name --format=csv,noheader) | awk '{print $2}')
+  echo "GPU type is $gpu_type"
+}
+
+check_hf_token() {
+  # check if HF_TOKEN is available and valid
+  if [[ -z "$HF_TOKEN" ]]; then
+    echo "Error: HF_TOKEN is not set."
+    exit 1
+  elif [[ ! "$HF_TOKEN" =~ ^hf_ ]]; then
+    echo "Error: HF_TOKEN does not start with 'hf_'."
+    exit 1
+  else
+    echo "HF_TOKEN is set and valid."
+  fi
+}
+
+json2args() {
+  # transforms the JSON string to command line args, and '_' is replaced to '-'
+  # example:
+  # input: { "model": "meta-llama/Llama-2-7b-chat-hf", "tensor_parallel_size": 1 }
+  # output: --model meta-llama/Llama-2-7b-chat-hf --tensor-parallel-size 1
+  local json_string=$1
+  local args=$(
+    echo "$json_string" | jq -r '
+      to_entries |
+      map("--" + (.key | gsub("_"; "-")) + " " + (.value | tostring)) |
+      join(" ")
+    '
+  )
+  echo "$args"
+}
+
+wait_for_server() {
+  # wait for vllm server to start
+  # return 1 if vllm server crashes
+  timeout 1200 bash -c '
+    until curl localhost:8000/v1/completions; do
+      sleep 1
+    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."
+  else
+      for pid in $pids; do
+          kill -9 "$pid"
+          echo "Killed process with PID: $pid"
+      done
+
+      echo "All GPU processes have been killed."
+  fi
+
+  # waiting for GPU processes to be fully killed
+  sleep 10
+
+  # remove vllm config file
+  rm -rf ~/.config/vllm
+
+  # Print the GPU memory usage
+  # so that we know if all GPU processes are killed.
+  gpu_memory_usage=$(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() {
+  # upload the benchmarking results to buildkite
+
+  # if the agent binary is not found, skip uploading the results, exit 0
+  if [ ! -f /workspace/buildkite-agent ]; then
+    echo "buildkite-agent binary not found. Skip uploading the results."
+    return 0
+  fi
+  /workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < $RESULTS_FOLDER/benchmark_results.md
+  /workspace/buildkite-agent artifact upload "$RESULTS_FOLDER/*"
+}
+
+run_latency_tests() {
+  # run latency tests using `benchmark_latency.py`
+  # $1: a json file specifying latency test cases
+
+  local latency_test_file
+  latency_test_file=$1
+
+  # Iterate over latency tests
+  jq -c '.[]' "$latency_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')
+    if [[ ! "$test_name" =~ ^latency_ ]]; then
+      echo "In latency-test.json, test_name must start with \"latency_\"."
+      exit 1
+    fi
+
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+
+    # get arguments
+    latency_params=$(echo "$params" | jq -r '.parameters')
+    latency_args=$(json2args "$latency_params")
+
+    # check if there is enough GPU to run the test
+    tp=$(echo "$latency_params" | jq -r '.tensor_parallel_size')
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $testname."
+      continue
+    fi
+
+    latency_command="python3 benchmark_latency.py \
+      --output-json $RESULTS_FOLDER/${test_name}.json \
+      $latency_args"
+
+    echo "Running test case $test_name"
+    echo "Latency command: $latency_command"
+
+    # recoding benchmarking command ang GPU command
+    jq_output=$(jq -n \
+      --arg latency "$latency_command" \
+      --arg gpu "$gpu_type" \
+      '{
+        latency_command: $latency,
+        gpu_type: $gpu
+      }')
+    echo "$jq_output" > "$RESULTS_FOLDER/$test_name.commands"
+
+    # run the benchmark
+    eval "$latency_command"
+
+    kill_gpu_processes
+
+  done
+}
+
+
+run_throughput_tests() {
+  # run throughput tests using `benchmark_throughput.py`
+  # $1: a json file specifying throughput test cases
+
+  local throughput_test_file
+  throughput_test_file=$1
+
+  # Iterate over throughput tests
+  jq -c '.[]' "$throughput_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')
+    if [[ ! "$test_name" =~ ^throughput_ ]]; then
+      echo "In throughput-test.json, test_name must start with \"throughput_\"."
+      exit 1
+    fi
+
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+
+    # get arguments
+    throughput_params=$(echo "$params" | jq -r '.parameters')
+    throughput_args=$(json2args "$throughput_params")
+
+    # check if there is enough GPU to run the test
+    tp=$(echo $throughput_params | jq -r '.tensor_parallel_size')
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $testname."
+      continue
+    fi
+
+    throughput_command="python3 benchmark_throughput.py \
+      --output-json $RESULTS_FOLDER/${test_name}.json \
+      $throughput_args"
+
+    echo "Running test case $test_name"
+    echo "Throughput command: $throughput_command"
+    # recoding benchmarking command ang GPU command
+    jq_output=$(jq -n \
+      --arg command "$throughput_command" \
+      --arg gpu "$gpu_type" \
+      '{
+        throughput_command: $command,
+        gpu_type: $gpu
+      }')
+    echo "$jq_output" > "$RESULTS_FOLDER/$test_name.commands"
+
+    # run the benchmark
+    eval "$throughput_command"
+
+    kill_gpu_processes
+
+  done
+}
+
+run_serving_tests() {
+  # run serving tests using `benchmark_serving.py`
+  # $1: a json file specifying serving test cases
+
+  local serving_test_file
+  serving_test_file=$1
+
+  # Iterate over serving tests
+  jq -c '.[]' "$serving_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')
+    if [[ ! "$test_name" =~ ^serving_ ]]; then
+      echo "In serving-test.json, test_name must start with \"serving_\"."
+      exit 1
+    fi
+
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r '.server_parameters')
+    client_params=$(echo "$params" | jq -r '.client_parameters')
+    server_args=$(json2args "$server_params")
+    client_args=$(json2args "$client_params")
+    qps_list=$(echo "$params" | jq -r '.qps_list')
+    qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
+    echo "Running over qps list $qps_list"
+
+    # check if there is enough GPU to run the test
+    tp=$(echo "$server_params" | jq -r '.tensor_parallel_size')
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $testname."
+      continue
+    fi
+
+    # check if server model and client model is aligned
+    server_model=$(echo "$server_params" | jq -r '.model')
+    client_model=$(echo "$client_params" | jq -r '.model')
+    if [[ $server_model != "$client_model" ]]; then
+      echo "Server model and client model must be the same. Skip testcase $testname."
+      continue
+    fi
+
+    server_command="python3 \
+      -m vllm.entrypoints.openai.api_server \
+      $server_args"
+
+    # run the server
+    echo "Running test case $test_name"
+    echo "Server command: $server_command"
+    eval "$server_command" &
+
+    # wait until the server is alive
+    wait_for_server
+    if [ $? -eq 0 ]; then
+      echo ""
+      echo "vllm server is up and running."
+    else
+      echo ""
+      echo "vllm failed to start within the timeout period."
+    fi
+
+    # iterate over different QPS
+    for qps in $qps_list; do
+      # remove the surrounding single quote from qps
+      if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
+        qps="inf"
+        echo "now qps is $qps"
+      fi
+
+      new_test_name=$test_name"_qps_"$qps
+
+      client_command="python3 benchmark_serving.py \
+        --save-result \
+        --result-dir $RESULTS_FOLDER \
+        --result-filename ${new_test_name}.json \
+        --request-rate $qps \
+        $client_args"
+
+      echo "Running test case $test_name with qps $qps"
+      echo "Client command: $client_command"
+
+      eval "$client_command"
+
+      # record the benchmarking commands
+      jq_output=$(jq -n \
+        --arg server "$server_command" \
+        --arg client "$client_command" \
+        --arg gpu "$gpu_type" \
+        '{
+          server_command: $server,
+          client_command: $client,
+          gpu_type: $gpu
+        }')
+      echo "$jq_output" > "$RESULTS_FOLDER/${new_test_name}.commands"
+
+    done
+
+    # clean up
+    kill_gpu_processes
+  done
+}
+
+main() {
+  check_gpus
+  check_hf_token
+
+  # dependencies
+  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+  (which jq) || (apt-get update && apt-get -y install jq)
+
+  # get the current IP address, required by benchmark_serving.py
+  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
+  # turn of the reporting of the status of each request, to clean up the terminal output
+  export VLLM_LOG_LEVEL="WARNING"
+
+  # prepare for benchmarking
+  cd benchmarks || exit 1
+  wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+  declare -g RESULTS_FOLDER=results/
+  mkdir -p $RESULTS_FOLDER
+  QUICK_BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
+
+  # benchmarking
+  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/serving-tests.json
+  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 $QUICK_BENCHMARK_ROOT/scripts/convert-results-json-to-markdown.py
+
+  upload_to_buildkite
+}
+
+main "$@"

.buildkite/nightly-benchmarks/sample.yaml

@@ -1,39 +0,0 @@
-steps:
-  # NOTE(simon): You can create separate blocks for different jobs
-  - label: "A100: NVIDIA SMI"
-    agents:
-      queue: A100
-    plugins:
-    - kubernetes:
-        podSpec:
-          containers:
-          # - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:$BUILDKITE_COMMIT
-          # TODO(simon): check latest main branch or use the PR image.
-          - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
-            command:
-            - bash -c 'nvidia-smi && nvidia-smi topo -m && pwd && ls'
-            resources:
-              limits:
-                nvidia.com/gpu: 8
-            volumeMounts:
-            - name: devshm
-              mountPath: /dev/shm
-          nodeSelector:
-            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
-          volumes:
-          - name: devshm
-            emptyDir:
-              medium: Memory
-  # TODO(simon): bring H100 online
-  # - label: "H100: NVIDIA SMI"
-  #   agents:
-  #     queue: H100
-  #   plugins:
-  #   - docker#v5.11.0:
-  #       image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
-  #       command:
-  #       - bash -c 'nvidia-smi && nvidia-smi topo -m'
-  #       propagate-environment: true
-  #       ipc: host
-  #       gpus: all
-

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

@@ -0,0 +1,192 @@
+import json
+import os
+from pathlib import Path
+
+import pandas as pd
+from tabulate import tabulate
+
+results_folder = Path("results/")
+
+# latency results and the keys that will be printed into markdown
+latency_results = []
+latency_column_mapping = {
+    "test_name": "Test name",
+    "gpu_type": "GPU",
+    "avg_latency": "Mean latency (ms)",
+    # "P10": "P10 (s)",
+    # "P25": "P25 (s)",
+    "P50": "Median latency (ms)",
+    # "P75": "P75 (s)",
+    # "P90": "P90 (s)",
+    "P99": "P99 latency (ms)",
+}
+
+# throughput tests and the keys that will be printed into markdown
+throughput_results = []
+throughput_results_column_mapping = {
+    "test_name": "Test name",
+    "gpu_type": "GPU",
+    # "num_requests": "# of req.",
+    # "total_num_tokens": "Total # of tokens",
+    # "elapsed_time": "Elapsed time (s)",
+    "requests_per_second": "Tput (req/s)",
+    # "tokens_per_second": "Tput (tok/s)",
+}
+
+# serving results and the keys that will be printed into markdown
+serving_results = []
+serving_column_mapping = {
+    "test_name": "Test name",
+    "gpu_type": "GPU",
+    # "completed": "# of req.",
+    "request_throughput": "Tput (req/s)",
+    # "input_throughput": "Input Tput (tok/s)",
+    # "output_throughput": "Output Tput (tok/s)",
+    "mean_ttft_ms": "Mean TTFT (ms)",
+    "median_ttft_ms": "Median TTFT (ms)",
+    "p99_ttft_ms": "P99 TTFT (ms)",
+    # "mean_tpot_ms": "Mean TPOT (ms)",
+    # "median_tpot_ms": "Median",
+    # "p99_tpot_ms": "P99",
+    "mean_itl_ms": "Mean ITL (ms)",
+    "median_itl_ms": "Median ITL (ms)",
+    "p99_itl_ms": "P99 ITL (ms)",
+}
+
+
+def read_markdown(file):
+    if os.path.exists(file):
+        with open(file, "r") as f:
+            return f.read() + "\n"
+    else:
+        return f"{file} not found.\n"
+
+
+def results_to_json(latency, throughput, serving):
+    return json.dumps({
+        'latency': latency.to_dict(),
+        'throughput': throughput.to_dict(),
+        'serving': serving.to_dict()
+    })
+
+
+if __name__ == "__main__":
+
+    # collect results
+    for test_file in results_folder.glob("*.json"):
+
+        with open(test_file, "r") as f:
+            raw_result = json.loads(f.read())
+
+        if "serving" in str(test_file):
+            # this result is generated via `benchmark_serving.py`
+
+            # attach the benchmarking command to raw_result
+            with open(test_file.with_suffix(".commands"), "r") as f:
+                command = json.loads(f.read())
+            raw_result.update(command)
+
+            # update the test name of this result
+            raw_result.update({"test_name": test_file.stem})
+
+            # add the result to raw_result
+            serving_results.append(raw_result)
+            continue
+
+        elif "latency" in f.name:
+            # this result is generated via `benchmark_latency.py`
+
+            # attach the benchmarking command to raw_result
+            with open(test_file.with_suffix(".commands"), "r") as f:
+                command = json.loads(f.read())
+            raw_result.update(command)
+
+            # update the test name of this result
+            raw_result.update({"test_name": test_file.stem})
+
+            # get different percentiles
+            for perc in [10, 25, 50, 75, 90, 99]:
+                # Multiply 1000 to convert the time unit from s to ms
+                raw_result.update(
+                    {f"P{perc}": 1000 * raw_result["percentiles"][str(perc)]})
+            raw_result["avg_latency"] = raw_result["avg_latency"] * 1000
+
+            # add the result to raw_result
+            latency_results.append(raw_result)
+            continue
+
+        elif "throughput" in f.name:
+            # this result is generated via `benchmark_throughput.py`
+
+            # attach the benchmarking command to raw_result
+            with open(test_file.with_suffix(".commands"), "r") as f:
+                command = json.loads(f.read())
+            raw_result.update(command)
+
+            # update the test name of this result
+            raw_result.update({"test_name": test_file.stem})
+
+            # add the result to raw_result
+            throughput_results.append(raw_result)
+            continue
+
+        print(f"Skipping {test_file}")
+
+    latency_results = pd.DataFrame.from_dict(latency_results)
+    serving_results = pd.DataFrame.from_dict(serving_results)
+    throughput_results = pd.DataFrame.from_dict(throughput_results)
+
+    raw_results_json = results_to_json(latency_results, throughput_results,
+                                       serving_results)
+
+    # remapping the key, for visualization purpose
+    if not latency_results.empty:
+        latency_results = latency_results[list(
+            latency_column_mapping.keys())].rename(
+                columns=latency_column_mapping)
+    if not serving_results.empty:
+        serving_results = serving_results[list(
+            serving_column_mapping.keys())].rename(
+                columns=serving_column_mapping)
+    if not throughput_results.empty:
+        throughput_results = throughput_results[list(
+            throughput_results_column_mapping.keys())].rename(
+                columns=throughput_results_column_mapping)
+
+    processed_results_json = results_to_json(latency_results,
+                                             throughput_results,
+                                             serving_results)
+
+    # get markdown tables
+    latency_md_table = tabulate(latency_results,
+                                headers='keys',
+                                tablefmt='pipe',
+                                showindex=False)
+    serving_md_table = tabulate(serving_results,
+                                headers='keys',
+                                tablefmt='pipe',
+                                showindex=False)
+    throughput_md_table = tabulate(throughput_results,
+                                   headers='keys',
+                                   tablefmt='pipe',
+                                   showindex=False)
+
+    # document the result
+    with open(results_folder / "benchmark_results.md", "w") as f:
+
+        results = read_markdown(
+            "../.buildkite/nightly-benchmarks/tests/descriptions.md")
+        results = results.format(
+            latency_tests_markdown_table=latency_md_table,
+            throughput_tests_markdown_table=throughput_md_table,
+            serving_tests_markdown_table=serving_md_table,
+            benchmarking_results_in_json_string=processed_results_json)
+        f.write(results)
+
+    # document benchmarking results in json
+    with open(results_folder / "benchmark_results.json", "w") as f:
+
+        results = latency_results.to_dict(
+            orient='records') + throughput_results.to_dict(
+                orient='records') + serving_results.to_dict(orient='records')
+        f.write(json.dumps(results))

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

@@ -0,0 +1,17 @@
+#!/bin/sh
+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"
+
+retries=0
+while [ $retries -lt 1000 ]; do
+    if [ $(curl -s -L -H "Authorization: Bearer $TOKEN" -o /dev/null -w "%{http_code}" $URL) -eq 200 ]; then
+        exit 0
+    fi
+
+    echo "Waiting for image to be available..."
+
+    retries=$((retries + 1))
+    sleep 5
+done
+
+exit 1

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

@@ -0,0 +1,67 @@
+
+## 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.
+- 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.
+
+- 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.
+- Evaluation metrics: throughput.
+
+### Throughput benchmarking results
+
+{throughput_tests_markdown_table}
+
+## Serving tests
+
+This test suite aims to test vllm's real serving metrics.
+
+- 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.
+- 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. 
+You can load the benchmarking tables into pandas dataframes as follows:
+
+```python
+import json
+import pandas as pd
+
+benchmarking_results_json = """The json string"""
+benchmarking_results = json.loads(benchmarking_results_json)
+latency_results = pd.DataFrame.from_dict(benchmarking_results["latency"])
+throughput_results = pd.DataFrame.from_dict(benchmarking_results["throughput"])
+serving_results = pd.DataFrame.from_dict(benchmarking_results["serving"])
+```
+
+The json string for all benchmarking tables:
+```json
+{benchmarking_results_in_json_string}
+```
+
+You can also check the raw experiment data in the Artifact tab of the Buildkite page.
+

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

@@ -0,0 +1,32 @@
+[
+    {
+        "test_name": "latency_llama8B_tp1",
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    },
+    {
+        "test_name": "latency_llama70B_tp4",
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "tensor_parallel_size": 4,
+            "load_format": "dummy",
+            "num-iters-warmup": 5,
+            "num-iters": 15
+        }
+    },
+    {
+        "test_name": "latency_mixtral8x7B_tp2",
+        "parameters": {
+            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "tensor_parallel_size": 2,
+            "load_format": "dummy",
+            "num-iters-warmup": 5,
+            "num-iters": 15
+        }
+    }
+]

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

@@ -0,0 +1,59 @@
+[
+    {
+        "test_name": "serving_llama8B_tp1_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "tensor_parallel_size": 1,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama70B_tp4_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "tensor_parallel_size": 4,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_mixtral8x7B_tp2_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_parameters": {
+            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "tensor_parallel_size": 2,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    }
+]

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

@@ -0,0 +1,35 @@
+[
+    {
+        "test_name": "throughput_llama8B_tp1",
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    },
+    {
+        "test_name": "throughput_llama70B_tp4",
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "tensor_parallel_size": 4,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    },
+    {
+        "test_name": "throughput_mixtral8x7B_tp2",
+        "parameters": {
+            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "tensor_parallel_size": 2,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    }
+]

.buildkite/release-pipeline.yaml

@@ -0,0 +1,21 @@
+steps:
+  - block: "Build wheels"
+
+  - label: "Build wheel - Python {{matrix.python_version}}, CUDA {{matrix.cuda_version}}" 
+    agents:
+      queue: cpu_queue
+    commands:
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION={{matrix.cuda_version}} --build-arg PYTHON_VERSION={{matrix.python_version}} --tag vllm-ci:build-image --target build --progress plain ."
+      - "mkdir artifacts"
+      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image cp -r dist /artifacts_host"
+      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
+    matrix:
+      setup:
+        cuda_version:
+          - "11.8.0"
+          - "12.1.0"
+        python_version:
+          - "3.8"
+          - "3.9"
+          - "3.10"
+          - "3.11"

.buildkite/run-cpu-test.sh

@@ -4,21 +4,25 @@ 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 .
 
 # Setup cleanup
-remove_docker_container() { docker rm -f cpu-test || true; }
+remove_docker_container() { docker rm -f cpu-test cpu-test-avx2 || true; }
 trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image
-docker run -itd -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
+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
+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
 
 # 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
-  bash ../.buildkite/download-images.sh
   cd ../
-  pytest -v -s tests/models --ignore=tests/models/test_llava.py --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py"
+  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

.buildkite/run-openvino-test.sh

@@ -0,0 +1,14 @@
+# This script build the OpenVINO 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 openvino-test -f Dockerfile.openvino .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f openvino-test || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image and launch offline inference
+docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/vllm/examples/offline_inference.py

.buildkite/run-xpu-test.sh

@@ -0,0 +1,14 @@
+# 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 xpu-test -f Dockerfile.xpu .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f xpu-test || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image and launch offline inference
+docker run --network host --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path xpu-test python3 examples/offline_inference.py

.buildkite/test-pipeline.yaml

@@ -1,7 +1,10 @@
 # In this file, you can add more tests to run either by adding a new step or
 # adding a new command to an existing step. See different options here for examples.
-# This script will be feed into Jinja template in `test-template.j2` to generate
-# the final pipeline yaml file.
+
+# This script will be feed into Jinja template in `test-template-aws.j2` at
+# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2 
+# to generate the final pipeline yaml file.
+
 
 steps:
 - label: Regression Test
@@ -24,38 +27,63 @@ steps:
 
 - label: Core Test
   mirror_hardwares: [amd]
-  command: pytest -v -s core
+  commands: 
+  - pytest -v -s core
+  - pytest -v -s distributed/test_parallel_state.py
 
 - label: Distributed Comm Ops Test
   #mirror_hardwares: [amd]
-  command: pytest -v -s distributed/test_comm_ops.py
   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: Distributed Tests
+- label: Distributed Tests (2 GPUs)
   mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   commands:
+  - bash ../.buildkite/download-images.sh
   - 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 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
-  - pytest -v -s spec_decode/e2e/test_integration_dist.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 (Multiple Groups)
+- label: Distributed Tests (4 GPUs)
   #mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   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=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:
+  - TP_SIZE=2 PP_SIZE=2 EAGER_MODE=1 CHUNKED_PREFILL=1 pytest -v -s distributed/test_pipeline_parallel.py
+  - TP_SIZE=2 PP_SIZE=2 EAGER_MODE=1 CHUNKED_PREFILL=0 pytest -v -s distributed/test_pipeline_parallel.py
+  - TP_SIZE=1 PP_SIZE=3 EAGER_MODE=1 CHUNKED_PREFILL=0 pytest -v -s distributed/test_pipeline_parallel.py
+  - PP_SIZE=4 EAGER_MODE=1 CHUNKED_PREFILL=1 pytest -v -s distributed/test_pipeline_parallel.py
+  - PP_SIZE=4 EAGER_MODE=1 CHUNKED_PREFILL=0 pytest -v -s distributed/test_pipeline_parallel.py
+
 
 - label: Engine Test
   mirror_hardwares: [amd]
@@ -65,8 +93,8 @@ steps:
   mirror_hardwares: [amd]
 
   commands:
-  - pytest -v -s entrypoints -m llm
-  - pytest -v -s entrypoints -m openai
+  - pytest -v -s entrypoints/llm
+  - pytest -v -s entrypoints/openai
 
 - label: Examples Test
   working_dir: "/vllm-workspace/examples"
@@ -90,19 +118,22 @@ steps:
 
 - label: Kernels Test %N
   #mirror_hardwares: [amd]
-  command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  commands:
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+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:
-    - pytest -v -s models -m \"not llava\"
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+    - pytest -v -s models -m \"not vlm\"
 
-- label: Llava Test
+- label: Vision Language Models Test
   mirror_hardwares: [amd]
   commands:
     - bash ../.buildkite/download-images.sh
-    - pytest -v -s models -m llava
+    - pytest -v -s models -m vlm
 
 - label: Prefix Caching Test
   mirror_hardwares: [amd]
@@ -141,6 +172,9 @@ steps:
   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
@@ -155,6 +189,15 @@ steps:
   #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]
@@ -162,9 +205,39 @@ steps:
   - 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
   working_dir: "/vllm-workspace/test_docs/docs"
   no_gpu: True
   commands:
   - pip install -r requirements-docs.txt
   - SPHINXOPTS=\"-W\" make html
+
+- label: Distributed Tests (A100)
+  gpu: a100
+  num_gpus: 4
+  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/v0.0.7/flashinfer-0.0.7+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
+  - pytest -v -s -x lora/test_mixtral.py

.buildkite/test-template-aws.j2

@@ -1,92 +0,0 @@
-{% set docker_image = "public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT" %}
-{% set default_working_dir = "/vllm-workspace/tests" %}
-
-steps:
-  - label: ":docker: build image"
-    agents:
-      queue: cpu_queue
-    commands:
-      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --tag {{ docker_image }} --target test --progress plain ."
-      - "docker push {{ docker_image }}"
-    env:
-      DOCKER_BUILDKIT: "1"
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 5
-        - exit_status: -10  # Agent was lost
-          limit: 5
-  - wait
-
-  - group: "AMD Tests"
-    depends_on: ~
-    steps:
-    {% for step in steps %}
-    {% if step.mirror_hardwares and "amd" in step.mirror_hardwares %}
-      - label: "AMD: {{ step.label }}"
-        agents:
-          queue: amd
-        command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
-        env:
-          DOCKER_BUILDKIT: "1"
-        soft_fail: true
-    {% endif %}
-    {% endfor %}
-
-  - label: "Neuron Test"
-    depends_on: ~
-    agents:
-      queue: neuron
-    command: bash .buildkite/run-neuron-test.sh
-    soft_fail: false
-
-  - label: "Intel Test"
-    depends_on: ~
-    agents:
-      queue: intel
-    command: bash .buildkite/run-cpu-test.sh
-
-  {% for step in steps %}
-  - label: "{{ step.label }}"
-    agents:
-      {% if step.label == "Documentation Build" %}
-      queue: small_cpu_queue
-      {% elif step.no_gpu %}
-      queue: cpu_queue
-      {% elif step.num_gpus == 2 or step.num_gpus == 4 %}
-      queue: gpu_4_queue
-      {% else %}
-      queue: gpu_1_queue
-      {% endif %}
-    soft_fail: {{ step.soft_fail or false }}
-    {% if step.parallelism %}
-    parallelism: {{ step.parallelism }}
-    {% endif %}
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 5
-        - exit_status: -10  # Agent was lost
-          limit: 5
-    plugins:
-      - docker#v5.2.0:
-          image: {{ docker_image }}
-          always-pull: true
-          propagate-environment: true
-          {% if not step.no_gpu %}
-          gpus: all
-          {% endif %}
-          {% if step.label == "Benchmarks" %}
-          mount-buildkite-agent: true
-          {% endif %}
-          command: ["bash", "-c", "cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}"]
-          environment:
-            - VLLM_USAGE_SOURCE=ci-test
-            - HF_TOKEN
-            {% if step.label == "Speculative decoding tests" %}
-            - VLLM_ATTENTION_BACKEND=XFORMERS
-            {% endif %}
-          volumes:
-            - /dev/shm:/dev/shm
-  {% endfor %}

.buildkite/test-template.j2

@@ -1,96 +0,0 @@
-{% set docker_image = "us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:$BUILDKITE_COMMIT" %}
-{% set default_num_gpu = 1 %}
-{% set default_working_dir = "/vllm-workspace/tests" %}
-
-steps:
-  - label: ":docker: build image"
-    commands:
-      - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
-      - "docker push {{ docker_image }}"
-    env:
-      DOCKER_BUILDKIT: "1"
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 5
-        - exit_status: -10  # Agent was lost
-          limit: 5
-  - wait
-
-  - group: "AMD Tests"
-    depends_on: ~
-    steps:
-    {% for step in steps %}
-    {% if step.mirror_hardwares and "amd" in step.mirror_hardwares %}
-      - label: "AMD: {{ step.label }}"
-        agents:
-          queue: amd
-        command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
-        env:
-          DOCKER_BUILDKIT: "1"
-        soft_fail: true
-    {% endif %}
-    {% endfor %}
-
-  - label: "Neuron Test"
-    depends_on: ~
-    agents:
-      queue: neuron
-    command: bash .buildkite/run-neuron-test.sh
-    soft_fail: false
-
-  - label: "Intel Test"
-    depends_on: ~
-    agents:
-      queue: intel
-    command: bash .buildkite/run-cpu-test.sh
-
-  {% for step in steps %}
-  - label: "{{ step.label }}"
-    agents:
-      queue: kubernetes
-    soft_fail: {{ step.soft_fail or false }}
-    {% if step.parallelism %}
-    parallelism: {{ step.parallelism }}
-    {% endif %}
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 5
-        - exit_status: -10  # Agent was lost
-          limit: 5
-    plugins:
-      - kubernetes:
-          podSpec:
-            {% if step.num_gpus %}
-            priorityClassName: gpu-priority-cls-{{ step.num_gpus }}
-            {% endif %}
-            volumes:
-              - name: dshm
-                emptyDir:
-                  medium: Memory
-            containers:
-              - image: "{{ docker_image }}"
-                command: ["bash"]
-                args:
-                - '-c'
-                - "'cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}'"
-                {% if not step.no_gpu %}
-                resources:
-                  requests:
-                    nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
-                  limits:
-                    nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
-                {% endif %}
-                env:
-                  - name: VLLM_USAGE_SOURCE
-                    value: ci-test
-                  - name: HF_TOKEN
-                    valueFrom:
-                      secretKeyRef:
-                        name: hf-token-secret
-                        key: token
-                volumeMounts:
-                  - mountPath: /dev/shm
-                    name: dshm
-  {% endfor %}

.github/workflows/mypy.yaml

@@ -47,5 +47,5 @@ jobs:
         mypy vllm/model_executor  --config-file pyproject.toml
         mypy vllm/lora --config-file pyproject.toml
         mypy vllm/logging --config-file pyproject.toml
-        mypy vllm/model_executor --config-file pyproject.toml
+        mypy tests --config-file pyproject.toml
 

CMakeLists.txt

@@ -2,7 +2,8 @@ cmake_minimum_required(VERSION 3.21)
 
 project(vllm_extensions LANGUAGES CXX)
 
-option(VLLM_TARGET_DEVICE "Target device backend for vLLM" "cuda")
+# CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
+set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
 
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
 message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
@@ -32,8 +33,7 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx11
 # versions are derived from Dockerfile.rocm
 #
 set(TORCH_SUPPORTED_VERSION_CUDA "2.3.0")
-set(TORCH_SUPPORTED_VERSION_ROCM_5X "2.0.1")
-set(TORCH_SUPPORTED_VERSION_ROCM_6X "2.1.1")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.4.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -98,18 +98,11 @@ elseif(HIP_FOUND)
   # .hip extension automatically, HIP must be enabled explicitly.
   enable_language(HIP)
 
-  # ROCm 5.x
-  if (ROCM_VERSION_DEV_MAJOR EQUAL 5 AND
-      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_5X})
-    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_5X} "
-      "expected for ROCMm 5.x build, saw ${Torch_VERSION} instead.")
-  endif()
-
-  # ROCm 6.x
-  if (ROCM_VERSION_DEV_MAJOR EQUAL 6 AND
-      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_6X})
-    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_6X} "
-      "expected for ROCMm 6.x build, saw ${Torch_VERSION} instead.")
+  # ROCm 5.X and 6.X
+  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
+      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
+    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM} "
+      "expected for ROCm build, saw ${Torch_VERSION} instead.")
   endif()
 else()
   message(FATAL_ERROR "Can't find CUDA or HIP installation.")
@@ -178,6 +171,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
+    "csrc/quantization/fp8/fp8_marlin.cu"
     "csrc/custom_all_reduce.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu"

Dockerfile

@@ -5,9 +5,26 @@
 # docs/source/dev/dockerfile/dockerfile.rst and
 # docs/source/assets/dev/dockerfile-stages-dependency.png
 
+ARG CUDA_VERSION=12.4.1
 #################### BASE BUILD IMAGE ####################
 # prepare basic build environment
-FROM nvidia/cuda:12.4.1-devel-ubuntu22.04 AS dev
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS base
+
+ARG CUDA_VERSION=12.4.1
+ARG PYTHON_VERSION=3
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
+    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
+    && apt-get update -y \
+    && apt-get install -y ccache software-properties-common \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv python3-pip \
+    && if [ "${PYTHON_VERSION}" != "3" ]; then update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; fi \
+    && python3 --version \
+    && python3 -m pip --version
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip git curl sudo
@@ -16,7 +33,7 @@ RUN apt-get update -y \
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
 # this won't be needed for future versions of this docker image
 # or future versions of triton.
-RUN ldconfig /usr/local/cuda-12.4/compat/
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
 WORKDIR /workspace
 
@@ -24,14 +41,11 @@ WORKDIR /workspace
 COPY requirements-common.txt requirements-common.txt
 COPY requirements-cuda.txt requirements-cuda.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements-cuda.txt
+    python3 -m pip install -r requirements-cuda.txt
 
-# install development dependencies
-COPY requirements-lint.txt requirements-lint.txt
-COPY requirements-test.txt requirements-test.txt
-COPY requirements-dev.txt requirements-dev.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements-dev.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`
@@ -41,14 +55,16 @@ 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}
 #################### BASE BUILD IMAGE ####################
 
-
 #################### WHEEL BUILD IMAGE ####################
-FROM dev AS build
+FROM base AS build
+
+ARG PYTHON_VERSION=3
 
 # install build dependencies
 COPY requirements-build.txt requirements-build.txt
+
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements-build.txt
+    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
@@ -101,9 +117,36 @@ RUN python3 check-wheel-size.py dist
 
 #################### EXTENSION Build IMAGE ####################
 
+#################### DEV IMAGE ####################
+FROM base as dev
+
+COPY requirements-lint.txt requirements-lint.txt
+COPY requirements-test.txt requirements-test.txt
+COPY requirements-dev.txt requirements-dev.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install -r requirements-dev.txt
+
+#################### DEV IMAGE ####################
+#################### MAMBA Build IMAGE ####################
+FROM dev as mamba-builder
+# max jobs used for build
+ARG max_jobs=2
+ENV MAX_JOBS=${max_jobs}
+
+WORKDIR /usr/src/mamba
+
+COPY requirements-mamba.txt requirements-mamba.txt
+
+# Download the wheel or build it if a pre-compiled release doesn't exist
+RUN pip --verbose wheel -r requirements-mamba.txt \
+    --no-build-isolation --no-deps --no-cache-dir
+
+#################### MAMBA Build IMAGE ####################
+
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
-FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS vllm-base
+FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
+ARG CUDA_VERSION=12.4.1
 WORKDIR /vllm-workspace
 
 RUN apt-get update -y \
@@ -113,12 +156,16 @@ RUN apt-get update -y \
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
 # this won't be needed for future versions of this docker image
 # or future versions of triton.
-RUN ldconfig /usr/local/cuda-12.4/compat/
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
 # install vllm wheel first, so that torch etc will be installed
 RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
     --mount=type=cache,target=/root/.cache/pip \
-    pip install dist/*.whl --verbose
+    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
 #################### vLLM installation IMAGE ####################
 
 
@@ -131,7 +178,7 @@ ADD . /vllm-workspace/
 
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements-dev.txt
+    python3 -m pip install -r requirements-dev.txt
 
 # doc requires source code
 # we hide them inside `test_docs/` , so that this source code
@@ -148,7 +195,7 @@ 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
+    pip install accelerate hf_transfer 'modelscope!=1.15.0'
 
 ENV VLLM_USAGE_SOURCE production-docker-image
 

Dockerfile.cpu

@@ -6,7 +6,13 @@ RUN apt-get update  -y \
     && apt-get install -y git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
 
-RUN echo 'export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD' >> ~/.bashrc
+# https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
+# intel-openmp provides additional performance improvement vs. openmp
+# tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
+RUN pip install intel-openmp
+
+ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so:$LD_PRELOAD"
+
 
 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
 
@@ -21,10 +27,14 @@ WORKDIR /workspace/vllm
 
 RUN pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
 
+# 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
 
 WORKDIR /workspace/
 
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
 
-CMD ["/bin/bash"]
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.openvino

@@ -0,0 +1,26 @@
+# The vLLM Dockerfile is used to construct vLLM image that can be directly used
+# to run the OpenAI compatible server.
+
+FROM ubuntu:22.04 AS dev
+
+RUN apt-get update -y && \
+    apt-get install -y python3-pip git
+WORKDIR /workspace
+
+# copy requirements
+COPY requirements-build.txt /workspace/vllm/
+COPY requirements-common.txt /workspace/vllm/
+COPY requirements-openvino.txt /workspace/vllm/
+
+COPY vllm/ /workspace/vllm/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/
+
+COPY examples/ /workspace/vllm/examples
+COPY benchmarks/ /workspace/vllm/benchmarks
+
+CMD ["/bin/bash"]

Dockerfile.ppc64le

@@ -0,0 +1,22 @@
+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
+
+# 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
+
+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 VLLM_TARGET_DEVICE=cpu python3 setup.py install
+
+WORKDIR /vllm-workspace
+ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.rocm

@@ -1,35 +1,35 @@
-# default base image
-ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+# Default ROCm 6.1 base image
+ARG BASE_IMAGE="rocm/pytorch:rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging"
 
-FROM $BASE_IMAGE
+# Tested and supported base rocm/pytorch images
+ARG ROCm_5_7_BASE="rocm/pytorch:rocm5.7_ubuntu20.04_py3.9_pytorch_2.0.1" \
+    ROCm_6_0_BASE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" \
+    ROCM_6_1_BASE="rocm/pytorch:rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging"
 
-ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+# Default ROCm ARCHes to build vLLM for.
+ARG PYTORCH_ROCM_ARCH="gfx908;gfx90a;gfx942;gfx1100"
 
-RUN echo "Base image is $BASE_IMAGE"
-
-# BASE_IMAGE for ROCm_5.7: "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1"
-# BASE_IMAGE for ROCm_6.0: "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
-
-
-ARG FA_GFX_ARCHS="gfx90a;gfx942"
-RUN echo "FA_GFX_ARCHS is $FA_GFX_ARCHS"
-
-ARG FA_BRANCH="ae7928c"
-RUN echo "FA_BRANCH is $FA_BRANCH"
-
-# whether to build flash-attention
-# if 0, will not build flash attention
-# this is useful for gfx target where flash-attention is not supported
-# In that case, we need to use the python reference attention implementation in vllm
+# Whether to build CK-based flash-attention
+# If 0, will not build flash attention
+# This is useful for gfx target where flash-attention is not supported
+# (i.e. those that do not appear in `FA_GFX_ARCHS`)
+# Triton FA is used by default on ROCm now so this is unnecessary.
 ARG BUILD_FA="1"
+ARG FA_GFX_ARCHS="gfx90a;gfx942"
+ARG FA_BRANCH="ae7928c"
 
-# whether to build triton on rocm
+# Whether to build triton on rocm
 ARG BUILD_TRITON="1"
+ARG TRITON_BRANCH="0ef1848"
+
+### Base image build stage
+FROM $BASE_IMAGE AS base
+
+# Import arg(s) defined before this build stage
+ARG PYTORCH_ROCM_ARCH
 
 # Install some basic utilities
 RUN apt-get update && apt-get install python3 python3-pip -y
-
-# Install some basic utilities
 RUN apt-get update && apt-get install -y \
     curl \
     ca-certificates \
@@ -40,76 +40,165 @@ RUN apt-get update && apt-get install -y \
     build-essential \
     wget \
     unzip \
-    nvidia-cuda-toolkit \
     tmux \
+    ccache \
  && rm -rf /var/lib/apt/lists/*
 
-### Mount Point ###
-# When launching the container, mount the code directory to /app
+# When launching the container, mount the code directory to /vllm-workspace
 ARG APP_MOUNT=/vllm-workspace
-VOLUME [ ${APP_MOUNT} ]
 WORKDIR ${APP_MOUNT}
 
-RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
+RUN 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; pip uninstall -y sccache; rm -f "$(which sccache)"
+# Install torch == 2.4.0 on ROCm
+RUN case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+        *"rocm-5.7"*) \
+            pip uninstall -y torch torchaudio torchvision \
+            && pip install --no-cache-dir --pre \
+                torch==2.4.0.dev20240612 torchaudio==2.4.0.dev20240612 \
+                torchvision==0.19.0.dev20240612 \
+               --index-url https://download.pytorch.org/whl/nightly/rocm5.7;; \
+        *"rocm-6.0"*) \
+            pip uninstall -y torch torchaudio torchvision \
+            && pip install --no-cache-dir --pre \
+                torch==2.4.0.dev20240612 torchaudio==2.4.0.dev20240612 \
+                torchvision==0.19.0.dev20240612 \
+               --index-url https://download.pytorch.org/whl/nightly/rocm6.0;; \
+        *"rocm-6.1"*) \
+            pip uninstall -y torch torchaudio torchvision \
+            && pip install --no-cache-dir --pre \
+                torch==2.4.0.dev20240612 torchaudio==2.4.0.dev20240612 \
+                torchvision==0.19.0.dev20240612 \
+               --index-url https://download.pytorch.org/whl/nightly/rocm6.1;; \
+        *) ;; esac
 
 ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer
 ENV PATH=$PATH:/opt/rocm/bin:/libtorch/bin:
 ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/lib/:/libtorch/lib:
 ENV CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/libtorch/include:/libtorch/include/torch/csrc/api/include/:/opt/rocm/include/:
 
-# Install ROCm flash-attention
-RUN if [ "$BUILD_FA" = "1" ]; then \
-    mkdir libs \
+ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
+ENV CCACHE_DIR=/root/.cache/ccache
+
+
+### AMD-SMI build stage
+FROM base AS build_amdsmi
+# Build amdsmi wheel always
+RUN cd /opt/rocm/share/amd_smi \
+    && pip wheel . --wheel-dir=/install
+
+
+### Flash-Attention wheel build stage
+FROM base AS build_fa
+ARG BUILD_FA
+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 \
+    mkdir -p libs \
     && cd libs \
     && git clone https://github.com/ROCm/flash-attention.git \
     && cd flash-attention \
-    && git checkout ${FA_BRANCH} \
+    && git checkout "${FA_BRANCH}" \
     && git submodule update --init \
-    && export GPU_ARCHS=${FA_GFX_ARCHS} \
-    && if [ "$BASE_IMAGE" = "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1" ]; then \
-        patch /opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/hipify/hipify_python.py hipify_patch.patch; fi \
-    && python3 setup.py install \
-    && cd ..; \
+    && case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+        *"rocm-5.7"*) \
+            export VLLM_TORCH_PATH="$(python3 -c 'import torch; print(torch.__path__[0])')" \
+            && patch "${VLLM_TORCH_PATH}"/utils/hipify/hipify_python.py hipify_patch.patch;; \
+        *) ;; esac \
+    && GPU_ARCHS="${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
 
-# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
-# Manually removed it so that later steps of numpy upgrade can continue
-RUN if [ "$BASE_IMAGE" = "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" ]; then \
-    rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/; fi
 
-# build triton
-RUN if [ "$BUILD_TRITON" = "1" ]; then \
+### Triton wheel build stage
+FROM base AS build_triton
+ARG BUILD_TRITON
+ARG TRITON_BRANCH
+# Build triton wheel if `BUILD_TRITON = 1`
+RUN --mount=type=cache,target=${CCACHE_DIR} \
+    if [ "$BUILD_TRITON" = "1" ]; then \
     mkdir -p libs \
     && cd libs \
-    && pip uninstall -y triton \
-    && git clone https://github.com/ROCm/triton.git \
-    && cd triton/python \
-    && pip3 install . \
-    && cd ../..; \
+    && git clone https://github.com/OpenAI/triton.git \
+    && cd triton \
+    && git checkout "${TRITON_BRANCH}" \
+    && cd python \
+    && python3 setup.py bdist_wheel --dist-dir=/install; \
+    # Create an empty directory otherwise as later build stages expect one
+    else mkdir -p /install; \
     fi
 
-WORKDIR /vllm-workspace
+
+### Final vLLM build stage
+FROM base AS final
+# Import the vLLM development directory from the build context
 COPY . .
 
-#RUN python3 -m pip install pynvml # to be removed eventually
-RUN python3 -m pip install --upgrade pip numba
+# 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
 
-# make sure punica kernels are built (for LoRA)
+# Package upgrades for useful functionality or to avoid dependency issues
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install --upgrade numba scipy huggingface-hub[cli]
+
+# 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
 
-ENV VLLM_NCCL_SO_PATH=/opt/rocm/lib/librccl.so
-
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=${CCACHE_DIR} \
+    --mount=type=cache,target=/root/.cache/pip \
     pip install -U -r requirements-rocm.txt \
-    && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
-    && python3 setup.py install \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.abi3.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.abi3.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.abi3.so vllm/ \
-    && cd ..
+    && case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+        *"rocm-6.0"*) \
+            patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h rocm_patch/rocm_bf16.patch;; \
+        *"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 rocm_patch \
+            && cp rocm_patch/libamdhip64.so.6 /opt/rocm/lib/libamdhip64.so.6 \
+            # 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
 
+# Copy amdsmi wheel into final image
+RUN --mount=type=bind,from=build_amdsmi,src=/install,target=/install \
+    mkdir -p libs \
+    && cp /install/*.whl libs \
+    # Preemptively uninstall to avoid same-version no-installs
+    && pip uninstall -y amdsmi;
+
+# Copy triton wheel(s) into final image if they were built
+RUN --mount=type=bind,from=build_triton,src=/install,target=/install \
+    mkdir -p libs \
+    && if ls /install/*.whl; then \
+        cp /install/*.whl libs \
+        # Preemptively uninstall to avoid same-version no-installs
+        && pip uninstall -y triton; fi
+
+# Copy flash-attn wheel(s) into final image if they were built
+RUN --mount=type=bind,from=build_fa,src=/install,target=/install \
+    mkdir -p libs \
+    && if ls /install/*.whl; then \
+        cp /install/*.whl libs \
+        # Preemptively uninstall to avoid same-version no-installs
+        && pip uninstall -y flash-attn; fi
+
+# Install wheels that were built to the final image
+RUN --mount=type=cache,target=/root/.cache/pip \
+    if ls libs/*.whl; then \
+    pip install libs/*.whl; fi
 
 CMD ["/bin/bash"]

Dockerfile.xpu

@@ -0,0 +1,22 @@
+FROM intel/oneapi-basekit:2024.1.0-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
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+RUN pip install -v -r requirements-xpu.txt
+
+RUN VLLM_TARGET_DEVICE=xpu python3 setup.py install
+
+CMD ["/bin/bash"]

README.md

@@ -59,7 +59,7 @@ vLLM is flexible and easy to use with:
 - Tensor parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs, AMD GPUs, and Intel CPUs
+- Support NVIDIA GPUs, AMD GPUs, Intel CPUs and GPUs
 - (Experimental) Prefix caching support
 - (Experimental) Multi-lora support
 
@@ -112,6 +112,7 @@ vLLM is a community project. Our compute resources for development and testing a
 - Trainy
 - UC Berkeley
 - UC San Diego
+- ZhenFund
 
 We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM.
 

benchmarks/backend_request_func.py

@@ -4,10 +4,13 @@ import sys
 import time
 import traceback
 from dataclasses import dataclass, field
-from typing import List, Optional
+from typing import List, Optional, Union
 
 import aiohttp
+import huggingface_hub.constants
 from tqdm.asyncio import tqdm
+from transformers import (AutoTokenizer, PreTrainedTokenizer,
+                          PreTrainedTokenizerFast)
 
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 
@@ -222,8 +225,8 @@ async def async_request_openai_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "v1/completions"
-    ), "OpenAI Completions API URL must end with 'v1/completions'."
+        "completions"
+    ), "OpenAI Completions API URL must end with 'completions'."
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search
@@ -262,6 +265,9 @@ async def async_request_openai_completions(
                         else:
                             data = json.loads(chunk)
 
+                            # NOTE: Some completion API might have a last
+                            # usage summary response without a token so we
+                            # want to check a token was generated
                             if data["choices"][0]["text"]:
                                 timestamp = time.perf_counter()
                                 # First token
@@ -270,12 +276,8 @@ async def async_request_openai_completions(
                                     output.ttft = ttft
 
                                 # Decoding phase
-                                # NOTE: Some completion API might have a last
-                                # usage summary response without a token so we
-                                # do not want to include as inter-token-latency
-                                elif data.get("usage", None) is None:
-                                    output.itl.append(timestamp -
-                                                      most_recent_timestamp)
+                                output.itl.append(timestamp -
+                                                  most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
                                 generated_text += data["choices"][0]["text"]
@@ -302,8 +304,8 @@ async def async_request_openai_chat_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "v1/chat/completions"
-    ), "OpenAI Chat Completions API URL must end with 'v1/chat/completions'."
+        "chat/completions"
+    ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search
@@ -388,6 +390,30 @@ def remove_prefix(text: str, prefix: str) -> str:
     return text
 
 
+def get_model(pretrained_model_name_or_path: str):
+    if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
+        from modelscope import snapshot_download
+    else:
+        from huggingface_hub import snapshot_download
+
+    model_path = snapshot_download(
+        model_id=pretrained_model_name_or_path,
+        local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+        ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"])
+    return model_path
+
+
+def get_tokenizer(
+    pretrained_model_name_or_path: str, trust_remote_code: bool
+) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
+    if pretrained_model_name_or_path is not None and not os.path.exists(
+            pretrained_model_name_or_path):
+        pretrained_model_name_or_path = get_model(
+            pretrained_model_name_or_path)
+    return AutoTokenizer.from_pretrained(pretrained_model_name_or_path,
+                                         trust_remote_code=trust_remote_code)
+
+
 ASYNC_REQUEST_FUNCS = {
     "tgi": async_request_tgi,
     "vllm": async_request_openai_completions,
@@ -396,4 +422,5 @@ ASYNC_REQUEST_FUNCS = {
     "openai": async_request_openai_completions,
     "openai-chat": async_request_openai_chat_completions,
     "tensorrt-llm": async_request_trt_llm,
+    "scalellm": async_request_openai_completions,
 }

benchmarks/benchmark_latency.py

@@ -10,8 +10,10 @@ import torch
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
+from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptStrictInputs
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.utils import FlexibleArgumentParser
 
 
 def main(args: argparse.Namespace):
@@ -19,25 +21,33 @@ def main(args: argparse.Namespace):
 
     # NOTE(woosuk): If the request cannot be processed in a single batch,
     # the engine will automatically process the request in multiple batches.
-    llm = LLM(model=args.model,
-              speculative_model=args.speculative_model,
-              num_speculative_tokens=args.num_speculative_tokens,
-              tokenizer=args.tokenizer,
-              quantization=args.quantization,
-              tensor_parallel_size=args.tensor_parallel_size,
-              trust_remote_code=args.trust_remote_code,
-              dtype=args.dtype,
-              enforce_eager=args.enforce_eager,
-              kv_cache_dtype=args.kv_cache_dtype,
-              quantization_param_path=args.quantization_param_path,
-              device=args.device,
-              ray_workers_use_nsight=args.ray_workers_use_nsight,
-              use_v2_block_manager=args.use_v2_block_manager,
-              enable_chunked_prefill=args.enable_chunked_prefill,
-              download_dir=args.download_dir,
-              block_size=args.block_size,
-              gpu_memory_utilization=args.gpu_memory_utilization,
-              distributed_executor_backend=args.distributed_executor_backend)
+    llm = LLM(
+        model=args.model,
+        speculative_model=args.speculative_model,
+        num_speculative_tokens=args.num_speculative_tokens,
+        speculative_draft_tensor_parallel_size=\
+            args.speculative_draft_tensor_parallel_size,
+        tokenizer=args.tokenizer,
+        quantization=args.quantization,
+        tensor_parallel_size=args.tensor_parallel_size,
+        trust_remote_code=args.trust_remote_code,
+        dtype=args.dtype,
+        max_model_len=args.max_model_len,
+        enforce_eager=args.enforce_eager,
+        kv_cache_dtype=args.kv_cache_dtype,
+        quantization_param_path=args.quantization_param_path,
+        device=args.device,
+        ray_workers_use_nsight=args.ray_workers_use_nsight,
+        use_v2_block_manager=args.use_v2_block_manager,
+        enable_chunked_prefill=args.enable_chunked_prefill,
+        download_dir=args.download_dir,
+        block_size=args.block_size,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        load_format=args.load_format,
+        distributed_executor_backend=args.distributed_executor_backend,
+        otlp_traces_endpoint=args.otlp_traces_endpoint,
+        enable_prefix_caching=args.enable_prefix_caching,
+    )
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -96,7 +106,7 @@ def main(args: argparse.Namespace):
     for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
         latencies.append(run_to_completion(profile_dir=None))
     latencies = np.array(latencies)
-    percentages = [10, 25, 50, 75, 90]
+    percentages = [10, 25, 50, 75, 90, 99]
     percentiles = np.percentile(latencies, percentages)
     print(f'Avg latency: {np.mean(latencies)} seconds')
     for percentage, percentile in zip(percentages, percentiles):
@@ -114,12 +124,16 @@ def main(args: argparse.Namespace):
 
 
 if __name__ == '__main__':
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description='Benchmark the latency of processing a single batch of '
         'requests till completion.')
     parser.add_argument('--model', type=str, default='facebook/opt-125m')
     parser.add_argument('--speculative-model', type=str, default=None)
     parser.add_argument('--num-speculative-tokens', type=int, default=None)
+    parser.add_argument('--speculative-draft-tensor-parallel-size',
+                        '-spec-draft-tp',
+                        type=int,
+                        default=None)
     parser.add_argument('--tokenizer', type=str, default=None)
     parser.add_argument('--quantization',
                         '-q',
@@ -145,6 +159,12 @@ if __name__ == '__main__':
     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,
@@ -188,9 +208,10 @@ if __name__ == '__main__':
     parser.add_argument(
         "--device",
         type=str,
-        default="cuda",
-        choices=["cuda", "cpu", "tpu"],
-        help='device type for vLLM execution, supporting CUDA and CPU.')
+        default="auto",
+        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
+        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
+        'CPU.')
     parser.add_argument('--block-size',
                         type=int,
                         default=16,
@@ -200,6 +221,9 @@ if __name__ == '__main__':
         action='store_true',
         help='If True, the prefill requests can be chunked based on the '
         'max_num_batched_tokens')
+    parser.add_argument("--enable-prefix-caching",
+                        action='store_true',
+                        help="Enable automatic prefix caching")
     parser.add_argument('--use-v2-block-manager', action='store_true')
     parser.add_argument(
         "--ray-workers-use-nsight",
@@ -222,6 +246,29 @@ if __name__ == '__main__':
                         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(
+        '--load-format',
+        type=str,
+        default=EngineArgs.load_format,
+        choices=[
+            'auto', 'pt', 'safetensors', 'npcache', 'dummy', 'tensorizer',
+            'bitsandbytes'
+        ],
+        help='The format of the model weights to load.\n\n'
+        '* "auto" will try to load the weights in the safetensors format '
+        'and fall back to the pytorch bin format if safetensors format '
+        'is not available.\n'
+        '* "pt" will load the weights in the pytorch bin format.\n'
+        '* "safetensors" will load the weights in the safetensors format.\n'
+        '* "npcache" will load the weights in pytorch format and store '
+        'a numpy cache to speed up the loading.\n'
+        '* "dummy" will initialize the weights with random values, '
+        'which is mainly for profiling.\n'
+        '* "tensorizer" will load the weights using tensorizer from '
+        'CoreWeave. See the Tensorize vLLM Model script in the Examples'
+        'section for more information.\n'
+        '* "bitsandbytes" will load the weights using bitsandbytes '
+        'quantization.\n')
     parser.add_argument(
         '--distributed-executor-backend',
         choices=['ray', 'mp'],
@@ -229,5 +276,10 @@ if __name__ == '__main__':
         help='Backend to use for distributed serving. When more than 1 GPU '
         'is used, will be automatically set to "ray" if installed '
         'or "mp" (multiprocessing) otherwise.')
+    parser.add_argument(
+        '--otlp-traces-endpoint',
+        type=str,
+        default=None,
+        help='Target URL to which OpenTelemetry traces will be sent.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prefix_caching.py

@@ -1,7 +1,7 @@
-import argparse
 import time
 
 from vllm import LLM, SamplingParams
+from vllm.utils import FlexibleArgumentParser
 
 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
 
@@ -44,7 +44,7 @@ def main(args):
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description='Benchmark the performance with or without automatic '
         'prefix caching.')
     parser.add_argument('--model',

benchmarks/benchmark_serving.py

@@ -31,7 +31,7 @@ import time
 import warnings
 from dataclasses import dataclass
 from datetime import datetime
-from typing import AsyncGenerator, List, Optional, Tuple
+from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple
 
 import numpy as np
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
@@ -39,7 +39,15 @@ from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 
-from vllm.transformers_utils.tokenizer import get_tokenizer
+try:
+    from vllm.transformers_utils.tokenizer import get_tokenizer
+except ImportError:
+    from backend_request_func import get_tokenizer
+
+try:
+    from vllm.utils import FlexibleArgumentParser
+except ImportError:
+    from argparse import ArgumentParser as FlexibleArgumentParser
 
 
 @dataclass
@@ -200,12 +208,12 @@ def calculate_metrics(
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
 ) -> Tuple[BenchmarkMetrics, List[int]]:
-    actual_output_lens = []
+    actual_output_lens: List[int] = []
     total_input = 0
     completed = 0
-    itls = []
-    tpots = []
-    ttfts = []
+    itls: List[float] = []
+    tpots: List[float] = []
+    ttfts: 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
@@ -265,7 +273,7 @@ async def benchmark(
     disable_tqdm: bool,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
-        request_func = ASYNC_REQUEST_FUNCS.get(backend)
+        request_func = ASYNC_REQUEST_FUNCS[backend]
     else:
         raise ValueError(f"Unknown backend: {backend}")
 
@@ -292,7 +300,7 @@ async def benchmark(
     pbar = None if disable_tqdm else tqdm(total=len(input_requests))
 
     benchmark_start_time = time.perf_counter()
-    tasks = []
+    tasks: List[asyncio.Task] = []
     async for request in get_request(input_requests, request_rate):
         prompt, prompt_len, output_len = request
         request_func_input = RequestFuncInput(
@@ -310,7 +318,7 @@ async def benchmark(
                              pbar=pbar)))
     outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
 
-    if not disable_tqdm:
+    if pbar is not None:
         pbar.close()
 
     benchmark_duration = time.perf_counter() - benchmark_start_time
@@ -466,7 +474,7 @@ def main(args: argparse.Namespace):
 
     # Save config and results to json
     if args.save_result:
-        result_json = {}
+        result_json: Dict[str, Any] = {}
 
         # Setup
         current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
@@ -499,6 +507,8 @@ def main(args: argparse.Namespace):
         # Save to file
         base_model_id = model_id.split("/")[-1]
         file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"  #noqa
+        if args.result_filename:
+            file_name = args.result_filename
         if args.result_dir:
             file_name = os.path.join(args.result_dir, file_name)
         with open(file_name, "w") as outfile:
@@ -506,7 +516,7 @@ def main(args: argparse.Namespace):
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput.")
     parser.add_argument(
         "--backend",
@@ -639,6 +649,15 @@ if __name__ == "__main__":
         help="Specify directory to save benchmark json results."
         "If not specified, results are saved in the current directory.",
     )
+    parser.add_argument(
+        "--result-filename",
+        type=str,
+        default=None,
+        help="Specify the filename to save benchmark json results."
+        "If not specified, results will be saved in "
+        "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
+        " format.",
+    )
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -10,7 +10,9 @@ from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                           PreTrainedTokenizerBase)
 
+from vllm.engine.arg_utils import EngineArgs
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.utils import FlexibleArgumentParser
 
 
 def sample_requests(
@@ -81,6 +83,7 @@ def run_vllm(
     distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
     download_dir: Optional[str] = None,
+    load_format: str = EngineArgs.load_format,
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(
@@ -102,11 +105,12 @@ def run_vllm(
         enable_chunked_prefill=enable_chunked_prefill,
         max_num_batched_tokens=max_num_batched_tokens,
         distributed_executor_backend=distributed_executor_backend,
+        load_format=load_format,
     )
 
     # Add the requests to the engine.
-    prompts = []
-    sampling_params = []
+    prompts: List[str] = []
+    sampling_params: List[SamplingParams] = []
     for prompt, _, output_len in requests:
         prompts.append(prompt)
         sampling_params.append(
@@ -228,7 +232,7 @@ 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.gpu_memory_utilization, args.download_dir, args.load_format)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -258,7 +262,7 @@ def main(args: argparse.Namespace):
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Benchmark the throughput.")
+    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument("--backend",
                         type=str,
                         choices=["vllm", "hf", "mii"],
@@ -345,9 +349,10 @@ if __name__ == "__main__":
     parser.add_argument(
         "--device",
         type=str,
-        default="cuda",
-        choices=["cuda", "cpu", "tpu"],
-        help='device type for vLLM execution, supporting CUDA and CPU.')
+        default="auto",
+        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
+        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
+        'CPU.')
     parser.add_argument(
         "--enable-prefix-caching",
         action='store_true',
@@ -377,6 +382,29 @@ if __name__ == "__main__":
         help='Backend to use for distributed serving. When more than 1 GPU '
         'is used, will be automatically set to "ray" if installed '
         'or "mp" (multiprocessing) otherwise.')
+    parser.add_argument(
+        '--load-format',
+        type=str,
+        default=EngineArgs.load_format,
+        choices=[
+            'auto', 'pt', 'safetensors', 'npcache', 'dummy', 'tensorizer',
+            'bitsandbytes'
+        ],
+        help='The format of the model weights to load.\n\n'
+        '* "auto" will try to load the weights in the safetensors format '
+        'and fall back to the pytorch bin format if safetensors format '
+        'is not available.\n'
+        '* "pt" will load the weights in the pytorch bin format.\n'
+        '* "safetensors" will load the weights in the safetensors format.\n'
+        '* "npcache" will load the weights in pytorch format and store '
+        'a numpy cache to speed up the loading.\n'
+        '* "dummy" will initialize the weights with random values, '
+        'which is mainly for profiling.\n'
+        '* "tensorizer" will load the weights using tensorizer from '
+        'CoreWeave. See the Tensorize vLLM Model script in the Examples'
+        'section for more information.\n'
+        '* "bitsandbytes" will load the weights using bitsandbytes '
+        'quantization.\n')
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -11,6 +11,7 @@ from torch.utils.benchmark import Measurement as TMeasurement
 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_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
@@ -46,7 +47,7 @@ def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
 # impl
 
 
-def pytorch_i8_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
+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)
@@ -115,14 +116,13 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
     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_i8_impl,
+                 torch.bfloat16, label, sub_label, pytorch_mm_impl,
                  "pytorch_bf16_bf16_bf16_matmul-no-scales"))
 
     # cutlass impl
     timers.append(
-        bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
-                 torch.bfloat16, label, sub_label, cutlass_impl,
-                 "cutlass_i8_i8_bf16_scaled_mm"))
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 cutlass_impl, "cutlass_i8_i8_bf16_scaled_mm"))
 
     return timers
 
@@ -136,6 +136,13 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
 
     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"))
+
     # pytorch impl: bf16 output, without fp8 fast accum
     timers.append(
         bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
@@ -160,14 +167,12 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
 
     # cutlass impl: bf16 output
     timers.append(
-        bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
-                 torch.bfloat16, label, sub_label, cutlass_impl,
-                 "cutlass_fp8_fp8_bf16_scaled_mm"))
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 cutlass_impl, "cutlass_fp8_fp8_bf16_scaled_mm"))
     # cutlass impl: fp16 output
     timers.append(
-        bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
-                 torch.float16, label, sub_label, cutlass_impl,
-                 "cutlass_fp8_fp8_fp16_scaled_mm"))
+        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
+                 cutlass_impl, "cutlass_fp8_fp8_fp16_scaled_mm"))
     return timers
 
 
@@ -289,7 +294,7 @@ if __name__ == '__main__':
             return torch.float8_e4m3fn
         raise ValueError("unsupported dtype")
 
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description="""
 Benchmark Cutlass GEMM.
 

benchmarks/cutlass_benchmarks/weight_shapes.py

@@ -22,6 +22,12 @@ WEIGHT_SHAPES = {
         ([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),

benchmarks/kernels/benchmark_aqlm.py

@@ -1,4 +1,3 @@
-import argparse
 import os
 import sys
 from typing import Optional
@@ -10,6 +9,7 @@ from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.aqlm import (
     dequantize_weight, generic_dequantize_gemm, get_int_dtype,
     optimized_dequantize_gemm)
+from vllm.utils import FlexibleArgumentParser
 
 os.environ['CUDA_VISIBLE_DEVICES'] = '0'
 
@@ -86,9 +86,9 @@ def dequant_no_scale(
 # Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against
 # the generic pytorch version.
 # Just visual comparison.
-def dequant_test(k: int, parts: torch.tensor, nbooks: int, bits: int) -> None:
+def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None:
 
-    n = parts.sum().item()
+    n = int(parts.sum().item())
 
     device = torch.device('cuda:0')
 
@@ -137,7 +137,7 @@ def dequant_test(k: int, parts: torch.tensor, nbooks: int, bits: int) -> None:
 
 def main():
 
-    parser = argparse.ArgumentParser(description="Benchmark aqlm performance.")
+    parser = FlexibleArgumentParser(description="Benchmark aqlm performance.")
 
     # Add arguments
     parser.add_argument("--nbooks",
@@ -204,7 +204,7 @@ def main():
         sys.stdout = sys.__stdout__
 
 
-def run_grid(m: int, k: int, parts: torch.tensor, nbooks: int, bits: int,
+def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int,
              methods):
 
     # I didn't see visible improvements from increasing these, but feel free :)
@@ -252,10 +252,10 @@ def run_grid(m: int, k: int, parts: torch.tensor, nbooks: int, bits: int,
     print('')
 
 
-def run_timing(num_calls: int, m: int, k: int, parts: torch.tensor,
+def run_timing(num_calls: int, m: int, k: int, parts: torch.Tensor,
                nbooks: int, bits: int, method) -> float:
 
-    n = parts.sum().item()
+    n = int(parts.sum().item())
 
     device = torch.device('cuda:0')
 

benchmarks/kernels/benchmark_marlin.py

@@ -1,4 +1,4 @@
-import argparse
+from typing import List
 
 import torch
 import torch.utils.benchmark as benchmark
@@ -15,6 +15,7 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     MarlinWorkspace, marlin_24_quantize, marlin_quantize)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     gptq_pack, quantize_weights, sort_weights)
+from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
@@ -23,8 +24,9 @@ ACT_ORDER_OPTS = [False, True]
 K_FULL_OPTS = [False, True]
 
 
-def bench_run(results, model, act_order, is_k_full, num_bits, group_size,
-              size_m, size_k, size_n):
+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):
     label = "Quant Matmul"
 
     sub_label = ("{}, act={} k_full={}, b={}, g={}, "
@@ -156,7 +158,7 @@ def main(args):
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
 
-    results = []
+    results: List[benchmark.Measurement] = []
 
     for model in args.models:
         for layer in WEIGHT_SHAPES[model]:
@@ -209,7 +211,7 @@ def main(args):
 #   python benchmark_marlin.py --batch-sizes 1 16 32 --limit-k 4096 --limit-n 4096 --limit-group-size 128 --limit-num-bits 4 --limit-act-order 0 --limit-k-full 1 # noqa E501
 #
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description="Benchmark Marlin across specified models/shapes/batches")
     parser.add_argument(
         "--models",

benchmarks/kernels/benchmark_moe.py

@@ -1,7 +1,7 @@
 import argparse
 import time
 from datetime import datetime
-from typing import Any, Dict, List, Tuple
+from typing import Any, Dict, List, Tuple, TypedDict
 
 import ray
 import torch
@@ -10,10 +10,20 @@ 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
+
+
+class BenchmarkConfig(TypedDict):
+    BLOCK_SIZE_M: int
+    BLOCK_SIZE_N: int
+    BLOCK_SIZE_K: int
+    GROUP_SIZE_M: int
+    num_warps: int
+    num_stages: int
 
 
 def benchmark_config(
-    config: Dict[str, int],
+    config: BenchmarkConfig,
     num_tokens: int,
     num_experts: int,
     shard_intermediate_size: int,
@@ -92,7 +102,7 @@ def benchmark_config(
     start_event = torch.cuda.Event(enable_timing=True)
     end_event = torch.cuda.Event(enable_timing=True)
 
-    latencies = []
+    latencies: List[float] = []
     for i in range(num_iters):
         prepare(i)
         torch.cuda.synchronize()
@@ -111,7 +121,7 @@ def get_configs_compute_bound() -> List[Dict[str, int]]:
     # Reduced search space for faster tuning.
     # TODO(woosuk): Increase the search space and use a performance model to
     # prune the search space.
-    configs = []
+    configs: List[BenchmarkConfig] = []
     for num_stages in [2, 3, 4, 5]:
         for block_m in [16, 32, 64, 128, 256]:
             for block_k in [64, 128, 256]:
@@ -175,8 +185,8 @@ class BenchmarkWorker:
         topk: int,
         dtype: torch.dtype,
         use_fp8: bool,
-        search_space: List[Dict[str, int]],
-    ) -> Dict[str, int]:
+        search_space: List[BenchmarkConfig],
+    ) -> BenchmarkConfig:
         best_config = None
         best_time = float("inf")
         for config in tqdm(search_space):
@@ -199,10 +209,11 @@ class BenchmarkWorker:
                 best_config = config
         now = datetime.now()
         print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
+        assert best_config is not None
         return best_config
 
 
-def sort_config(config: Dict[str, int]) -> Dict[str, int]:
+def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     return {
         "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
         "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
@@ -214,7 +225,7 @@ def sort_config(config: Dict[str, int]) -> Dict[str, int]:
 
 
 def save_configs(
-    configs: Dict[int, Dict[str, int]],
+    configs: Dict[int, BenchmarkConfig],
     num_experts: int,
     shard_intermediate_size: int,
     hidden_size: int,
@@ -305,7 +316,7 @@ def main(args: argparse.Namespace):
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
+    parser = FlexibleArgumentParser()
     parser.add_argument("--model",
                         type=str,
                         default="mistralai/Mixtral-8x7B-Instruct-v0.1")

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,12 +1,12 @@
-import argparse
 import random
 import time
-from typing import Optional
+from typing import List, Optional
 
 import torch
 
 from vllm import _custom_ops as ops
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, create_kv_caches_with_random
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
+                        create_kv_caches_with_random)
 
 NUM_BLOCKS = 1024
 PARTITION_SIZE = 512
@@ -54,14 +54,17 @@ def main(
 
     # Create the block tables.
     max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
-    block_tables = []
+    block_tables_lst: List[List[int]] = []
     for _ in range(num_seqs):
         block_table = [
             random.randint(0, NUM_BLOCKS - 1)
             for _ in range(max_num_blocks_per_seq)
         ]
-        block_tables.append(block_table)
-    block_tables = torch.tensor(block_tables, dtype=torch.int, device=device)
+        block_tables_lst.append(block_table)
+
+    block_tables = torch.tensor(block_tables_lst,
+                                dtype=torch.int,
+                                device=device)
 
     # Create the KV cache.
     key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
@@ -158,14 +161,14 @@ def main(
 
 
 if __name__ == '__main__':
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description="Benchmark the paged attention kernel.")
     parser.add_argument("--version",
                         type=str,
                         choices=["v1", "v2"],
                         default="v2")
     parser.add_argument("--batch-size", type=int, default=8)
-    parser.add_argument("--seq_len", type=int, default=4096)
+    parser.add_argument("--seq-len", type=int, default=4096)
     parser.add_argument("--num-query-heads", type=int, default=64)
     parser.add_argument("--num-kv-heads", type=int, default=8)
     parser.add_argument("--head-size",

benchmarks/kernels/benchmark_rope.py

@@ -1,11 +1,12 @@
-import argparse
 from itertools import accumulate
-from typing import Optional
+from typing import List, Optional
 
 import nvtx
 import torch
 
-from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
+                                                         get_rope)
+from vllm.utils import FlexibleArgumentParser
 
 
 def benchmark_rope_kernels_multi_lora(
@@ -37,7 +38,7 @@ def benchmark_rope_kernels_multi_lora(
                             })
     # non-batched RoPE takes only one scaling factor, we create multiple
     # instances to simulate the same behavior
-    non_batched_ropes = []
+    non_batched_ropes: List[RotaryEmbedding] = []
     for scaling_factor in scaling_factors:
         non_batched_ropes.append(
             get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
@@ -85,7 +86,7 @@ def benchmark_rope_kernels_multi_lora(
 
 
 if __name__ == '__main__':
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description="Benchmark the rotary embedding kernels.")
     parser.add_argument("--is-neox-style", type=bool, default=True)
     parser.add_argument("--batch-size", type=int, default=16)

benchmarks/overheads/benchmark_hashing.py

@@ -1,8 +1,8 @@
-import argparse
 import cProfile
 import pstats
 
 from vllm import LLM, SamplingParams
+from vllm.utils import FlexibleArgumentParser
 
 # A very long prompt, total number of tokens is about 15k.
 LONG_PROMPT = ["You are an expert in large language models, aren't you?"
@@ -47,7 +47,7 @@ def main(args):
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description='Benchmark the performance of hashing function in'
         'automatic prefix caching.')
     parser.add_argument('--model', type=str, default='lmsys/longchat-7b-16k')

cmake/cpu_extension.cmake

@@ -33,9 +33,23 @@ function (find_isa CPUINFO TARGET OUT)
     endif()
 endfunction()
 
-find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
+function (is_avx512_disabled OUT)
+    set(DISABLE_AVX512 $ENV{VLLM_CPU_DISABLE_AVX512})
+    if(DISABLE_AVX512 AND DISABLE_AVX512 STREQUAL "true")
+        set(${OUT} ON PARENT_SCOPE)
+    else()
+        set(${OUT} OFF PARENT_SCOPE)
+    endif()
+endfunction()
 
-if (AVX512_FOUND)
+is_avx512_disabled(AVX512_DISABLED)
+
+find_isa(${CPUINFO} "avx2" AVX2_FOUND)
+find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
+find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
+find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
+
+if (AVX512_FOUND AND NOT AVX512_DISABLED)
     list(APPEND CXX_COMPILE_FLAGS
         "-mavx512f"
         "-mavx512vl"
@@ -53,8 +67,18 @@ if (AVX512_FOUND)
     else()
         message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
     endif()
+elseif (AVX2_FOUND)
+    list(APPEND CXX_COMPILE_FLAGS "-mavx2")
+    message(WARNING "vLLM CPU backend using AVX2 ISA")
+elseif (POWER9_FOUND OR POWER10_FOUND)
+    message(STATUS "PowerPC detected")
+    # Check for PowerPC VSX support
+    list(APPEND CXX_COMPILE_FLAGS
+        "-mvsx"
+        "-mcpu=native"
+        "-mtune=native")
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512 ISA support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512 or AVX2 or Power9+ ISA support.")
 endif()
 
 message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")

cmake/utils.cmake

@@ -147,16 +147,23 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
   if (${GPU_LANG} STREQUAL "HIP")
     #
     # `GPU_ARCHES` controls the `--offload-arch` flags.
-    # `CMAKE_HIP_ARCHITECTURES` is set up by torch and can be controlled
-    # via the `PYTORCH_ROCM_ARCH` env variable.
     #
-
+    # If PYTORCH_ROCM_ARCH env variable exists, then we take it as a list,
+    # if not, then we use CMAKE_HIP_ARCHITECTURES which was generated by calling
+    # "rocm_agent_enumerator" in "enable_language(HIP)"
+    # (in file Modules/CMakeDetermineHIPCompiler.cmake)
+    #
+    if(DEFINED ENV{PYTORCH_ROCM_ARCH})
+      set(HIP_ARCHITECTURES $ENV{PYTORCH_ROCM_ARCH})
+    else()
+      set(HIP_ARCHITECTURES ${CMAKE_HIP_ARCHITECTURES})
+    endif()
     #
     # Find the intersection of the supported + detected architectures to
     # set the module architecture flags.
     #
     set(${GPU_ARCHES})
-    foreach (_ARCH ${CMAKE_HIP_ARCHITECTURES})
+    foreach (_ARCH ${HIP_ARCHITECTURES})
       if (_ARCH IN_LIST _GPU_SUPPORTED_ARCHES_LIST)
         list(APPEND ${GPU_ARCHES} ${_ARCH})
       endif()
@@ -164,7 +171,7 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
 
     if(NOT ${GPU_ARCHES})
       message(FATAL_ERROR
-        "None of the detected ROCm architectures: ${CMAKE_HIP_ARCHITECTURES} is"
+        "None of the detected ROCm architectures: ${HIP_ARCHITECTURES} is"
         " supported. Supported ROCm architectures are: ${_GPU_SUPPORTED_ARCHES_LIST}.")
     endif()
 

csrc/activation_kernels.cu

@@ -135,6 +135,12 @@ __device__ __forceinline__ T gelu_fast_kernel(const T& x) {
   return ((T)0.5) * x * (((T)1.0) + t);
 }
 
+template <typename T>
+__device__ __forceinline__ T gelu_quick_kernel(const T& x) {
+  // x * sigmoid(1.702 * x)
+  return (T)(((float)x) / (1.0f + expf(-1.702f * (float)x)));
+}
+
 }  // namespace vllm
 
 void gelu_new(torch::Tensor& out,    // [..., d]
@@ -148,3 +154,9 @@ void gelu_fast(torch::Tensor& out,    // [..., d]
 {
   LAUNCH_ACTIVATION_KERNEL(vllm::gelu_fast_kernel);
 }
+
+void gelu_quick(torch::Tensor& out,    // [..., d]
+                torch::Tensor& input)  // [..., d]
+{
+  LAUNCH_ACTIVATION_KERNEL(vllm::gelu_quick_kernel);
+}

csrc/cpu/activation.cpp

@@ -59,6 +59,13 @@ FORCE_INLINE vec_op::FP32Vec8 gelu_fast_act(const vec_op::FP32Vec8& x) {
   return w3 * x * (ones + t);
 }
 
+FORCE_INLINE vec_op::FP32Vec8 gelu_quick_act(const vec_op::FP32Vec8& x) {
+  const vec_op::FP32Vec8 zeros(0.0);
+  const vec_op::FP32Vec8 ones(1.0);
+  const vec_op::FP32Vec8 w1(1.702f);
+  return x / (ones + (zeros - w1 * x).exp());
+}
+
 FORCE_INLINE vec_op::FP32Vec8 gelu_act(const vec_op::FP32Vec8& x) {
   const vec_op::FP32Vec8 ones(1.0);
   const vec_op::FP32Vec8 w1(M_SQRT1_2);
@@ -142,3 +149,15 @@ void gelu_fast(torch::Tensor& out, torch::Tensor& input) {
     CPU_KERNEL_GUARD_OUT(gelu_fast_impl)
   });
 }
+
+void gelu_quick(torch::Tensor& out, torch::Tensor& input) {
+  int num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "gelu_quick_impl", [&] {
+    CPU_KERNEL_GUARD_IN(gelu_quick_impl)
+    activation_kernel<scalar_t, gelu_quick_act, false>(
+        num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
+    CPU_KERNEL_GUARD_OUT(gelu_quick_impl)
+  });
+}

csrc/cpu/cpu_types.hpp

@@ -2,351 +2,14 @@
 #ifndef CPU_TYPES_HPP
 #define CPU_TYPES_HPP
 
-#include <immintrin.h>
-#include <torch/all.h>
-
-namespace vec_op {
-
-// FIXME: FP16 is not fully supported in Torch-CPU
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
-
-#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
-  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
-
-#ifndef CPU_OP_GUARD
-#define CPU_KERNEL_GUARD_IN(NAME)
-#define CPU_KERNEL_GUARD_OUT(NAME)
+#if defined(__x86_64__)
+  //x86 implementation
+  #include "cpu_types_x86.hpp"
+#elif defined(__POWER9_VECTOR__)
+  //ppc implementation
+  #include "cpu_types_vsx.hpp"
 #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;
+  #warning "unsupported vLLM cpu implementation"
 #endif
 
-#define FORCE_INLINE __attribute__((always_inline)) inline
-
-namespace {
-template <typename T, T... indexes, typename F>
-constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
-  (f(std::integral_constant<T, indexes>{}), ...);
-}
-}; // namespace
-
-template <typename T, T count, typename F,
-          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
-constexpr void unroll_loop(F &&f) {
-  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
-}
-
-template <typename T> struct Vec {
-  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
-};
-
-struct FP32Vec8;
-struct FP32Vec16;
-
-#ifdef __AVX512FP16__
-struct FP16Vec8 : public Vec<FP16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-
-  __m128h reg;
-
-  explicit FP16Vec8(_Float16 v) : reg(_mm_set1_ph(v)) {}
-
-  explicit FP16Vec8(const void *ptr) : reg(_mm_loadu_ph(ptr)) {}
-
-  explicit FP16Vec8(__m128h data) : reg(data) {}
-
-  FP16Vec8 operator*(const FP16Vec8 &b) const {
-    return FP16Vec8(_mm_mul_ph(reg, b.reg));
-  }
-
-  FP16Vec8 operator+(const FP16Vec8 &b) const {
-    return FP16Vec8(_mm_add_ph(reg, b.reg));
-  }
-
-  FP16Vec8 operator-(const FP16Vec8 &b) const {
-    return FP16Vec8(_mm_sub_ph(reg, b.reg));
-  }
-
-  FP16Vec8 operator/(const FP16Vec8 &b) const {
-    return FP16Vec8(_mm_div_ph(reg, b.reg));
-  }
-
-  void save(void *ptr) const { _mm_storeu_ph(ptr, reg); }
-};
-#endif
-
-struct BF16Vec8 : public Vec<BF16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-
-  __m128i reg;
-
-  explicit BF16Vec8(const void *ptr)
-      : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {}
-
-  explicit BF16Vec8(const FP32Vec8 &);
-
-  void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; }
-};
-
-struct BF16Vec16 : public Vec<BF16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-
-  __m256i reg;
-
-  explicit BF16Vec16(const void *ptr)
-      : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {}
-
-  explicit BF16Vec16(const FP32Vec16 &);
-
-  void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; }
-};
-
-struct BF16Vec32 : public Vec<BF16Vec32> {
-  constexpr static int VEC_ELEM_NUM = 32;
-
-  __m512i reg;
-
-  explicit BF16Vec32(const void *ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
-
-  explicit BF16Vec32(__m512i data) : reg(data) {}
-
-  explicit BF16Vec32(BF16Vec8 &vec8_data)
-      : reg((__m512i)_mm512_inserti32x4(
-            _mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(
-                                                      (__m128i)vec8_data.reg),
-                                                  (__m128i)vec8_data.reg, 1),
-                               (__m128i)vec8_data.reg, 2),
-            (__m128i)vec8_data.reg, 3)) {}
-
-  void save(void *ptr) const { *reinterpret_cast<__m512i *>(ptr) = reg; }
-};
-
-struct FP32Vec4 : public Vec<FP32Vec4> {
-  constexpr static int VEC_ELEM_NUM = 4;
-  union AliasReg {
-    __m128 reg;
-    float values[VEC_ELEM_NUM];
-  };
-
-  __m128 reg;
-
-  explicit FP32Vec4(float v) : reg(_mm_set1_ps(v)) {}
-
-  explicit FP32Vec4() : reg(_mm_set1_ps(0.0)) {}
-
-  explicit FP32Vec4(const float *ptr) : reg(_mm_loadu_ps(ptr)) {}
-
-  explicit FP32Vec4(__m128 data) : reg(data) {}
-
-  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {}
-};
-
-struct FP32Vec8 : public Vec<FP32Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  union AliasReg {
-    __m256 reg;
-    float values[VEC_ELEM_NUM];
-  };
-
-  __m256 reg;
-
-  explicit FP32Vec8(float v) : reg(_mm256_set1_ps(v)) {}
-
-  explicit FP32Vec8() : reg(_mm256_set1_ps(0.0)) {}
-
-  explicit FP32Vec8(const float *ptr) : reg(_mm256_loadu_ps(ptr)) {}
-
-  explicit FP32Vec8(__m256 data) : reg(data) {}
-
-  explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {}
-
-#ifdef __AVX512FP16__
-  explicit FP32Vec8(__m128h v) : reg(_mm256_cvtph_ps(_mm_castph_si128(v))) {}
-#endif
-
-  explicit FP32Vec8(const BF16Vec8 &v)
-      : reg(_mm256_castsi256_ps(
-            _mm256_bslli_epi128(_mm256_cvtepu16_epi32(v.reg), 2))) {}
-
-  float reduce_sum() const {
-    AliasReg ar;
-    ar.reg = reg;
-    float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
-
-    return result;
-  }
-
-  FP32Vec8 exp() const {
-    AliasReg ar;
-    ar.reg = reg;
-    return FP32Vec8(_mm256_set_ps(expf(ar.values[7]), expf(ar.values[6]),
-                                  expf(ar.values[5]), expf(ar.values[4]),
-                                  expf(ar.values[3]), expf(ar.values[2]),
-                                  expf(ar.values[1]), expf(ar.values[0])));
-  }
-
-  FP32Vec8 tanh() const {
-    AliasReg ar;
-    ar.reg = reg;
-    return FP32Vec8(_mm256_set_ps(tanhf(ar.values[7]), tanhf(ar.values[6]),
-                                  tanhf(ar.values[5]), tanhf(ar.values[4]),
-                                  tanhf(ar.values[3]), tanhf(ar.values[2]),
-                                  tanhf(ar.values[1]), tanhf(ar.values[0])));
-  }
-
-  FP32Vec8 er() const {
-    AliasReg ar;
-    ar.reg = reg;
-    return FP32Vec8(_mm256_set_ps(erf(ar.values[7]), erf(ar.values[6]),
-                                  erf(ar.values[5]), erf(ar.values[4]),
-                                  erf(ar.values[3]), erf(ar.values[2]),
-                                  erf(ar.values[1]), erf(ar.values[0])));
-  }
-
-  FP32Vec8 operator*(const FP32Vec8 &b) const {
-    return FP32Vec8(_mm256_mul_ps(reg, b.reg));
-  }
-
-  FP32Vec8 operator+(const FP32Vec8 &b) const {
-    return FP32Vec8(_mm256_add_ps(reg, b.reg));
-  }
-
-  FP32Vec8 operator-(const FP32Vec8 &b) const {
-    return FP32Vec8(_mm256_sub_ps(reg, b.reg));
-  }
-
-  FP32Vec8 operator/(const FP32Vec8 &b) const {
-    return FP32Vec8(_mm256_div_ps(reg, b.reg));
-  }
-
-  void save(float *ptr) const { _mm256_storeu_ps(ptr, reg); }
-};
-
-struct FP32Vec16 : public Vec<FP32Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  union AliasReg {
-    __m512 reg;
-    float values[VEC_ELEM_NUM];
-  };
-
-  __m512 reg;
-
-  explicit FP32Vec16(float v) : reg(_mm512_set1_ps(v)) {}
-
-  explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
-
-  explicit FP32Vec16(const float *ptr) : reg(_mm512_loadu_ps(ptr)) {}
-
-  explicit FP32Vec16(__m512 data) : reg(data) {}
-
-  explicit FP32Vec16(const FP32Vec16 &data) : reg(data.reg) {}
-
-  explicit FP32Vec16(const FP32Vec4 &data)
-      : reg((__m512)_mm512_inserti32x4(
-            _mm512_inserti32x4(
-                _mm512_inserti32x4(_mm512_castsi128_si512((__m128i)data.reg),
-                                   (__m128i)data.reg, 1),
-                (__m128i)data.reg, 2),
-            (__m128i)data.reg, 3)) {}
-
-  explicit FP32Vec16(const FP32Vec8 &data)
-      : reg((__m512)_mm512_inserti32x8(
-            _mm512_castsi256_si512((__m256i)data.reg), (__m256i)data.reg, 1)) {}
-
-  explicit FP32Vec16(const BF16Vec16 &v)
-      : reg(_mm512_castsi512_ps(
-            _mm512_bslli_epi128(_mm512_cvtepu16_epi32(v.reg), 2))) {}
-
-  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
-
-  FP32Vec16 operator*(const FP32Vec16 &b) const {
-    return FP32Vec16(_mm512_mul_ps(reg, b.reg));
-  }
-
-  FP32Vec16 operator+(const FP32Vec16 &b) const {
-    return FP32Vec16(_mm512_add_ps(reg, b.reg));
-  }
-
-  FP32Vec16 operator-(const FP32Vec16 &b) const {
-    return FP32Vec16(_mm512_sub_ps(reg, b.reg));
-  }
-
-  FP32Vec16 operator/(const FP32Vec16 &b) const {
-    return FP32Vec16(_mm512_div_ps(reg, b.reg));
-  }
-
-  float reduce_sum() const { return _mm512_reduce_add_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));
-    __mmask16 mask = _cvtu32_mask16(base_mask << (idx * group_size));
-    return _mm512_mask_reduce_add_ps(mask, reg);
-  }
-
-  void save(float *ptr) const { _mm512_storeu_ps(ptr, reg); }
-};
-
-template <typename T> struct VecType { using vec_type = void; };
-
-template <typename T> using vec_t = typename VecType<T>::vec_type;
-
-template <> struct VecType<float> { using vec_type = FP32Vec8; };
-
-#ifdef __AVX512FP16__
-template <> struct VecType<c10::Half> { using vec_type = FP16Vec16; };
-#endif
-
-template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
-
-template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
-
-#ifdef __AVX512FP16__
-template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
-  *reinterpret_cast<_Float16 *>(ptr) = v;
-}
-#endif
-
-inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
-  acc = acc + a * b;
-}
-
-#ifdef __AVX512BF16__
-template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
-  *reinterpret_cast<__bfloat16 *>(ptr) = _mm_cvtness_sbh(v);
-}
-
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
-    : reg((__m128i)_mm256_cvtneps_pbh(v.reg)) {}
-
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
-    : reg((__m256i)_mm512_cvtneps_pbh(v.reg)) {}
-
-inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
-  acc.reg = _mm512_dpbf16_ps(acc.reg, (__m512bh)a.reg, (__m512bh)b.reg);
-}
-#else
-template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
-  c10::BFloat16 __attribute__((__may_alias__)) *v_ptr =
-      reinterpret_cast<c10::BFloat16 *>(&v);
-  *ptr = *(v_ptr + 1);
-}
-
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
-    : reg(_mm256_cvtepi32_epi16(
-          _mm256_bsrli_epi128(_mm256_castps_si256(v.reg), 2))) {}
-
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
-    : reg(_mm512_cvtepi32_epi16(
-          _mm512_bsrli_epi128(_mm512_castps_si512(v.reg), 2))) {}
-#endif
-
-inline void prefetch(const void *addr) { _mm_prefetch(addr, _MM_HINT_T1); }
-
-}; // namespace vec_op
-
 #endif

csrc/cpu/cpu_types_vsx.hpp

@@ -0,0 +1,491 @@
+
+#ifndef CPU_TYPES_VSX_HPP
+#define CPU_TYPES_VSX_HPP
+
+#include <altivec.h>
+#include <cmath>
+#include <torch/all.h>
+
+namespace vec_op {
+
+// FIXME: FP16 is not fully supported in Torch-CPU
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#ifndef CPU_OP_GUARD
+#define CPU_KERNEL_GUARD_IN(NAME)
+#define CPU_KERNEL_GUARD_OUT(NAME)
+#else
+#define CPU_KERNEL_GUARD_IN(NAME)                                              \
+  std::cout << #NAME << " invoked." << std::endl;
+#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl;
+#endif
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+namespace {
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+}
+}; // namespace
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
+constexpr void unroll_loop(F &&f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T> struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
+};
+
+typedef struct ss16x8x2_t {
+  __vector signed short val[2];
+} ss16x8x2_t;
+
+typedef struct ss16x8x4_t {
+  __vector signed short val[4];
+} ss16x8x4_t;
+
+typedef struct f32x4x2_t {
+  __vector float val[2];
+} f32x4x2_t;
+
+typedef struct f32x4x4_t {
+  __vector float val[4];
+} f32x4x4_t;
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __vector signed short reg;
+
+  explicit BF16Vec8(const void *ptr)
+      : reg((__vector signed short)vec_xl(0, (__vector signed short *)ptr)) {}
+
+  explicit BF16Vec8(const FP32Vec8 &);
+
+  void save(void *ptr) const { *reinterpret_cast<__vector signed short *>(ptr) = reg; }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  ss16x8x2_t reg;
+
+  explicit BF16Vec16(const void *ptr) {
+    // Load 256 bits in two parts
+    reg.val[0] = (__vector signed short)vec_xl(0,  (signed short *)ptr);
+    reg.val[1] = (__vector signed short)vec_xl(16, (signed short *)ptr);
+  }
+
+  explicit BF16Vec16(const FP32Vec16 &);
+
+  void save(void *ptr) const {
+    // Save 256 bits in two parts
+    vec_xst(reg.val[0], 0, (signed short *)ptr);
+    vec_xst(reg.val[1], 16, (signed short *)ptr);
+  }
+};
+
+const static __vector signed short zero = vec_splats((signed short)0);
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+
+  ss16x8x4_t reg;
+  explicit BF16Vec32(const void *ptr)
+      : reg(*reinterpret_cast<const ss16x8x4_t *>(ptr)) {}
+
+  explicit BF16Vec32(ss16x8x4_t data) : reg(data) {}
+
+  explicit BF16Vec32(const BF16Vec8 &vec8_data) : reg({
+    vec8_data.reg,
+    vec8_data.reg,
+    vec8_data.reg,
+    vec8_data.reg
+  }) {}
+
+  void save(void *ptr) const { *reinterpret_cast<ss16x8x4_t *>(ptr) = reg; }
+};
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+  union AliasReg {
+    __vector float reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __vector float reg;
+
+  explicit FP32Vec4(float v) : reg(vec_splats(v)) {}
+
+  explicit FP32Vec4() : reg(vec_splats(0.0f)) {}
+
+  explicit FP32Vec4(const float *ptr) : reg(vec_xl(0, ptr)) {}
+
+  explicit FP32Vec4(__vector float data) : reg(data) {}
+
+  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {}
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  union AliasReg {
+    f32x4x2_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  f32x4x2_t reg;
+
+  explicit FP32Vec8(float v) {
+    reg.val[0] = vec_splats(v);
+    reg.val[1] = vec_splats(v);
+  }
+
+  explicit FP32Vec8() {
+    reg.val[0] = vec_splats(0.0f);
+    reg.val[1] = vec_splats(0.0f);
+  }
+
+  explicit FP32Vec8(const float *ptr) {
+    reg.val[0] = vec_xl(0, ptr);
+    reg.val[1] = vec_xl(16, ptr);
+  }
+
+  explicit FP32Vec8(f32x4x2_t data) : reg(data) {}
+
+  explicit FP32Vec8(const FP32Vec8 &data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+  }
+
+  explicit FP32Vec8(const BF16Vec8 &v) {
+    reg.val[0] = (__vector float)vec_mergeh(zero, v.reg);
+    reg.val[1] = (__vector float)vec_mergel(zero, v.reg);
+  }
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
+
+    return result;
+  }
+
+  FP32Vec8 exp() const {
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::exp(ar.values[0]);
+    ret.val[0][1] = std::exp(ar.values[1]);
+    ret.val[0][2] = std::exp(ar.values[2]);
+    ret.val[0][3] = std::exp(ar.values[3]);
+    ret.val[1][0] = std::exp(ar.values[4]);
+    ret.val[1][1] = std::exp(ar.values[5]);
+    ret.val[1][2] = std::exp(ar.values[6]);
+    ret.val[1][3] = std::exp(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+  }
+
+  FP32Vec8 tanh() const {
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::tanh(ar.values[0]);
+    ret.val[0][1] = std::tanh(ar.values[1]);
+    ret.val[0][2] = std::tanh(ar.values[2]);
+    ret.val[0][3] = std::tanh(ar.values[3]);
+    ret.val[1][0] = std::tanh(ar.values[4]);
+    ret.val[1][1] = std::tanh(ar.values[5]);
+    ret.val[1][2] = std::tanh(ar.values[6]);
+    ret.val[1][3] = std::tanh(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+  }
+
+  FP32Vec8 er() const {
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::erf(ar.values[0]);
+    ret.val[0][1] = std::erf(ar.values[1]);
+    ret.val[0][2] = std::erf(ar.values[2]);
+    ret.val[0][3] = std::erf(ar.values[3]);
+    ret.val[1][0] = std::erf(ar.values[4]);
+    ret.val[1][1] = std::erf(ar.values[5]);
+    ret.val[1][2] = std::erf(ar.values[6]);
+    ret.val[1][3] = std::erf(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+  }
+
+  FP32Vec8 operator*(const FP32Vec8 &b) const {
+    return FP32Vec8({vec_mul(reg.val[0], b.reg.val[0]), vec_mul(reg.val[1], b.reg.val[1])});
+  }
+
+  FP32Vec8 operator+(const FP32Vec8 &b) const {
+    return FP32Vec8({vec_add(reg.val[0], b.reg.val[0]), vec_add(reg.val[1], b.reg.val[1])});
+  }
+
+  FP32Vec8 operator-(const FP32Vec8 &b) const {
+    return FP32Vec8({vec_sub(reg.val[0], b.reg.val[0]), vec_sub(reg.val[1], b.reg.val[1])});
+  }
+
+  FP32Vec8 operator/(const FP32Vec8 &b) const {
+    return FP32Vec8({vec_div(reg.val[0], b.reg.val[0]), vec_div(reg.val[1], b.reg.val[1])});
+  }
+
+  void save(float *ptr) const {
+    vec_xst(reg.val[0], 0, ptr);
+    vec_xst(reg.val[1], 16, ptr);
+  }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    f32x4x4_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  f32x4x4_t reg;
+
+  explicit FP32Vec16(float v) {
+    reg.val[0] = vec_splats(v);
+    reg.val[1] = vec_splats(v);
+    reg.val[2] = vec_splats(v);
+    reg.val[3] = vec_splats(v);
+  }
+
+  explicit FP32Vec16() {
+    reg.val[0] = vec_splats(0.0f);
+    reg.val[1] = vec_splats(0.0f);
+    reg.val[2] = vec_splats(0.0f);
+    reg.val[3] = vec_splats(0.0f);
+  }
+
+  explicit FP32Vec16(const float *ptr) {
+    reg.val[0] = vec_xl(0, ptr);
+    reg.val[1] = vec_xl(16, ptr);
+    reg.val[2] = vec_xl(32, ptr);
+    reg.val[3] = vec_xl(48, ptr);
+  }
+
+  explicit FP32Vec16(f32x4x4_t data) : reg(data) {}
+
+  explicit FP32Vec16(const FP32Vec16 &data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+    reg.val[2] = data.reg.val[2];
+    reg.val[3] = data.reg.val[3];
+  }
+
+  explicit FP32Vec16(const FP32Vec4 &data) {
+    reg.val[0] = data.reg;
+    reg.val[1] = data.reg;
+    reg.val[2] = data.reg;
+    reg.val[3] = data.reg;
+  }
+
+  explicit FP32Vec16(const FP32Vec8 &data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+    reg.val[2] = data.reg.val[0];
+    reg.val[3] = data.reg.val[1];
+  }
+
+  explicit FP32Vec16(const BF16Vec16 &v) {
+    reg.val[0] = (__vector float)vec_mergeh(zero, v.reg.val[0]);
+    reg.val[1] = (__vector float)vec_mergel(zero, v.reg.val[0]);
+    reg.val[2] = (__vector float)vec_mergeh(zero, v.reg.val[1]);
+    reg.val[3] = (__vector float)vec_mergel(zero, v.reg.val[1]);
+  }
+
+  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+
+  FP32Vec16 operator*(const FP32Vec16 &b) const {
+    return FP32Vec16(f32x4x4_t({
+        vec_mul(reg.val[0], b.reg.val[0]),
+        vec_mul(reg.val[1], b.reg.val[1]),
+        vec_mul(reg.val[2], b.reg.val[2]),
+        vec_mul(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 operator+(const FP32Vec16 &b) const {
+    return FP32Vec16(f32x4x4_t({
+        vec_add(reg.val[0], b.reg.val[0]),
+        vec_add(reg.val[1], b.reg.val[1]),
+        vec_add(reg.val[2], b.reg.val[2]),
+        vec_add(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 operator-(const FP32Vec16 &b) const {
+    return FP32Vec16(f32x4x4_t({
+        vec_sub(reg.val[0], b.reg.val[0]),
+        vec_sub(reg.val[1], b.reg.val[1]),
+        vec_sub(reg.val[2], b.reg.val[2]),
+        vec_sub(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 operator/(const FP32Vec16 &b) const {
+    return FP32Vec16(f32x4x4_t({
+        vec_div(reg.val[0], b.reg.val[0]),
+        vec_div(reg.val[1], b.reg.val[1]),
+        vec_div(reg.val[2], b.reg.val[2]),
+        vec_div(reg.val[3], b.reg.val[3])}));
+  }
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
+
+    return result;
+  }
+
+  template <int group_size> float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    const int start = idx * group_size;
+    unroll_loop<int, group_size>(
+        [&result, &start, ar](int i) { result += ar.values[start + i]; });
+
+    return result;
+  }
+
+  void save(float *ptr) const {
+    vec_xst(reg.val[0], 0, ptr);
+    vec_xst(reg.val[1], 16, ptr);
+    vec_xst(reg.val[2], 32, ptr);
+    vec_xst(reg.val[3], 48, ptr);
+  }
+};
+
+template <typename T> struct VecType { using vec_type = void; };
+
+template <typename T> using vec_t = typename VecType<T>::vec_type;
+
+template <> struct VecType<float> { using vec_type = FP32Vec8; };
+
+template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+
+template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
+
+inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
+  acc = acc + a * b;
+}
+
+template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
+  c10::BFloat16 __attribute__((__may_alias__)) *v_ptr =
+      reinterpret_cast<c10::BFloat16 *>(&v);
+  *ptr = *(v_ptr + 1);
+}
+
+#ifndef __VEC_CLASS_FP_NAN
+#define __VEC_CLASS_FP_NAN (1 << 6)
+#endif
+
+const static __vector unsigned char omask = { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 };
+#ifndef _ARCH_PWR10
+const static __vector unsigned int bias = { 0x00007fff, 0x00007fff, 0x00007fff, 0x00007fff };
+const static __vector unsigned int nan  = { 0x7fc00000, 0x7fc00000, 0x7fc00000, 0x7fc00000 };
+const static __vector unsigned int sh16 = { 16, 16, 16, 16 };
+const static __vector unsigned int one  = { 1, 1, 1, 1 };
+#endif
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) {
+#ifdef _ARCH_PWR10
+  __vector signed short ret[2];
+  ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[0]);
+  ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[1]);
+  reg = vec_perm(ret[0], ret[1], omask);
+#elif defined(_ARCH_PWR9)
+  __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]);
+  __vector unsigned int inp1 = (__vector unsigned int)(v.reg.val[1]);
+  __vector unsigned int lsb0 = vec_sr(inp0, sh16);
+  __vector unsigned int lsb1 = vec_sr(inp1, sh16);
+  lsb0 = vec_and(lsb0, one);
+  lsb1 = vec_and(lsb1, one);
+  __vector unsigned int rnd0 = vec_add(lsb0, bias);
+  __vector unsigned int rnd1 = vec_add(lsb1, bias);
+  inp0 = vec_add(inp0, rnd0);
+  inp1 = vec_add(inp1, rnd1);
+  __vector __bool int sel0 = vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel1 = vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN);
+  inp0 = vec_sel(inp0, nan, sel0);
+  inp1 = vec_sel(inp1, nan, sel1);
+  inp0 = vec_sr(inp0, sh16);
+  inp1 = vec_sr(inp1, sh16);
+  reg = (__vector signed short)vec_perm(inp0, inp1, omask);
+#endif
+}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) {
+#ifdef _ARCH_PWR10
+  __vector signed short ret[4];
+  ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[0]);
+  ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[1]);
+  ret[2] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[2]);
+  ret[3] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[3]);
+  reg.val[0] = vec_perm(ret[0], ret[1], omask);
+  reg.val[1] = vec_perm(ret[2], ret[3], omask);
+#elif defined(_ARCH_PWR9)
+  __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]);
+  __vector unsigned int inp1 = (__vector unsigned int)(v.reg.val[1]);
+  __vector unsigned int inp2 = (__vector unsigned int)(v.reg.val[2]);
+  __vector unsigned int inp3 = (__vector unsigned int)(v.reg.val[3]);
+  __vector unsigned int lsb0 = vec_sr(inp0, sh16);
+  __vector unsigned int lsb1 = vec_sr(inp1, sh16);
+  __vector unsigned int lsb2 = vec_sr(inp2, sh16);
+  __vector unsigned int lsb3 = vec_sr(inp3, sh16);
+  lsb0 = vec_and(lsb0, one);
+  lsb1 = vec_and(lsb1, one);
+  lsb2 = vec_and(lsb2, one);
+  lsb3 = vec_and(lsb3, one);
+  __vector unsigned int rnd0 = vec_add(lsb0, bias);
+  __vector unsigned int rnd1 = vec_add(lsb1, bias);
+  __vector unsigned int rnd2 = vec_add(lsb2, bias);
+  __vector unsigned int rnd3 = vec_add(lsb3, bias);
+  inp0 = vec_add(inp0, rnd0);
+  inp1 = vec_add(inp1, rnd1);
+  inp2 = vec_add(inp2, rnd2);
+  inp3 = vec_add(inp3, rnd3);
+  __vector __bool int sel0 = vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel1 = vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel2 = vec_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel3 = vec_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN);
+  inp0 = vec_sel(inp0, nan, sel0);
+  inp1 = vec_sel(inp1, nan, sel1);
+  inp2 = vec_sel(inp2, nan, sel2);
+  inp3 = vec_sel(inp3, nan, sel3);
+  inp0 = vec_sr(inp0, sh16);
+  inp1 = vec_sr(inp1, sh16);
+  inp2 = vec_sr(inp2, sh16);
+  inp3 = vec_sr(inp3, sh16);
+  reg.val[0] = (__vector signed short)vec_perm(inp0, inp1, omask);
+  reg.val[1] = (__vector signed short)vec_perm(inp2, inp3, omask);
+#endif
+}
+
+inline void prefetch(const void *addr) {
+  __asm__ __volatile__("dcbt 0, %0" : : "r"(addr) : "memory");
+}
+
+}; // namespace vec_op
+
+#endif

csrc/cpu/cpu_types_x86.hpp

@@ -0,0 +1,515 @@
+
+#ifndef CPU_TYPES_X86_HPP
+#define CPU_TYPES_X86_HPP
+
+#include <immintrin.h>
+#include <torch/all.h>
+
+#ifndef __AVX2__
+static_assert(false, "AVX2 must be supported for the current implementation.");
+#endif
+
+namespace vec_op {
+
+// FIXME: FP16 is not fully supported in Torch-CPU
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#ifndef CPU_OP_GUARD
+#define CPU_KERNEL_GUARD_IN(NAME)
+#define CPU_KERNEL_GUARD_OUT(NAME)
+#else
+#define CPU_KERNEL_GUARD_IN(NAME)                                              \
+  std::cout << #NAME << " invoked." << std::endl;
+#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl;
+#endif
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+namespace {
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+}
+}; // namespace
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
+constexpr void unroll_loop(F &&f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T> struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
+};
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+#ifdef __AVX512FP16__
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __m128h reg;
+
+  explicit FP16Vec8(_Float16 v) : reg(_mm_set1_ph(v)) {}
+
+  explicit FP16Vec8(const void *ptr) : reg(_mm_loadu_ph(ptr)) {}
+
+  explicit FP16Vec8(__m128h data) : reg(data) {}
+
+  FP16Vec8 operator*(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_mul_ph(reg, b.reg));
+  }
+
+  FP16Vec8 operator+(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_add_ph(reg, b.reg));
+  }
+
+  FP16Vec8 operator-(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_sub_ph(reg, b.reg));
+  }
+
+  FP16Vec8 operator/(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_div_ph(reg, b.reg));
+  }
+
+  void save(void *ptr) const { _mm_storeu_ph(ptr, reg); }
+};
+#endif
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __m128i reg;
+
+  explicit BF16Vec8(const void *ptr)
+      : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {}
+
+  explicit BF16Vec8(const FP32Vec8 &);
+
+  void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  __m256i reg;
+
+  explicit BF16Vec16(const void *ptr)
+      : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {}
+
+  explicit BF16Vec16(const FP32Vec16 &);
+
+  void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; }
+};
+
+#ifdef __AVX512F__
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+
+  __m512i reg;
+
+  explicit BF16Vec32(const void *ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
+
+  explicit BF16Vec32(__m512i data) : reg(data) {}
+
+  explicit BF16Vec32(BF16Vec8 &vec8_data)
+      : reg((__m512i)_mm512_inserti32x4(
+            _mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(
+                                                      (__m128i)vec8_data.reg),
+                                                  (__m128i)vec8_data.reg, 1),
+                               (__m128i)vec8_data.reg, 2),
+            (__m128i)vec8_data.reg, 3)) {}
+
+  void save(void *ptr) const { *reinterpret_cast<__m512i *>(ptr) = reg; }
+};
+#else
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+
+  __m256i reg_low;
+  __m256i reg_high;
+
+  explicit BF16Vec32(const void *ptr)
+      : reg_low(_mm256_loadu_si256((__m256i const *)ptr)),
+        reg_high(_mm256_loadu_si256((__m256i const *)ptr + 1)) {}
+
+  explicit BF16Vec32(__m256i low, __m256i high) : reg_low(low),
+                                                  reg_high(high) {}
+
+  explicit BF16Vec32(BF16Vec8 &vec8_data)
+      : reg_low((__m256i)_mm256_inserti32x4(
+                _mm256_castsi128_si256((__m128i)vec8_data.reg),
+                                       (__m128i)vec8_data.reg, 1)),
+        reg_high((__m256i)_mm256_inserti32x4(
+                _mm256_castsi128_si256((__m128i)vec8_data.reg),
+                                       (__m128i)vec8_data.reg, 1)) {}
+
+  void save(void *ptr) const {
+    *reinterpret_cast<__m256i *>(ptr) = reg_low;
+    *reinterpret_cast<__m256i *>((__m256i *)ptr + 1) = reg_high;
+  }
+};
+#endif
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+  union AliasReg {
+    __m128 reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __m128 reg;
+
+  explicit FP32Vec4(float v) : reg(_mm_set1_ps(v)) {}
+
+  explicit FP32Vec4() : reg(_mm_set1_ps(0.0)) {}
+
+  explicit FP32Vec4(const float *ptr) : reg(_mm_loadu_ps(ptr)) {}
+
+  explicit FP32Vec4(__m128 data) : reg(data) {}
+
+  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {}
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  union AliasReg {
+    __m256 reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __m256 reg;
+
+  explicit FP32Vec8(float v) : reg(_mm256_set1_ps(v)) {}
+
+  explicit FP32Vec8() : reg(_mm256_set1_ps(0.0)) {}
+
+  explicit FP32Vec8(const float *ptr) : reg(_mm256_loadu_ps(ptr)) {}
+
+  explicit FP32Vec8(__m256 data) : reg(data) {}
+
+  explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {}
+
+#ifdef __AVX512FP16__
+  explicit FP32Vec8(__m128h v) : reg(_mm256_cvtph_ps(_mm_castph_si128(v))) {}
+#endif
+
+  explicit FP32Vec8(const BF16Vec8 &v)
+      : reg(_mm256_castsi256_ps(
+            _mm256_bslli_epi128(_mm256_cvtepu16_epi32(v.reg), 2))) {}
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
+
+    return result;
+  }
+
+  FP32Vec8 exp() const {
+    AliasReg ar;
+    ar.reg = reg;
+    return FP32Vec8(_mm256_set_ps(expf(ar.values[7]), expf(ar.values[6]),
+                                  expf(ar.values[5]), expf(ar.values[4]),
+                                  expf(ar.values[3]), expf(ar.values[2]),
+                                  expf(ar.values[1]), expf(ar.values[0])));
+  }
+
+  FP32Vec8 tanh() const {
+    AliasReg ar;
+    ar.reg = reg;
+    return FP32Vec8(_mm256_set_ps(tanhf(ar.values[7]), tanhf(ar.values[6]),
+                                  tanhf(ar.values[5]), tanhf(ar.values[4]),
+                                  tanhf(ar.values[3]), tanhf(ar.values[2]),
+                                  tanhf(ar.values[1]), tanhf(ar.values[0])));
+  }
+
+  FP32Vec8 er() const {
+    AliasReg ar;
+    ar.reg = reg;
+    return FP32Vec8(_mm256_set_ps(erf(ar.values[7]), erf(ar.values[6]),
+                                  erf(ar.values[5]), erf(ar.values[4]),
+                                  erf(ar.values[3]), erf(ar.values[2]),
+                                  erf(ar.values[1]), erf(ar.values[0])));
+  }
+
+  FP32Vec8 operator*(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_mul_ps(reg, b.reg));
+  }
+
+  FP32Vec8 operator+(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_add_ps(reg, b.reg));
+  }
+
+  FP32Vec8 operator-(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_sub_ps(reg, b.reg));
+  }
+
+  FP32Vec8 operator/(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_div_ps(reg, b.reg));
+  }
+
+  void save(float *ptr) const { _mm256_storeu_ps(ptr, reg); }
+};
+
+#ifdef __AVX512F__
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    __m512 reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __m512 reg;
+
+  explicit FP32Vec16(float v) : reg(_mm512_set1_ps(v)) {}
+
+  explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
+
+  explicit FP32Vec16(const float *ptr) : reg(_mm512_loadu_ps(ptr)) {}
+
+  explicit FP32Vec16(__m512 data) : reg(data) {}
+
+  explicit FP32Vec16(const FP32Vec16 &data) : reg(data.reg) {}
+
+  explicit FP32Vec16(const FP32Vec4 &data)
+      : reg((__m512)_mm512_inserti32x4(
+            _mm512_inserti32x4(
+                _mm512_inserti32x4(_mm512_castsi128_si512((__m128i)data.reg),
+                                   (__m128i)data.reg, 1),
+                (__m128i)data.reg, 2),
+            (__m128i)data.reg, 3)) {}
+
+  explicit FP32Vec16(const FP32Vec8 &data)
+      : reg((__m512)_mm512_inserti32x8(
+            _mm512_castsi256_si512((__m256i)data.reg), (__m256i)data.reg, 1)) {}
+
+  explicit FP32Vec16(const BF16Vec16 &v)
+      : reg(_mm512_castsi512_ps(
+            _mm512_bslli_epi128(_mm512_cvtepu16_epi32(v.reg), 2))) {}
+
+  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+
+  FP32Vec16 operator*(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_mul_ps(reg, b.reg));
+  }
+
+  FP32Vec16 operator+(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_add_ps(reg, b.reg));
+  }
+
+  FP32Vec16 operator-(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_sub_ps(reg, b.reg));
+  }
+
+  FP32Vec16 operator/(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_div_ps(reg, b.reg));
+  }
+
+  float reduce_sum() const { return _mm512_reduce_add_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));
+    __mmask16 mask = _cvtu32_mask16(base_mask << (idx * group_size));
+    return _mm512_mask_reduce_add_ps(mask, reg);
+  }
+
+  void save(float *ptr) const { _mm512_storeu_ps(ptr, reg); }
+};
+#else
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  union AliasReg {
+    __m256 reg;
+    float values[8];
+  };
+
+  __m256 reg_low;
+  __m256 reg_high;
+
+  explicit FP32Vec16(float v) : reg_low(_mm256_set1_ps(v)),
+                                reg_high(_mm256_set1_ps(v)) {}
+
+  explicit FP32Vec16() : reg_low(_mm256_set1_ps(0.0)),
+                         reg_high(_mm256_set1_ps(0.0)) {}
+
+  explicit FP32Vec16(const float *ptr) : reg_low(_mm256_loadu_ps(ptr)),
+                                         reg_high(_mm256_loadu_ps(ptr + 8)) {}
+
+  explicit FP32Vec16(__m256 low, __m256 high) : reg_low(low), reg_high(high) {}
+
+  explicit FP32Vec16(const FP32Vec16 &data) : reg_low(data.reg_low),
+                                              reg_high(data.reg_high) {}
+
+  explicit FP32Vec16(const FP32Vec4 &data)
+      : reg_low((__m256)_mm256_inserti128_si256(
+                _mm256_castsi128_si256((__m128i)data.reg),
+                                       (__m128i)data.reg, 1)),
+        reg_high((__m256)_mm256_inserti128_si256(
+                 _mm256_castsi128_si256((__m128i)data.reg),
+                                       (__m128i)data.reg, 1)) {}
+
+  explicit FP32Vec16(const FP32Vec8 &data)
+      : reg_low(data.reg), reg_high(data.reg) {}
+
+  explicit FP32Vec16(const BF16Vec16 &v) {
+    __m128i low = _mm256_extractf128_si256(v.reg, 0);
+    __m128i high = _mm256_extractf128_si256(v.reg, 1);
+
+    __m256i v_low_epi32 = _mm256_cvtepu16_epi32(low);
+    __m256i v_high_epi32 = _mm256_cvtepu16_epi32(high);
+
+    __m256i v_low_shifted = _mm256_bslli_epi128(v_low_epi32, 2);
+    __m256i v_high_shifted = _mm256_bslli_epi128(v_high_epi32, 2);
+
+    reg_low = _mm256_castsi256_ps(v_low_shifted);
+    reg_high = _mm256_castsi256_ps(v_high_shifted);
+  }
+
+  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+
+  FP32Vec16 operator*(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm256_mul_ps(reg_low, b.reg_low),
+                     _mm256_mul_ps(reg_high, b.reg_high));
+  }
+
+  FP32Vec16 operator+(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm256_add_ps(reg_low, b.reg_low),
+                     _mm256_add_ps(reg_high, b.reg_high));
+  }
+
+  FP32Vec16 operator-(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm256_sub_ps(reg_low, b.reg_low),
+                     _mm256_sub_ps(reg_high, b.reg_high));
+  }
+
+  FP32Vec16 operator/(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm256_div_ps(reg_low, b.reg_low),
+                     _mm256_div_ps(reg_high, b.reg_high));
+  }
+
+  float reduce_sum() const {
+    FP32Vec8 low = FP32Vec8(reg_low);
+    FP32Vec8 high = FP32Vec8(reg_high);
+    return low.reduce_sum() + high.reduce_sum();
+  }
+
+  template <int group_size> float reduce_sub_sum(int idx) {
+    float sum = 0.0;
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+    constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size));
+    uint32_t mask = base_mask << (idx * group_size);
+
+    AliasReg ar;
+
+    auto func = [&sum, &mask, &ar](int i) {
+      int flag = mask & 0x1;
+      mask = mask >> 1;
+      if (flag != 0) sum += ar.values[i];
+    };
+
+    ar.reg = reg_low;
+    unroll_loop<int, 8>(func);
+
+    ar.reg = reg_high;
+    unroll_loop<int, 8>(func);
+
+    return sum;
+  }
+
+  void save(float *ptr) const {
+    _mm256_storeu_ps(ptr, reg_low);
+    _mm256_storeu_ps(ptr + 8, reg_high);
+  }
+};
+#endif
+
+template <typename T> struct VecType { using vec_type = void; };
+
+template <typename T> using vec_t = typename VecType<T>::vec_type;
+
+template <> struct VecType<float> { using vec_type = FP32Vec8; };
+
+#ifdef __AVX512FP16__
+template <> struct VecType<c10::Half> { using vec_type = FP16Vec16; };
+#endif
+
+template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+
+template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
+
+#ifdef __AVX512FP16__
+template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
+  *reinterpret_cast<_Float16 *>(ptr) = v;
+}
+#endif
+
+inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
+  acc = acc + a * b;
+}
+
+#ifdef __AVX512BF16__
+template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
+  *reinterpret_cast<__bfloat16 *>(ptr) = _mm_cvtness_sbh(v);
+}
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+    : reg((__m128i)_mm256_cvtneps_pbh(v.reg)) {}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
+    : reg((__m256i)_mm512_cvtneps_pbh(v.reg)) {}
+
+inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
+  acc.reg = _mm512_dpbf16_ps(acc.reg, (__m512bh)a.reg, (__m512bh)b.reg);
+}
+#else
+template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
+  c10::BFloat16 __attribute__((__may_alias__)) *v_ptr =
+      reinterpret_cast<c10::BFloat16 *>(&v);
+  *ptr = *(v_ptr + 1);
+}
+
+#ifdef __AVX512F__
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+    : reg(_mm256_cvtepi32_epi16(
+          _mm256_bsrli_epi128(_mm256_castps_si256(v.reg), 2))) {}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
+    : reg(_mm512_cvtepi32_epi16(
+          _mm512_bsrli_epi128(_mm512_castps_si512(v.reg), 2))) {}
+#else
+namespace{
+__m128i FP32Vec8_to_BF16Vec8_avx2(__m256 a) {
+  __m256i ai = _mm256_castps_si256(a);
+  ai = _mm256_srli_epi32(ai, 16);
+  ai = _mm256_packus_epi32(ai, ai);
+  ai = _mm256_permute4x64_epi64(ai, 0b00111001);
+  return _mm256_extracti128_si256(ai, 0);
+}
+}
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+    : reg(FP32Vec8_to_BF16Vec8_avx2(v.reg)) {}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) {
+  BF16Vec8 low = BF16Vec8(FP32Vec8(v.reg_low));
+  BF16Vec8 high = BF16Vec8(FP32Vec8(v.reg_high));
+  reg = _mm256_insertf128_si256(_mm256_castsi128_si256(low.reg), high.reg, 1);
+}
+#endif // __AVX512F__
+#endif // __AVX512BF16__
+
+inline void prefetch(const void *addr) { _mm_prefetch(addr, _MM_HINT_T1); }
+
+}; // namespace vec_op
+
+#endif

csrc/cpu/torch_bindings.cpp

@@ -58,6 +58,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("gelu_fast(Tensor! out, Tensor input) -> ()");
   ops.impl("gelu_fast", torch::kCPU, &gelu_fast);
 
+  // Quick GELU implementation.
+  ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
+  ops.impl("gelu_quick", torch::kCPU, &gelu_quick);
+
   // Layernorm
   // Apply Root Mean Square (RMS) Normalization to the input tensor.
   ops.def(

csrc/ops.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <optional>
 #include <torch/library.h>
 
 void paged_attention_v1(
@@ -49,6 +50,8 @@ void gelu_new(torch::Tensor& out, torch::Tensor& input);
 
 void gelu_fast(torch::Tensor& out, torch::Tensor& input);
 
+void gelu_quick(torch::Tensor& out, torch::Tensor& input);
+
 #ifndef USE_ROCM
 torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
                         const torch::Tensor& codebooks,
@@ -90,9 +93,17 @@ torch::Tensor gptq_marlin_repack(torch::Tensor& b_q_weight, torch::Tensor& perm,
                                  int64_t size_k, int64_t size_n,
                                  int64_t num_bits);
 
+torch::Tensor fp8_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
+                              torch::Tensor& b_scales, torch::Tensor& workspace,
+                              int64_t num_bits, int64_t size_m, int64_t size_n,
+                              int64_t size_k);
+
+bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
+
 void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,
                        torch::Tensor const& b, torch::Tensor const& a_scales,
-                       torch::Tensor const& b_scales);
+                       torch::Tensor const& b_scales,
+                       c10::optional<torch::Tensor> const& bias);
 
 #endif
 

csrc/punica/bgmv/bgmv_config.h

@@ -16,14 +16,20 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 512) \
     f(in_T, out_T, W_T, narrow, 640) \
     f(in_T, out_T, W_T, narrow, 768) \
+    f(in_T, out_T, W_T, narrow, 896) \
     f(in_T, out_T, W_T, narrow, 1024) \
     f(in_T, out_T, W_T, narrow, 1152) \
+    f(in_T, out_T, W_T, narrow, 1216) \
     f(in_T, out_T, W_T, narrow, 1280) \
     f(in_T, out_T, W_T, narrow, 1536) \
+    f(in_T, out_T, W_T, narrow, 1664) \
     f(in_T, out_T, W_T, narrow, 1728) \
     f(in_T, out_T, W_T, narrow, 1792) \
     f(in_T, out_T, W_T, narrow, 2048) \
+    f(in_T, out_T, W_T, narrow, 2240) \
     f(in_T, out_T, W_T, narrow, 2304) \
+    f(in_T, out_T, W_T, narrow, 2368) \
+    f(in_T, out_T, W_T, narrow, 2432) \
     f(in_T, out_T, W_T, narrow, 2560) \
     f(in_T, out_T, W_T, narrow, 2752) \
     f(in_T, out_T, W_T, narrow, 2816) \
@@ -31,32 +37,47 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 3328) \
     f(in_T, out_T, W_T, narrow, 3456) \
     f(in_T, out_T, W_T, narrow, 3584) \
+    f(in_T, out_T, W_T, narrow, 3712) \
     f(in_T, out_T, W_T, narrow, 4096) \
+    f(in_T, out_T, W_T, narrow, 4480) \
     f(in_T, out_T, W_T, narrow, 4608) \
+    f(in_T, out_T, W_T, narrow, 4736) \
+    f(in_T, out_T, W_T, narrow, 4864) \
     f(in_T, out_T, W_T, narrow, 5120) \
     f(in_T, out_T, W_T, narrow, 5504) \
     f(in_T, out_T, W_T, narrow, 5632) \
+    f(in_T, out_T, W_T, narrow, 5888) \
     f(in_T, out_T, W_T, narrow, 6144) \
     f(in_T, out_T, W_T, narrow, 6400) \
     f(in_T, out_T, W_T, narrow, 6848) \
     f(in_T, out_T, W_T, narrow, 6912) \
     f(in_T, out_T, W_T, narrow, 7168) \
+    f(in_T, out_T, W_T, narrow, 7424) \
     f(in_T, out_T, W_T, narrow, 8192) \
+    f(in_T, out_T, W_T, narrow, 8960) \
     f(in_T, out_T, W_T, narrow, 9216) \
+    f(in_T, out_T, W_T, narrow, 9472) \
     f(in_T, out_T, W_T, narrow, 10240) \
     f(in_T, out_T, W_T, narrow, 11008) \
+    f(in_T, out_T, W_T, narrow, 11264) \
     f(in_T, out_T, W_T, narrow, 12288) \
     f(in_T, out_T, W_T, narrow, 13696) \
     f(in_T, out_T, W_T, narrow, 13824) \
     f(in_T, out_T, W_T, narrow, 14336) \
+    f(in_T, out_T, W_T, narrow, 14784) \
+    f(in_T, out_T, W_T, narrow, 14848) \
     f(in_T, out_T, W_T, narrow, 15360) \
     f(in_T, out_T, W_T, narrow, 16384) \
+    f(in_T, out_T, W_T, narrow, 18944) \
     f(in_T, out_T, W_T, narrow, 20480) \
     f(in_T, out_T, W_T, narrow, 22016) \
+    f(in_T, out_T, W_T, narrow, 22528) \
     f(in_T, out_T, W_T, narrow, 24576) \
     f(in_T, out_T, W_T, narrow, 27392) \
     f(in_T, out_T, W_T, narrow, 27648) \
     f(in_T, out_T, W_T, narrow, 28672) \
+    f(in_T, out_T, W_T, narrow, 29568) \
+    f(in_T, out_T, W_T, narrow, 29696) \
     f(in_T, out_T, W_T, narrow, 32000) \
     f(in_T, out_T, W_T, narrow, 32256) \
     f(in_T, out_T, W_T, narrow, 32512) \
@@ -65,6 +86,9 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 36864) \
     f(in_T, out_T, W_T, narrow, 43264) \
     f(in_T, out_T, W_T, narrow, 49152) \
+    f(in_T, out_T, W_T, narrow, 49408) \
+    f(in_T, out_T, W_T, narrow, 60544) \
+    f(in_T, out_T, W_T, narrow, 60672) \
     f(in_T, out_T, W_T, narrow, 64000) \
     f(in_T, out_T, W_T, narrow, 64256) \
     f(in_T, out_T, W_T, narrow, 64512) \
@@ -74,12 +98,14 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 128000) \
     f(in_T, out_T, W_T, narrow, 128256) \
     f(in_T, out_T, W_T, narrow, 128512) \
+    
+    
 // Keep above in sync with vllm/lora/layers::LogitsProcessorWithLoRA
 // and vllm/tests/lora/test_punica.py
 
-// Used for defining kernels going from the variety of 
+// Used for defining kernels going from the variety of
 // dim in to the narrow dim out
-    // Using it for the fully sharded column 
+    // Using it for the fully sharded column
     // parallel LoRA A which splits the rank dim
 #define FOR_INST_BGMV_NARROW(f, in_T, out_T, W_T, narrow) \
     f(in_T, out_T, W_T, 128, narrow) \
@@ -87,14 +113,20 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 512, narrow) \
     f(in_T, out_T, W_T, 640, narrow) \
     f(in_T, out_T, W_T, 768, narrow) \
+    f(in_T, out_T, W_T, 896, narrow) \
     f(in_T, out_T, W_T, 1024, narrow) \
     f(in_T, out_T, W_T, 1152, narrow) \
+    f(in_T, out_T, W_T, 1216, narrow) \
     f(in_T, out_T, W_T, 1280, narrow) \
     f(in_T, out_T, W_T, 1536, narrow) \
+    f(in_T, out_T, W_T, 1664, narrow) \
     f(in_T, out_T, W_T, 1728, narrow) \
     f(in_T, out_T, W_T, 1792, narrow) \
     f(in_T, out_T, W_T, 2048, narrow) \
+    f(in_T, out_T, W_T, 2240, narrow) \
     f(in_T, out_T, W_T, 2304, narrow) \
+    f(in_T, out_T, W_T, 2368, narrow) \
+    f(in_T, out_T, W_T, 2432, narrow) \
     f(in_T, out_T, W_T, 2560, narrow) \
     f(in_T, out_T, W_T, 2752, narrow) \
     f(in_T, out_T, W_T, 2816, narrow) \
@@ -102,32 +134,47 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 3328, narrow) \
     f(in_T, out_T, W_T, 3456, narrow) \
     f(in_T, out_T, W_T, 3584, narrow) \
+    f(in_T, out_T, W_T, 3712, narrow) \
     f(in_T, out_T, W_T, 4096, narrow) \
+    f(in_T, out_T, W_T, 4480, narrow) \
     f(in_T, out_T, W_T, 4608, narrow) \
+    f(in_T, out_T, W_T, 4736, narrow) \
+    f(in_T, out_T, W_T, 4864, narrow) \
     f(in_T, out_T, W_T, 5120, narrow) \
     f(in_T, out_T, W_T, 5504, narrow) \
     f(in_T, out_T, W_T, 5632, narrow) \
+    f(in_T, out_T, W_T, 5888, narrow) \
     f(in_T, out_T, W_T, 6144, narrow) \
     f(in_T, out_T, W_T, 6400, narrow) \
     f(in_T, out_T, W_T, 6848, narrow) \
     f(in_T, out_T, W_T, 6912, narrow) \
     f(in_T, out_T, W_T, 7168, narrow) \
+    f(in_T, out_T, W_T, 7424, narrow) \
     f(in_T, out_T, W_T, 8192, narrow) \
+    f(in_T, out_T, W_T, 8960, narrow) \
     f(in_T, out_T, W_T, 9216, narrow) \
+    f(in_T, out_T, W_T, 9472, narrow) \
     f(in_T, out_T, W_T, 10240, narrow) \
     f(in_T, out_T, W_T, 11008, narrow) \
+    f(in_T, out_T, W_T, 11264, narrow) \
     f(in_T, out_T, W_T, 12288, narrow) \
     f(in_T, out_T, W_T, 13696, narrow) \
     f(in_T, out_T, W_T, 13824, narrow) \
     f(in_T, out_T, W_T, 14336, narrow) \
+    f(in_T, out_T, W_T, 14784, narrow) \
+    f(in_T, out_T, W_T, 14848, narrow) \
     f(in_T, out_T, W_T, 15360, narrow) \
     f(in_T, out_T, W_T, 16384, narrow) \
+    f(in_T, out_T, W_T, 18944, narrow) \
     f(in_T, out_T, W_T, 20480, narrow) \
     f(in_T, out_T, W_T, 22016, narrow) \
+    f(in_T, out_T, W_T, 22528, narrow) \
     f(in_T, out_T, W_T, 24576, narrow) \
     f(in_T, out_T, W_T, 27392, narrow) \
     f(in_T, out_T, W_T, 27648, narrow) \
     f(in_T, out_T, W_T, 28672, narrow) \
+    f(in_T, out_T, W_T, 29568, narrow) \
+    f(in_T, out_T, W_T, 29696, narrow) \
     f(in_T, out_T, W_T, 32000, narrow) \
     f(in_T, out_T, W_T, 32256, narrow) \
     f(in_T, out_T, W_T, 32512, narrow) \
@@ -136,6 +183,9 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 36864, narrow) \
     f(in_T, out_T, W_T, 43264, narrow) \
     f(in_T, out_T, W_T, 49152, narrow) \
+    f(in_T, out_T, W_T, 49408, narrow) \
+    f(in_T, out_T, W_T, 60544, narrow) \
+    f(in_T, out_T, W_T, 60672, narrow) \
     f(in_T, out_T, W_T, 64000, narrow) \
     f(in_T, out_T, W_T, 64256, narrow) \
     f(in_T, out_T, W_T, 64512, narrow) \

csrc/quantization/cutlass_w8a8/broadcast_load_epilogue_c3x.hpp

@@ -153,7 +153,7 @@ struct Sm90RowOrScalarBroadcast {
 
     CUTLASS_DEVICE void
     begin(uint64_t* full_mbarrier_ptr, int load_iteration, bool issue_tma_load) {
-      if (params.ptr_row == nullptr) {
+      if (!params.row_broadcast) {
         return;
       }
 

csrc/quantization/cutlass_w8a8/common.hpp

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "cutlass/cutlass.h"
+#include <climits>
 
 /**
  * Helper function for checking CUTLASS errors
@@ -10,3 +11,17 @@
     TORCH_CHECK(status == cutlass::Status::kSuccess, \
                 cutlassGetStatusString(status))      \
   }
+
+inline uint32_t next_pow_2(uint32_t const num) {
+  if (num <= 1) return num;
+  return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
+}
+
+inline int get_cuda_max_shared_memory_per_block_opt_in(int const device) {
+  int max_shared_mem_per_block_opt_in = 0;
+  cudaDeviceGetAttribute(&max_shared_mem_per_block_opt_in,
+                        cudaDevAttrMaxSharedMemoryPerBlockOptin,
+                        device);
+  return max_shared_mem_per_block_opt_in;
+}
+

csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu

@@ -77,24 +77,12 @@ struct enable_sm89_to_sm90 : Kernel {
 };
 
 /*
-   This epilogue function defines a quantized GEMM operation similar to
-   torch._scaled_mm.
-
-   A and B may be both either int8 or fp8_e4m3. A can be quantized per-tensor or
-   per-row. B can be quantized per-tensor or per-column.
-   Any combination of per-tensor and per-row or column is supported.
-   A and B must have symmetric quantization (zero point == 0).
-
-   So the GEMM operation is D = (a_scales * A) (b_scales * B), where the
-   scales are applied elementwise with numpy-style broadcasting.
-
-   ScaleA and ScaleB define the epilogue functions that apply the scales for
-   the A and B operands respectively. These scales may be either per-tensor or
-   per row or column.
-*/
+ * This class provides the common ScaleA and ScaleB descriptors for the
+ * ScaledEpilogue and ScaledEpilogueBias classes.
+ */
 template <typename ElementD, typename OutputTileThreadMap>
-struct ScaledEpilogue {
- private:
+struct ScaledEpilogueBase {
+ protected:
   using Accum = cutlass::epilogue::threadblock::VisitorAccFetch;
 
   using ScaleA = cutlass::epilogue::threadblock::VisitorColOrScalarBroadcast<
@@ -102,6 +90,32 @@ struct ScaledEpilogue {
 
   using ScaleB = cutlass::epilogue::threadblock::VisitorRowOrScalarBroadcast<
       OutputTileThreadMap, float, Stride<Int<0>, Int<1>, Int<0>>>;
+};
+
+/*
+ This epilogue function defines a quantized GEMM operation similar to
+ torch._scaled_mm.
+
+ A and B may be both either int8 or fp8_e4m3. A can be quantized per-tensor or
+ per-row. B can be quantized per-tensor or per-column.
+ Any combination of per-tensor and per-row or column is supported.
+ A and B must have symmetric quantization (zero point == 0).
+
+ So the GEMM operation is D = (a_scales * A) (b_scales * B), where the
+ scales are applied elementwise with numpy-style broadcasting.
+
+ ScaleA and ScaleB define the epilogue functions that apply the scales for
+ the A and B operands respectively. These scales may be either per-tensor or
+ per row or column.
+*/
+template <typename ElementD, typename OutputTileThreadMap>
+struct ScaledEpilogue
+    : private ScaledEpilogueBase<ElementD, OutputTileThreadMap> {
+ private:
+  using SUPER = ScaledEpilogueBase<ElementD, OutputTileThreadMap>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::ScaleA;
+  using ScaleB = typename SUPER::ScaleB;
 
   using Compute0 = cutlass::epilogue::threadblock::VisitorCompute<
       cutlass::multiplies, float, float,
@@ -134,6 +148,53 @@ struct ScaledEpilogue {
   }
 };
 
+template <typename ElementD, typename OutputTileThreadMap>
+struct ScaledEpilogueBias
+    : private ScaledEpilogueBase<ElementD, OutputTileThreadMap> {
+ private:
+  using SUPER = ScaledEpilogueBase<ElementD, OutputTileThreadMap>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::ScaleA;
+  using ScaleB = typename SUPER::ScaleB;
+
+  using Compute0 = cutlass::epilogue::threadblock::VisitorCompute<
+      cutlass::multiplies, float, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using EVTCompute0 =
+      cutlass::epilogue::threadblock::Sm80EVT<Compute0, ScaleB, Accum>;
+
+  using Compute1 = cutlass::epilogue::threadblock::VisitorCompute<
+      cutlass::multiply_add, ElementD, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using Bias = cutlass::epilogue::threadblock::VisitorRowBroadcast<
+      OutputTileThreadMap, ElementD, Stride<Int<0>, Int<1>, Int<0>>>;
+
+ public:
+  using EVTCompute = cutlass::epilogue::threadblock::Sm80EVT<Compute1, ScaleA,
+                                                             EVTCompute0, Bias>;
+  using ArgumentType = typename EVTCompute::Arguments;
+
+  static ArgumentType prepare_args(torch::Tensor const& a_scales,
+                                   torch::Tensor const& b_scales,
+                                   torch::Tensor const& bias) {
+    using ScaleAArgs = typename ScaleA::Arguments;
+    using ScaleBArgs = typename ScaleB::Arguments;
+    using BiasArgs = typename Bias::Arguments;
+
+    ScaleBArgs b_args{b_scales.data_ptr<float>(), b_scales.numel() != 1, {}};
+    ScaleAArgs a_args{a_scales.data_ptr<float>(), a_scales.numel() != 1, {}};
+    BiasArgs bias_args{static_cast<ElementD*>(bias.data_ptr()), {}};
+
+    typename EVTCompute0::Arguments evt0_compute_args{b_args};
+
+    typename EVTCompute::Arguments evt_compute_args{a_args, evt0_compute_args,
+                                                    bias_args};
+    return evt_compute_args;
+  }
+};
+
 template <typename Arch, template <typename> typename ArchGuard,
           typename ElementAB_, typename ElementD_,
           template <typename, typename> typename Epilogue_, typename TileShape,
@@ -168,13 +229,13 @@ struct cutlass_2x_gemm {
   // clang-format off
   using RowMajor = typename cutlass::layout::RowMajor;
   using ColumnMajor = typename cutlass::layout::ColumnMajor;
-  using KernelType = 
+  using KernelType =
     ArchGuard<typename cutlass::gemm::kernel::DefaultGemmWithVisitor<
-      ElementAB, RowMajor, cutlass::ComplexTransform::kNone, 16, 
-      ElementAB, ColumnMajor, cutlass::ComplexTransform::kNone, 16, 
+      ElementAB, RowMajor, cutlass::ComplexTransform::kNone, 16,
+      ElementAB, ColumnMajor, cutlass::ComplexTransform::kNone, 16,
       float, cutlass::layout::RowMajor, 4,
-      ElementAcc, float, cutlass::arch::OpClassTensorOp, 
-      Arch, 
+      ElementAcc, float, cutlass::arch::OpClassTensorOp,
+      Arch,
       TileShape, WarpShape, InstructionShape,
       EVTD,
       cutlass::gemm::threadblock::ThreadblockSwizzleStreamK,
@@ -250,16 +311,167 @@ void cutlass_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
   CUTLASS_CHECK(status);
 }
 
+template <typename Gemm, typename FallbackGemm, typename... EpilogueArgs>
+void fallback_cutlass_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
+                                  torch::Tensor const& b,
+                                  EpilogueArgs&&... args) {
+  // In some cases, the GPU isn't able to accommodate the
+  // shared memory requirements of the Gemm. In such cases, use
+  // the FallbackGemm instead.
+  static const int max_shared_mem_per_block_opt_in =
+      get_cuda_max_shared_memory_per_block_opt_in(0);
+
+  size_t const gemm_shared_mem_size =
+      sizeof(typename Gemm::KernelType::SharedStorage);
+  size_t const fallback_gemm_shared_mem_size =
+      sizeof(typename FallbackGemm::KernelType::SharedStorage);
+
+  if (gemm_shared_mem_size <= max_shared_mem_per_block_opt_in) {
+    return cutlass_gemm_caller<Gemm>(out, a, b,
+                                     std::forward<EpilogueArgs>(args)...);
+  } else {
+    TORCH_CHECK(fallback_gemm_shared_mem_size <=
+                max_shared_mem_per_block_opt_in);
+    return cutlass_gemm_caller<FallbackGemm>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+}
+
+template <typename InType, typename OutType,
+          template <typename, typename> typename Epilogue>
+struct sm80_config_default {
+  // This config is used in 2 cases,
+  //  - M in (128, inf)
+  //  - M in (64, 128] and N >= 8192
+  // Shared Memory required by this Gemm - 81920 bytes
+  static_assert(std::is_same<InType, int8_t>());
+  using TileShape = typename cutlass::gemm::GemmShape<128, 128, 64>;
+  using WarpShape = typename cutlass::gemm::GemmShape<64, 64, 64>;
+  using InstructionShape = typename cutlass::gemm::GemmShape<16, 8, 32>;
+  using Cutlass2xGemm =
+      cutlass_2x_gemm<cutlass::arch::Sm80, enable_sm80_to_sm89, InType, OutType,
+                      Epilogue, TileShape, WarpShape, InstructionShape, 5>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename> typename Epilogue>
+struct sm80_config_M64 {
+  // This config is used in 2 cases,
+  // - M in (32, 64]
+  // - M in (64, 128] and N < 8192
+  // Shared Memory required by this Gemm - 122880 bytes
+  static_assert(std::is_same<InType, int8_t>());
+  using TileShape = typename cutlass::gemm::GemmShape<64, 128, 128>;
+  using WarpShape = typename cutlass::gemm::GemmShape<64, 64, 64>;
+  using InstructionShape = typename cutlass::gemm::GemmShape<16, 8, 32>;
+  using Cutlass2xGemm =
+      cutlass_2x_gemm<cutlass::arch::Sm80, enable_sm80_to_sm89, InType, OutType,
+                      Epilogue, TileShape, WarpShape, InstructionShape, 5>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename> typename Epilogue>
+struct sm80_config_M32 {
+  // M in (16, 32]
+  // Shared Memory required by this Gemm - 61440 bytes
+  static_assert(std::is_same<InType, int8_t>());
+  using TileShape = typename cutlass::gemm::GemmShape<32, 64, 128>;
+  using WarpShape = typename cutlass::gemm::GemmShape<32, 64, 64>;
+  using InstructionShape = typename cutlass::gemm::GemmShape<16, 8, 32>;
+  using Cutlass2xGemm =
+      cutlass_2x_gemm<cutlass::arch::Sm80, enable_sm80_to_sm89, InType, OutType,
+                      Epilogue, TileShape, WarpShape, InstructionShape, 5>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename> typename Epilogue>
+struct sm80_config_M16 {
+  // M in [1, 16]
+  // Shared Memory required by this Gemm - 51200 bytes
+  static_assert(std::is_same<InType, int8_t>());
+  using TileShape = typename cutlass::gemm::GemmShape<16, 64, 128>;
+  using WarpShape = typename cutlass::gemm::GemmShape<16, 64, 64>;
+  using InstructionShape = typename cutlass::gemm::GemmShape<16, 8, 32>;
+  using Cutlass2xGemm =
+      cutlass_2x_gemm<cutlass::arch::Sm80, enable_sm80_to_sm89, InType, OutType,
+                      Epilogue, TileShape, WarpShape, InstructionShape, 5>;
+};
+
 }  // namespace
 
-void cutlass_scaled_mm_sm75(torch::Tensor& out, torch::Tensor const& a,
-                            torch::Tensor const& b,
-                            torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales) {
+template <typename InType, typename OutType,
+          template <typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_gemm_sm80_dispatch(torch::Tensor& out, torch::Tensor const& a,
+                                torch::Tensor const& b,
+                                EpilogueArgs&&... args) {
+  static_assert(std::is_same<InType, int8_t>());
+  TORCH_CHECK(a.dtype() == torch::kInt8);
+  TORCH_CHECK(b.dtype() == torch::kInt8);
+
+  using Cutlass2xGemmDefault =
+      typename sm80_config_default<InType, OutType, Epilogue>::Cutlass2xGemm;
+  using Cutlass2xGemmM128BigN =
+      typename sm80_config_default<InType, OutType, Epilogue>::Cutlass2xGemm;
+  using Cutlass2xGemmM128SmallN =
+      typename sm80_config_M64<InType, OutType, Epilogue>::Cutlass2xGemm;
+  using Cutlass2xGemmM64 =
+      typename sm80_config_M64<InType, OutType, Epilogue>::Cutlass2xGemm;
+  using Cutlass2xGemmM32 =
+      typename sm80_config_M32<InType, OutType, Epilogue>::Cutlass2xGemm;
+  using Cutlass2xGemmM16 =
+      typename sm80_config_M16<InType, OutType, Epilogue>::Cutlass2xGemm;
+
+  // Due to shared memory requirements, some Gemms may fail to run on some
+  // GPUs. As the name indicates, the Fallback Gemm is used as an alternative
+  // in such cases.
+  // sm80_config_M16 has the least shared-memory requirement. However,
+  // based on some profiling, we select sm80_config_M32 as a better alternative
+  // performance wise.
+  using FallbackGemm =
+      typename sm80_config_M32<InType, OutType, Epilogue>::Cutlass2xGemm;
+
+  uint32_t const m = a.size(0);
+  uint32_t const mp2 =
+      std::max(static_cast<uint32_t>(16), next_pow_2(m));  // next power of 2
+  if (mp2 <= 16) {
+    // M in [1, 16]
+    return fallback_cutlass_gemm_caller<Cutlass2xGemmM16, FallbackGemm>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  } else if (mp2 <= 32) {
+    // M in (16, 32]
+    return fallback_cutlass_gemm_caller<Cutlass2xGemmM32, FallbackGemm>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  } else if (mp2 <= 64) {
+    // M in (32, 64]
+    return fallback_cutlass_gemm_caller<Cutlass2xGemmM64, FallbackGemm>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  } else if (mp2 <= 128) {
+    // M in (64, 128]
+    uint32_t const n = out.size(1);
+    bool const small_n = n < 8192;
+    if (small_n) {
+      return fallback_cutlass_gemm_caller<Cutlass2xGemmM128SmallN,
+                                          FallbackGemm>(
+          out, a, b, std::forward<EpilogueArgs>(args)...);
+    } else {
+      return fallback_cutlass_gemm_caller<Cutlass2xGemmM128BigN, FallbackGemm>(
+          out, a, b, std::forward<EpilogueArgs>(args)...);
+    }
+  } else {
+    // M in (128, inf)
+    return fallback_cutlass_gemm_caller<Cutlass2xGemmDefault, FallbackGemm>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+}
+
+template <template <typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_scaled_mm_sm75_epilogue(torch::Tensor& out, torch::Tensor const& a,
+                                     torch::Tensor const& b,
+                                     EpilogueArgs&&... epilogue_args) {
   TORCH_CHECK(a.dtype() == torch::kInt8);
   TORCH_CHECK(b.dtype() == torch::kInt8);
-  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
-  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
 
   using TileShape = typename cutlass::gemm::GemmShape<128, 128, 64>;
   using WarpShape = typename cutlass::gemm::GemmShape<64, 64, 64>;
@@ -268,85 +480,130 @@ void cutlass_scaled_mm_sm75(torch::Tensor& out, torch::Tensor const& a,
   if (out.dtype() == torch::kBFloat16) {
     return cutlass_gemm_caller<cutlass_2x_gemm<
         cutlass::arch::Sm75, enable_sm75_to_sm80, int8_t, cutlass::bfloat16_t,
-        ScaledEpilogue, TileShape, WarpShape, InstructionShape, 2>>(
-        out, a, b, a_scales, b_scales);
+        Epilogue, TileShape, WarpShape, InstructionShape, 2>>(
+        out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
   } else {
     TORCH_CHECK(out.dtype() == torch::kFloat16);
     return cutlass_gemm_caller<cutlass_2x_gemm<
         cutlass::arch::Sm75, enable_sm75_to_sm80, int8_t, cutlass::half_t,
-        ScaledEpilogue, TileShape, WarpShape, InstructionShape, 2>>(
-        out, a, b, a_scales, b_scales);
+        Epilogue, TileShape, WarpShape, InstructionShape, 2>>(
+        out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
+  }
+}
+
+void cutlass_scaled_mm_sm75(torch::Tensor& out, torch::Tensor const& a,
+                            torch::Tensor const& b,
+                            torch::Tensor const& a_scales,
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias) {
+  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
+  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+  if (bias) {
+    TORCH_CHECK(bias->dtype() == out.dtype(),
+                "currently bias dtype must match output dtype ", out.dtype());
+    return cutlass_scaled_mm_sm75_epilogue<ScaledEpilogueBias>(
+        out, a, b, a_scales, b_scales, *bias);
+  } else {
+    return cutlass_scaled_mm_sm75_epilogue<ScaledEpilogue>(out, a, b, a_scales,
+                                                           b_scales);
+  }
+}
+
+template <template <typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_scaled_mm_sm80_epilogue(torch::Tensor& out, torch::Tensor const& a,
+                                     torch::Tensor const& b,
+                                     EpilogueArgs&&... epilogue_args) {
+  TORCH_CHECK(a.dtype() == torch::kInt8);
+  TORCH_CHECK(b.dtype() == torch::kInt8);
+
+  if (out.dtype() == torch::kBFloat16) {
+    return cutlass_gemm_sm80_dispatch<int8_t, cutlass::bfloat16_t, Epilogue>(
+        out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
+  } else {
+    TORCH_CHECK(out.dtype() == torch::kFloat16);
+    return cutlass_gemm_sm80_dispatch<int8_t, cutlass::half_t, Epilogue>(
+        out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
   }
 }
 
 void cutlass_scaled_mm_sm80(torch::Tensor& out, torch::Tensor const& a,
                             torch::Tensor const& b,
                             torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales) {
-  TORCH_CHECK(a.dtype() == torch::kInt8);
-  TORCH_CHECK(b.dtype() == torch::kInt8);
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias) {
   TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
   TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
-
-  using TileShape = typename cutlass::gemm::GemmShape<128, 128, 64>;
-  using WarpShape = typename cutlass::gemm::GemmShape<64, 64, 64>;
-  using InstructionShape = typename cutlass::gemm::GemmShape<16, 8, 32>;
-
-  if (out.dtype() == torch::kBFloat16) {
-    return cutlass_gemm_caller<cutlass_2x_gemm<
-        cutlass::arch::Sm80, enable_sm80_to_sm89, int8_t, cutlass::bfloat16_t,
-        ScaledEpilogue, TileShape, WarpShape, InstructionShape, 5>>(
-        out, a, b, a_scales, b_scales);
+  if (bias) {
+    TORCH_CHECK(bias->dtype() == out.dtype(),
+                "currently bias dtype must match output dtype ", out.dtype());
+    return cutlass_scaled_mm_sm80_epilogue<ScaledEpilogueBias>(
+        out, a, b, a_scales, b_scales, *bias);
   } else {
-    TORCH_CHECK(out.dtype() == torch::kFloat16);
-    return cutlass_gemm_caller<cutlass_2x_gemm<
-        cutlass::arch::Sm80, enable_sm80_to_sm89, int8_t, cutlass::half_t,
-        ScaledEpilogue, TileShape, WarpShape, InstructionShape, 5>>(
-        out, a, b, a_scales, b_scales);
+    return cutlass_scaled_mm_sm80_epilogue<ScaledEpilogue>(out, a, b, a_scales,
+                                                           b_scales);
   }
 }
 
-void cutlass_scaled_mm_sm89(torch::Tensor& out, torch::Tensor const& a,
-                            torch::Tensor const& b,
-                            torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales) {
+template <template <typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_scaled_mm_sm89_epilogue(torch::Tensor& out, torch::Tensor const& a,
+                                     torch::Tensor const& b,
+                                     EpilogueArgs&&... epilogue_args) {
   using TileShape = typename cutlass::gemm::GemmShape<128, 128, 64>;
   using WarpShape = typename cutlass::gemm::GemmShape<64, 64, 64>;
   using InstructionShape = typename cutlass::gemm::GemmShape<16, 8, 32>;
 
-  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
-  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
-
   if (a.dtype() == torch::kInt8) {
     TORCH_CHECK(b.dtype() == torch::kInt8);
 
     if (out.dtype() == torch::kBFloat16) {
       return cutlass_gemm_caller<cutlass_2x_gemm<
           cutlass::arch::Sm89, enable_sm89_to_sm90, int8_t, cutlass::bfloat16_t,
-          ScaledEpilogue, TileShape, WarpShape, InstructionShape, 5>>(
-          out, a, b, a_scales, b_scales);
+          Epilogue, TileShape, WarpShape, InstructionShape, 5>>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     } else {
       assert(out.dtype() == torch::kFloat16);
       return cutlass_gemm_caller<cutlass_2x_gemm<
           cutlass::arch::Sm89, enable_sm89_to_sm90, int8_t, cutlass::half_t,
-          ScaledEpilogue, TileShape, WarpShape, InstructionShape, 5>>(
-          out, a, b, a_scales, b_scales);
+          Epilogue, TileShape, WarpShape, InstructionShape, 5>>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     }
   } else {
     TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
     TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
 
     if (out.dtype() == torch::kBFloat16) {
-      return cutlass_gemm_caller<cutlass_2x_gemm<
-          cutlass::arch::Sm89, enable_sm89_to_sm90, cutlass::float_e4m3_t,
-          cutlass::bfloat16_t, ScaledEpilogue, TileShape, WarpShape,
-          InstructionShape, 5>>(out, a, b, a_scales, b_scales);
+      return cutlass_gemm_caller<
+          cutlass_2x_gemm<cutlass::arch::Sm89, enable_sm89_to_sm90,
+                          cutlass::float_e4m3_t, cutlass::bfloat16_t, Epilogue,
+                          TileShape, WarpShape, InstructionShape, 5>>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     } else {
       TORCH_CHECK(out.dtype() == torch::kFloat16);
-      return cutlass_gemm_caller<cutlass_2x_gemm<
-          cutlass::arch::Sm89, enable_sm89_to_sm90, cutlass::float_e4m3_t,
-          cutlass::half_t, ScaledEpilogue, TileShape, WarpShape,
-          InstructionShape, 5>>(out, a, b, a_scales, b_scales);
+      return cutlass_gemm_caller<
+          cutlass_2x_gemm<cutlass::arch::Sm89, enable_sm89_to_sm90,
+                          cutlass::float_e4m3_t, cutlass::half_t, Epilogue,
+                          TileShape, WarpShape, InstructionShape, 5>>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     }
   }
 }
+
+void cutlass_scaled_mm_sm89(torch::Tensor& out, torch::Tensor const& a,
+                            torch::Tensor const& b,
+                            torch::Tensor const& a_scales,
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias) {
+  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
+  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+  if (bias) {
+    TORCH_CHECK(bias->dtype() == out.dtype(),
+                "currently bias dtype must match output dtype ", out.dtype());
+    return cutlass_scaled_mm_sm89_epilogue<ScaledEpilogueBias>(
+        out, a, b, a_scales, b_scales, *bias);
+  } else {
+    return cutlass_scaled_mm_sm89_epilogue<ScaledEpilogue>(out, a, b, a_scales,
+                                                           b_scales);
+  }
+}

csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu

@@ -44,11 +44,6 @@ using namespace cute;
 
 namespace {
 
-uint32_t next_pow_2(uint32_t const num) {
-  if (num <= 1) return num;
-  return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
-}
-
 // A wrapper for the GEMM kernel that is used to guard against compilation on
 // architectures that will never use the kernel. The purpose of this is to
 // reduce the size of the compiled binary.
@@ -64,6 +59,28 @@ struct enable_sm90_or_later : Kernel {
   }
 };
 
+/*
+ * This class provides the common ScaleA and ScaleB descriptors for the
+ * ScaledEpilogue and ScaledEpilogueBias classes.
+ */
+template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
+struct ScaledEpilogueBase {
+ protected:
+  using Accum = cutlass::epilogue::fusion::Sm90AccFetch;
+
+  using ScaleA = cutlass::epilogue::fusion::Sm90ColOrScalarBroadcast<
+      0 /*Stages*/, typename EpilogueDescriptor::TileShape, float,
+      Stride<Int<1>, Int<0>, Int<0>>>;
+
+  using ScaleBDescriptor =
+      cutlass::epilogue::collective::detail::RowBroadcastDescriptor<
+          EpilogueDescriptor, float>;
+
+  using ScaleB = cutlass::epilogue::fusion::Sm90RowOrScalarBroadcast<
+      ScaleBDescriptor::Stages, typename EpilogueDescriptor::TileShape,
+      typename ScaleBDescriptor::Element, Stride<Int<0>, Int<1>, Int<0>>>;
+};
+
 /*
    This epilogue function defines a quantized GEMM operation similar to
    torch.scaled_mm_.
@@ -81,21 +98,13 @@ struct enable_sm90_or_later : Kernel {
    per row or column.
 */
 template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
-struct ScaledEpilogue {
+struct ScaledEpilogue
+    : private ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor> {
  private:
-  using Accum = cutlass::epilogue::fusion::Sm90AccFetch;
-
-  using ScaleA = cutlass::epilogue::fusion::Sm90ColOrScalarBroadcast<
-      0 /*Stages*/, typename EpilogueDescriptor::TileShape, float,
-      Stride<Int<1>, Int<0>, Int<0>>>;
-
-  using ScaleBDescriptor =
-      cutlass::epilogue::collective::detail::RowBroadcastDescriptor<
-          EpilogueDescriptor, float>;
-
-  using ScaleB = cutlass::epilogue::fusion::Sm90RowOrScalarBroadcast<
-      ScaleBDescriptor::Stages, typename EpilogueDescriptor::TileShape,
-      typename ScaleBDescriptor::Element, Stride<Int<0>, Int<1>, Int<0>>>;
+  using SUPER = ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::ScaleA;
+  using ScaleB = typename SUPER::ScaleB;
 
   using Compute0 = cutlass::epilogue::fusion::Sm90Compute<
       cutlass::multiplies, float, float,
@@ -125,6 +134,54 @@ struct ScaledEpilogue {
   }
 };
 
+template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
+struct ScaledEpilogueBias
+    : private ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor> {
+ private:
+  using SUPER = ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::ScaleA;
+  using ScaleB = typename SUPER::ScaleB;
+
+  using Compute0 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiplies, float, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using EVTCompute0 =
+      cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
+
+  using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiply_add, ElementD, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using BiasDescriptor =
+      cutlass::epilogue::collective::detail::RowBroadcastDescriptor<
+          EpilogueDescriptor, ElementD>;
+
+  using Bias = cutlass::epilogue::fusion::Sm90RowBroadcast<
+      BiasDescriptor::Stages, typename EpilogueDescriptor::TileShape, ElementD,
+      Stride<Int<0>, Int<1>, Int<0>>, 128 / sizeof_bits_v<ElementD>, false>;
+
+ public:
+  using EVTCompute =
+      cutlass::epilogue::fusion::Sm90EVT<Compute1, ScaleA, EVTCompute0, Bias>;
+  using ArgumentType = typename EVTCompute::Arguments;
+
+  static ArgumentType prepare_args(torch::Tensor const& a_scales,
+                                   torch::Tensor const& b_scales,
+                                   torch::Tensor const& bias) {
+    using ScaleA_Args = typename ScaleA::Arguments;
+    using ScaleB_Args = typename ScaleB::Arguments;
+    using Bias_Args = typename Bias::Arguments;
+
+    ScaleA_Args a_args{a_scales.data_ptr<float>(), a_scales.numel() != 1, {}};
+    ScaleB_Args b_args{b_scales.data_ptr<float>(), b_scales.numel() != 1, {}};
+    Bias_Args bias_args{static_cast<ElementD*>(bias.data_ptr())};
+
+    return ArgumentType{a_args, {b_args}, bias_args};
+  }
+};
+
 template <typename ElementAB_, typename ElementD_,
           template <typename, typename, typename> typename Epilogue_,
           typename TileShape, typename ClusterShape, typename KernelSchedule,
@@ -234,15 +291,15 @@ void cutlass_gemm_caller(torch::Tensor& out, torch::Tensor const& a,
 }
 
 template <typename InType, typename OutType,
-          template <typename, typename, typename> typename Epilogue, int32_t M>
-struct sm90_fp8_config {
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_default {
+  // M in (128, inf)
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule =
       cutlass::gemm::KernelTmaWarpSpecializedPingpongFP8FastAccum;
   using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
   using TileShape = Shape<_128, _128, _128>;
   using ClusterShape = Shape<_2, _1, _1>;
-
   using Cutlass3xGemm =
       cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
                       KernelSchedule, EpilogueSchedule>;
@@ -250,14 +307,14 @@ struct sm90_fp8_config {
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
-struct sm90_fp8_config<InType, OutType, Epilogue, 128> {
+struct sm90_fp8_config_M128 {
+  // M in (64, 128]
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule =
       cutlass::gemm::KernelTmaWarpSpecializedPingpongFP8FastAccum;
   using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
   using TileShape = Shape<_64, _128, _128>;
   using ClusterShape = Shape<_2, _1, _1>;
-
   using Cutlass3xGemm =
       cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
                       KernelSchedule, EpilogueSchedule>;
@@ -265,7 +322,8 @@ struct sm90_fp8_config<InType, OutType, Epilogue, 128> {
 
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
-struct sm90_fp8_config<InType, OutType, Epilogue, 64> {
+struct sm90_fp8_config_M64 {
+  // M in [1, 64]
   static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
   using KernelSchedule =
       cutlass::gemm::KernelTmaWarpSpecializedPingpongFP8FastAccum;
@@ -278,6 +336,78 @@ struct sm90_fp8_config<InType, OutType, Epilogue, 64> {
                       KernelSchedule, EpilogueSchedule>;
 };
 
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_int8_config_default {
+  // For M > 128 and any N
+  static_assert(std::is_same<InType, int8_t>());
+  using KernelSchedule =
+      typename cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
+  using TileShape = Shape<_128, _128, _128>;
+  using ClusterShape = Shape<_2, _1, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                      KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_int8_config_M128 {
+  // For M in (64, 128] and any N
+  static_assert(std::is_same<InType, int8_t>());
+  using KernelSchedule =
+      typename cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
+  using TileShape = Shape<_64, _128, _128>;
+  using ClusterShape = Shape<_2, _1, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                      KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_int8_config_M64 {
+  // For M in (32, 64] and any N
+  static_assert(std::is_same<InType, int8_t>());
+  using KernelSchedule = typename cutlass::gemm::KernelTmaWarpSpecialized;
+  using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
+  using TileShape = Shape<_64, _64, _256>;
+  using ClusterShape = Shape<_1, _1, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                      KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_int8_config_M32_NBig {
+  // For M in [1, 32] and N >= 8192
+  static_assert(std::is_same<InType, int8_t>());
+  using KernelSchedule = typename cutlass::gemm::KernelTmaWarpSpecialized;
+  using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
+  using TileShape = Shape<_64, _128, _256>;
+  using ClusterShape = Shape<_1, _4, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                      KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_int8_config_M32_NSmall {
+  // For M in [1, 32] and N < 8192
+  static_assert(std::is_same<InType, int8_t>());
+  using KernelSchedule = typename cutlass::gemm::KernelTmaWarpSpecialized;
+  using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
+  using TileShape = Shape<_64, _64, _256>;
+  using ClusterShape = Shape<_1, _8, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                      KernelSchedule, EpilogueSchedule>;
+};
+
 }  // namespace
 
 template <typename InType, typename OutType,
@@ -291,11 +421,12 @@ void cutlass_gemm_sm90_fp8_dispatch(torch::Tensor& out, torch::Tensor const& a,
   TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
 
   using Cutlass3xGemmDefault =
-      typename sm90_fp8_config<InType, OutType, Epilogue, 0>::Cutlass3xGemm;
+      typename sm90_fp8_config_default<InType, OutType,
+                                       Epilogue>::Cutlass3xGemm;
   using Cutlass3xGemmM64 =
-      typename sm90_fp8_config<InType, OutType, Epilogue, 64>::Cutlass3xGemm;
+      typename sm90_fp8_config_M64<InType, OutType, Epilogue>::Cutlass3xGemm;
   using Cutlass3xGemmM128 =
-      typename sm90_fp8_config<InType, OutType, Epilogue, 128>::Cutlass3xGemm;
+      typename sm90_fp8_config_M128<InType, OutType, Epilogue>::Cutlass3xGemm;
 
   uint32_t const m = a.size(0);
   uint32_t const mp2 =
@@ -316,49 +447,111 @@ void cutlass_gemm_sm90_fp8_dispatch(torch::Tensor& out, torch::Tensor const& a,
   }
 }
 
-void cutlass_scaled_mm_sm90(torch::Tensor& out, torch::Tensor const& a,
-                            torch::Tensor const& b,
-                            torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales) {
-  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
-  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_gemm_sm90_int8_dispatch(torch::Tensor& out, torch::Tensor const& a,
+                                     torch::Tensor const& b,
+                                     EpilogueArgs&&... args) {
+  static_assert(std::is_same<InType, int8_t>());
+  TORCH_CHECK(a.dtype() == torch::kInt8);
+  TORCH_CHECK(b.dtype() == torch::kInt8);
 
+  using Cutlass3xGemmDefault =
+      typename sm90_int8_config_default<InType, OutType,
+                                        Epilogue>::Cutlass3xGemm;
+  using Cutlass3xGemmM128 =
+      typename sm90_int8_config_M128<InType, OutType, Epilogue>::Cutlass3xGemm;
+  using Cutlass3xGemmM64 =
+      typename sm90_int8_config_M64<InType, OutType, Epilogue>::Cutlass3xGemm;
+  using Cutlass3xGemmM32NBig =
+      typename sm90_int8_config_M32_NBig<InType, OutType,
+                                         Epilogue>::Cutlass3xGemm;
+  using Cutlass3xGemmM32NSmall =
+      typename sm90_int8_config_M32_NSmall<InType, OutType,
+                                           Epilogue>::Cutlass3xGemm;
+
+  uint32_t const n = out.size(1);
+  bool const is_small_n = n < 8192;
+
+  uint32_t const m = a.size(0);
+  uint32_t const mp2 =
+      std::max(static_cast<uint32_t>(32), next_pow_2(m));  // next power of 2
+
+  if (mp2 <= 32) {
+    // m in [1, 32]
+    if (is_small_n) {
+      return cutlass_gemm_caller<Cutlass3xGemmM32NSmall>(
+          out, a, b, std::forward<EpilogueArgs>(args)...);
+    } else {
+      return cutlass_gemm_caller<Cutlass3xGemmM32NBig>(
+          out, a, b, std::forward<EpilogueArgs>(args)...);
+    }
+  } else if (mp2 <= 64) {
+    // m in (32, 64]
+    return cutlass_gemm_caller<Cutlass3xGemmM64>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  } else if (mp2 <= 128) {
+    // m in (64, 128]
+    return cutlass_gemm_caller<Cutlass3xGemmM128>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  } else {
+    // m in (128, inf)
+    return cutlass_gemm_caller<Cutlass3xGemmDefault>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+}
+
+template <template <typename, typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_scaled_mm_sm90_epilogue(torch::Tensor& out, torch::Tensor const& a,
+                                     torch::Tensor const& b,
+                                     EpilogueArgs&&... epilogue_args) {
   if (a.dtype() == torch::kInt8) {
     TORCH_CHECK(b.dtype() == torch::kInt8);
 
-    using TileShape = Shape<_128, _128, _128>;
-    using ClusterShape = Shape<_1, _2, _1>;
-    using KernelSchedule =
-        typename cutlass::gemm::KernelTmaWarpSpecializedPingpong;
-    using EpilogueSchedule = typename cutlass::epilogue::TmaWarpSpecialized;
-
     if (out.dtype() == torch::kBFloat16) {
-      return cutlass_gemm_caller<cutlass_3x_gemm<
-          int8_t, cutlass::bfloat16_t, ScaledEpilogue, TileShape, ClusterShape,
-          KernelSchedule, EpilogueSchedule>>(out, a, b, a_scales, b_scales);
+      return cutlass_gemm_sm90_int8_dispatch<int8_t, cutlass::bfloat16_t,
+                                             Epilogue>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     } else {
       TORCH_CHECK(out.dtype() == torch::kFloat16);
-
-      return cutlass_gemm_caller<
-          cutlass_3x_gemm<int8_t, cutlass::half_t, ScaledEpilogue, TileShape,
-                          ClusterShape, KernelSchedule, EpilogueSchedule>>(
-          out, a, b, a_scales, b_scales);
+      return cutlass_gemm_sm90_int8_dispatch<int8_t, cutlass::half_t, Epilogue>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     }
   } else {
     TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
     TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
 
     if (out.dtype() == torch::kBFloat16) {
-      return cutlass_gemm_sm90_fp8_dispatch<
-          cutlass::float_e4m3_t, cutlass::bfloat16_t, ScaledEpilogue>(
-          out, a, b, a_scales, b_scales);
+      return cutlass_gemm_sm90_fp8_dispatch<cutlass::float_e4m3_t,
+                                            cutlass::bfloat16_t, Epilogue>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     } else {
       TORCH_CHECK(out.dtype() == torch::kFloat16);
       return cutlass_gemm_sm90_fp8_dispatch<cutlass::float_e4m3_t,
-                                            cutlass::half_t, ScaledEpilogue>(
-          out, a, b, a_scales, b_scales);
+                                            cutlass::half_t, Epilogue>(
+          out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
     }
   }
 }
 
+void cutlass_scaled_mm_sm90(torch::Tensor& c, torch::Tensor const& a,
+                            torch::Tensor const& b,
+                            torch::Tensor const& a_scales,
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias) {
+  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
+  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+  if (bias) {
+    TORCH_CHECK(bias->dtype() == c.dtype(),
+                "currently bias dtype must match output dtype ", c.dtype());
+    return cutlass_scaled_mm_sm90_epilogue<ScaledEpilogueBias>(
+        c, a, b, a_scales, b_scales, *bias);
+  } else {
+    return cutlass_scaled_mm_sm90_epilogue<ScaledEpilogue>(c, a, b, a_scales,
+                                                           b_scales);
+  }
+}
+
 #endif

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -6,28 +6,49 @@
 void cutlass_scaled_mm_sm75(torch::Tensor& c, torch::Tensor const& a,
                             torch::Tensor const& b,
                             torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales);
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias);
 
 void cutlass_scaled_mm_sm80(torch::Tensor& c, torch::Tensor const& a,
                             torch::Tensor const& b,
                             torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales);
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias);
 
 void cutlass_scaled_mm_sm89(torch::Tensor& c, torch::Tensor const& a,
                             torch::Tensor const& b,
                             torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales);
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias);
 
 #if defined CUDA_VERSION && CUDA_VERSION >= 12000
 void cutlass_scaled_mm_sm90(torch::Tensor& c, torch::Tensor const& a,
                             torch::Tensor const& b,
                             torch::Tensor const& a_scales,
-                            torch::Tensor const& b_scales);
+                            torch::Tensor const& b_scales,
+                            c10::optional<torch::Tensor> const& bias);
 #endif
 
+bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability) {
+  // CUTLASS FP8 kernels need at least
+  //   CUDA 12.0 on SM90 systems (Hopper)
+  //   CUDA 12.4 on SM89 systems (Lovelace)
+
+#if defined CUDA_VERSION
+  if (cuda_device_capability >= 90) {
+    return CUDA_VERSION >= 12000;
+  } else if (cuda_device_capability >= 89) {
+    return CUDA_VERSION >= 12040;
+  }
+#endif
+
+  return false;
+}
+
 void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
                        torch::Tensor const& b, torch::Tensor const& a_scales,
-                       torch::Tensor const& b_scales) {
+                       torch::Tensor const& b_scales,
+                       c10::optional<torch::Tensor> const& bias) {
   int32_t major_capability;
   int32_t minor_capability;
   cudaDeviceGetAttribute(&major_capability, cudaDevAttrComputeCapabilityMajor,
@@ -50,6 +71,11 @@ void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
               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);
+  }
+
   at::cuda::OptionalCUDAGuard const device_guard(device_of(a));
 
   if (version_num >= 90) {
@@ -57,19 +83,19 @@ void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
 
     // Guard against compilation issues for sm90 kernels
 #if defined CUDA_VERSION && CUDA_VERSION >= 12000
-    cutlass_scaled_mm_sm90(c, a, b, a_scales, b_scales);
+    cutlass_scaled_mm_sm90(c, a, b, a_scales, b_scales, bias);
 #else
-    cutlass_scaled_mm_sm80(c, a, b, a_scales, b_scales);
+    cutlass_scaled_mm_sm80(c, a, b, a_scales, b_scales, bias);
 #endif
   } else if (version_num == 89) {
     // Ada Lovelace
-    cutlass_scaled_mm_sm89(c, a, b, a_scales, b_scales);
+    cutlass_scaled_mm_sm89(c, a, b, a_scales, b_scales, bias);
   } else if (version_num >= 80) {
     // Ampere
-    cutlass_scaled_mm_sm80(c, a, b, a_scales, b_scales);
+    cutlass_scaled_mm_sm80(c, a, b, a_scales, b_scales, bias);
   } else {
     // Turing
     TORCH_CHECK(version_num >= 75);
-    cutlass_scaled_mm_sm75(c, a, b, a_scales, b_scales);
+    cutlass_scaled_mm_sm75(c, a, b, a_scales, b_scales, bias);
   }
-}
+}

csrc/quantization/marlin/sparse/common/mma.h

@@ -17,9 +17,23 @@
 
 #pragma once
 #include "base.h"
+#include <cudaTypedefs.h>
 
 namespace marlin_24 {
 
+// On CUDA earlier than 12.5, the ordered_metadata version of this instruction
+// is not supported. On later versions of CUDA the version without ordered
+// metadata results in the following warning:
+//  | Advisory: Modifier ‘.sp::ordered_metadata’ should be used on instruction
+//  | ‘mma’ instead of modifier ‘.sp’ as it is expected to have substantially
+//  | reduced performance on some future architectures
+#if defined CUDA_VERSION && CUDA_VERSION >= 12050
+  #define MMA_SP_INST \
+    "mma.sp::ordered_metadata.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
+#else
+  #define MMA_SP_INST "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
+#endif
+
 // m16n8k32 sparse tensor core mma instruction with fp16 inputs and fp32
 // output/accumulation.
 __device__ inline void mma_sp(const FragB& a_frag0, const FragB& a_frag1,
@@ -29,41 +43,38 @@ __device__ inline void mma_sp(const FragB& a_frag0, const FragB& a_frag1,
   const uint32_t* a1 = reinterpret_cast<const uint32_t*>(&a_frag1);
   const uint32_t* b = reinterpret_cast<const uint32_t*>(&frag_b);
   const uint32_t* e = reinterpret_cast<const uint32_t*>(&frag_m);
+
   float* c = reinterpret_cast<float*>(&frag_c);
   if (psel == 0) {
-    asm volatile(
-        "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
-        "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
-        "{%12,%13,%14,%15}, %16, 0x0;\n"
-        : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
-        : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[0]), "r"(b[2]),
-          "r"(b[4]), "r"(b[6]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]),
-          "r"(e[0]));
-    asm volatile(
-        "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
-        "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
-        "{%12,%13,%14,%15}, %16, 0x0;\n"
-        : "=f"(c[4]), "=f"(c[5]), "=f"(c[6]), "=f"(c[7])
-        : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[1]), "r"(b[3]),
-          "r"(b[5]), "r"(b[7]), "f"(c[4]), "f"(c[5]), "f"(c[6]), "f"(c[7]),
-          "r"(e[0]));
+    asm volatile(MMA_SP_INST
+                 "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
+                 "{%12,%13,%14,%15}, %16, 0x0;\n"
+                 : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
+                 : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[0]),
+                   "r"(b[2]), "r"(b[4]), "r"(b[6]), "f"(c[0]), "f"(c[1]),
+                   "f"(c[2]), "f"(c[3]), "r"(e[0]));
+    asm volatile(MMA_SP_INST
+                 "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
+                 "{%12,%13,%14,%15}, %16, 0x0;\n"
+                 : "=f"(c[4]), "=f"(c[5]), "=f"(c[6]), "=f"(c[7])
+                 : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[1]),
+                   "r"(b[3]), "r"(b[5]), "r"(b[7]), "f"(c[4]), "f"(c[5]),
+                   "f"(c[6]), "f"(c[7]), "r"(e[0]));
   } else {
-    asm volatile(
-        "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
-        "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
-        "{%12,%13,%14,%15}, %16, 0x1;\n"
-        : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
-        : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[0]), "r"(b[2]),
-          "r"(b[4]), "r"(b[6]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]),
-          "r"(e[0]));
-    asm volatile(
-        "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
-        "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
-        "{%12,%13,%14,%15}, %16, 0x1;\n"
-        : "=f"(c[4]), "=f"(c[5]), "=f"(c[6]), "=f"(c[7])
-        : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[1]), "r"(b[3]),
-          "r"(b[5]), "r"(b[7]), "f"(c[4]), "f"(c[5]), "f"(c[6]), "f"(c[7]),
-          "r"(e[0]));
+    asm volatile(MMA_SP_INST
+                 "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
+                 "{%12,%13,%14,%15}, %16, 0x1;\n"
+                 : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
+                 : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[0]),
+                   "r"(b[2]), "r"(b[4]), "r"(b[6]), "f"(c[0]), "f"(c[1]),
+                   "f"(c[2]), "f"(c[3]), "r"(e[0]));
+    asm volatile(MMA_SP_INST
+                 "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
+                 "{%12,%13,%14,%15}, %16, 0x1;\n"
+                 : "=f"(c[4]), "=f"(c[5]), "=f"(c[6]), "=f"(c[7])
+                 : "r"(a0[0]), "r"(a1[0]), "r"(a0[1]), "r"(a1[1]), "r"(b[1]),
+                   "r"(b[3]), "r"(b[5]), "r"(b[7]), "f"(c[4]), "f"(c[5]),
+                   "f"(c[6]), "f"(c[7]), "r"(e[0]));
   }
 }
 

csrc/torch_bindings.cpp

@@ -68,6 +68,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("gelu_fast(Tensor! out, Tensor input) -> ()");
   ops.impl("gelu_fast", torch::kCUDA, &gelu_fast);
 
+  // Quick GELU implementation.
+  ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
+  ops.impl("gelu_quick", torch::kCUDA, &gelu_quick);
+
   // Layernorm
   // Apply Root Mean Square (RMS) Normalization to the input tensor.
   ops.def(
@@ -133,13 +137,23 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("gptq_marlin_repack", &gptq_marlin_repack);
   ops.impl("gptq_marlin_repack", torch::kCUDA, &gptq_marlin_repack);
 
+  // fp8_marlin Optimized Quantized GEMM for FP8 weight-only.
+  ops.def("fp8_marlin_gemm", &fp8_marlin_gemm);
+  ops.impl("fp8_marlin_gemm", torch::kCUDA, &fp8_marlin_gemm);
+
   // CUTLASS 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 b_scales, Tensor? bias) -> ()");
   ops.impl("cutlass_scaled_mm", torch::kCUDA, &cutlass_scaled_mm);
+
+  // Check if cutlass scaled_mm is supported for CUDA devices of the given
+  // capability
+  ops.def("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);
+  ops.impl("cutlass_scaled_mm_supports_fp8", torch::kCUDA,
+           &cutlass_scaled_mm_supports_fp8);
 #endif
 
   // Quantized GEMM for GPTQ.

docs/requirements-docs.txt

@@ -1,7 +1,7 @@
-sphinx == 6.2.1
-sphinx-book-theme == 1.0.1
-sphinx-copybutton == 0.5.2
-myst-parser == 2.0.0
+sphinx==6.2.1
+sphinx-book-theme==1.0.1
+sphinx-copybutton==0.5.2
+myst-parser==2.0.0
 sphinx-argparse
 
 # packages to install to build the documentation

docs/source/_templates/sections/header.html

@@ -0,0 +1,38 @@
+<style>
+  .notification-bar {
+    width: 100vw;
+    display: flex;
+    justify-content: center;
+    align-items: center;
+    font-size: 16px;
+  }
+  .notification-bar p {
+    margin: 0;
+  }
+  .notification-bar a {
+    font-weight: bold;
+    text-decoration: none;
+  }
+
+  /* Light mode styles (default) */
+  .notification-bar {
+    background-color: #fff3cd;
+    color: #856404;
+  }
+  .notification-bar a {
+    color: #d97706;
+  }
+
+  /* Dark mode styles */
+  html[data-theme=dark] .notification-bar {
+    background-color: #333;
+    color: #ddd;
+  }
+  html[data-theme=dark] .notification-bar a {
+    color: #ffa500; /* Brighter color for visibility */
+  }
+</style>
+
+<div class="notification-bar">
+  <p>You are viewing the latest developer preview docs. <a href="https://docs.vllm.ai/en/stable/">Click here</a> to view docs for the latest stable release.</p>
+</div>

docs/source/community/sponsors.md

@@ -22,5 +22,6 @@ vLLM is a community project. Our compute resources for development and testing a
 - Trainy
 - UC Berkeley
 - UC San Diego
+- ZhenFund
 
 We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM.

docs/source/conf.py

@@ -66,8 +66,20 @@ html_theme_options = {
     'path_to_docs': 'docs/source',
     'repository_url': 'https://github.com/vllm-project/vllm',
     'use_repository_button': True,
+    'use_edit_page_button': True,
 }
 
+# see https://docs.readthedocs.io/en/stable/reference/environment-variables.html # noqa
+READTHEDOCS_VERSION_TYPE = os.environ.get('READTHEDOCS_VERSION_TYPE')
+if READTHEDOCS_VERSION_TYPE == "tag":
+    # remove the warning banner if the version is a tagged release
+    header_file = os.path.join(os.path.dirname(__file__),
+                               "_templates/sections/header.html")
+    # The file might be removed already if the build is triggered multiple times
+    # (readthedocs build both HTML and PDF versions separately)
+    if os.path.exists(header_file):
+        os.remove(header_file)
+
 # Add any paths that contain custom static files (such as style sheets) here,
 # relative to this directory. They are copied after the builtin static files,
 # so a file named "default.css" will overwrite the builtin "default.css".
@@ -95,6 +107,7 @@ autodoc_mock_imports = [
     'triton',
     "tqdm",
     "tensorizer",
+    "pynvml",
 ]
 
 for mock_target in autodoc_mock_imports:

docs/source/dev/dockerfile/dockerfile.rst

@@ -1,20 +1,20 @@
 Dockerfile
 ====================
 
-See `here <https://github.com/vllm-project/vllm/blob/main/Dockerfile>`_ for the main Dockerfile to construct 
-the image for running an OpenAI compatible server with vLLM.
+See `here <https://github.com/vllm-project/vllm/blob/main/Dockerfile>`__ for the main Dockerfile to construct 
+the image for running an OpenAI compatible server with vLLM. More information about deploying with Docker can be found `here <https://docs.vllm.ai/en/stable/serving/deploying_with_docker.html>`__.
 
--  Below is a visual representation of the multi-stage Dockerfile. The build graph contains the following nodes:
+Below is a visual representation of the multi-stage Dockerfile. The build graph contains the following nodes:
 
-   - All build stages
-   - The default build target (highlighted in grey)
-   - External images (with dashed borders)
+- All build stages
+- The default build target (highlighted in grey)
+- External images (with dashed borders)
    
-   The edges of the build graph represent:
-   
-   - FROM ... dependencies (with a solid line and a full arrow head)
-   - COPY --from=... dependencies (with a dashed line and an empty arrow head)
-   - RUN --mount=(.*)from=... dependencies (with a dotted line and an empty diamond arrow head)
+The edges of the build graph represent:
+
+- FROM ... dependencies (with a solid line and a full arrow head)
+- COPY --from=... dependencies (with a dashed line and an empty arrow head)
+- RUN --mount=(.*)from=... dependencies (with a dotted line and an empty diamond arrow head)
 
    .. figure:: ../../assets/dev/dockerfile-stages-dependency.png
       :alt: query

docs/source/dev/input_processing/input_processing_pipeline.rst

@@ -0,0 +1,20 @@
+.. _input_processing_pipeline:
+
+Input Processing Pipeline
+=========================
+
+1. Input data is passed to :class:`~vllm.LLMEngine` (or :class:`~vllm.AsyncLLMEngine`).
+
+2. Tokenize the data if necessary.
+
+3. Process the inputs using :meth:`INPUT_REGISTRY.process_input <vllm.inputs.registry.InputRegistry.process_input>`.
+
+   - For example, add placeholder tokens to reserve KV cache for multi-modal embeddings.
+
+4. Send the processed inputs to :class:`~vllm.executor.executor_base.ExecutorBase`.
+
+5. Distribute the inputs via :class:`~vllm.worker.worker_base.WorkerBase` to :class:`~vllm.worker.model_runner_base.ModelRunnerBase`.
+
+6. If the data contains multi-modal data, convert it into keyword arguments using :meth:`MULTIMODAL_REGISTRY.map_input <vllm.multimodal.MultiModalRegistry.map_input>`.
+
+   - For example, convert a :class:`PIL.Image.Image` input to its pixel values for a vision language model.

docs/source/dev/input_processing/model_inputs_index.rst

@@ -0,0 +1,39 @@
+.. _input_processing:
+
+Input Processing
+================
+
+.. currentmodule:: vllm.inputs
+
+vLLM provides a mechanism for defining input processors for each model so that the inputs are processed
+in :class:`~vllm.LLMEngine` before they are passed to model executors. 
+
+Currently, this mechanism is only utilized in :ref:`multi-modal models <multi_modality>` for preprocessing multi-modal input 
+data in addition to input prompt, but it can be extended to text-only language models when needed.
+
+Guides
+++++++
+
+.. toctree::
+   :maxdepth: 1
+
+   input_processing_pipeline
+
+Module Contents
++++++++++++++++
+
+LLM Engine Inputs
+-----------------
+
+.. autoclass:: vllm.inputs.LLMInputs
+    :members:
+    :show-inheritance:
+
+Registry
+--------
+
+.. autodata:: vllm.inputs.INPUT_REGISTRY
+
+.. automodule:: vllm.inputs.registry
+    :members:
+    :show-inheritance:

docs/source/dev/multimodal/adding_multimodal_model.rst

@@ -0,0 +1,152 @@
+.. _adding_a_new_multimodal_model:
+
+Adding a New Multimodal Model
+=============================
+
+This document provides a high-level guide on integrating a :ref:`multi-modal model <multi_modality>` into vLLM.
+
+.. note::
+    The complexity of adding a new model depends heavily on the model's architecture.
+    The process is considerably straightforward if the model shares a similar architecture with an existing model in vLLM.
+    However, for models that include new operators (e.g., a new attention mechanism), the process can be a bit more complex.
+
+.. tip::
+    If you are encountering issues while integrating your model into vLLM, feel free to open an issue on our `GitHub <https://github.com/vllm-project/vllm/issues>`_ repository.
+    We will be happy to help you out!
+
+
+1. Set up the base vLLM model
+-----------------------------
+
+As usual, follow :ref:`these steps <adding_a_new_model>` to implement the model in vLLM, but note the following:
+
+- You should additionally implement the :class:`~vllm.model_executor.models.interfaces.SupportsVision` interface.
+
+  .. code-block:: diff
+
+      + from vllm.model_executor.models.interfaces import SupportsVision
+
+      - class YourModelForImage2Seq(nn.Module):
+      + class YourModelForImage2Seq(nn.Module, SupportsVision):
+
+  .. note::
+      The model class does not have to be named :code:`*ForCausalLM`.
+      Check out `the HuggingFace Transformers documentation <https://huggingface.co/docs/transformers/model_doc/auto#multimodal>`__ for some examples.
+
+- While implementing the :meth:`~torch.nn.Module.forward` method, reserve a keyword parameter
+  for each input tensor that corresponds to a multi-modal input, as shown in the following example:
+
+  .. code-block:: diff
+
+      def forward(
+          self,
+          input_ids: torch.Tensor,
+          positions: torch.Tensor,
+          kv_caches: List[torch.Tensor],
+          attn_metadata: AttentionMetadata,
+      +   pixel_values: torch.Tensor,
+      ) -> SamplerOutput:
+
+
+2. Register input mappers
+-------------------------
+
+For each modality type that the model accepts as input, decorate the model class with :meth:`MULTIMODAL_REGISTRY.register_input_mapper <vllm.multimodal.MultiModalRegistry.register_input_mapper>`.
+This decorator accepts a function that maps multi-modal inputs to the keyword arguments you have previously defined in :meth:`~torch.nn.Module.forward`.
+
+.. code-block:: diff
+
+      from vllm.model_executor.models.interfaces import SupportsVision
+    + from vllm.multimodal import MULTIMODAL_REGISTRY
+
+    + @MULTIMODAL_REGISTRY.register_image_input_mapper()
+      class YourModelForImage2Seq(nn.Module, SupportsVision):
+
+A default mapper is available for each modality in the core vLLM library. This input mapper will be used if you do not provide your own function.
+
+.. seealso::
+    :ref:`input_processing_pipeline`
+
+
+3. Register maximum number of multimodal tokens
+----------------------------------------------------------
+
+For each modality type that the model accepts as input, calculate the maximum possible number of tokens
+and register it via :meth:`INPUT_REGISTRY.register_dummy_data <vllm.inputs.registry.InputRegistry.register_max_multimodal_tokens>`.
+
+.. code-block:: diff
+
+      from vllm.inputs import INPUT_REGISTRY
+      from vllm.model_executor.models.interfaces import SupportsVision
+      from vllm.multimodal import MULTIMODAL_REGISTRY
+
+      @MULTIMODAL_REGISTRY.register_image_input_mapper()
+    + @MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
+      @INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
+      class YourModelForImage2Seq(nn.Module, SupportsVision):
+
+Here are some examples:
+
+- Image inputs (static feature size): `LLaVA-1.5 Model <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/llava.py>`__
+- Image inputs (dynamic feature size): `LLaVA-NeXT Model <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/llava_next.py>`__
+
+.. seealso::
+    :ref:`input_processing_pipeline`
+
+
+4. (Optional) Register dummy data
+---------------------------------
+
+During startup, dummy data is passed to the vLLM model to allocate memory. This only consists of text input by default, which may not be applicable to multi-modal models.
+In such cases, you can define your own dummy data by registering a factory method via :meth:`INPUT_REGISTRY.register_dummy_data <vllm.inputs.registry.InputRegistry.register_dummy_data>`.
+
+.. code-block:: diff
+
+      from vllm.inputs import INPUT_REGISTRY
+      from vllm.model_executor.models.interfaces import SupportsVision
+      from vllm.multimodal import MULTIMODAL_REGISTRY
+
+      @MULTIMODAL_REGISTRY.register_image_input_mapper()
+      @MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
+    + @INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
+      class YourModelForImage2Seq(nn.Module, SupportsVision):
+
+.. note::
+    The dummy data should have the maximum possible number of multi-modal tokens, as described in the previous step.
+
+Here are some examples:
+
+- Image inputs (static feature size): `LLaVA-1.5 Model <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/llava.py>`__
+- Image inputs (dynamic feature size): `LLaVA-NeXT Model <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/llava_next.py>`__
+
+.. seealso::
+    :ref:`input_processing_pipeline`
+
+
+5. (Optional) Register input processor
+--------------------------------------
+
+Sometimes, there is a need to process inputs at the :class:`~vllm.LLMEngine` level before they are passed to the model executor. 
+This is often due to the fact that unlike implementations in HuggingFace Transformers, the reshaping and/or expansion of multi-modal embeddings needs to take place outside model's :meth:`~torch.nn.Module.forward` call.
+You can register input processors via :meth:`INPUT_REGISTRY.register_input_processor <vllm.inputs.registry.InputRegistry.register_input_processor>`.
+
+.. code-block:: diff
+
+      from vllm.inputs import INPUT_REGISTRY
+      from vllm.model_executor.models.interfaces import SupportsVision
+      from vllm.multimodal import MULTIMODAL_REGISTRY
+
+      @MULTIMODAL_REGISTRY.register_image_input_mapper()
+      @MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
+      @INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
+    + @INPUT_REGISTRY.register_input_processor(<your_input_processor>)
+      class YourModelForImage2Seq(nn.Module, SupportsVision):
+
+A common use case of input processors is inserting placeholder tokens to leverage the vLLM framework for attention mask generation.
+Here are some examples:
+
+- Insert static number of image tokens: `LLaVA-1.5 Model <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/llava.py>`__
+- Insert dynamic number of image tokens: `LLaVA-NeXT Model <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/llava_next.py>`__
+
+.. seealso::
+    :ref:`input_processing_pipeline`

docs/source/dev/multimodal/multimodal_index.rst

@@ -1,3 +1,5 @@
+.. _multi_modality:
+
 Multi-Modality
 ==============
 
@@ -8,13 +10,22 @@ vLLM provides experimental support for multi-modal models through the :mod:`vllm
 :class:`vllm.inputs.PromptStrictInputs` accepts an additional attribute ``multi_modal_data``
 which allows you to pass in multi-modal input alongside text and token prompts.
 
-By default, vLLM models do not support multi-modal inputs. To enable multi-modal support for a model,
-you must decorate the model class with :meth:`MULTIMODAL_REGISTRY.register_dummy_data <MultiModalRegistry.register_dummy_data>`,
-as well as :meth:`MULTIMODAL_REGISTRY.register_input <MultiModalRegistry.register_input>` for each modality type to support.
+.. note::
+   ``multi_modal_data`` can accept keys and values beyond the builtin ones, as long as a customized plugin is registered through 
+    :class:`vllm.multimodal.MULTIMODAL_REGISTRY`.
 
-.. contents::
-   :local:
-   :backlinks: none
+By default, vLLM models do not support multi-modal inputs. To enable multi-modal support for a model, please follow :ref:`the guide for adding a new multimodal model. <adding_a_new_multimodal_model>`.
+
+
+# TODO: Add more instructions on how to do that once embeddings is in.
+
+Guides
+++++++
+
+.. toctree::
+   :maxdepth: 1
+
+   adding_multimodal_model
 
 Module Contents
 +++++++++++++++
@@ -24,9 +35,7 @@ Module Contents
 Registry
 --------
 
-.. data:: vllm.multimodal.MULTIMODAL_REGISTRY
-
-    The global :class:`MultiModalRegistry` which is used by model runners.
+.. autodata:: vllm.multimodal.MULTIMODAL_REGISTRY
 
 .. autoclass:: vllm.multimodal.MultiModalRegistry
     :members:
@@ -35,7 +44,11 @@ Registry
 Base Classes
 ------------
 
-.. autoclass:: vllm.multimodal.MultiModalData
+.. autoclass:: vllm.multimodal.MultiModalDataDict
+    :members:
+    :show-inheritance:
+
+.. autoclass:: vllm.multimodal.MultiModalInputs
     :members:
     :show-inheritance:
 

docs/source/getting_started/amd-installation.rst

@@ -88,7 +88,7 @@ Option 2: Build from source
 - `Pytorch <https://pytorch.org/>`_
 - `hipBLAS <https://rocm.docs.amd.com/projects/hipBLAS/en/latest/install.html>`_
 
-For installing PyTorch, you can start from a fresh docker image, e.g, `rocm6.0.2_ubuntu22.04_py3.10_pytorch_2.1.2`, `rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1`, `rocm/pytorch-nightly`.
+For installing PyTorch, you can start from a fresh docker image, e.g, `rocm/pytorch:rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging`, `rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1`, `rocm/pytorch-nightly`.
 
 Alternatively, you can install pytorch using pytorch wheels. You can check Pytorch installation guild in Pytorch `Getting Started <https://pytorch.org/get-started/locally/>`_
 
@@ -126,12 +126,12 @@ Install ROCm's flash attention (v2.0.4) following the instructions from `ROCm/fl
 
     $ cd vllm
     $ pip install -U -r requirements-rocm.txt
-    $ python setup.py install # This may take 5-10 minutes. Currently, `pip install .`` does not work for ROCm installation
+    $ python setup.py develop # This may take 5-10 minutes. Currently, `pip install .`` does not work for ROCm installation
 
 
 .. tip::
 
     - You may need to turn on the ``--enforce-eager`` flag if you experience process hang when running the `benchmark_thoughput.py` script to test your installation.
     - Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm up step before collecting perf numbers.
-    - To use CK flash-attention, please use this flag ``export VLLM_USE_FLASH_ATTN_TRITON=0`` to turn off triton flash attention. 
+    - To use CK flash-attention, please use this flag ``export VLLM_USE_TRITON_FLASH_ATTN=0`` to turn off triton flash attention. 
     - The ROCm version of pytorch, ideally, should match the ROCm driver version.

docs/source/getting_started/debugging.rst

@@ -24,22 +24,42 @@ If you have already taken care of the above issues, but the vLLM instance still
 
 With more logging, hopefully you can find the root cause of the issue.
 
+If it crashes, and the error trace shows somewhere around ``self.graph.replay()`` in ``vllm/worker/model_runner.py``, it is a cuda error inside cudagraph. To know the particular cuda operation that causes the error, you can add ``--enforce-eager`` to the command line, or ``enforce_eager=True`` to the ``LLM`` class, to disable the cudagraph optimization. This way, you can locate the exact cuda operation that causes the error.
+
 Here are some common issues that can cause hangs:
 
-- **Incorrect network setup**: The vLLM instance cannot get the correct IP address. You can find the log such as ``DEBUG 06-10 21:32:17 parallel_state.py:88] world_size=8 rank=0 local_rank=0 distributed_init_method=tcp://xxx.xxx.xxx.xxx:54641 backend=nccl``. The IP address should be the correct one. If not, override the IP address by setting the environment variable ``export VLLM_HOST_IP=your_ip_address``.
-- **Incorrect hardware/driver**: GPU communication cannot be established. You can run the following sanity check script to see if the GPU communication is working correctly.
+- **Incorrect network setup**: The vLLM instance cannot get the correct IP address if you have complicated network config. You can find the log such as ``DEBUG 06-10 21:32:17 parallel_state.py:88] world_size=8 rank=0 local_rank=0 distributed_init_method=tcp://xxx.xxx.xxx.xxx:54641 backend=nccl``. The IP address should be the correct one. If not, override the IP address by setting the environment variable ``export VLLM_HOST_IP=your_ip_address``. You might also need to set ``export NCCL_SOCKET_IFNAME=your_network_interface`` and ``export GLOO_SOCKET_IFNAME=your_network_interface`` to specify the network interface for the IP address.
+- **Incorrect hardware/driver**: GPU/CPU communication cannot be established. You can run the following sanity check script to see if the GPU/CPU communication is working correctly.
 
 .. code-block:: python
 
-    # save it as `test.py` , and run it with `NCCL_DEBUG=TRACE torchrun --nproc-per-node=8 test.py`
-    # adjust `--nproc-per-node` to the number of GPUs you want to use.
     import torch
     import torch.distributed as dist
     dist.init_process_group(backend="nccl")
-    data = torch.FloatTensor([1,] * 128).to(f"cuda:{dist.get_rank()}")
+    local_rank = dist.get_rank() % torch.cuda.device_count()
+    data = torch.FloatTensor([1,] * 128).to(f"cuda:{local_rank}")
     dist.all_reduce(data, op=dist.ReduceOp.SUM)
     torch.cuda.synchronize()
     value = data.mean().item()
-    assert value == dist.get_world_size()
+    world_size = dist.get_world_size()
+    assert value == world_size, f"Expected {world_size}, got {value}"
+
+    gloo_group = dist.new_group(ranks=list(range(world_size)), backend="gloo")
+    cpu_data = torch.FloatTensor([1,] * 128)
+    dist.all_reduce(cpu_data, op=dist.ReduceOp.SUM, group=gloo_group)
+    value = cpu_data.mean().item()
+    assert value == world_size, f"Expected {world_size}, got {value}"
+
+.. tip::
+
+    Save the script as ``test.py``.
+    
+    If you are testing in a single-node, run it with ``NCCL_DEBUG=TRACE torchrun --nproc-per-node=8 test.py``, adjust ``--nproc-per-node`` to the number of GPUs you want to use.
+    
+    If you are testing with multi-nodes, run it with ``NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR test.py``. Adjust ``--nproc-per-node`` and ``--nnodes`` according to your setup. Make sure ``MASTER_ADDR``:
+  
+    - is the correct IP address of the master node
+    - is reachable from all nodes
+    - is set before running the script.
 
 If the problem persists, feel free to `open an issue on GitHub <https://github.com/vllm-project/vllm/issues/new/choose>`_, with a detailed description of the issue, your environment, and the logs.

docs/source/getting_started/installation.rst

@@ -20,7 +20,7 @@ You can install vLLM using pip:
 .. code-block:: console
 
     $ # (Recommended) Create a new conda environment.
-    $ conda create -n myenv python=3.9 -y
+    $ conda create -n myenv python=3.10 -y
     $ conda activate myenv
 
     $ # Install vLLM with CUDA 12.1.
@@ -35,7 +35,7 @@ You can install vLLM using pip:
 
         $ # Install vLLM with CUDA 11.8.
         $ export VLLM_VERSION=0.4.0
-        $ export PYTHON_VERSION=39
+        $ export PYTHON_VERSION=310
         $ pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu118-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu118
 
     In order to be performant, vLLM has to compile many cuda kernels. The compilation unfortunately introduces binary incompatibility with other CUDA versions and PyTorch versions, even for the same PyTorch version with different building configurations.

docs/source/getting_started/openvino-installation.rst

@@ -0,0 +1,95 @@
+.. _installation_openvino:
+
+Installation with OpenVINO
+==========================
+
+vLLM powered by OpenVINO supports all LLM models from :doc:`vLLM supported models list <../models/supported_models>` and can perform optimal model serving on all x86-64 CPUs with, at least, AVX2 support. OpenVINO vLLM backend supports the following advanced vLLM features:
+
+- Prefix caching (``--enable-prefix-caching``)
+- Chunked prefill (``--enable-chunked-prefill``)
+
+**Table of contents**:
+
+- :ref:`Requirements <openvino_backend_requirements>`
+- :ref:`Quick start using Dockerfile <openvino_backend_quick_start_dockerfile>`
+- :ref:`Build from source <install_openvino_backend_from_source>`
+- :ref:`Performance tips <openvino_backend_performance_tips>`
+- :ref:`Limitations <openvino_backend_limitations>`
+
+.. _openvino_backend_requirements:
+
+Requirements
+------------
+
+* OS: Linux
+* Instruction set architecture (ISA) requirement: at least AVX2.
+
+.. _openvino_backend_quick_start_dockerfile:
+
+Quick start using Dockerfile
+----------------------------
+
+.. code-block:: console
+
+    $ docker build -f Dockerfile.openvino -t vllm-openvino-env .
+    $ docker run -it --rm vllm-openvino-env
+
+.. _install_openvino_backend_from_source:
+
+Install from source
+-------------------
+
+- First, install Python. For example, on Ubuntu 22.04, you can run:
+
+  .. code-block:: console
+
+      $ sudo apt-get update  -y
+      $ sudo apt-get install python3
+
+- Second, install prerequisites vLLM OpenVINO backend installation:
+
+  .. code-block:: console
+
+      $ pip install --upgrade pip
+      $ pip install -r requirements-build.txt --extra-index-url https://download.pytorch.org/whl/cpu
+
+- Finally, install vLLM with OpenVINO backend: 
+
+  .. code-block:: console
+
+      $ PIP_PRE=1 PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu https://storage.openvinotoolkit.org/simple/wheels/nightly/" VLLM_TARGET_DEVICE=openvino python -m pip install -v .
+
+.. _openvino_backend_performance_tips:
+
+Performance tips
+----------------
+
+vLLM OpenVINO backend uses the following environment variables to control behavior:
+
+- ``VLLM_OPENVINO_KVCACHE_SPACE`` to specify the KV Cache size (e.g, ``VLLM_OPENVINO_KVCACHE_SPACE=40`` means 40 GB space for KV cache), larger setting will allow vLLM running more requests in parallel. This parameter should be set based on the hardware configuration and memory management pattern of users.
+
+- ``VLLM_OPENVINO_CPU_KV_CACHE_PRECISION=u8`` to control KV cache precision. By default, FP16 / BF16 is used depending on platform.
+
+- ``VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON`` to enable U8 weights compression during model loading stage. By default, compression is turned off.
+
+To enable better TPOT / TTFT latency, you can use vLLM's chunked prefill feature (``--enable-chunked-prefill``). Based on the experiments, the recommended batch size is ``256`` (``--max-num-batched-tokens``)
+
+OpenVINO best known configuration is:
+
+.. code-block:: console
+
+    $ VLLM_OPENVINO_KVCACHE_SPACE=100 VLLM_OPENVINO_CPU_KV_CACHE_PRECISION=u8 VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON \
+        python3 vllm/benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-chat-hf --dataset vllm/benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json --enable-chunked-prefill --max-num-batched-tokens 256
+
+.. _openvino_backend_limitations:
+
+Limitations
+-----------
+
+- LoRA serving is not supported.
+
+- Only LLM models are currently supported. LLaVa and encoder-decoder models are not currently enabled in vLLM OpenVINO integration.
+
+- Tensor and pipeline parallelism are not currently enabled in vLLM integration.
+
+- Speculative sampling is not tested within vLLM integration.

docs/source/getting_started/tpu-installation.rst

@@ -73,3 +73,21 @@ Next, build vLLM from source. This will only take a few seconds:
 .. code-block:: console
 
     $ VLLM_TARGET_DEVICE="tpu" python setup.py develop
+
+
+.. tip::
+
+    If you encounter the following error:
+
+    .. code-block:: console
+
+        from torch._C import *  # noqa: F403
+        ImportError: libopenblas.so.0: cannot open shared object file: No such file or directory
+
+
+    You can install OpenBLAS with the following command:
+
+    .. code-block:: console
+
+        $ sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev
+

docs/source/getting_started/xpu-installation.rst

@@ -0,0 +1,61 @@
+.. _installation_xpu:
+
+Installation with XPU
+========================
+
+vLLM initially supports basic model inferencing and serving on Intel GPU platform.
+
+Table of contents:
+
+#. :ref:`Requirements <xpu_backend_requirements>`
+#. :ref:`Quick start using Dockerfile <xpu_backend_quick_start_dockerfile>`
+#. :ref:`Build from source <build_xpu_backend_from_source>`
+
+.. _xpu_backend_requirements:
+
+Requirements
+------------
+
+* OS: Linux
+* Supported Hardware: Intel Data Center GPU (Intel ARC GPU WIP)
+* OneAPI requirements: oneAPI 2024.1 
+
+.. _xpu_backend_quick_start_dockerfile:
+
+Quick start using Dockerfile
+----------------------------
+
+.. code-block:: console
+
+    $ docker build -f Dockerfile.xpu -t vllm-xpu-env --shm-size=4g .
+    $ docker run -it \
+                 --rm \
+                 --network=host \
+                 --device /dev/dri \
+                 -v /dev/dri/by-path:/dev/dri/by-path \
+                 vllm-xpu-env
+
+.. _build_xpu_backend_from_source:
+
+Build from source
+-----------------
+
+- First, install required driver and intel OneAPI 2024.1.
+
+- Second, install Python packages for vLLM XPU backend building:
+
+.. code-block:: console
+
+    $ pip install --upgrade pip
+    $ pip install -v -r requirements-xpu.txt 
+
+- Finally, build and install vLLM XPU backend: 
+
+.. code-block:: console
+
+    $ VLLM_TARGET_DEVICE=xpu python setup.py install
+
+.. note::
+    - FP16 is the default data type in the current XPU backend. The BF16 data
+      type will be supported in the future.
+

docs/source/index.rst

@@ -63,9 +63,11 @@ Documentation
 
    getting_started/installation
    getting_started/amd-installation
+   getting_started/openvino-installation
    getting_started/cpu-installation
    getting_started/neuron-installation
    getting_started/tpu-installation
+   getting_started/xpu-installation
    getting_started/quickstart
    getting_started/debugging
    getting_started/examples/examples_index
@@ -81,6 +83,8 @@ Documentation
    serving/env_vars
    serving/usage_stats
    serving/integrations
+   serving/tensorizer
+   serving/faq
 
 .. toctree::
    :maxdepth: 1
@@ -98,6 +102,7 @@ Documentation
    :maxdepth: 1
    :caption: Quantization
 
+   quantization/supported_hardware
    quantization/auto_awq
    quantization/fp8
    quantization/fp8_e5m2_kvcache
@@ -117,6 +122,7 @@ Documentation
    dev/offline_inference/offline_index
    dev/engine/engine_index
    dev/kernel/paged_attention
+   dev/input_processing/model_inputs_index
    dev/multimodal/multimodal_index
    dev/dockerfile/dockerfile
 

docs/source/models/adding_model.rst

@@ -37,7 +37,7 @@ For instance, vLLM's `OPT model <https://github.com/vllm-project/vllm/blob/main/
 2. Rewrite the :code:`forward` methods
 --------------------------------------
 
-Next, you need to rewrite the :code:`forward` methods of your model by following these steps:
+Next, you need to rewrite the :meth:`~torch.nn.Module.forward` method of your model by following these steps:
 
 1. Remove any unnecessary code, such as the code only used for training.
 2. Change the input parameters:
@@ -75,7 +75,7 @@ Next, you need to rewrite the :code:`forward` methods of your model by following
 
 If your model is too large to fit into a single GPU, you can use tensor parallelism to manage it.
 To do this, substitute your model's linear and embedding layers with their tensor-parallel versions.
-For the embedding layer, you can simply replace :code:`nn.Embedding` with :code:`VocabParallelEmbedding`. For the output LM head, you can use :code:`ParallelLMHead`.
+For the embedding layer, you can simply replace :class:`torch.nn.Embedding` with :code:`VocabParallelEmbedding`. For the output LM head, you can use :code:`ParallelLMHead`.
 When it comes to the linear layers, we provide the following options to parallelize them:
 
 * :code:`ReplicatedLinear`: Replicates the inputs and weights across multiple GPUs. No memory saving.

docs/source/models/lora.rst

@@ -4,6 +4,9 @@ Using LoRA adapters
 ===================
 
 This document shows you how to use `LoRA adapters <https://arxiv.org/abs/2106.09685>`_ with vLLM on top of a base model.
+
+LoRA adapters can be used with any vLLM model that implements :class:`~vllm.model_executor.models.interfaces.SupportsLoRA`.
+
 Adapters can be efficiently served on a per request basis with minimal overhead. First we download the adapter(s) and save
 them locally with
 

docs/source/models/supported_models.rst

@@ -55,6 +55,10 @@ Alongside each architecture, we include some popular models that use it.
     - Gemma
     - :code:`google/gemma-2b`, :code:`google/gemma-7b`, etc.
     - ✅︎
+  * - :code:`Gemma2ForCausalLM`
+    - Gemma2
+    - :code:`google/gemma-2-9b`, :code:`google/gemma-2-27b`, etc.
+    - ✅︎
   * - :code:`GPT2LMHeadModel`
     - GPT-2
     - :code:`gpt2`, :code:`gpt2-xl`, etc.
@@ -83,6 +87,10 @@ Alongside each architecture, we include some popular models that use it.
     - Jais
     - :code:`core42/jais-13b`, :code:`core42/jais-13b-chat`, :code:`core42/jais-30b-v3`, :code:`core42/jais-30b-chat-v3`, etc.
     -
+  * - :code:`JambaForCausalLM`
+    - Jamba
+    - :code:`ai21labs/Jamba-v0.1`, etc.
+    - ✅︎
   * - :code:`LlamaForCausalLM`
     - LLaMA, Llama 2, Meta Llama 3, Vicuna, Alpaca, Yi
     - :code:`meta-llama/Meta-Llama-3-8B-Instruct`, :code:`meta-llama/Meta-Llama-3-70B-Instruct`, :code:`meta-llama/Llama-2-13b-hf`, :code:`meta-llama/Llama-2-70b-hf`, :code:`openlm-research/open_llama_13b`, :code:`lmsys/vicuna-13b-v1.3`, :code:`01-ai/Yi-6B`, :code:`01-ai/Yi-34B`, etc.
@@ -129,12 +137,16 @@ Alongside each architecture, we include some popular models that use it.
     - ✅︎
   * - :code:`Phi3ForCausalLM`
     - Phi-3
-    - :code:`microsoft/Phi-3-mini-4k-instruct`, :code:`microsoft/Phi-3-mini-128k-instruct`, etc.
+    - :code:`microsoft/Phi-3-mini-4k-instruct`, :code:`microsoft/Phi-3-mini-128k-instruct`, :code:`microsoft/Phi-3-medium-128k-instruct`, etc.
     -
   * - :code:`Phi3SmallForCausalLM`
     - Phi-3-Small
     - :code:`microsoft/Phi-3-small-8k-instruct`, :code:`microsoft/Phi-3-small-128k-instruct`, etc.
     -
+  * - :code:`Phi3VForCausalLM`
+    - Phi-3-Vision
+    - :code:`microsoft/Phi-3-vision-128k-instruct`, etc.
+    -
   * - :code:`QWenLMHeadModel`
     - Qwen
     - :code:`Qwen/Qwen-7B`, :code:`Qwen/Qwen-7B-Chat`, etc.

docs/source/models/vlm.rst

@@ -5,22 +5,12 @@ Using VLMs
 
 vLLM provides experimental support for Vision Language Models (VLMs). This document shows you how to run and serve these models using vLLM.
 
-Engine Arguments
-----------------
-
-The following :ref:`engine arguments <engine_args>` are specific to VLMs:
-
-.. argparse::
-    :module: vllm.engine.arg_utils
-    :func: _vlm_engine_args_parser
-    :prog: -m vllm.entrypoints.openai.api_server
-    :nodefaultconst:
-
 .. important::
+    We are actively iterating on VLM support. Expect breaking changes to VLM usage and development in upcoming releases without prior deprecation.
+
     Currently, the support for vision language models on vLLM has the following limitations:
 
     * Only single image input is supported per text prompt.
-    * Dynamic ``image_input_shape`` is not supported: the input image will be resized to the static ``image_input_shape``. This means our LLaVA-NeXT output may not exactly match the huggingface implementation.
 
     We are continuously improving user & developer experience for VLMs. Please `open an issue on GitHub <https://github.com/vllm-project/vllm/issues/new/choose>`_ if you have any feedback or feature requests.
 
@@ -31,38 +21,60 @@ To initialize a VLM, the aforementioned arguments must be passed to the ``LLM``
 
 .. code-block:: python
 
-    llm = LLM(
-        model="llava-hf/llava-1.5-7b-hf",
-        image_input_type="pixel_values",
-        image_token_id=32000,
-        image_input_shape="1,3,336,336",
-        image_feature_size=576,
-    )
+    llm = LLM(model="llava-hf/llava-1.5-7b-hf")
+
+.. important::
+    We have removed all vision language related CLI args in the ``0.5.1`` release. **This is a breaking change**, so please update your code to follow
+    the above snippet. Specifically, ``image_feature_size`` is no longer required to be specified as we now calculate that
+    internally for each model.
+
 
 To pass an image to the model, note the following in :class:`vllm.inputs.PromptStrictInputs`:
 
-* ``prompt``: The prompt should have a number of ``<image>`` tokens equal to ``image_feature_size``.
-* ``multi_modal_data``: This should be an instance of :class:`~vllm.multimodal.image.ImagePixelData` or :class:`~vllm.multimodal.image.ImageFeatureData`.
+* ``prompt``: The prompt should follow the format that is documented on HuggingFace.
+* ``multi_modal_data``: This is a dictionary that follows the schema defined in :class:`vllm.multimodal.MultiModalDataDict`. 
 
 .. code-block:: python
 
-    prompt = "<image>" * 576 + (
-        "\nUSER: What is the content of this image?\nASSISTANT:")
+    # Refer to the HuggingFace repo for the correct format to use
+    prompt = "USER: <image>\nWhat is the content of this image?\nASSISTANT:"
 
     # Load the image using PIL.Image
-    image = ...
-
+    image = PIL.Image.open(...)
+    
+    # Single prompt inference
     outputs = llm.generate({
         "prompt": prompt,
-        "multi_modal_data": ImagePixelData(image),
+        "multi_modal_data": {"image": image},
     })
 
     for o in outputs:
         generated_text = o.outputs[0].text
         print(generated_text)
+    
+    # Batch inference
+    image_1 = PIL.Image.open(...)
+    image_2 = PIL.Image.open(...)
+    outputs = llm.generate(
+        [
+            {
+                "prompt": "USER: <image>\nWhat is the content of this image?\nASSISTANT:",
+                "multi_modal_data": {"image": image_1},
+            },
+            {
+                "prompt": "USER: <image>\nWhat's the color of this image?\nASSISTANT:",
+                "multi_modal_data": {"image": image_2},
+            }
+        ]
+    )
+
+    for o in outputs:
+        generated_text = o.outputs[0].text
+        print(generated_text)
 
 A code example can be found in `examples/llava_example.py <https://github.com/vllm-project/vllm/blob/main/examples/llava_example.py>`_.
 
+
 Online OpenAI Vision API Compatible Inference
 ----------------------------------------------
 
@@ -83,12 +95,13 @@ Below is an example on how to launch the same ``llava-hf/llava-1.5-7b-hf`` with
 
     python -m vllm.entrypoints.openai.api_server \
         --model llava-hf/llava-1.5-7b-hf \
-        --image-input-type pixel_values \
-        --image-token-id 32000 \
-        --image-input-shape 1,3,336,336 \
-        --image-feature-size 576 \
         --chat-template template_llava.jinja
 
+.. important::
+    We have removed all vision language related CLI args in the ``0.5.1`` release. **This is a breaking change**, so please update your code to follow
+    the above snippet. Specifically, ``image_feature_size`` is no longer required to be specified as we now calculate that
+    internally for each model.
+
 To consume the server, you can use the OpenAI client like in the example below:
 
 .. code-block:: python
@@ -105,6 +118,8 @@ To consume the server, you can use the OpenAI client like in the example below:
         messages=[{
             "role": "user",
             "content": [
+                # NOTE: The prompt formatting with the image token `<image>` is not needed
+                # since the prompt will be processed automatically by the API server.
                 {"type": "text", "text": "What's in this image?"},
                 {
                     "type": "image_url",
@@ -117,6 +132,8 @@ To consume the server, you can use the OpenAI client like in the example below:
     )
     print("Chat response:", chat_response)
 
+A full code example can be found in `examples/openai_vision_api_client.py <https://github.com/vllm-project/vllm/blob/main/examples/openai_vision_api_client.py>`_.
+
 .. note::
 
     By default, the timeout for fetching images through http url is ``5`` seconds. You can override this by setting the environment variable:
@@ -126,5 +143,4 @@ To consume the server, you can use the OpenAI client like in the example below:
         export VLLM_IMAGE_FETCH_TIMEOUT=<timeout>
 
 .. note::
-    The prompt formatting with the image token ``<image>`` is not needed when serving VLMs with the API server since the prompt will be 
-    processed automatically by the server.
+    There is no need to format the prompt in the API request since it will be handled by the server.

docs/source/quantization/fp8.rst

@@ -3,7 +3,10 @@
 FP8
 ==================
 
-vLLM supports FP8 (8-bit floating point) computation using hardware acceleration on GPUs such as Nvidia H100 and AMD MI300x. Currently, only Hopper and Ada Lovelace GPUs are supported. Quantization of models with FP8 allows for a 2x reduction in model memory requirements and up to a 1.6x improvement in throughput with minimal impact on accuracy.
+vLLM supports FP8 (8-bit floating point) weight and activation quantization using hardware acceleration on GPUs such as Nvidia H100 and AMD MI300x. 
+Currently, only Hopper and Ada Lovelace GPUs are officially supported for W8A8. 
+Ampere GPUs are supported for W8A16 (weight-only FP8) utilizing Marlin kernels.
+Quantization of models with FP8 allows for a 2x reduction in model memory requirements and up to a 1.6x improvement in throughput with minimal impact on accuracy.
 
 Please visit the HF collection of `quantized FP8 checkpoints of popular LLMs ready to use with vLLM <https://huggingface.co/collections/neuralmagic/fp8-llms-for-vllm-666742ed2b78b7ac8df13127>`_.
 

docs/source/quantization/supported_hardware.rst

@@ -0,0 +1,30 @@
+.. _supported_hardware_for_quantization:
+
+Supported Hardware for Quantization Kernels
+===========================================
+
+The table below shows the compatibility of various quantization implementations with different hardware platforms in vLLM:
+
+==============  ======  =======  =======  =====  ======  =======  =========  =======  ==============  ==========
+Implementation  Volta   Turing   Ampere   Ada    Hopper  AMD GPU  Intel GPU  x86 CPU  AWS Inferentia  Google TPU
+==============  ======  =======  =======  =====  ======  =======  =========  =======  ==============  ==========
+AQLM            ✅      ✅       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+AWQ             ❌      ✅       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+DeepSpeedFP     ✅      ✅       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+FP8             ❌      ❌       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+Marlin          ❌      ❌       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+GPTQ            ✅      ✅       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+SqueezeLLM      ✅      ✅       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+bitsandbytes    ✅      ✅       ✅       ✅     ✅      ❌        ❌         ❌       ❌              ❌
+==============  ======  =======  =======  =====  ======  =======  =========  =======  ==============  ==========
+
+Notes:
+^^^^^^
+
+- Volta refers to SM 7.0, Turing to SM 7.5, Ampere to SM 8.0/8.6, Ada to SM 8.9, and Hopper to SM 9.0.
+- "✅" indicates that the quantization method is supported on the specified hardware.
+- "❌" indicates that the quantization method is not supported on the specified hardware.
+
+Please note that this compatibility chart may be subject to change as vLLM continues to evolve and expand its support for different hardware platforms and quantization methods.
+
+For the most up-to-date information on hardware support and quantization methods, please check the `quantization directory <https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization>`_ or consult with the vLLM development team.

docs/source/serving/deploying_with_cerebrium.rst

@@ -0,0 +1,109 @@
+.. _deploying_with_cerebrium:
+
+Deploying with Cerebrium
+============================
+
+.. raw:: html
+
+    <p align="center">
+        <img src="https://i.ibb.co/hHcScTT/Screenshot-2024-06-13-at-10-14-54.png" alt="vLLM_plus_cerebrium"/>
+    </p>
+
+vLLM can be run on a cloud based GPU machine with `Cerebrium <https://www.cerebrium.ai/>`__, a serverless AI infrastructure platform that makes it easier for companies to build and deploy AI based applications.
+
+To install the Cerebrium client, run:
+
+.. code-block:: console
+
+    $ pip install cerebrium
+    $ cerebrium login
+
+Next, create your Cerebrium project, run:
+    
+.. code-block:: console
+
+    $ cerebrium init vllm-project
+
+Next, to install the required packages, add the following to your cerebrium.toml:
+
+.. code-block:: toml
+
+    [cerebrium.dependencies.pip]
+    vllm = "latest"
+
+Next, let us add our code to handle inference for the LLM of your choice(`mistralai/Mistral-7B-Instruct-v0.1` for this example), add the following code to your main.py`:
+    
+.. code-block:: python
+
+    from vllm import LLM, SamplingParams
+
+    llm = LLM(model="mistralai/Mistral-7B-Instruct-v0.1")
+
+    def run(prompts: list[str], temperature: float = 0.8, top_p: float = 0.95):
+    
+        sampling_params = SamplingParams(temperature=temperature, top_p=top_p)
+        outputs = llm.generate(prompts, sampling_params)
+
+        # Print the outputs.
+        results = []
+        for output in outputs:
+            prompt = output.prompt
+            generated_text = output.outputs[0].text
+            results.append({"prompt": prompt, "generated_text": generated_text})
+
+        return {"results": results}
+
+
+Then, run the following code to deploy it to the cloud
+
+.. code-block:: console
+
+    $ cerebrium deploy
+
+If successful, you should be returned a CURL command that you can call inference against. Just remember to end the url with the function name you are calling (in our case /run)
+
+.. code-block:: python
+
+    curl -X POST https://api.cortex.cerebrium.ai/v4/p-xxxxxx/vllm/run \
+     -H 'Content-Type: application/json' \
+     -H 'Authorization: <JWT TOKEN>' \
+     --data '{
+       "prompts": [
+         "Hello, my name is",
+         "The president of the United States is",
+         "The capital of France is",
+         "The future of AI is"
+       ]
+     }'
+
+You should get a response like:
+
+.. code-block:: python
+    
+    {
+        "run_id": "52911756-3066-9ae8-bcc9-d9129d1bd262",
+        "result": {
+            "result": [
+                {
+                    "prompt": "Hello, my name is",
+                    "generated_text": " Sarah, and I'm a teacher. I teach elementary school students. One of"
+                },
+                {
+                    "prompt": "The president of the United States is",
+                    "generated_text": " elected every four years. This is a democratic system.\n\n5. What"
+                },
+                {
+                    "prompt": "The capital of France is",
+                    "generated_text": " Paris.\n"
+                },
+                {
+                    "prompt": "The future of AI is",
+                    "generated_text": " bright, but it's important to approach it with a balanced and nuanced perspective."
+                }
+            ]
+        },
+        "run_time_ms": 152.53663063049316
+    }
+
+You now have an autoscaling endpoint where you only pay for the compute you use!
+

docs/source/serving/deploying_with_docker.rst

@@ -3,9 +3,8 @@
 Deploying with Docker
 ============================
 
-vLLM offers official docker image for deployment.
-The image can be used to run OpenAI compatible server.
-The image is available on Docker Hub as `vllm/vllm-openai <https://hub.docker.com/r/vllm/vllm-openai/tags>`_.
+vLLM offers an official Docker image for deployment.
+The image can be used to run OpenAI compatible server and is available on Docker Hub as `vllm/vllm-openai <https://hub.docker.com/r/vllm/vllm-openai/tags>`_.
 
 .. code-block:: console
 
@@ -25,7 +24,7 @@ The image is available on Docker Hub as `vllm/vllm-openai <https://hub.docker.co
         memory to share data between processes under the hood, particularly for tensor parallel inference.
 
 
-You can build and run vLLM from source via the provided dockerfile. To build vLLM:
+You can build and run vLLM from source via the provided `Dockerfile <https://github.com/vllm-project/vllm/blob/main/Dockerfile>`_. To build vLLM:
 
 .. code-block:: console
 

docs/source/serving/distributed_serving.rst

@@ -19,7 +19,7 @@ To run multi-GPU serving, pass in the :code:`--tensor-parallel-size` argument wh
 
 .. code-block:: console
 
-    $ python -m vllm.entrypoints.api_server \
+    $ python -m vllm.entrypoints.openai.api_server \
     $     --model facebook/opt-13b \
     $     --tensor-parallel-size 4
 
@@ -35,4 +35,7 @@ To scale vLLM beyond a single machine, install and start a `Ray runtime <https:/
     $ # On worker nodes
     $ ray start --address=<ray-head-address>
 
-After that, you can run inference and serving on multiple machines by launching the vLLM process on the head node by setting :code:`tensor_parallel_size` to the number of GPUs to be the total number of GPUs across all machines.
+After that, you can run inference and serving on multiple machines by launching the vLLM process on the head node by setting :code:`tensor_parallel_size` to the number of GPUs to be the total number of GPUs across all machines.
+
+.. warning::
+    Please make sure you downloaded the model to all the nodes, or the model is downloaded to some distributed file system that is accessible by all nodes.

docs/source/serving/env_vars.rst

@@ -3,6 +3,11 @@ Environment Variables
 
 vLLM uses the following environment variables to configure the system:
 
+.. warning::
+    Please note that ``VLLM_PORT`` and ``VLLM_HOST_IP`` set the port and ip for vLLM's **internal usage**. It is not the port and ip for the API server. If you use ``--host $VLLM_HOST_IP`` and ``--port $VLLM_PORT`` to start the API server, it will not work.
+
+    All environment variables used by vLLM are prefixed with ``VLLM_``. **Special care should be taken for Kubernetes users**: please do not name the service as ``vllm``, otherwise environment variables set by Kubernetes might conflict with vLLM's environment variables, because `Kubernetes sets environment variables for each service with the capitalized service name as the prefix <https://kubernetes.io/docs/concepts/services-networking/service/#environment-variables>`_.
+
 .. literalinclude:: ../../../vllm/envs.py
     :language: python
     :start-after: begin-env-vars-definition

docs/source/serving/faq.rst

@@ -0,0 +1,12 @@
+Frequently Asked Questions
+===========================
+
+    Q: How can I serve multiple models on a single port using the OpenAI API?
+
+A: Assuming that you're referring to using OpenAI compatible server to serve multiple models at once, that is not currently supported, you can run multiple instances of the server (each serving a different model) at the same time, and have another layer to route the incoming request to the correct server accordingly.
+
+----------------------------------------
+
+    Q: Which model to use for offline inference embedding?
+
+A: If you want to use an embedding model, try: https://huggingface.co/intfloat/e5-mistral-7b-instruct. Instead models, such as Llama-3-8b, Mistral-7B-Instruct-v0.3, are generation models rather than an embedding model

docs/source/serving/integrations.rst

@@ -8,6 +8,7 @@ Integrations
    deploying_with_kserve
    deploying_with_triton
    deploying_with_bentoml
+   deploying_with_cerebrium
    deploying_with_lws
    deploying_with_dstack
    serving_with_langchain