Bump version to 0.5.3.post1 (#6696)

[doc][distributed] add more doc for setting up multi-node environment (#6529)

.buildkite/download-images.sh

@@ -1,18 +0,0 @@
-#!/bin/bash
-
-set -ex
-set -o pipefail
-
-(which wget && which curl) || (apt-get update && apt-get install -y wget curl)
-
-# 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
-
-cd -

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

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m deepseek-ai/DeepSeek-V2-Lite-Chat -b "auto" -l 1000 -f 5 -t 2
+model_name: "deepseek-ai/DeepSeek-V2-Lite-Chat"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.671
+  - name: "exact_match,flexible-extract"
+    value: 0.664
+limit: 1000
+num_fewshot: 5

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

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform -b auto -l 1000 -f 5
+model_name: "nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.905
+  - name: "exact_match,flexible-extract"
+    value: 0.905
+limit: 1000
+num_fewshot: 5

.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-Channelwise-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-W8A8-FP8-Channelwise-compressed-tensors -b auto -l 1000 -f 5 -t 1
+model_name: "nm-testing/Meta-Llama-3-8B-Instruct-W8A8-FP8-Channelwise-compressed-tensors"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.752
+  - name: "exact_match,flexible-extract"
+    value: 0.754
+limit: 1000
+num_fewshot: 5

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

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform -b auto -l 1000 -f 5 -t 1
+model_name: "nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.753
+  - name: "exact_match,flexible-extract"
+    value: 0.753
+limit: 1000
+num_fewshot: 5

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

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test -b 32 -l 1000 -f 5 -t 1
+model_name: "nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.755
+  - name: "exact_match,flexible-extract"
+    value: 0.755
+limit: 1000
+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-vllm-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: 0.753
+  - name: "exact_match,flexible-extract"
+    value: 0.753
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-INT8-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Per-Token-Test -b "auto" -l 250 -f 5 -t 1
+model_name: "nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Per-Token-Test"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.728
+  - name: "exact_match,flexible-extract"
+    value: 0.728
+limit: 250
+num_fewshot: 5

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

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-nonuniform-test -b auto -l 1000 -f 5 -t 1
+model_name: "nm-testing/Meta-Llama-3-8B-Instruct-nonuniform-test"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.758
+  - name: "exact_match,flexible-extract"
+    value: 0.759
+limit: 1000
+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/Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Qwen2-1.5B-Instruct-quantized.w8a8 -b "auto" -l 1000 -f 5 -t 1
+model_name: "neuralmagic/Qwen2-1.5B-Instruct-quantized.w8a8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.593
+  - name: "exact_match,flexible-extract"
+    value: 0.588
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-W8A16-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise -b "auto" -l 1000 -f 5 -t 1
+model_name: "nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.595
+  - name: "exact_match,flexible-extract"
+    value: 0.582
+limit: 1000
+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 +1,5 @@
+Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform.yaml
+Meta-Llama-3-70B-Instruct.yaml
+Mixtral-8x7B-Instruct-v0.1.yaml
+Qwen2-57B-A14-Instruct.yaml
+DeepSeek-V2-Lite-Chat.yaml

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

@@ -0,0 +1,7 @@
+Meta-Llama-3-8B-Instruct.yaml
+Meta-Llama-3-8B-Instruct-FP8.yaml
+Meta-Llama-3-8B-Instruct-FP8-compressed-tensors.yaml
+Meta-Llama-3-8B-Instruct-INT8-compressed-tensors.yaml
+Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
+Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
+Qwen2-1.5B-Instruct-INT8-compressed-tensors.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==0.4.3
+
+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,distributed_executor_backend="ray",trust_remote_code=true,max_model_len=4096 \
+  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --batch_size $BATCH_SIZE

.buildkite/lm-eval-harness/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 +1,55 @@
+"""
+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}," \
+                 f"add_bos_token=true"
+
+    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 +1,104 @@
+# 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 +1,61 @@
+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"
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          priorityClassName: perf-benchmark
+          containers:
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+            command:
+            - bash .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+            env:
+            - name: VLLM_USAGE_SOURCE
+              value: ci-test
+            - name: HF_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: hf-token-secret
+                  key: token
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+  - label: "H100"
+    agents:
+      queue: H100
+    plugins:
+    - docker#v5.11.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: all
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
+

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

@@ -0,0 +1,45 @@
+
+# Nightly benchmark
+
+The main goal of this benchmarking is two-fold:
+- Performance clarity: Provide clarity on which one (vllm, tensorrt-llm, lmdeploy and tgi) leads in performance in what workload.
+- Reproducible: one can run the exact same set of benchmarking commands inside the exact same docker by following reproducing instructions in [reproduce.md]().
+
+
+## Docker images
+
+We benchmark vllm, tensorrt-llm, lmdeploy and tgi using the following docker images:
+- vllm/vllm-openai:v0.5.0.post1
+- nvcr.io/nvidia/tritonserver:24.04-trtllm-python-py3
+- openmmlab/lmdeploy:v0.5.0
+- ghcr.io/huggingface/text-generation-inference:2.1
+
+<!-- Please check <a href="artifact://workspace/build/buildkite/vllm/performance-benchmark/.buildkite/nightly-benchmarks/nightly-pipeline.yaml">nightly-pipeline.yaml</a> artifact for more details on how we deploy the docker images. -->
+
+
+## Hardware
+
+One AWS node with 8x NVIDIA A100 GPUs.
+
+
+## Workload description
+
+We benchmark vllm, tensorrt-llm, lmdeploy and tgi using the following workload:
+
+- Input length: randomly sample 500 prompts from ShareGPT dataset (with fixed random seed).
+- Output length: the corresponding output length of these 500 prompts.
+- Models: llama-3 8B, llama-3 70B, mixtral 8x7B.
+- Average QPS (query per second): 4 for the small model (llama-3 8B) and 2 for other two models. For each QPS, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
+- Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).
+
+<!-- Check <a href="artifact://workspace/build/buildkite/vllm/performance-benchmark/.buildkite/nightly-benchmarks/tests/nightly-tests.json">nightly-tests.json</a> artifact for more details. -->
+
+## Plots
+
+In the following plots, the dot shows the mean and the error bar shows the standard error of the mean. Value 0 means that the corresponding benchmark crashed.
+
+<img src="artifact://nightly_results.png" alt="Benchmarking results" height=250 >
+
+## Results
+
+{nightly_results_benchmarking_table}

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

@@ -0,0 +1,120 @@
+common_pod_spec: &common_pod_spec
+  priorityClassName: perf-benchmark
+  nodeSelector:
+    nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+  volumes:
+    - name: devshm
+      emptyDir:
+        medium: Memory
+    - name: hf-cache
+      hostPath:
+        path: /root/.cache/huggingface
+        type: Directory
+
+common_container_settings: &common_container_settings
+  command:
+    - bash .buildkite/nightly-benchmarks/run-nightly-suite.sh
+  resources:
+    limits:
+      nvidia.com/gpu: 8
+  volumeMounts:
+    - name: devshm
+      mountPath: /dev/shm
+    - name: hf-cache
+      mountPath: /root/.cache/huggingface
+  env:
+    - name: VLLM_USAGE_SOURCE
+      value: ci-test
+    - name: HF_HOME
+      value: /root/.cache/huggingface
+    - name: VLLM_SOURCE_CODE_LOC
+      value: /workspace/build/buildkite/vllm/performance-benchmark
+    - name: HF_TOKEN
+      valueFrom:
+        secretKeyRef:
+          name: hf-token-secret
+          key: token
+
+steps:
+  - block: ":rocket: Ready for comparing vllm against alternatives? This will take 4 hours."
+  - label: "A100 trt benchmark"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: nvcr.io/nvidia/tritonserver:24.04-trtllm-python-py3
+                <<: *common_container_settings
+
+  - label: "A100 lmdeploy benchmark"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: openmmlab/lmdeploy:v0.5.0
+                <<: *common_container_settings
+  
+
+  - label: "A100 vllm benchmark"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: vllm/vllm-openai:latest 
+                <<: *common_container_settings
+
+  - label: "A100 tgi benchmark"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: ghcr.io/huggingface/text-generation-inference:2.1 
+                <<: *common_container_settings
+        
+  - wait
+
+  - label: "Plot"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+            - image: vllm/vllm-openai:v0.5.0.post1
+              command:
+              - bash .buildkite/nightly-benchmarks/scripts/nightly-annotate.sh
+              resources:
+                limits:
+                  nvidia.com/gpu: 8
+              volumeMounts:
+              - name: devshm
+                mountPath: /dev/shm
+              env:
+              - name: VLLM_USAGE_SOURCE
+                value: ci-test
+              - name: VLLM_SOURCE_CODE_LOC
+                value: /workspace/build/buildkite/vllm/performance-benchmark
+              - name: HF_TOKEN
+                valueFrom:
+                  secretKeyRef:
+                    name: hf-token-secret
+                    key: token
+
+  - wait

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

@@ -0,0 +1,376 @@
+#!/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 -X POST 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: $pid"
+      done
+
+      echo "All GPU processes have been killed."
+  fi
+
+  # Sometimes kill with pid doesn't work properly, we can also kill all process running python or python3
+  # since we are in container anyway
+  pkill -9 -f python
+  pkill -9 -f python3
+
+  # waiting for GPU processes to be fully killed
+  # loop while nvidia-smi returns any processes
+  while [ -n "$(nvidia-smi --query-compute-apps=pid --format=csv,noheader)" ]; do
+    sleep 1
+    echo "Waiting for GPU processes to be killed"
+  done
+
+  # remove vllm config file
+  rm -rf ~/.config/vllm
+
+  # Print the GPU memory usage
+  # so that we know if all GPU processes are killed.
+  gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0)
+  # The memory usage should be 0 MB.
+  echo "GPU 0 Memory Usage: $gpu_memory_usage MB"
+}
+
+upload_to_buildkite() {
+  # upload the benchmarking results to buildkite
+
+  # if the agent binary is not found, skip uploading the results, exit 0
+  # Check if buildkite-agent is available in the PATH or at /workspace/buildkite-agent
+  if command -v buildkite-agent >/dev/null 2>&1; then
+    BUILDKITE_AGENT_COMMAND="buildkite-agent"
+  elif [ -f /workspace/buildkite-agent ]; then
+    BUILDKITE_AGENT_COMMAND="/workspace/buildkite-agent"
+  else
+    echo "buildkite-agent binary not found. Skip uploading the results."
+    return 0
+  fi
+
+  # Use the determined command to annotate and upload artifacts
+  $BUILDKITE_AGENT_COMMAND annotate --style "info" --context "$BUILDKITE_LABEL-benchmark-results" < $RESULTS_FOLDER/benchmark_results.md
+  $BUILDKITE_AGENT_COMMAND artifact upload "$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" &
+    server_pid=$!
+
+    # 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 -9 $server_pid
+    kill_gpu_processes
+  done
+}
+
+main() {
+  check_gpus
+  check_hf_token
+
+  # dependencies
+  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+  (which jq) || (apt-get update && apt-get -y install jq)
+
+  # get the current IP address, required by benchmark_serving.py
+  export VLLM_HOST_IP=$(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/run-nightly-suite.sh

@@ -0,0 +1,76 @@
+#!/bin/bash
+
+set -o pipefail
+set -x
+
+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
+}
+
+main() {
+
+    check_gpus
+    check_hf_token
+
+    df -h
+
+    (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+    (which jq) || (apt-get update && apt-get -y install jq)
+
+    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+    
+
+    # run lmdeploy
+    if which lmdeploy >/dev/null; then
+        echo "lmdeploy is available, redirect to run-lmdeploy-nightly.sh"
+        bash ../.buildkite/nightly-benchmarks/scripts/run-lmdeploy-nightly.sh
+        exit 0
+    fi
+
+    # run tgi
+    if [ -e /tgi-entrypoint.sh ]; then
+        echo "tgi is available, redirect to run-tgi-nightly.sh"
+        bash ../.buildkite/nightly-benchmarks/scripts/run-tgi-nightly.sh
+        exit 0
+    fi
+
+    # run trt
+    if which trtllm-build >/dev/null; then
+        echo "trtllm is available, redirect to run-trt-nightly.sh"
+        bash ../.buildkite/nightly-benchmarks/scripts/run-trt-nightly.sh
+        exit 0
+    fi
+
+    # run vllm
+    if [ -e /vllm-workspace ]; then
+        echo "vllm is available, redirect to run-vllm-nightly.sh"
+        bash ../.buildkite/nightly-benchmarks/scripts/run-vllm-nightly.sh
+        exit 0
+    fi
+
+}
+
+main "$@"

.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/download-tokenizer.py

@@ -0,0 +1,26 @@
+import argparse
+
+from transformers import AutoTokenizer
+
+
+def main(model, cachedir):
+    # Load the tokenizer and save it to the specified directory
+    tokenizer = AutoTokenizer.from_pretrained(model)
+    tokenizer.save_pretrained(cachedir)
+    print(f"Tokenizer saved to {cachedir}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Download and save Hugging Face tokenizer")
+    parser.add_argument("--model",
+                        type=str,
+                        required=True,
+                        help="Name of the model")
+    parser.add_argument("--cachedir",
+                        type=str,
+                        required=True,
+                        help="Directory to save the tokenizer")
+
+    args = parser.parse_args()
+    main(args.model, args.cachedir)

.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py

@@ -0,0 +1,6 @@
+from lmdeploy.serve.openai.api_client import APIClient
+
+api_client = APIClient("http://localhost:8000")
+model_name = api_client.available_models[0]
+
+print(model_name)

.buildkite/nightly-benchmarks/scripts/launch-trt-server.sh

@@ -0,0 +1,102 @@
+#!/bin/bash
+
+
+server_params=$1
+common_params=$2
+
+
+
+model_path=$(echo "$common_params" | jq -r '.model')
+model_name="${model_path#*/}"
+model_type=$(echo "$server_params" | jq -r '.model_type')
+model_dtype=$(echo "$server_params" | jq -r '.model_dtype')
+model_tp_size=$(echo "$common_params" | jq -r '.tp')
+max_batch_size=$(echo "$server_params" | jq -r '.max_batch_size')
+max_input_len=$(echo "$server_params" | jq -r '.max_input_len')
+max_output_len=$(echo "$server_params" | jq -r '.max_output_len')
+trt_llm_version=$(echo "$server_params" | jq -r '.trt_llm_version')
+
+cd ~
+rm -rf models
+mkdir -p models
+cd models
+models_dir=$(pwd)
+trt_model_path=${models_dir}/${model_name}-trt-ckpt
+trt_engine_path=${models_dir}/${model_name}-trt-engine
+
+cd ~
+rm -rf tensorrt-demo
+git clone https://github.com/neuralmagic/tensorrt-demo.git
+cd tensorrt-demo
+tensorrt_demo_dir=$(pwd)
+
+# make sure the parameter inside tensorrt_demo is consistent to envvar
+sed -i.bak "/key: \"tokenizer_dir\"/,/string_value:/s|string_value: \".*\"|string_value: \"$model_path\"|" ./triton_model_repo/postprocessing/config.pbtxt
+sed -i.bak "/key: \"tokenizer_dir\"/,/string_value:/s|string_value: \".*\"|string_value: \"$model_path\"|" ./triton_model_repo/preprocessing/config.pbtxt
+sed -i.bak "s|\(max_batch_size:\s*\)[0-9]*|\1$max_batch_size|g" ./triton_model_repo/ensemble/config.pbtxt
+sed -i.bak "s|\(max_batch_size:\s*\)[0-9]*|\1$max_batch_size|g" ./triton_model_repo/preprocessing/config.pbtxt
+sed -i.bak "s|\(max_batch_size:\s*\)[0-9]*|\1$max_batch_size|g" ./triton_model_repo/postprocessing/config.pbtxt
+sed -i.bak "s|\(max_batch_size:\s*\)[0-9]*|\1$max_batch_size|g" ./triton_model_repo/tensorrt_llm_bls/config.pbtxt
+
+
+cd /
+rm -rf tensorrtllm_backend
+git clone https://github.com/triton-inference-server/tensorrtllm_backend.git
+git lfs install
+cd tensorrtllm_backend
+git checkout $trt_llm_version
+tensorrtllm_backend_dir=$(pwd)
+git submodule update --init --recursive
+cp -r ${tensorrt_demo_dir}/triton_model_repo ${tensorrtllm_backend_dir}/
+
+cd /tensorrtllm_backend
+cd ./tensorrt_llm/examples/${model_type}
+
+
+if echo "$common_params" | jq -e 'has("fp8")' > /dev/null; then
+
+    echo "Key 'fp8' exists in common params. Use quantize.py instead of convert_checkpoint.py"
+    echo "Reference: https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/llama/README.md"
+    python ../quantization/quantize.py \
+        --model_dir ${model_path} \
+        --dtype ${model_dtype} \
+        --tp_size ${model_tp_size} \
+        --output_dir ${trt_model_path} \
+        --qformat fp8 \
+        --kv_cache_dtype fp8 \
+        --calib_size 2
+
+else
+
+    echo "Key 'fp8' does not exist in common params. Use convert_checkpoint.py"
+    python3 convert_checkpoint.py \
+        --model_dir ${model_path} \
+        --dtype ${model_dtype} \
+        --tp_size ${model_tp_size} \
+        --output_dir ${trt_model_path}
+
+fi
+
+
+
+trtllm-build \
+--checkpoint_dir=${trt_model_path} \
+--gpt_attention_plugin=${model_dtype} \
+--gemm_plugin=${model_dtype} \
+--remove_input_padding=enable \
+--paged_kv_cache=enable \
+--tp_size=${model_tp_size} \
+--max_batch_size=${max_batch_size} \
+--max_input_len=${max_input_len} \
+--max_output_len=${max_output_len} \
+--max_num_tokens=${max_output_len} \
+--opt_num_tokens=${max_output_len} \
+--output_dir=${trt_engine_path} 
+
+cd /tensorrtllm_backend/triton_model_repo
+rm -rf ./tensorrt_llm/1/*
+cp -r ${trt_engine_path}/* ./tensorrt_llm/1
+cd /tensorrtllm_backend
+python3 scripts/launch_triton_server.py \
+--world_size=${model_tp_size} \
+--model_repo=/tensorrtllm_backend/triton_model_repo &

.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+set -ex
+set -o pipefail
+
+
+main() {
+
+    (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+    (which jq) || (apt-get update && apt-get -y install jq)
+
+    if [ ! -f /workspace/buildkite-agent ]; then
+        echo "buildkite-agent binary not found. Skip plotting the results."
+        exit 0
+    fi
+
+    # initial annotation
+    description="$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/nightly-descriptions.md"
+
+    # download results
+    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    mkdir -p results/
+    /workspace/buildkite-agent artifact download 'results/*nightly_results.json' results/
+    ls
+    ls results/
+
+    # generate figures
+    python3 -m pip install tabulate pandas matplotlib
+    python3 $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/plot-nightly-results.py \
+        --description $description \
+        --results-folder results/
+    
+    # upload results and figures
+    /workspace/buildkite-agent artifact upload "nightly_results.png"
+    /workspace/buildkite-agent artifact upload $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/nightly-pipeline.yaml
+    /workspace/buildkite-agent artifact upload $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/tests/nightly-tests.json
+    /workspace/buildkite-agent annotate --style "success" --context "nightly-benchmarks-results" --append < nightly_results.md
+}
+
+main "$@"

.buildkite/nightly-benchmarks/scripts/plot-nightly-results.py

@@ -0,0 +1,135 @@
+import argparse
+import json
+import math
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import pandas as pd
+from tabulate import tabulate
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        description=
+        'Parse command line arguments for summary-nightly-results script.')
+    parser.add_argument('--results-folder',
+                        type=str,
+                        required=True,
+                        help='The folder where the results are stored.')
+    parser.add_argument('--description',
+                        type=str,
+                        required=True,
+                        help='Description of the results.')
+
+    args = parser.parse_args()
+    return args
+
+
+def main(args):
+    bar_colors = ['#56B4E9', '#009E73', '#D55E00', '#E69F00']
+    results_folder = Path(args.results_folder)
+
+    results = []
+
+    # collect results
+    for test_file in results_folder.glob("*_nightly_results.json"):
+        with open(test_file, "r") as f:
+            results = results + json.loads(f.read())
+
+    # generate markdown table
+    df = pd.DataFrame.from_dict(results)
+
+    md_table = tabulate(df, headers='keys', tablefmt='pipe', showindex=False)
+
+    with open(args.description, "r") as f:
+        description = f.read()
+
+    description = description.format(
+        nightly_results_benchmarking_table=md_table)
+
+    with open("nightly_results.md", "w") as f:
+        f.write(description)
+
+    plt.rcParams.update({'font.size': 20})
+
+    # plot results
+    fig, axes = plt.subplots(3, 3, figsize=(16, 14))
+    fig.subplots_adjust(hspace=1)
+    methods = ["vllm", "trt", "lmdeploy", "tgi"]
+    for i, model in enumerate(["llama8B", "llama70B", "mixtral8x7B"]):
+        for j, metric in enumerate(["TTFT", "ITL"]):
+            means, stds = [], []
+            for method in methods:
+                target = df['Test name'].str.contains(model)
+                target = target & df['Engine'].str.contains(method)
+                filtered_df = df[target]
+
+                if filtered_df.empty:
+                    means.append(0.)
+                    stds.append(0.)
+                else:
+                    means.append(filtered_df[f"Mean {metric} (ms)"].values[0])
+                    std = filtered_df[f"Std {metric} (ms)"].values[0]
+                    success = filtered_df["Successful req."].values[0]
+                    stds.append(std / math.sqrt(success))
+
+            print(model, metric)
+            print(means, stds)
+
+            ax = axes[i, j + 1]
+
+            bars = ax.bar(
+                ["vllm", "trt", "lmdeploy", "tgi"],
+                means,
+                yerr=stds,
+                capsize=10,
+            )
+            for idx, bar in enumerate(bars):
+                bar.set_color(bar_colors[idx])
+            ax.set_ylim(bottom=0)
+
+            ax.set_ylabel(f"{metric} (ms)")
+            ax.set_title(f"{model} {metric}")
+            ax.grid(axis='y')
+
+        metric = "Tput"
+        j = 0
+        if True:
+            tputs = []
+            for method in methods:
+                target = df['Test name'].str.contains(model)
+                target = target & df['Engine'].str.contains(method)
+                filtered_df = df[target]
+
+                if filtered_df.empty:
+                    tputs.append(0.)
+                else:
+                    input_tput = filtered_df["Input Tput (tok/s)"].values[0]
+                    output_tput = filtered_df["Output Tput (tok/s)"].values[0]
+                    tputs.append(input_tput + output_tput)
+
+            print(model, metric)
+            print(tputs)
+
+            ax = axes[i, j]
+
+            bars = ax.bar(
+                ["vllm", "trt", "lmdeploy", "tgi"],
+                tputs,
+            )
+            for idx, bar in enumerate(bars):
+                bar.set_color(bar_colors[idx])
+
+            ax.set_ylim(bottom=0)
+
+            ax.set_ylabel("Tput (token/s)")
+            ax.set_title(f"{model} {metric}")
+            ax.grid(axis='y')
+
+    fig.tight_layout()
+    fig.savefig("nightly_results.png", bbox_inches='tight', dpi=400)
+
+
+if __name__ == '__main__':
+    args = parse_arguments()
+    main(args)

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

@@ -0,0 +1,218 @@
+#!/bin/bash
+
+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 $gpu_type"
+}
+
+kill_gpu_processes() {
+  pkill lmdeploy || true
+  # waiting for GPU processes to be fully killed
+  sleep 10
+  # Print the GPU memory usage
+  # so that we know if all GPU processes are killed.
+  gpu_memory_usage=$(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"
+}
+
+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 -s localhost:8000/v1/completions > /dev/null; do
+      sleep 1
+    done' && return 0 || return 1
+}
+
+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_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 $test_name."
+      continue
+    fi
+    
+    # append lmdeploy to the test name
+    test_name=lmdeploy_$test_name
+    
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+
+    
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r '.lmdeploy_server_parameters')
+    client_params=$(echo "$params" | jq -r '.lmdeploy_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
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    # prepare tokenizer
+    rm -rf /tokenizer_cache
+    mkdir /tokenizer_cache
+    python ../.buildkite/nightly-benchmarks/scripts/download-tokenizer.py \
+      --model "$model" \
+      --cachedir /tokenizer_cache
+
+    server_command="lmdeploy serve api_server $model \
+      --tp $tp \
+      --server-port $port \
+      $server_args"
+
+    # run the server
+    echo "Running test case $test_name"
+    echo "Server command: $server_command"
+    bash -c "$server_command" &
+
+    # wait until the server is alive
+    wait_for_server
+    if [ $? -eq 0 ]; then
+      echo ""
+      echo "lmdeploy server is up and running."
+    else
+      echo ""
+      echo "lmdeploy failed to start within the timeout period."
+      break
+    fi
+
+    # get model name
+    model_name=$(python ../.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py)
+
+    # 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 \
+        --backend lmdeploy \
+        --tokenizer /tokenizer_cache \
+        --dataset-name $dataset_name \
+        --dataset-path $dataset_path \
+        --num-prompts $num_prompts \
+        --port $port \
+        --save-result \
+        --result-dir $RESULTS_FOLDER \
+        --result-filename ${new_test_name}.json \
+        --request-rate $qps \
+        --model \"$model_name\" \
+        $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" \
+        --arg engine "lmdeploy" \
+        '{
+          server_command: $server,
+          client_command: $client,
+          gpu_type: $gpu,
+          engine: $engine
+        }')
+      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
+
+    done
+
+    # clean up
+    kill_gpu_processes
+    rm -rf /root/.cache/huggingface/*
+  done
+}
+
+
+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 "success" --context "benchmark-results" --append < $RESULTS_FOLDER/${CURRENT_LLM_SERVING_ENGINE}_nightly_results.md
+  /workspace/buildkite-agent artifact upload "$RESULTS_FOLDER/*"
+}
+
+
+main() {
+
+  check_gpus
+  # enter vllm directory
+  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+
+  declare -g RESULTS_FOLDER=results/
+  mkdir -p $RESULTS_FOLDER
+  BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
+
+  python -m pip install transformers==4.41.2
+
+  export CURRENT_LLM_SERVING_ENGINE=lmdeploy
+  run_serving_tests $BENCHMARK_ROOT/tests/nightly-tests.json
+  python -m pip install tabulate pandas
+  python $BENCHMARK_ROOT/scripts/summary-nightly-results.py
+  upload_to_buildkite
+
+}
+
+main "$@"

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

@@ -0,0 +1,216 @@
+#!/bin/bash
+
+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 $gpu_type"
+}
+
+kill_gpu_processes() {
+  pkill text-generation || true
+  # waiting for GPU processes to be fully killed
+  sleep 10
+  # Print the GPU memory usage
+  # so that we know if all GPU processes are killed.
+  gpu_memory_usage=$(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"
+}
+
+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() {
+  timeout 1200 bash -c '
+    until curl -s localhost:8000/generate_stream > /dev/null; do
+      sleep 1
+    done' && return 0 || return 1
+}
+
+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_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 $test_name."
+      continue
+    fi
+
+    # append tgi to the test name
+    test_name=tgi_$test_name
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r '.tgi_server_parameters')
+    client_params=$(echo "$params" | jq -r '.tgi_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
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if echo "$common_params" | jq -e 'has("fp8")' > /dev/null; then
+      echo "Key 'fp8' exists in common params."
+      server_command="/tgi-entrypoint.sh \
+        --model-id $model \
+        --num-shard $tp \
+        --port $port \
+        --quantize fp8 \
+        $server_args"
+    else
+      echo "Key 'fp8' does not exist in common params."
+      server_command="/tgi-entrypoint.sh \
+        --model-id $model \
+        --num-shard $tp \
+        --port $port \
+        $server_args"
+    fi
+
+
+    
+
+    # 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 "tgi server is up and running."
+    else
+      echo ""
+      echo "tgi failed to start within the timeout period."
+      break
+    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 \
+        --backend tgi \
+        --model $model \
+        --dataset-name $dataset_name \
+        --dataset-path $dataset_path \
+        --num-prompts $num_prompts \
+        --port $port \
+        --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" \
+        --arg engine "tgi" \
+        '{
+          server_command: $server,
+          client_command: $client,
+          gpu_type: $gpu,
+          engine: $engine
+        }')
+      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
+
+    done
+
+    # clean up
+    kill_gpu_processes
+    rm -rf /root/.cache/huggingface/*
+  done
+}
+
+
+
+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 "success" --context "benchmark-results" --append < $RESULTS_FOLDER/${CURRENT_LLM_SERVING_ENGINE}_nightly_results.md
+  /workspace/buildkite-agent artifact upload "$RESULTS_FOLDER/*"
+}
+
+main() {
+
+  check_gpus
+  # enter vllm directory
+  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+  declare -g RESULTS_FOLDER=results/
+  mkdir -p $RESULTS_FOLDER
+  BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
+
+  export CURRENT_LLM_SERVING_ENGINE=tgi
+  run_serving_tests $BENCHMARK_ROOT/tests/nightly-tests.json
+  python -m pip install tabulate pandas
+  python $BENCHMARK_ROOT/scripts/summary-nightly-results.py
+  upload_to_buildkite
+
+}
+
+main "$@"

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

@@ -0,0 +1,214 @@
+#!/bin/bash
+
+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 $gpu_type"
+}
+
+kill_gpu_processes() {
+  pkill tritonserver || true
+  # waiting for GPU processes to be fully killed
+  sleep 20
+  # Print the GPU memory usage
+  # so that we know if all GPU processes are killed.
+  gpu_memory_usage=$(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"
+}
+
+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() {
+  timeout 1200 bash -c '
+    until curl -s localhost:8000/generate_stream > /dev/null; do
+      sleep 1
+    done' && return 0 || return 1
+}
+
+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_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 $test_name."
+      continue
+    fi
+
+    # append trt to the test name
+    test_name=trt_$test_name
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r '.trt_server_parameters')
+    client_params=$(echo "$params" | jq -r '.trt_client_parameters')
+    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
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required model_tp_size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+
+
+    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+
+
+    echo "Running test case $test_name"
+    bash ../.buildkite/nightly-benchmarks/scripts/launch-trt-server.sh "$server_params" "$common_params"
+
+    # wait until the server is alive
+    wait_for_server
+    if [ $? -eq 0 ]; then
+      echo ""
+      echo "trt server is up and running."
+    else
+      echo ""
+      echo "trt failed to start within the timeout period."
+      break
+    fi
+
+    # prepare tokenizer
+    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    rm -rf /tokenizer_cache
+    mkdir /tokenizer_cache
+    python ../.buildkite/nightly-benchmarks/scripts/download-tokenizer.py \
+      --model "$model" \
+      --cachedir /tokenizer_cache
+    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    
+
+    # 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 \
+        --backend tensorrt-llm \
+        --tokenizer /tokenizer_cache \
+        --model $model \
+        --dataset-name $dataset_name \
+        --dataset-path $dataset_path \
+        --num-prompts $num_prompts \
+        --port $port \
+        --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"
+
+      server_command=""
+      # record the benchmarking commands
+      jq_output=$(jq -n \
+        --arg server "$server_command" \
+        --arg client "$client_command" \
+        --arg gpu "$gpu_type" \
+        --arg engine "trt" \
+        '{
+          server_command: $server,
+          client_command: $client,
+          gpu_type: $gpu,
+          engine: $engine
+        }')
+      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
+
+    done
+
+    # clean up
+    kill_gpu_processes
+    rm -rf /root/.cache/huggingface/*
+  done
+}
+
+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 "success" --context "benchmark-results" --append < $RESULTS_FOLDER/${CURRENT_LLM_SERVING_ENGINE}_nightly_results.md
+  /workspace/buildkite-agent artifact upload "$RESULTS_FOLDER/*"
+}
+
+
+main() {
+
+  check_gpus
+
+
+  # enter vllm directory
+  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+
+  declare -g RESULTS_FOLDER=results/
+  mkdir -p $RESULTS_FOLDER
+  BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
+
+  # update transformers package, to make sure mixtral tokenizer is available
+  python -m pip install transformers -U
+
+  export CURRENT_LLM_SERVING_ENGINE=trt
+  run_serving_tests $BENCHMARK_ROOT/tests/nightly-tests.json
+  python -m pip install tabulate pandas
+  python $BENCHMARK_ROOT/scripts/summary-nightly-results.py
+  upload_to_buildkite
+
+}
+
+main "$@"

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

@@ -0,0 +1,221 @@
+#!/bin/bash
+
+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 $gpu_type"
+}
+
+kill_gpu_processes() {
+  # kill all processes on GPU.
+  pkill pt_main_thread
+  sleep 10
+
+  # remove vllm config file
+  rm -rf ~/.config/vllm
+
+  # Print the GPU memory usage
+  # so that we know if all GPU processes are killed.
+  gpu_memory_usage=$(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"
+}
+
+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 -s localhost:8000/v1/completions > /dev/null; do
+      sleep 1
+    done' && return 0 || return 1
+}
+
+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_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 $test_name."
+      continue
+    fi
+
+    # append vllm to the test name
+    test_name=vllm_$test_name
+
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r '.vllm_server_parameters')
+    client_params=$(echo "$params" | jq -r '.vllm_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
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if echo "$common_params" | jq -e 'has("fp8")' > /dev/null; then
+      echo "Key 'fp8' exists in common params. Use neuralmagic fp8 model for convenience."
+      model=$(echo "$common_params" | jq -r '.neuralmagic_quantized_model')
+      server_command="python3 \
+        -m vllm.entrypoints.openai.api_server \
+        -tp $tp \
+        --model $model \
+        --port $port \
+        $server_args"
+    else
+      echo "Key 'fp8' does not exist in common params."
+      server_command="python3 \
+        -m vllm.entrypoints.openai.api_server \
+        -tp $tp \
+        --model $model \
+        --port $port \
+        $server_args"
+    fi
+
+    # 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."
+      break
+    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 \
+        --backend vllm \
+        --model $model \
+        --dataset-name $dataset_name \
+        --dataset-path $dataset_path \
+        --num-prompts $num_prompts \
+        --port $port \
+        --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" \
+        --arg engine "vllm" \
+        '{
+          server_command: $server,
+          client_command: $client,
+          gpu_type: $gpu,
+          engine: $engine
+        }')
+      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
+
+    done
+
+    # clean up
+    kill_gpu_processes
+    rm -rf /root/.cache/huggingface/*
+  done
+}
+
+
+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 "success" --context "benchmark-results" --append < $RESULTS_FOLDER/${CURRENT_LLM_SERVING_ENGINE}_nightly_results.md
+  /workspace/buildkite-agent artifact upload "$RESULTS_FOLDER/*"
+}
+
+main() {
+
+  check_gpus
+  # enter vllm directory
+  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+  declare -g RESULTS_FOLDER=results/
+  mkdir -p $RESULTS_FOLDER
+  BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
+
+  export CURRENT_LLM_SERVING_ENGINE=vllm
+  run_serving_tests $BENCHMARK_ROOT/tests/nightly-tests.json
+
+  python3 -m pip install tabulate pandas
+  python3 $BENCHMARK_ROOT/scripts/summary-nightly-results.py
+  upload_to_buildkite
+
+}
+
+main "$@"

.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py

@@ -0,0 +1,76 @@
+import datetime
+import json
+import os
+from pathlib import Path
+
+import pandas as pd
+from tabulate import tabulate
+
+results_folder = Path("results/")
+
+# serving results and the keys that will be printed into markdown
+serving_results = []
+serving_column_mapping = {
+    "test_name": "Test name",
+    "gpu_type": "GPU",
+    "completed": "Successful req.",
+    "request_throughput": "Tput (req/s)",
+    "mean_ttft_ms": "Mean TTFT (ms)",
+    "std_ttft_ms": "Std TTFT (ms)",
+    "mean_itl_ms": "Mean ITL (ms)",
+    "std_itl_ms": "Std ITL (ms)",
+    "input_throughput": "Input Tput (tok/s)",
+    "output_throughput": "Output Tput (tok/s)",
+    "engine": "Engine",
+}
+
+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())
+
+        # 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
+
+    serving_results = pd.DataFrame.from_dict(serving_results)
+
+    if not serving_results.empty:
+        serving_results = serving_results[list(
+            serving_column_mapping.keys())].rename(
+                columns=serving_column_mapping)
+
+    serving_md_table_with_headers = tabulate(serving_results,
+                                             headers='keys',
+                                             tablefmt='pipe',
+                                             showindex=False)
+    # remove the first line of header
+    serving_md_table_lines = serving_md_table_with_headers.split('\n')
+    serving_md_table_without_header = '\n'.join(serving_md_table_lines[2:])
+
+    prefix = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+    prefix = prefix + "_" + os.environ.get("CURRENT_LLM_SERVING_ENGINE")
+
+    # document benchmarking results in markdown
+    with open(results_folder / f"{prefix}_nightly_results.md", "w") as f:
+        # document results with header.
+        # for those who wants to reproduce our benchmark.
+        f.write(serving_md_table_with_headers)
+        f.write('\n')
+
+    # document benchmarking results in json
+    with open(results_folder / f"{prefix}_nightly_results.json", "w") as f:
+
+        results = 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/nightly-tests.json

@@ -0,0 +1,116 @@
+[
+    {
+        "test_name": "llama8B_tp1",
+        "qps_list": [4],
+        "common_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B",
+            "tp": 1,
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 500,
+            "port": 8000
+        },
+        "lmdeploy_server_parameters": {
+        },
+        "lmdeploy_client_parameters": {
+        },
+        "tgi_server_parameters": {
+        },
+        "tgi_client_parameters": {
+            "endpoint": "/generate_stream"
+        },
+        "trt_server_parameters": {
+            "model_type": "llama",
+            "model_dtype": "float16",
+            "max_batch_size": 256,
+            "max_input_len": 4096,
+            "max_output_len": 4096,
+            "trt_llm_version": "r24.04"
+        },
+        "trt_client_parameters": {
+            "endpoint": "/v2/models/ensemble/generate_stream"
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "disable_log_requests": ""
+        },
+        "vllm_client_parameters": {
+        }
+    },
+    {
+        "test_name": "llama70B_tp4",
+        "qps_list": [2],
+        "common_parameters": {
+            "model": "meta-llama/Meta-Llama-3-70B-Instruct",
+            "tp": 4,
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 500,
+            "port": 8000
+        },
+        "lmdeploy_server_parameters": {
+        },
+        "lmdeploy_client_parameters": {
+        },
+        "tgi_server_parameters": {
+        },
+        "tgi_client_parameters": {
+            "endpoint": "/generate_stream"
+        },
+        "trt_server_parameters": {
+            "model_type": "llama",
+            "model_dtype": "float16",
+            "max_batch_size": 256,
+            "max_input_len": 4096,
+            "max_output_len": 4096,
+            "trt_llm_version": "r24.04"
+        },
+        "trt_client_parameters": {
+            "endpoint": "/v2/models/ensemble/generate_stream"
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "disable_log_requests": ""
+        },
+        "vllm_client_parameters": {
+        }
+    },
+    {
+        "test_name": "mixtral8x7B_tp2",
+        "qps_list": [2],
+        "common_parameters": {
+            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "tp": 2,
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 500,
+            "port": 8000
+        },
+        "lmdeploy_server_parameters": {
+        },
+        "lmdeploy_client_parameters": {
+        },
+        "tgi_server_parameters": {
+        },
+        "tgi_client_parameters": {
+            "endpoint": "/generate_stream"
+        },
+        "trt_server_parameters": {
+            "model_type": "llama",
+            "model_dtype": "float16",
+            "max_batch_size": 256,
+            "max_input_len": 4096,
+            "max_output_len": 4096,
+            "trt_llm_version": "r24.04"
+        },
+        "trt_client_parameters": {
+            "endpoint": "/v2/models/ensemble/generate_stream"
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "disable_log_requests": ""
+        },
+        "vllm_client_parameters": {
+        }
+    }
+]

.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 +1,19 @@
+steps:
+  - label: "Build wheel - CUDA {{matrix.cuda_version}}"
+    agents:
+      queue: cpu_queue
+    commands:
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg buildkite_commit=$BUILDKITE_COMMIT --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION={{matrix.cuda_version}} --tag vllm-ci:build-image --target build --progress plain ."
+      - "mkdir artifacts"
+      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+      # rename the files to change linux -> manylinux1
+      - "for f in artifacts/dist/*.whl; do mv -- \"$$f\" \"$${f/linux/manylinux1}\"; done"
+      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
+      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/nightly/"
+    env:
+      DOCKER_BUILDKIT: "1"
+    matrix:
+      setup:
+        cuda_version:
+          - "11.8.0"
+          - "12.1.0"

.buildkite/run-amd-test.sh

@@ -2,6 +2,15 @@
 set -ex
 
 # Print ROCm version
+echo "--- Confirming Clean Initial State"
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
+
 echo "--- ROCm info"
 rocminfo
 
@@ -45,15 +54,10 @@ while true; do
         fi
 done
 
-echo "--- Building container"
+echo "--- Pulling container" 
-sha=$(git rev-parse --short HEAD)
+image_name="rocmshared/vllm-ci:${BUILDKITE_COMMIT}"
-image_name=rocm_${sha}
+container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-container_name=rocm_${sha}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)
+docker pull ${image_name}
-docker build \
-        -t ${image_name} \
-        -f Dockerfile.rocm \
-        --progress plain \
-        .
 
 remove_docker_container() {
    docker rm -f ${container_name} || docker image rm -f ${image_name} || true
@@ -62,11 +66,18 @@ trap remove_docker_container EXIT
 
 echo "--- Running container"
 
+HF_CACHE="$(realpath ~)/huggingface"
+mkdir -p ${HF_CACHE}
+HF_MOUNT="/root/.cache/huggingface"
+
 docker run \
         --device /dev/kfd --device /dev/dri \
         --network host \
+        --shm-size=16gb \
         --rm \
         -e HF_TOKEN \
+        -v ${HF_CACHE}:${HF_MOUNT} \
+        -e HF_HOME=${HF_MOUNT} \
         --name ${container_name} \
         ${image_name} \
         /bin/bash -c "${@}"

.buildkite/run-benchmarks.sh

@@ -50,16 +50,16 @@ echo "### Serving Benchmarks" >> benchmark_results.md
 sed -n '1p' benchmark_serving.txt >> benchmark_results.md # first line
 echo "" >> benchmark_results.md
 echo '```' >> benchmark_results.md
-tail -n 20 benchmark_serving.txt >> benchmark_results.md # last 20 lines
+tail -n 24 benchmark_serving.txt >> benchmark_results.md # last 24 lines
 echo '```' >> benchmark_results.md
 
 # if the agent binary is not found, skip uploading the results, exit 0
-if [ ! -f /workspace/buildkite-agent ]; then
+if [ ! -f /usr/bin/buildkite-agent ]; then
     exit 0
 fi
 
 # upload the results to buildkite
-/workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
+buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
 
 # exit with the exit code of the benchmarks
 if [ $bench_latency_exit_code -ne 0 ]; then
@@ -75,4 +75,4 @@ if [ $bench_serving_exit_code -ne 0 ]; then
 fi
 
 rm ShareGPT_V3_unfiltered_cleaned_split.json
-/workspace/buildkite-agent artifact upload "*.json"
+buildkite-agent artifact upload "*.json"

.buildkite/run-cpu-test.sh

@@ -4,11 +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 and launch offline inference
+# Run the image
-docker run --network host --env VLLM_CPU_KVCACHE_SPACE=1 --name cpu-test cpu-test python3 vllm/examples/offline_inference.py
+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
+  cd ../
+  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py" # Mamba on CPU is not supported

.buildkite/run-multi-node-test.sh

@@ -0,0 +1,105 @@
+#!/bin/bash
+
+set -euox pipefail
+
+if [[ $# -lt 4 ]]; then
+    echo "Usage: .buildkite/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
+    exit 1
+fi
+
+WORKING_DIR=$1
+NUM_NODES=$2
+NUM_GPUS=$3
+DOCKER_IMAGE=$4
+
+shift 4
+COMMANDS=("$@")
+if [ ${#COMMANDS[@]} -ne $NUM_NODES ]; then
+    echo "The number of commands must be equal to the number of nodes."
+    echo "Number of nodes: $NUM_NODES"
+    echo "Number of commands: ${#COMMANDS[@]}"
+    exit 1
+fi
+
+echo "List of commands"
+for command in "${COMMANDS[@]}"; do
+    echo $command
+done
+
+start_network() {
+    docker network create --subnet=192.168.10.0/24 docker-net
+}
+
+start_nodes() {
+    for node in $(seq 0 $(($NUM_NODES-1))); do
+        GPU_DEVICES='"device='
+        for node_gpu in $(seq 0 $(($NUM_GPUS - 1))); do
+            DEVICE_NUM=$(($node * $NUM_GPUS + $node_gpu))
+            GPU_DEVICES+=$(($DEVICE_NUM))
+            if [ $node_gpu -lt $(($NUM_GPUS - 1)) ]; then
+                GPU_DEVICES+=','
+            fi
+        done
+        GPU_DEVICES+='"'
+
+        # start the container in detached mode
+        # things to note:
+        # 1. --shm-size=10.24gb is required. don't use --ipc=host
+        # 2. pass HF_TOKEN to the container
+        # 3. map the huggingface cache directory to the container
+        # 3. assign ip addresses to the containers (head node: 192.168.10.10, worker nodes:
+        #    starting from 192.168.10.11)
+        docker run -d --gpus "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN -v ~/.cache/huggingface:/root/.cache/huggingface --name node$node --network docker-net --ip 192.168.10.$((10 + $node)) --rm $DOCKER_IMAGE /bin/bash -c "tail -f /dev/null"
+
+        # organize containers into a ray cluster
+        if [ $node -eq 0 ]; then
+            # start the ray head node
+            docker exec -d node$node /bin/bash -c "ray start --head --port=6379 --block"
+            # wait for the head node to be ready
+            sleep 10
+        else
+            # start the ray worker nodes, and connect them to the head node
+            docker exec -d node$node /bin/bash -c "ray start --address=192.168.10.10:6379 --block"
+        fi
+    done
+
+    # wait for the cluster to be ready
+    sleep 10
+
+    # print the cluster status
+    docker exec node0 /bin/bash -c "ray status"
+}
+
+run_nodes() {
+    # important: iterate in reverse order to start the head node last
+    # we start the worker nodes first, in detached mode, and then start the head node
+    # in the foreground, so that the output of the head node is visible in the buildkite logs
+    for node in $(seq $(($NUM_NODES - 1)) -1 0); do
+        GPU_DEVICES='"device='
+        for node_gpu in $(seq 0 $(($NUM_GPUS - 1))); do
+            DEVICE_NUM=$(($node * $NUM_GPUS + $node_gpu))
+            GPU_DEVICES+=$(($DEVICE_NUM))
+            if [ $node_gpu -lt $(($NUM_GPUS - 1)) ]; then
+                GPU_DEVICES+=','
+            fi
+        done
+        GPU_DEVICES+='"'
+        echo "Running node$node with GPU devices: $GPU_DEVICES"
+        if [ $node -ne 0 ]; then
+            docker exec -d node$node /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
+        else
+            docker exec node$node /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
+        fi
+    done
+}
+cleanup() {
+    for node in $(seq 0 $(($NUM_NODES-1))); do
+        docker stop node$node
+    done
+    docker network rm docker-net
+}
+trap cleanup EXIT
+start_network
+start_nodes
+run_nodes
+

.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-tpu-test.sh

@@ -0,0 +1,16 @@
+set -e
+
+# Build the docker image.
+docker build -f Dockerfile.tpu -t vllm-tpu .
+
+# Set up cleanup.
+remove_docker_container() { docker rm -f tpu-test || true; }
+trap remove_docker_container EXIT
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
+# For HF_TOKEN.
+source /etc/environment
+# Run a simple end-to-end example.
+docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu \
+    python3 /workspace/vllm/examples/offline_inference_tpu.py

.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

@@ -1,11 +1,38 @@
 # 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: Async Engine, Inputs, Utils, Worker Test
+  fast_check: true
+  fast_check_only: true
+  commands:
+  - pytest -v -s async_engine # Async Engine
+  - pytest -v -s test_inputs.py
+  - pytest -v -s multimodal
+  - pytest -v -s test_utils.py # Utils
+  - pytest -v -s worker # Worker
+
+- label: Tensorizer, Metrics, Tracing Test
+  fast_check: true
+  fast_check_only: true
+  commands:
+  - apt-get install -y curl libsodium23 && pytest -v -s tensorizer_loader # Tensorizer
+  - pytest -v -s metrics # Metrics
+  - "pip install \
+      opentelemetry-sdk \
+      opentelemetry-api \
+      opentelemetry-exporter-otlp \
+      opentelemetry-semantic-conventions-ai" # Tracing
+  - pytest -v -s tracing
+
 - label: Regression Test
   mirror_hardwares: [amd]
+  fast_check: true
   command: pytest -v -s test_regression.py
   working_dir: "/vllm-workspace/tests" # optional
 
@@ -15,56 +42,100 @@ steps:
 
 - label: Basic Correctness Test
   mirror_hardwares: [amd]
+  fast_check: true
   commands:
-  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_basic_correctness.py
+  # This flashinfer installation will fail on AMD ROCm, so it is set as optional.
-  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_basic_correctness.py
+  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl || true
+  - pytest -v -s basic_correctness/test_basic_correctness.py
+  - pytest -v -s basic_correctness/test_cpu_offload.py
   - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
   - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Core Test
   mirror_hardwares: [amd]
-  command: pytest -v -s core
+  fast_check: true
+  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: 2 Node Tests (4 GPUs in total)
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  num_nodes: 2
+  commands:
+  - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
+    - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
+  - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
+
+- label: Distributed Tests (2 GPUs)
   mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   commands:
+  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
+  - TEST_DIST_MODEL=microsoft/Phi-3-vision-128k-instruct DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - 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
+  fast_check: true
   commands:
   - pytest -v -s distributed/test_pynccl.py
+  # We want to test that models which use 2 GPUs work with 4 GPUs, which is why we duplicate them here.
+  # See https://github.com/vllm-project/vllm/pull/5473#issuecomment-2166601837 for context.
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
+  - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
+
+- label: Pipeline Parallelism Test
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  commands:
+  - pytest -v -s distributed/test_pipeline_parallel.py
 
 - label: Engine Test
   mirror_hardwares: [amd]
-  command: pytest -v -s engine tokenization test_sequence.py test_config.py test_logger.py
+  commands:
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
+  # OOM in the CI unless we run this separately
+  - pytest -v -s tokenization
 
 - label: Entrypoints Test
+  fast_check: true
   mirror_hardwares: [amd]
 
   commands:
-  - pytest -v -s test_inputs.py
+  - pytest -v -s entrypoints/llm
-  - pytest -v -s entrypoints -m llm
+  - pytest -v -s entrypoints/openai
-  - pytest -v -s entrypoints -m openai
 
 - label: Examples Test
   working_dir: "/vllm-workspace/examples"
@@ -74,27 +145,35 @@ steps:
     # install tensorizer for tensorize_vllm_model.py
     - pip install awscli tensorizer
     - python3 offline_inference.py
+    - python3 cpu_offload.py
     - python3 offline_inference_with_prefix.py
     - python3 llm_engine_example.py
     - python3 llava_example.py
     - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
 
+- label: Inputs Test
+  #mirror_hardwares: [amd]
+  commands:
+    - pytest -v -s test_inputs.py
+    - pytest -v -s multimodal
+
 - label: Kernels Test %N
   #mirror_hardwares: [amd]
-  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/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 4
 
 - label: Models Test
   #mirror_hardwares: [amd]
   commands:
-    - bash ../.buildkite/download-images.sh
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
-    - pytest -v -s models --ignore=models/test_llava.py
+    - 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 vlm
-    - pytest -v -s models/test_llava.py
 
 - label: Prefix Caching Test
   mirror_hardwares: [amd]
@@ -110,7 +189,9 @@ steps:
   command: pytest -v -s test_logits_processor.py
 
 - label: Utils Test
-  command: pytest -v -s test_utils.py
+  commands:
+    - pytest -v -s test_utils.py
+    - pytest -v -s test_embedded_commit.py
 
 - label: Worker Test
   mirror_hardwares: [amd]
@@ -118,7 +199,10 @@ steps:
 
 - label: Speculative decoding tests
   #mirror_hardwares: [amd]
-  command: pytest -v -s spec_decode
+  commands:
+    # See https://github.com/vllm-project/vllm/issues/5152
+    - export VLLM_ATTENTION_BACKEND=XFORMERS
+    - pytest -v -s spec_decode
 
 - label: LoRA Test %N
   #mirror_hardwares: [amd]
@@ -130,18 +214,17 @@ steps:
   num_gpus: 4
   # This test runs llama 13B, so it is required to run on 4 GPUs.
   commands:
-    # Temporarily run this way because we cannot clean up GPU mem usage
+    # FIXIT: find out which code initialize cuda before running the test
-    # for multi GPU tests.
+    # before the fix, we need to use spawn to test it
-    # TODO(sang): Fix it.
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -v -s lora/test_long_context.py::test_rotary_emb_replaced
+    - pytest -v -s -x lora/test_long_context.py
-    - pytest -v -s lora/test_long_context.py::test_batched_rope_kernel
-    - pytest -v -s lora/test_long_context.py::test_self_consistency
-    - pytest -v -s lora/test_long_context.py::test_quality
-    - pytest -v -s lora/test_long_context.py::test_max_len
 
 - label: Tensorizer Test
   #mirror_hardwares: [amd]
-  command: apt-get install curl libsodium23 && pytest -v -s tensorizer_loader
+  commands:
+    - apt-get install -y curl libsodium23
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s tensorizer_loader
 
 - label: Metrics Test
   mirror_hardwares: [amd]
@@ -151,6 +234,15 @@ 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]
@@ -158,9 +250,40 @@ 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"
+  fast_check: true
   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/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+  - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=meta-llama/Meta-Llama-3-8B DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - pytest -v -s -x lora/test_mixtral.py

.buildkite/test-template.j2

@@ -1,93 +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"
-    {% endif %}
-    {% endfor %}
-
-  - label: "Neuron Test"
-    depends_on: ~
-    agents:
-      queue: neuron
-    command: bash .buildkite/run-neuron-test.sh
-    soft_fail: true
-
-  - label: "Intel Test"
-    depends_on: ~
-    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/FUNDING.yml

@@ -0,0 +1,2 @@
+github: [vllm-project]
+open_collective: [vllm]

.github/workflows/add_label_automerge.yml

@@ -0,0 +1,21 @@
+name: Add label on auto-merge enabled
+on:
+    pull_request_target:
+        types:
+            - auto_merge_enabled
+jobs:
+    add-label-on-auto-merge:
+        runs-on: ubuntu-latest
+        steps:
+            -   name: Add label
+                uses: actions/github-script@v5
+                with:
+                    script: |
+                        github.rest.issues.addLabels({
+                            owner: context.repo.owner,
+                            repo: context.repo.repo,
+                            issue_number: context.issue.number,
+                            labels: ['ready']
+                        })
+                env:
+                    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/add_label_ready_comment.yml

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

.github/workflows/mypy.yaml

@@ -32,19 +32,22 @@ jobs:
         pip install types-setuptools
     - name: Mypy
       run: |
+        mypy tests --config-file pyproject.toml
+        mypy vllm/*.py --config-file pyproject.toml
         mypy vllm/attention --config-file pyproject.toml
         mypy vllm/core --config-file pyproject.toml
         mypy vllm/distributed --config-file pyproject.toml
+        mypy vllm/engine  --config-file pyproject.toml
         mypy vllm/entrypoints --config-file pyproject.toml
         mypy vllm/executor --config-file pyproject.toml
-        mypy vllm/usage --config-file pyproject.toml
+        mypy vllm/inputs --config-file pyproject.toml
-        mypy vllm/*.py --config-file pyproject.toml
-        mypy vllm/transformers_utils --config-file pyproject.toml
-        mypy vllm/engine  --config-file pyproject.toml
-        mypy vllm/worker --config-file pyproject.toml
-        mypy vllm/spec_decode --config-file pyproject.toml
-        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 vllm/lora --config-file pyproject.toml
+        mypy vllm/model_executor  --config-file pyproject.toml
+        mypy vllm/multimodal --config-file pyproject.toml
+        mypy vllm/platforms --config-file pyproject.toml
+        mypy vllm/spec_decode --config-file pyproject.toml
+        mypy vllm/transformers_utils --config-file pyproject.toml
+        mypy vllm/usage --config-file pyproject.toml
+        mypy vllm/worker --config-file pyproject.toml
 

.github/workflows/publish.yml

@@ -49,7 +49,7 @@ jobs:
       matrix:
           os: ['ubuntu-20.04']
           python-version: ['3.8', '3.9', '3.10', '3.11']
-          pytorch-version: ['2.3.0']  # Must be the most recent version that meets requirements-cuda.txt.
+          pytorch-version: ['2.3.1']  # Must be the most recent version that meets requirements-cuda.txt.
           cuda-version: ['11.8', '12.1']
 
     steps:

.github/workflows/reminder_comment.yml

@@ -0,0 +1,21 @@
+name: PR Reminder Comment Bot
+on:
+  pull_request_target:
+    types: [opened]
+
+jobs:
+  pr_reminder:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Remind to run full CI on PR
+        uses: actions/github-script@v6
+        with:
+          script: |
+            github.rest.issues.createComment({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: context.issue.number,
+              body: '👋 Hi! Thank you for contributing to the vLLM project.\n Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your `fast-check` build on Buildkite UI. \n\nOnce the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).\n\n To run full CI, you can do one of these:\n- Comment `/ready` on the PR\n- Add `ready` label to the PR\n- Enable auto-merge.\n\n🚀'
+            })
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/ruff.yml

@@ -25,7 +25,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install ruff==0.1.5 codespell==2.2.6 tomli==2.0.1 isort==5.13.2
+        pip install ruff==0.1.5 codespell==2.3.0 tomli==2.0.1 isort==5.13.2
     - name: Analysing the code with ruff
       run: |
         ruff .

.gitignore

@@ -1,3 +1,6 @@
+# vllm commit id, generated by setup.py
+vllm/commit_id.py
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

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}")
@@ -31,9 +32,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx11
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.3.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.3.1")
-set(TORCH_SUPPORTED_VERSION_ROCM_5X "2.0.1")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.5.0")
-set(TORCH_SUPPORTED_VERSION_ROCM_6X "2.1.1")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -66,19 +66,6 @@ endif()
 #
 find_package(Torch REQUIRED)
 
-#
-# Normally `torch.utils.cpp_extension.CUDAExtension` would add
-# `libtorch_python.so` for linking against an extension. Torch's cmake
-# configuration does not include this library (presumably since the cmake
-# config is used for standalone C++ binaries that link against torch).
-# The `libtorch_python.so` library defines some of the glue code between
-# torch/python via pybind and is required by VLLM extensions for this
-# reason. So, add it by manually with `find_library` using torch's
-# installed library path.
-#
-find_library(torch_python_LIBRARY torch_python PATHS
-  "${TORCH_INSTALL_PREFIX}/lib")
-
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -111,18 +98,11 @@ elseif(HIP_FOUND)
   # .hip extension automatically, HIP must be enabled explicitly.
   enable_language(HIP)
 
-  # ROCm 5.x
+  # ROCm 5.X and 6.X
-  if (ROCM_VERSION_DEV_MAJOR EQUAL 5 AND
+  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
-      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_5X})
+      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
-    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_5X} "
+    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
-      "expected for ROCMm 5.x build, saw ${Torch_VERSION} instead.")
+      "expected for ROCm 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.")
   endif()
 else()
   message(FATAL_ERROR "Can't find CUDA or HIP installation.")
@@ -171,7 +151,8 @@ set(VLLM_EXT_SRC
   "csrc/quantization/fp8/common.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
-  "csrc/pybind.cpp")
+  "csrc/prepare_inputs/advance_step.cu"
+  "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   include(FetchContent)
@@ -191,10 +172,12 @@ 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/gptq_marlin/awq_marlin_repack.cu"
+    "csrc/quantization/fp8/fp8_marlin.cu"
     "csrc/custom_all_reduce.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c2x.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu")
+    "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu")
 
   #
   # The CUTLASS kernels for Hopper require sm90a to be enabled.
@@ -202,7 +185,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # That adds an extra 17MB to compiled binary, so instead we selectively enable it.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0)
     set_source_files_properties(
-          "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu"
+          "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
           PROPERTIES
           COMPILE_FLAGS
           "-gencode arch=compute_90a,code=sm_90a")
@@ -218,6 +201,7 @@ define_gpu_extension_target(
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
   INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  USE_SABI 3
   WITH_SOABI)
 
 #
@@ -225,7 +209,7 @@ define_gpu_extension_target(
 #
 
 set(VLLM_MOE_EXT_SRC
-  "csrc/moe/moe_ops.cpp"
+  "csrc/moe/torch_bindings.cpp"
   "csrc/moe/topk_softmax_kernels.cu")
 
 define_gpu_extension_target(
@@ -235,6 +219,7 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  USE_SABI 3
   WITH_SOABI)
 
 #
@@ -249,7 +234,7 @@ set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
   "csrc/punica/punica_ops.cu"
-  "csrc/punica/punica_pybind.cpp")
+  "csrc/punica/torch_bindings.cpp")
 
 #
 # Copy GPU compilation flags+update for punica
@@ -286,6 +271,7 @@ if (VLLM_PUNICA_GPU_ARCHES)
     SOURCES ${VLLM_PUNICA_EXT_SRC}
     COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
     ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    USE_SABI 3
     WITH_SOABI)
 else()
   message(WARNING "Unable to create _punica_C target because none of the "
@@ -311,6 +297,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   message(STATUS "Enabling C extension.")
   add_dependencies(default _C)
 
+  message(STATUS "Enabling moe extension.")
+  add_dependencies(default _moe_C)
+
   # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
   # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
   # there are supported target arches.
@@ -320,8 +309,3 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     add_dependencies(default _punica_C)
   endif()
 endif()
-
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  message(STATUS "Enabling moe extension.")
-  add_dependencies(default _moe_C)
-endif()

Dockerfile

@@ -5,18 +5,38 @@
 # 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:${CUDA_VERSION}-devel-ubuntu20.04 AS base
+
+ARG CUDA_VERSION=12.4.1
+ARG PYTHON_VERSION=3.10
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
+    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
+    && apt-get update -y \
+    && apt-get install -y ccache software-properties-common \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+    && if [ "${PYTHON_VERSION}" != "3" ]; then update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; fi \
+    && python3 --version
 
 RUN apt-get update -y \
-    && apt-get install -y python3-pip git
+    && apt-get install -y git curl sudo
+
+# Install pip s.t. it will be compatible with our PYTHON_VERSION
+RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}
+RUN python3 -m pip --version
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # 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
 
@@ -24,12 +44,11 @@ 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-mamba.txt requirements-mamba.txt
-COPY requirements-dev.txt requirements-dev.txt
+RUN python3 -m pip install packaging
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN python3 -m pip install -r requirements-mamba.txt
-    pip install -r requirements-dev.txt
 
 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
@@ -39,14 +58,16 @@ 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=3.10
 
 # 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
@@ -70,10 +91,36 @@ ENV NVCC_THREADS=$nvcc_threads
 # make sure punica kernels are built (for LoRA)
 ENV VLLM_INSTALL_PUNICA_KERNELS=1
 
+ARG buildkite_commit
+ENV BUILDKITE_COMMIT=${buildkite_commit}
+
+ARG USE_SCCACHE
+# if USE_SCCACHE is set, use sccache to speed up compilation
+RUN --mount=type=cache,target=/root/.cache/pip \
+    if [ "$USE_SCCACHE" = "1" ]; then \
+        echo "Installing sccache..." \
+        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
+        && tar -xzf sccache.tar.gz \
+        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
+        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
+        && if [ "$CUDA_VERSION" = "11.8.0" ]; then \
+            export SCCACHE_BUCKET=vllm-build-sccache-2; \
+           else \
+            export SCCACHE_BUCKET=vllm-build-sccache; \
+           fi \
+        && export SCCACHE_REGION=us-west-2 \
+        && export CMAKE_BUILD_TYPE=Release \
+        && sccache --show-stats \
+        && python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38 \
+        && sccache --show-stats; \
+    fi
+
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/pip \
-    python3 setup.py bdist_wheel --dist-dir=dist
+    if [ "$USE_SCCACHE" != "1" ]; then \
+        python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
+    fi
 
 # check the size of the wheel, we cannot upload wheels larger than 100MB
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
@@ -81,24 +128,73 @@ 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-ubuntu20.04 AS vllm-base
+ARG CUDA_VERSION=12.4.1
+ARG PYTHON_VERSION=3.10
 WORKDIR /vllm-workspace
 
+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${PYTHON_VERSION}-venv \
+    && if [ "${PYTHON_VERSION}" != "3" ]; then update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; fi \
+    && python3 --version
+
 RUN apt-get update -y \
-    && apt-get install -y python3-pip git vim
+    && apt-get install -y python3-pip git vim curl libibverbs-dev
+
+# Install pip s.t. it will be compatible with our PYTHON_VERSION
+RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}
+RUN python3 -m pip --version
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # 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
+
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.9/flashinfer-0.0.9+cu121torch2.3-cp310-cp310-linux_x86_64.whl
 #################### vLLM installation IMAGE ####################
 
 
@@ -111,7 +207,7 @@ 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
@@ -128,7 +224,7 @@ 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

@@ -1,13 +1,25 @@
 # This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
 
-FROM ubuntu:22.04
+FROM ubuntu:22.04 AS cpu-test-1
 
 RUN apt-get update  -y \
-    && apt-get install -y git wget vim numactl gcc-12 g++-12 python3 python3-pip \
+    && 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
 
+# 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
+
 RUN pip install --upgrade pip \
-    && pip install wheel packaging ninja setuptools>=49.4.0 numpy
+    && pip install wheel packaging ninja "setuptools>=49.4.0" numpy
+
+FROM cpu-test-1 AS build
 
 COPY ./ /workspace/vllm
 
@@ -15,8 +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/
 
-CMD ["/bin/bash"]
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.neuron

@@ -28,7 +28,7 @@ COPY ./requirements-neuron.txt /app/vllm/requirements-neuron.txt
 RUN cd /app/vllm \
     && python3 -m pip install -U -r requirements-neuron.txt
 
-ENV VLLM_BUILD_WITH_NEURON 1
+ENV VLLM_TARGET_DEVICE neuron
 RUN cd /app/vllm \
     && pip install -e . \
     && cd ..

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:20.04 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 +1,33 @@
-# default base image
+# Default ROCm 6.1 base image
-ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+ARG BASE_IMAGE="rocm/pytorch:rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging"
 
-FROM $BASE_IMAGE
+# Default ROCm ARCHes to build vLLM for.
+ARG PYTORCH_ROCM_ARCH="gfx908;gfx90a;gfx942;gfx1100"
 
-ARG BASE_IMAGE="rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
+# Whether to install CK-based flash-attention
-
+# If 0, will not install flash-attention
-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
 ARG BUILD_FA="1"
+# If `TRY_FA_WHEEL=1`, we will try installing flash-attention from `FA_WHEEL_URL`
+# If this succeeds, we use the downloaded wheel and skip building flash-attention.
+# Otherwise, ROCm flash-attention from `FA_BRANCH` will be built for the
+# architectures specified in `FA_GFX_ARCHS`
+ARG TRY_FA_WHEEL="1"
+ARG FA_WHEEL_URL="https://github.com/ROCm/flash-attention/releases/download/v2.5.9post1-cktile-vllm/flash_attn-2.5.9.post1-cp39-cp39-linux_x86_64.whl"
+ARG FA_GFX_ARCHS="gfx90a;gfx942"
+ARG FA_BRANCH="23a2b1c2"
 
-# whether to build triton on rocm
+# Whether to build triton on rocm
 ARG BUILD_TRITON="1"
+ARG TRITON_BRANCH="e0fc12c"
+
+### 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 \
@@ -40,75 +38,152 @@ 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 /vllm-workspace
-# When launching the container, mount the code directory to /app
 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
+# Remove sccache so it doesn't interfere with ccache
+# TODO: implement sccache support across components
+RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)"
+# Install torch == 2.5.0 on ROCm
+RUN case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+        *"rocm-6.1"*) \
+            python3 -m pip uninstall -y torch torchaudio torchvision \
+            && python3 -m pip install --no-cache-dir --pre \
+                torch==2.5.0.dev20240710 torchaudio==2.4.0.dev20240710 \
+                torchvision==0.20.0.dev20240710 \
+               --index-url https://download.pytorch.org/whl/nightly/rocm6.1;; \
+        *) ;; 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
+ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
-RUN if [ "$BUILD_FA" = "1" ]; then \
+ENV CCACHE_DIR=/root/.cache/ccache
-    mkdir libs \
+
-    && cd libs \
+
-    && git clone https://github.com/ROCm/flash-attention.git \
+### AMD-SMI build stage
-    && cd flash-attention \
+FROM base AS build_amdsmi
-    && git checkout ${FA_BRANCH} \
+# Build amdsmi wheel always
-    && git submodule update --init \
+RUN cd /opt/rocm/share/amd_smi \
-    && export GPU_ARCHS=${FA_GFX_ARCHS} \
+    && python3 -m pip wheel . --wheel-dir=/install
-    && if [ "$BASE_IMAGE" = "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1" ]; then \
+
-        patch /opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/hipify/hipify_python.py hipify_patch.patch; fi \
+
-    && python3 setup.py install \
+### Flash-Attention wheel build stage
-    && cd ..; \
+FROM base AS build_fa
+ARG BUILD_FA
+ARG TRY_FA_WHEEL
+ARG FA_WHEEL_URL
+ARG FA_GFX_ARCHS
+ARG FA_BRANCH
+# Build ROCm flash-attention wheel if `BUILD_FA = 1`
+RUN --mount=type=cache,target=${CCACHE_DIR} \
+    if [ "$BUILD_FA" = "1" ]; then \
+        if [ "${TRY_FA_WHEEL}" = "1" ] && python3 -m pip install "${FA_WHEEL_URL}"; then \
+            # If a suitable wheel exists, we download it instead of building FA
+            mkdir -p /install && wget -N "${FA_WHEEL_URL}" -P /install; \
+        else \
+            mkdir -p libs \
+            && cd libs \
+            && git clone https://github.com/ROCm/flash-attention.git \
+            && cd flash-attention \
+            && git checkout "${FA_BRANCH}" \
+            && git submodule update --init \
+            && GPU_ARCHS="${FA_GFX_ARCHS}" python3 setup.py bdist_wheel --dist-dir=/install; \
+        fi; \
+    # 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
+### Triton wheel build stage
-RUN if [ "$BUILD_TRITON" = "1" ]; then \
+FROM base AS build_triton
+ARG BUILD_TRITON
+ARG TRITON_BRANCH
+# Build triton wheel if `BUILD_TRITON = 1`
+RUN --mount=type=cache,target=${CCACHE_DIR} \
+    if [ "$BUILD_TRITON" = "1" ]; then \
     mkdir -p libs \
     && cd libs \
-    && pip uninstall -y triton \
+    && git clone https://github.com/OpenAI/triton.git \
-    && git clone https://github.com/ROCm/triton.git \
+    && cd triton \
-    && cd triton/python \
+    && git checkout "${TRITON_BRANCH}" \
-    && pip3 install . \
+    && cd python \
-    && cd ../..; \
+    && 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
+# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
-RUN python3 -m pip install --upgrade pip numba
+# 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 \
+    python3 -m 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=${CCACHE_DIR} \
+    --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install -Ur requirements-rocm.txt \
+    && case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+        *"rocm-6.1"*) \
+            # Bring in upgrades to HIP graph earlier than ROCm 6.2 for vLLM
+            wget -N https://github.com/ROCm/vllm/raw/fa78403/rocm_patch/libamdhip64.so.6 -P /opt/rocm/lib \
+            # Prevent interference if torch bundles its own HIP runtime
+            && rm -f "$(python3 -c 'import torch; print(torch.__path__[0])')"/lib/libamdhip64.so* || true;; \
+        *) ;; esac \
+    && python3 setup.py clean --all \
+    && python3 setup.py develop
 
+# 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
+    && python3 -m 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
+        && python3 -m 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
+        && python3 -m pip uninstall -y flash-attn; fi
+
+# Install wheels that were built to the final image
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -U -r requirements-rocm.txt \
+    if ls libs/*.whl; then \
-    && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
+    python3 -m pip install libs/*.whl; fi
-    && python3 setup.py install \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cd ..
-
 
 CMD ["/bin/bash"]

Dockerfile.tpu

@@ -0,0 +1,20 @@
+ARG NIGHTLY_DATE="20240713"
+ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
+
+FROM $BASE_IMAGE
+WORKDIR /workspace
+
+# Install aiohttp separately to avoid build errors.
+RUN pip install aiohttp
+# Install NumPy 1 instead of NumPy 2.
+RUN pip install "numpy<2"
+# Install the TPU and Pallas dependencies.
+RUN pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
+RUN pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
+
+# Build vLLM.
+COPY . /workspace/vllm
+ENV VLLM_TARGET_DEVICE="tpu"
+RUN cd /workspace/vllm && python setup.py develop
+
+CMD ["/bin/bash"]

Dockerfile.xpu

@@ -0,0 +1,22 @@
+FROM intel/oneapi-basekit:2024.1.0-devel-ubuntu20.04
+
+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

@@ -16,26 +16,22 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
-**The Fourth vLLM Bay Area Meetup (June 11th 5:30pm-8pm PT)**
+**The Fifth vLLM Bay Area Meetup (July 24th 5pm-8pm PT)**
 
-We are thrilled to announce our fourth vLLM Meetup!
+We are excited to announce our fifth vLLM Meetup!
-The vLLM team will share recent updates and roadmap.
+Join us to hear the vLLM's recent updates and the upcoming roadmap.
-We will also have vLLM collaborators from BentoML and Cloudflare coming up to the stage to discuss their experience in deploying LLMs with vLLM.
+Additionally, our collaborators from AWS will be presenting their insights and experiences in deploying vLLM.
-Please register [here](https://lu.ma/agivllm) and join us!
+Register now [here](https://lu.ma/lp0gyjqr) and be part of the event!
 
 ---
 
 *Latest News* 🔥
+- [2024/07] In partnership with Meta, vLLM officially supports Llama 3.1 with FP8 quantization and pipeline parallelism! Please check out our blog post [here](https://blog.vllm.ai/2024/07/23/llama31.html).
+- [2024/06] We hosted [the fourth vLLM meetup](https://lu.ma/agivllm) with Cloudflare and BentoML! Please find the meetup slides [here](https://docs.google.com/presentation/d/1iJ8o7V2bQEi0BFEljLTwc5G1S10_Rhv3beed5oB0NJ4/edit?usp=sharing).
 - [2024/04] We hosted [the third vLLM meetup](https://robloxandvllmmeetup2024.splashthat.com/) with Roblox! Please find the meetup slides [here](https://docs.google.com/presentation/d/1A--47JAK4BJ39t954HyTkvtfwn0fkqtsL8NGFuslReM/edit?usp=sharing).
-- [2024/01] We hosted [the second vLLM meetup](https://lu.ma/ygxbpzhl) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing).
+- [2024/01] We hosted [the second vLLM meetup](https://lu.ma/ygxbpzhl) with IBM! Please find the meetup slides [here](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing).
-- [2024/01] Added ROCm 6.0 support to vLLM.
+- [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) with a16z! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing).
-- [2023/12] Added ROCm 5.7 support to vLLM.
-- [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing).
-- [2023/09] We created our [Discord server](https://discord.gg/jz7wjKhh6g)! Join us to discuss vLLM and LLM serving! We will also post the latest announcements and updates there.
-- [2023/09] We released our [PagedAttention paper](https://arxiv.org/abs/2309.06180) on arXiv!
 - [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
-- [2023/07] Added support for LLaMA-2! You can run and serve 7B/13B/70B LLaMA-2s on vLLM with a single command!
-- [2023/06] Serving vLLM On any Cloud with SkyPilot. Check out a 1-click [example](https://github.com/skypilot-org/skypilot/blob/master/llm/vllm) to start the vLLM demo, and the [blog post](https://blog.skypilot.co/serving-llm-24x-faster-on-the-cloud-with-vllm-and-skypilot/) for the story behind vLLM development on the clouds.
 - [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
 
 ---
@@ -51,14 +47,16 @@ vLLM is fast with:
 - Quantization: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [SqueezeLLM](https://arxiv.org/abs/2306.07629), FP8 KV Cache
 - Optimized CUDA kernels
 
+**Performance benchmark**: We include a [performance benchmark](https://buildkite.com/vllm/performance-benchmark/builds/3924) that compares the performance of vllm against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [text-generation-inference](https://github.com/huggingface/text-generation-inference) and [lmdeploy](https://github.com/InternLM/lmdeploy)).
+
 vLLM is flexible and easy to use with:
 
 - Seamless integration with popular Hugging Face models
 - High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
-- Tensor parallelism support for distributed inference
+- Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs and AMD GPUs
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs
 - (Experimental) Prefix caching support
 - (Experimental) Multi-lora support
 
@@ -102,14 +100,17 @@ vLLM is a community project. Our compute resources for development and testing a
 - Databricks
 - DeepInfra
 - Dropbox
+- Google Cloud
 - Lambda Lab
 - NVIDIA
 - Replicate
 - Roblox
 - RunPod
+- Sequoia Capital
 - 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)
 
@@ -68,9 +71,13 @@ async def async_request_tgi(
                         chunk_bytes = chunk_bytes.strip()
                         if not chunk_bytes:
                             continue
+                        chunk_bytes = chunk_bytes.decode("utf-8")
 
-                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                        #NOTE: Sometimes TGI returns a ping response without
-                                              "data:")
+                        # any data, we should skip it.
+                        if chunk_bytes.startswith(":"):
+                            continue
+                        chunk = remove_prefix(chunk_bytes, "data:")
 
                         data = json.loads(chunk)
                         timestamp = time.perf_counter()
@@ -218,8 +225,8 @@ async def async_request_openai_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "v1/completions"
+        "completions"
-    ), "OpenAI Completions API URL must end with 'v1/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
@@ -258,6 +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
@@ -266,12 +276,8 @@ async def async_request_openai_completions(
                                     output.ttft = ttft
 
                                 # Decoding phase
-                                # NOTE: Some completion API might have a last
+                                output.itl.append(timestamp -
-                                # usage summary response without a token so we
+                                                  most_recent_timestamp)
-                                # do not want to include as inter-token-latency
-                                elif data.get("usage", None) is None:
-                                    output.itl.append(timestamp -
-                                                      most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
                                 generated_text += data["choices"][0]["text"]
@@ -298,8 +304,8 @@ async def async_request_openai_chat_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "v1/chat/completions"
+        "chat/completions"
-    ), "OpenAI Chat Completions API URL must end with 'v1/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
@@ -384,6 +390,30 @@ def remove_prefix(text: str, prefix: str) -> str:
     return text
 
 
+def get_model(pretrained_model_name_or_path: str) -> str:
+    if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
+        from modelscope 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
+    return pretrained_model_name_or_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,
@@ -392,4 +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.inputs import PromptStrictInputs
+from vllm.engine.arg_utils import EngineArgs
+from vllm.inputs import PromptInputs
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.utils import FlexibleArgumentParser
 
 
 def main(args: argparse.Namespace):
@@ -19,24 +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,
+    llm = LLM(
-              speculative_model=args.speculative_model,
+        model=args.model,
-              num_speculative_tokens=args.num_speculative_tokens,
+        speculative_model=args.speculative_model,
-              tokenizer=args.tokenizer,
+        num_speculative_tokens=args.num_speculative_tokens,
-              quantization=args.quantization,
+        speculative_draft_tensor_parallel_size=\
-              tensor_parallel_size=args.tensor_parallel_size,
+            args.speculative_draft_tensor_parallel_size,
-              trust_remote_code=args.trust_remote_code,
+        tokenizer=args.tokenizer,
-              dtype=args.dtype,
+        quantization=args.quantization,
-              enforce_eager=args.enforce_eager,
+        tensor_parallel_size=args.tensor_parallel_size,
-              kv_cache_dtype=args.kv_cache_dtype,
+        trust_remote_code=args.trust_remote_code,
-              quantization_param_path=args.quantization_param_path,
+        dtype=args.dtype,
-              device=args.device,
+        max_model_len=args.max_model_len,
-              ray_workers_use_nsight=args.ray_workers_use_nsight,
+        enforce_eager=args.enforce_eager,
-              use_v2_block_manager=args.use_v2_block_manager,
+        kv_cache_dtype=args.kv_cache_dtype,
-              enable_chunked_prefill=args.enable_chunked_prefill,
+        quantization_param_path=args.quantization_param_path,
-              download_dir=args.download_dir,
+        device=args.device,
-              block_size=args.block_size,
+        ray_workers_use_nsight=args.ray_workers_use_nsight,
-              gpu_memory_utilization=args.gpu_memory_utilization)
+        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,
@@ -50,7 +61,7 @@ def main(args: argparse.Namespace):
     dummy_prompt_token_ids = np.random.randint(10000,
                                                size=(args.batch_size,
                                                      args.input_len))
-    dummy_inputs: List[PromptStrictInputs] = [{
+    dummy_inputs: List[PromptInputs] = [{
         "prompt_token_ids": batch
     } for batch in dummy_prompt_token_ids.tolist()]
 
@@ -95,7 +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):
@@ -113,12 +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',
@@ -144,6 +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,
@@ -187,9 +208,10 @@ if __name__ == '__main__':
     parser.add_argument(
         "--device",
         type=str,
-        default="cuda",
+        default="auto",
-        choices=["cuda", "cpu"],
+        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
-        help='device type for vLLM execution, supporting CUDA and CPU.')
+        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
+        'CPU.')
     parser.add_argument('--block-size',
                         type=int,
                         default=16,
@@ -199,6 +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",
@@ -221,5 +246,40 @@ 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'],
+        default=None,
+        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

@@ -2,8 +2,8 @@
 
 On the server side, run one of the following commands:
     vLLM OpenAI API server
-    python -m vllm.entrypoints.openai.api_server \
+    vllm serve <your_model> \
-        --model <your_model> --swap-space 16 \
+        --swap-space 16 \
         --disable-log-requests
 
     (TGI backend)
@@ -17,7 +17,7 @@ On the client side, run:
         --dataset-path <path to dataset> \
         --request-rate <request_rate> \ # By default <request_rate> is inf
         --num-prompts <num_prompts> # By default <num_prompts> is 1000
-        
+
     when using tgi backend, add
         --endpoint /generate_stream
     to the end of the command above.
@@ -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
@@ -52,10 +60,16 @@ class BenchmarkMetrics:
     output_throughput: float
     mean_ttft_ms: float
     median_ttft_ms: float
+    std_ttft_ms: float
     p99_ttft_ms: float
     mean_tpot_ms: float
     median_tpot_ms: float
+    std_tpot_ms: float
     p99_tpot_ms: float
+    mean_itl_ms: float
+    median_itl_ms: float
+    std_itl_ms: float
+    p99_itl_ms: float
 
 
 def sample_sharegpt_requests(
@@ -66,7 +80,6 @@ def sample_sharegpt_requests(
 ) -> List[Tuple[str, int, int]]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
-
     # Load the dataset.
     with open(dataset_path) as f:
         dataset = json.load(f)
@@ -174,6 +187,31 @@ def sample_sonnet_requests(
     return sampled_requests
 
 
+def sample_random_requests(
+        input_len: int, output_len: int, num_prompts: int, range_ratio: float,
+        tokenizer: PreTrainedTokenizerBase) -> List[Tuple[str, int, int]]:
+
+    input_lens = np.random.randint(
+        int(input_len * range_ratio),
+        input_len + 1,
+        size=num_prompts,
+    )
+    output_lens = np.random.randint(
+        int(output_len * range_ratio),
+        output_len + 1,
+        size=num_prompts,
+    )
+    offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)
+    input_requests = []
+    for i in range(num_prompts):
+        prompt = tokenizer.decode([(offsets[i] + i + j) % tokenizer.vocab_size
+                                   for j in range(input_lens[i])])
+        input_requests.append(
+            (prompt, int(input_lens[i]), int(output_lens[i])))
+
+    return input_requests
+
+
 async def get_request(
     input_requests: List[Tuple[str, int, int]],
     request_rate: float,
@@ -185,6 +223,7 @@ async def get_request(
         if request_rate == float("inf"):
             # If the request rate is infinity, then we don't need to wait.
             continue
+
         # Sample the request interval from the exponential distribution.
         interval = np.random.exponential(1.0 / request_rate)
         # The next request will be sent after the interval.
@@ -197,19 +236,27 @@ 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
-    tpots = []
+    itls: List[float] = []
-    ttfts = []
+    tpots: List[float] = []
+    ttfts: List[float] = []
     for i in range(len(outputs)):
         if outputs[i].success:
-            output_len = len(tokenizer(outputs[i].generated_text).input_ids)
+            # We use the tokenizer to count the number of output tokens for all
+            # serving backends instead of looking at len(outputs[i].itl) since
+            # multiple output tokens may be bundled together
+            # Note : this may inflate the output token count slightly
+            output_len = len(
+                tokenizer(outputs[i].generated_text,
+                          add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
             total_input += input_requests[i][1]
             if output_len > 1:
                 tpots.append(
                     (outputs[i].latency - outputs[i].ttft) / (output_len - 1))
+            itls += outputs[i].itl
             ttfts.append(outputs[i].ttft)
             completed += 1
         else:
@@ -230,10 +277,16 @@ def calculate_metrics(
         mean_ttft_ms=np.mean(ttfts or 0) *
         1000,  # ttfts is empty if streaming is not supported by backend
         median_ttft_ms=np.median(ttfts or 0) * 1000,
+        std_ttft_ms=np.std(ttfts or 0) * 1000,
         p99_ttft_ms=np.percentile(ttfts or 0, 99) * 1000,
         mean_tpot_ms=np.mean(tpots or 0) * 1000,
         median_tpot_ms=np.median(tpots or 0) * 1000,
+        std_tpot_ms=np.std(tpots or 0) * 1000,
         p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
+        mean_itl_ms=np.mean(itls or 0) * 1000,
+        median_itl_ms=np.median(itls or 0) * 1000,
+        std_itl_ms=np.std(itls or 0) * 1000,
+        p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
     )
 
     return metrics, actual_output_lens
@@ -251,7 +304,7 @@ 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}")
 
@@ -278,7 +331,7 @@ 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(
@@ -296,7 +349,7 @@ 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
@@ -333,6 +386,10 @@ async def benchmark(
     print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
                                     metrics.median_tpot_ms))
     print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
+    print("{s:{c}^{n}}".format(s='Inter-token Latency', n=50, c='-'))
+    print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
+    print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
+    print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
     print("=" * 50)
 
     result = {
@@ -345,10 +402,16 @@ async def benchmark(
         "output_throughput": metrics.output_throughput,
         "mean_ttft_ms": metrics.mean_ttft_ms,
         "median_ttft_ms": metrics.median_ttft_ms,
+        "std_ttft_ms": metrics.std_ttft_ms,
         "p99_ttft_ms": metrics.p99_ttft_ms,
         "mean_tpot_ms": metrics.mean_tpot_ms,
         "median_tpot_ms": metrics.median_tpot_ms,
+        "std_tpot_ms": metrics.std_tpot_ms,
         "p99_tpot_ms": metrics.p99_tpot_ms,
+        "mean_itl_ms": metrics.mean_itl_ms,
+        "median_itl_ms": metrics.median_itl_ms,
+        "std_itl_ms": metrics.std_itl_ms,
+        "p99_itl_ms": metrics.p99_itl_ms,
         "input_lens": [output.prompt_len for output in outputs],
         "output_lens": actual_output_lens,
         "ttfts": [output.ttft for output in outputs],
@@ -427,6 +490,15 @@ def main(args: argparse.Namespace):
                               for prompt, prompt_formatted, prompt_len,
                               output_len in input_requests]
 
+    elif args.dataset_name == "random":
+        input_requests = sample_random_requests(
+            input_len=args.random_input_len,
+            output_len=args.random_output_len,
+            num_prompts=args.num_prompts,
+            range_ratio=args.random_range_ratio,
+            tokenizer=tokenizer,
+        )
+
     else:
         raise ValueError(f"Unknown dataset: {args.dataset_name}")
 
@@ -445,7 +517,7 @@ 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")
@@ -478,6 +550,8 @@ 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:
@@ -485,7 +559,7 @@ def main(args: argparse.Namespace):
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
+    parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput.")
     parser.add_argument(
         "--backend",
@@ -518,7 +592,7 @@ if __name__ == "__main__":
         "--dataset-name",
         type=str,
         default="sharegpt",
-        choices=["sharegpt", "sonnet"],
+        choices=["sharegpt", "sonnet", "random"],
         help="Name of the dataset to benchmark on.",
     )
     parser.add_argument("--dataset-path",
@@ -535,7 +609,7 @@ if __name__ == "__main__":
         "--tokenizer",
         type=str,
         help=
-        "Name or path of the tokenizer, if not using the default tokenizer.",
+        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
     parser.add_argument(
         "--best-of",
@@ -578,6 +652,27 @@ if __name__ == "__main__":
         help=
         "Number of prefix tokens per request, used only for sonnet dataset.",
     )
+    parser.add_argument(
+        "--random-input-len",
+        type=int,
+        default=1024,
+        help=
+        "Number of input tokens per request, used only for random sampling.",
+    )
+    parser.add_argument(
+        "--random-output-len",
+        type=int,
+        default=128,
+        help=
+        "Number of output tokens per request, used only for random sampling.",
+    )
+    parser.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=1.0,
+        help="Range of sampled ratio of input/output length, "
+        "used only for random sampling.",
+    )
     parser.add_argument(
         "--request-rate",
         type=float,
@@ -618,6 +713,15 @@ 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(
@@ -78,8 +80,10 @@ def run_vllm(
     enable_prefix_caching: bool,
     enable_chunked_prefill: bool,
     max_num_batched_tokens: int,
+    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(
@@ -100,11 +104,13 @@ def run_vllm(
         download_dir=download_dir,
         enable_chunked_prefill=enable_chunked_prefill,
         max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
+        load_format=load_format,
     )
 
     # Add the requests to the engine.
-    prompts = []
+    prompts: List[str] = []
-    sampling_params = []
+    sampling_params: List[SamplingParams] = []
     for prompt, _, output_len in requests:
         prompts.append(prompt)
         sampling_params.append(
@@ -225,8 +231,8 @@ def main(args: argparse.Namespace):
             args.enforce_eager, args.kv_cache_dtype,
             args.quantization_param_path, args.device,
             args.enable_prefix_caching, args.enable_chunked_prefill,
-            args.max_num_batched_tokens, args.gpu_memory_utilization,
+            args.max_num_batched_tokens, args.distributed_executor_backend,
-            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,
@@ -256,7 +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"],
@@ -343,9 +349,10 @@ if __name__ == "__main__":
     parser.add_argument(
         "--device",
         type=str,
-        default="cuda",
+        default="auto",
-        choices=["cuda", "cpu"],
+        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
-        help='device type for vLLM execution, supporting CUDA and CPU.')
+        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
+        'CPU.')
     parser.add_argument(
         "--enable-prefix-caching",
         action='store_true',
@@ -368,6 +375,36 @@ if __name__ == "__main__":
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        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

@@ -0,0 +1,353 @@
+import argparse
+import copy
+import itertools
+import pickle as pkl
+import time
+from typing import Callable, Iterable, List, Tuple
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())[1:]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+# helpers
+
+
+def to_fp8(tensor: torch.Tensor) -> torch.Tensor:
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def to_int8(tensor: torch.Tensor) -> torch.Tensor:
+    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
+
+
+def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
+                      k: int) -> Tuple[torch.Tensor, torch.Tensor]:
+
+    a = torch.randn((m, k), device='cuda') * 5
+    b = torch.randn((n, k), device='cuda').t() * 5
+
+    if dtype == torch.int8:
+        return to_int8(a), to_int8(b)
+    if dtype == torch.float8_e4m3fn:
+        return to_fp8(a), to_fp8(b)
+
+    raise ValueError("unsupported dtype")
+
+
+# impl
+
+
+def pytorch_mm_impl(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
+                    scale_b: torch.Tensor,
+                    out_dtype: torch.dtype) -> torch.Tensor:
+    return torch.mm(a, b)
+
+
+def pytorch_fp8_impl(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
+                     scale_b: torch.Tensor,
+                     out_dtype: torch.dtype) -> torch.Tensor:
+    return torch._scaled_mm(a,
+                            b,
+                            scale_a=scale_a,
+                            scale_b=scale_b,
+                            out_dtype=out_dtype)
+
+
+def pytorch_fp8_impl_fast_accum(a: torch.Tensor, b: torch.Tensor,
+                                scale_a: torch.Tensor, scale_b: torch.Tensor,
+                                out_dtype: torch.dtype) -> torch.Tensor:
+    return torch._scaled_mm(a,
+                            b,
+                            scale_a=scale_a,
+                            scale_b=scale_b,
+                            out_dtype=out_dtype,
+                            use_fast_accum=True)
+
+
+def cutlass_impl(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
+                 scale_b: torch.Tensor,
+                 out_dtype: torch.dtype) -> torch.Tensor:
+    return ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype=out_dtype)
+
+
+# bench
+def bench_fn(a: torch.Tensor, b: torch.Tensor, scale_a: torch.Tensor,
+             scale_b: torch.Tensor, out_dtype: torch.dtype, label: str,
+             sub_label: str, fn: Callable, description: str) -> TMeasurement:
+
+    min_run_time = 1
+
+    globals = {
+        "a": a,
+        "b": b,
+        "scale_a": scale_a,
+        "scale_b": scale_b,
+        "out_dtype": out_dtype,
+        "fn": fn,
+    }
+    return TBenchmark.Timer(
+        stmt="fn(a, b, scale_a, scale_b, out_dtype)",
+        globals=globals,
+        label=label,
+        sub_label=sub_label,
+        description=description,
+    ).blocked_autorange(min_run_time=min_run_time)
+
+
+def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+               sub_label: str) -> Iterable[TMeasurement]:
+    assert dtype == torch.int8
+    a, b = make_rand_tensors(torch.int8, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+
+    timers = []
+    # pytorch impl
+    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"))
+
+    # cutlass impl
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 cutlass_impl, "cutlass_i8_i8_bf16_scaled_mm"))
+
+    return timers
+
+
+def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+              sub_label: str) -> Iterable[TMeasurement]:
+    assert dtype == torch.float8_e4m3fn
+    a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+
+    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,
+                 pytorch_fp8_impl, "pytorch_fp8_fp8_bf16_scaled_mm"))
+
+    # pytorch impl: bf16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 pytorch_fp8_impl_fast_accum,
+                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum"))
+
+    # pytorch impl: fp16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
+                 pytorch_fp8_impl, "pytorch_fp8_fp8_fp16_scaled_mm"))
+
+    # pytorch impl: fp16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
+                 pytorch_fp8_impl_fast_accum,
+                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum"))
+
+    # cutlass impl: bf16 output
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
+                 cutlass_impl, "cutlass_fp8_fp8_bf16_scaled_mm"))
+    # cutlass impl: fp16 output
+    timers.append(
+        bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
+                 cutlass_impl, "cutlass_fp8_fp8_fp16_scaled_mm"))
+    return timers
+
+
+def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+          sub_label: str) -> Iterable[TMeasurement]:
+    if dtype == torch.int8:
+        return bench_int8(dtype, m, k, n, label, sub_label)
+    if dtype == torch.float8_e4m3fn:
+        return bench_fp8(dtype, m, k, n, label, sub_label)
+    raise ValueError("unsupported type")
+
+
+# runner
+def print_timers(timers: Iterable[TMeasurement]):
+    compare = TBenchmark.Compare(timers)
+    compare.print()
+
+
+def run(dtype: torch.dtype,
+        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+
+    results = []
+    for m, k, n in MKNs:
+        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
+                       f"MKN=({m}x{k}x{n})")
+        print_timers(timers)
+        results.extend(timers)
+
+    return results
+
+
+# output makers
+def make_output(data: Iterable[TMeasurement],
+                MKNs: Iterable[Tuple[int, int, int]],
+                base_description: str,
+                timestamp=None):
+
+    print(f"== All Results {base_description} ====")
+    print_timers(data)
+
+    # pickle all the results
+    timestamp = int(time.time()) if timestamp is None else timestamp
+    with open(f"{base_description}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(data, f)
+
+
+# argparse runners
+
+
+def run_square_bench(args):
+    dim_sizes = list(
+        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
+    data = run(args.dtype, MKNs)
+
+    make_output(data, MKNs, f"square_bench-{args.dtype}")
+
+
+def run_range_bench(args):
+    dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment))
+    n = len(dim_sizes)
+    Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes
+    Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes
+    Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes
+    MKNs = list(zip(Ms, Ks, Ns))
+    data = run(args.dtype, MKNs)
+
+    make_output(data, MKNs, f"range_bench-{args.dtype}")
+
+
+def run_model_bench(args):
+
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+        KNs = []
+        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
+            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
+            KNs.append(KN)
+        return KNs
+
+    model_bench_data = []
+    models_tps = list(itertools.product(args.models, args.tp_sizes))
+    for model, tp_size in models_tps:
+        Ms = args.batch_sizes
+        KNs = model_shapes(model, tp_size)
+        MKNs = []
+        for m in Ms:
+            for k, n in KNs:
+                MKNs.append((m, k, n))
+
+        data = run(args.dtype, MKNs)
+        model_bench_data.append(data)
+
+    # Print all results
+    for data, model_tp in zip(model_bench_data, models_tps):
+        model, tp_size = model_tp
+        print(f"== Results {args.dtype} {model}-TP{tp_size} ====")
+        print_timers(data)
+
+    timestamp = int(time.time())
+
+    all_data = []
+    for d in model_bench_data:
+        all_data.extend(d)
+    # pickle all data
+    with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(all_data, f)
+
+
+if __name__ == '__main__':
+
+    def to_torch_dtype(dt):
+        if dt == "int8":
+            return torch.int8
+        if dt == "fp8":
+            return torch.float8_e4m3fn
+        raise ValueError("unsupported dtype")
+
+    parser = FlexibleArgumentParser(
+        description="""
+Benchmark Cutlass GEMM.
+
+    To run square GEMMs:
+        python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64
+    
+    To run constant N and K and sweep M:
+        python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384
+    
+    To run dimensions from a model:
+        python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1
+    
+    Output:
+        - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
+            """,  # noqa: E501
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--dtype",
+                        type=to_torch_dtype,
+                        required=True,
+                        help="Available options are ['int8', 'fp8']")
+    subparsers = parser.add_subparsers(dest="cmd")
+
+    square_parser = subparsers.add_parser("square_bench")
+    square_parser.add_argument("--dim-start", type=int, required=True)
+    square_parser.add_argument("--dim-end", type=int, required=True)
+    square_parser.add_argument("--dim-increment", type=int, required=True)
+    square_parser.set_defaults(func=run_square_bench)
+
+    range_parser = subparsers.add_parser("range_bench")
+    range_parser.add_argument("--dim-start", type=int, required=True)
+    range_parser.add_argument("--dim-end", type=int, required=True)
+    range_parser.add_argument("--dim-increment", type=int, required=True)
+    range_parser.add_argument("--m-constant", type=int, default=None)
+    range_parser.add_argument("--n-constant", type=int, default=None)
+    range_parser.add_argument("--k-constant", type=int, default=None)
+    range_parser.set_defaults(func=run_range_bench)
+
+    model_parser = subparsers.add_parser("model_bench")
+    model_parser.add_argument("--models",
+                              nargs="+",
+                              type=str,
+                              default=DEFAULT_MODELS,
+                              choices=WEIGHT_SHAPES.keys())
+    model_parser.add_argument("--tp-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_TP_SIZES)
+    model_parser.add_argument("--batch-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_BATCH_SIZES)
+    model_parser.set_defaults(func=run_model_bench)
+
+    args = parser.parse_args()
+    args.func(args)

benchmarks/cutlass_benchmarks/weight_shapes.py

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

benchmarks/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

@@ -1,20 +1,23 @@
-import argparse
+from typing import List
 
 import torch
 import torch.utils.benchmark as benchmark
 from benchmark_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
-from vllm.model_executor.layers.quantization.gptq_marlin import (
-    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
-    GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS)
 from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
     GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
     GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_NUM_BITS)
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
-    MarlinWorkspace, marlin_24_quantize, marlin_quantize)
+    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
+    GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
+    MarlinWorkspace, marlin_quantize)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
+    marlin_24_quantize)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     gptq_pack, quantize_weights, sort_weights)
+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 +26,9 @@ ACT_ORDER_OPTS = [False, True]
 K_FULL_OPTS = [False, True]
 
 
-def bench_run(results, model, act_order, is_k_full, num_bits, group_size,
+def bench_run(results: List[benchmark.Measurement], model: str,
-              size_m, size_k, size_n):
+              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 +160,7 @@ 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 +213,7 @@ 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_mixtral_moe.py

@@ -1,239 +0,0 @@
-import argparse
-import json
-import os
-import sys
-
-import torch
-import torch.nn.functional as F
-import triton
-from tqdm import tqdm
-
-from vllm.model_executor.layers.fused_moe import (fused_moe,
-                                                  get_config_file_name)
-
-
-def main(model, tp_size, gpu, dtype: str):
-    os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu)
-    method = fused_moe
-    for bs in [
-            1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
-            2048, 3072, 4096
-    ]:
-        run_grid(bs,
-                 model=model,
-                 method=method,
-                 gpu=gpu,
-                 tp_size=tp_size,
-                 dtype=dtype)
-
-
-def run_grid(bs, model, method, gpu, tp_size, dtype: str):
-    if model == '8x7B':
-        d_model = 4096
-        model_intermediate_size = 14336
-        num_layers = 32
-    elif model == '8x22B':
-        d_model = 6144
-        model_intermediate_size = 16384
-        num_layers = 56
-    else:
-        raise ValueError(f'Unsupported Mixtral model {model}')
-    num_total_experts = 8
-    top_k = 2
-    # tp_size = 2
-    num_calls = 100
-
-    num_warmup_trials = 1
-    num_trials = 1
-
-    configs = []
-
-    for block_size_n in [32, 64, 128, 256]:
-        for block_size_m in [16, 32, 64, 128, 256]:
-            for block_size_k in [64, 128, 256]:
-                for group_size_m in [1, 16, 32, 64]:
-                    for num_warps in [4, 8]:
-                        for num_stages in [2, 3, 4, 5]:
-                            configs.append({
-                                "BLOCK_SIZE_M": block_size_m,
-                                "BLOCK_SIZE_N": block_size_n,
-                                "BLOCK_SIZE_K": block_size_k,
-                                "GROUP_SIZE_M": group_size_m,
-                                "num_warps": num_warps,
-                                "num_stages": num_stages,
-                            })
-
-    best_config = None
-    best_time_us = 1e20
-
-    print(f'{tp_size=} {bs=}')
-
-    for config in tqdm(configs):
-        # warmup
-        try:
-            for _ in range(num_warmup_trials):
-                run_timing(
-                    num_calls=num_calls,
-                    bs=bs,
-                    d_model=d_model,
-                    num_total_experts=num_total_experts,
-                    top_k=top_k,
-                    tp_size=tp_size,
-                    model_intermediate_size=model_intermediate_size,
-                    method=method,
-                    config=config,
-                    dtype=dtype,
-                )
-        except triton.runtime.autotuner.OutOfResources:
-            continue
-
-        # trial
-        for _ in range(num_trials):
-            kernel_dur_ms = run_timing(
-                num_calls=num_calls,
-                bs=bs,
-                d_model=d_model,
-                num_total_experts=num_total_experts,
-                top_k=top_k,
-                tp_size=tp_size,
-                model_intermediate_size=model_intermediate_size,
-                method=method,
-                config=config,
-                dtype=dtype,
-            )
-
-            kernel_dur_us = 1000 * kernel_dur_ms
-            model_dur_ms = kernel_dur_ms * num_layers
-
-            if kernel_dur_us < best_time_us:
-                best_config = config
-                best_time_us = kernel_dur_us
-
-                tqdm.write(
-                    f'{kernel_dur_us=:.1f} {model_dur_ms=:.1f}'
-                    f' {bs=} {tp_size=} {top_k=} {num_total_experts=} '
-                    f'{d_model=} {model_intermediate_size=} {num_layers=}')
-
-    print("best_time_us", best_time_us)
-    print("best_config", best_config)
-
-    # holds Dict[str, Dict[str, int]]
-    filename = get_config_file_name(num_total_experts,
-                                    model_intermediate_size // tp_size,
-                                    "float8" if dtype == "float8" else None)
-    print(f"writing config to file {filename}")
-    existing_content = {}
-    if os.path.exists(filename):
-        with open(filename, "r") as f:
-            existing_content = json.load(f)
-    existing_content[str(bs)] = best_config
-    with open(filename, "w") as f:
-        json.dump(existing_content, f, indent=4)
-        f.write("\n")
-
-
-def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
-               top_k: int, tp_size: int, model_intermediate_size: int, method,
-               config, dtype: str) -> float:
-    shard_intermediate_size = model_intermediate_size // tp_size
-
-    hidden_states = torch.rand(
-        (bs, d_model),
-        device="cuda:0",
-        dtype=torch.float16,
-    )
-
-    w1 = torch.rand(
-        (num_total_experts, 2 * shard_intermediate_size, d_model),
-        device=hidden_states.device,
-        dtype=hidden_states.dtype,
-    )
-
-    w2 = torch.rand(
-        (num_total_experts, d_model, shard_intermediate_size),
-        device=hidden_states.device,
-        dtype=hidden_states.dtype,
-    )
-
-    w1_scale = None
-    w2_scale = None
-    a1_scale = None
-    a2_scale = None
-
-    if dtype == "float8":
-        w1 = w1.to(torch.float8_e4m3fn)
-        w2 = w2.to(torch.float8_e4m3fn)
-        w1_scale = torch.ones(num_total_experts,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-        w2_scale = torch.ones(num_total_experts,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-        a1_scale = torch.ones(1,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-        a2_scale = torch.ones(1,
-                              device=hidden_states.device,
-                              dtype=torch.float32)
-
-    gating_output = F.softmax(torch.rand(
-        (num_calls, bs, num_total_experts),
-        device=hidden_states.device,
-        dtype=torch.float32,
-    ),
-                              dim=-1)
-
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-
-    start_event.record()
-    for i in range(num_calls):
-        hidden_states = method(
-            hidden_states=hidden_states,
-            w1=w1,
-            w2=w2,
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            gating_output=gating_output[i],
-            topk=2,
-            renormalize=True,
-            inplace=True,
-            override_config=config,
-            use_fp8=dtype == "float8",
-        )
-    end_event.record()
-    end_event.synchronize()
-
-    dur_ms = start_event.elapsed_time(end_event) / num_calls
-    return dur_ms
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        prog='benchmark_mixtral_moe',
-        description='Benchmark and tune the fused_moe kernel',
-    )
-    parser.add_argument(
-        '--dtype',
-        type=str,
-        default='auto',
-        choices=['float8', 'float16'],
-        help='Data type used for fused_moe kernel computations',
-    )
-    parser.add_argument('--model',
-                        type=str,
-                        default='8x7B',
-                        choices=['8x7B', '8x22B'],
-                        help='The Mixtral model to benchmark')
-    parser.add_argument('--tp-size',
-                        type=int,
-                        default=2,
-                        help='Tensor paralleli size')
-    parser.add_argument('--gpu',
-                        type=int,
-                        default=0,
-                        help="GPU ID for benchmarking")
-    args = parser.parse_args()
-    sys.exit(main(args.model, args.tp_size, args.gpu, args.dtype))

benchmarks/kernels/benchmark_moe.py

@@ -0,0 +1,333 @@
+import argparse
+import time
+from datetime import datetime
+from typing import Any, Dict, List, Tuple, TypedDict
+
+import ray
+import torch
+import triton
+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: BenchmarkConfig,
+    num_tokens: int,
+    num_experts: int,
+    shard_intermediate_size: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8: bool,
+    num_iters: int = 100,
+) -> float:
+    init_dtype = torch.float16 if use_fp8 else dtype
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    w1 = torch.randn(num_experts,
+                     shard_intermediate_size,
+                     hidden_size,
+                     dtype=init_dtype)
+    w2 = torch.randn(num_experts,
+                     hidden_size,
+                     shard_intermediate_size // 2,
+                     dtype=init_dtype)
+    gating_output = torch.randn(num_iters,
+                                num_tokens,
+                                num_experts,
+                                dtype=torch.float32)
+
+    w1_scale = None
+    w2_scale = None
+    a1_scale = None
+    a2_scale = None
+    if use_fp8:
+        w1_scale = torch.randn(num_experts, dtype=torch.float32)
+        w2_scale = torch.randn(num_experts, dtype=torch.float32)
+        a1_scale = torch.randn(1, dtype=torch.float32)
+        a2_scale = torch.randn(1, dtype=torch.float32)
+
+        w1 = w1.to(torch.float8_e4m3fn)
+        w2 = w2.to(torch.float8_e4m3fn)
+
+    input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)
+
+    def prepare(i: int):
+        input_gating.copy_(gating_output[i])
+
+    def run():
+        fused_moe(
+            x,
+            w1,
+            w2,
+            input_gating,
+            topk,
+            renormalize=True,
+            inplace=True,
+            override_config=config,
+            use_fp8=use_fp8,
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+            a1_scale=a1_scale,
+            a2_scale=a2_scale,
+        )
+
+    # JIT compilation & warmup
+    run()
+    torch.cuda.synchronize()
+
+    # Capture 10 invocations with CUDA graph
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for _ in range(10):
+            run()
+    torch.cuda.synchronize()
+
+    # Warmup
+    for _ in range(5):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: List[float] = []
+    for i in range(num_iters):
+        prepare(i)
+        torch.cuda.synchronize()
+
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    graph.reset()
+    return avg
+
+
+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: 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]:
+                for block_n in [32, 64, 128, 256]:
+                    for num_warps in [4, 8]:
+                        for group_size in [1, 16, 32, 64]:
+                            configs.append({
+                                "BLOCK_SIZE_M": block_m,
+                                "BLOCK_SIZE_N": block_n,
+                                "BLOCK_SIZE_K": block_k,
+                                "GROUP_SIZE_M": group_size,
+                                "num_warps": num_warps,
+                                "num_stages": num_stages,
+                            })
+    return configs
+
+
+@ray.remote(num_gpus=1)
+class BenchmarkWorker:
+
+    def __init__(self, seed: int) -> None:
+        torch.set_default_device("cuda")
+        torch.cuda.manual_seed_all(seed)
+        self.seed = seed
+
+    def benchmark(
+        self,
+        num_tokens: int,
+        num_experts: int,
+        shard_intermediate_size: int,
+        hidden_size: int,
+        topk: int,
+        dtype: torch.dtype,
+        use_fp8: bool,
+    ) -> Tuple[Dict[str, int], float]:
+        torch.cuda.manual_seed_all(self.seed)
+
+        dtype_str = "float8" if use_fp8 else None
+        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
+        # is the intermediate size after silu_and_mul.
+        op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
+                                    dtype_str)
+        if op_config is None:
+            config = get_default_config(num_tokens, num_experts,
+                                        shard_intermediate_size, hidden_size,
+                                        topk, dtype_str)
+        else:
+            config = op_config[min(op_config.keys(),
+                                   key=lambda x: abs(x - num_tokens))]
+        kernel_time = benchmark_config(config, num_tokens, num_experts,
+                                       shard_intermediate_size, hidden_size,
+                                       topk, dtype, use_fp8)
+        return config, kernel_time
+
+    def tune(
+        self,
+        num_tokens: int,
+        num_experts: int,
+        shard_intermediate_size: int,
+        hidden_size: int,
+        topk: int,
+        dtype: torch.dtype,
+        use_fp8: bool,
+        search_space: List[BenchmarkConfig],
+    ) -> BenchmarkConfig:
+        best_config = None
+        best_time = float("inf")
+        for config in tqdm(search_space):
+            try:
+                kernel_time = benchmark_config(config,
+                                               num_tokens,
+                                               num_experts,
+                                               shard_intermediate_size,
+                                               hidden_size,
+                                               topk,
+                                               dtype,
+                                               use_fp8,
+                                               num_iters=10)
+            except triton.runtime.autotuner.OutOfResources:
+                # Some configurations may be invalid and fail to compile.
+                continue
+
+            if kernel_time < best_time:
+                best_time = kernel_time
+                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: BenchmarkConfig) -> BenchmarkConfig:
+    return {
+        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
+        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
+        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
+        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
+        "num_warps": config["num_warps"],
+        "num_stages": config["num_stages"],
+    }
+
+
+def save_configs(
+    configs: Dict[int, BenchmarkConfig],
+    num_experts: int,
+    shard_intermediate_size: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8: bool,
+) -> None:
+    dtype_str = "float8" if use_fp8 else None
+    # NOTE(woosuk): The current naming convention uses w2.shape[2], which
+    # is the intermediate size after silu_and_mul.
+    filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
+                                    dtype_str)
+    print(f"Writing best config to {filename}...")
+    with open(filename, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def main(args: argparse.Namespace):
+    print(args)
+
+    config = AutoConfig.from_pretrained(args.model)
+    if config.architectures[0] == "DbrxForCausalLM":
+        E = config.ffn_config.moe_num_experts
+        topk = config.ffn_config.moe_top_k
+        intermediate_size = config.ffn_config.ffn_hidden_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    else:
+        # Default: Mixtral.
+        E = config.num_local_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
+
+    hidden_size = config.hidden_size
+    dtype = config.torch_dtype
+    use_fp8 = args.dtype == "fp8"
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
+            2048, 3072, 4096
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+
+    ray.init()
+    num_gpus = int(ray.available_resources()["GPU"])
+    workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
+
+    def _distribute(method: str, inputs: List[Any]) -> List[Any]:
+        outputs = []
+        worker_idx = 0
+        for input_args in inputs:
+            worker = workers[worker_idx]
+            worker_method = getattr(worker, method)
+            output = worker_method.remote(*input_args)
+            outputs.append(output)
+            worker_idx = (worker_idx + 1) % num_gpus
+        return ray.get(outputs)
+
+    if args.tune:
+        search_space = get_configs_compute_bound()
+        print(f"Start tuning over {len(search_space)} configurations...")
+
+        start = time.time()
+        configs = _distribute(
+            "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
+                      topk, dtype, use_fp8, search_space)
+                     for batch_size in batch_sizes])
+        best_configs = {
+            M: sort_config(config)
+            for M, config in zip(batch_sizes, configs)
+        }
+        save_configs(best_configs, E, shard_intermediate_size, hidden_size,
+                     topk, dtype, use_fp8)
+        end = time.time()
+        print(f"Tuning took {end - start:.2f} seconds")
+    else:
+        outputs = _distribute("benchmark",
+                              [(batch_size, E, shard_intermediate_size,
+                                hidden_size, topk, dtype, use_fp8)
+                               for batch_size in batch_sizes])
+
+        for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
+            print(f"Batch size: {batch_size}, config: {config}")
+            print(f"Kernel time: {kernel_time:.2f} us")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument("--model",
+                        type=str,
+                        default="mistralai/Mixtral-8x7B-Instruct-v0.1")
+    parser.add_argument("--tp-size", "-tp", type=int, default=2)
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["auto", "fp8"],
+                        default="auto")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--tune", action="store_true")
+    args = parser.parse_args()
+
+    main(args)

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_lst.append(block_table)
-    block_tables = torch.tensor(block_tables, dtype=torch.int, device=device)
+
+    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,
@@ -97,7 +100,7 @@ def main(
         start_time = time.perf_counter()
 
         # Using default kv_scale
-        kv_scale = 1.0
+        k_scale = v_scale = 1.0
 
         for _ in range(num_iters):
             if version == "v1":
@@ -114,7 +117,8 @@ def main(
                     max_seq_len,
                     alibi_slopes,
                     kv_cache_dtype,
-                    kv_scale,
+                    k_scale,
+                    v_scale,
                 )
             elif version == "v2":
                 ops.paged_attention_v2(
@@ -133,7 +137,8 @@ def main(
                     max_seq_len,
                     alibi_slopes,
                     kv_cache_dtype,
-                    kv_scale,
+                    k_scale,
+                    v_scale,
                 )
             else:
                 raise ValueError(f"Invalid version: {version}")
@@ -158,14 +163,14 @@ 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

@@ -12,7 +12,7 @@ include_directories("${CMAKE_SOURCE_DIR}/csrc")
 #
 # Check the compile flags
 #
-list(APPEND CXX_COMPILE_FLAGS 
+list(APPEND CXX_COMPILE_FLAGS
     "-fopenmp"
     "-DVLLM_CPU_EXTENSION")
 
@@ -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"
@@ -44,8 +58,8 @@ if (AVX512_FOUND)
 
     find_isa(${CPUINFO} "avx512_bf16" AVX512BF16_FOUND)
     if (AVX512BF16_FOUND OR ENABLE_AVX512BF16)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND 
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3) 
+            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
             list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
         else()
             message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
@@ -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}")
@@ -73,7 +97,7 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cache.cpp"
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/cpu/pybind.cpp")
+    "csrc/cpu/torch_bindings.cpp")
 
 define_gpu_extension_target(
     _C
@@ -81,10 +105,10 @@ define_gpu_extension_target(
     LANGUAGE CXX
     SOURCES ${VLLM_EXT_SRC}
     COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
-    WITH_SOABI 
+    USE_SABI 3
+    WITH_SOABI
 )
 
 add_custom_target(default)
 message(STATUS "Enabling C extension.")
 add_dependencies(default _C)
-

cmake/utils.cmake

@@ -5,7 +5,7 @@
 macro (find_python_from_executable EXECUTABLE SUPPORTED_VERSIONS)
   file(REAL_PATH ${EXECUTABLE} EXECUTABLE)
   set(Python_EXECUTABLE ${EXECUTABLE})
-  find_package(Python COMPONENTS Interpreter Development.Module)
+  find_package(Python COMPONENTS Interpreter Development.Module Development.SABIModule)
   if (NOT Python_FOUND)
     message(FATAL_ERROR "Unable to find python matching: ${EXECUTABLE}.")
   endif()
@@ -147,16 +147,23 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
   if (${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()
 
@@ -294,6 +301,7 @@ endmacro()
 # INCLUDE_DIRECTORIES <dirs> - Extra include directories.
 # LIBRARIES <libraries>      - Extra link libraries.
 # WITH_SOABI                 - Generate library with python SOABI suffix name.
+# USE_SABI <version>         - Use python stable api <version>
 #
 # Note: optimization level/debug info is set via cmake build type.
 #
@@ -301,7 +309,7 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   cmake_parse_arguments(PARSE_ARGV 1
     GPU
     "WITH_SOABI"
-    "DESTINATION;LANGUAGE"
+    "DESTINATION;LANGUAGE;USE_SABI"
     "SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
 
   # Add hipify preprocessing step when building with HIP/ROCm.
@@ -315,7 +323,11 @@ function (define_gpu_extension_target GPU_MOD_NAME)
     set(GPU_WITH_SOABI)
   endif()
 
-  Python_add_library(${GPU_MOD_NAME} MODULE "${GPU_SOURCES}" ${GPU_WITH_SOABI})
+  if (GPU_USE_SABI)
+    Python_add_library(${GPU_MOD_NAME} MODULE USE_SABI ${GPU_USE_SABI} ${GPU_WITH_SOABI} "${GPU_SOURCES}")
+  else()
+    Python_add_library(${GPU_MOD_NAME} MODULE ${GPU_WITH_SOABI} "${GPU_SOURCES}")
+  endif()
 
   if (GPU_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.

collect_env.py

@@ -64,6 +64,7 @@ DEFAULT_CONDA_PATTERNS = {
     "triton",
     "optree",
     "nccl",
+    "transformers",
 }
 
 DEFAULT_PIP_PATTERNS = {
@@ -75,6 +76,7 @@ DEFAULT_PIP_PATTERNS = {
     "optree",
     "onnx",
     "nccl",
+    "transformers",
 }
 
 
@@ -601,6 +603,11 @@ Versions of relevant libraries:
 {conda_packages}
 """.strip()
 
+# both the above code and the following code use `strip()` to
+# remove leading/trailing whitespaces, so we need to add a newline
+# in between to separate the two sections
+env_info_fmt += "\n"
+
 env_info_fmt += """
 ROCM Version: {rocm_version}
 Neuron SDK Version: {neuron_sdk_version}

csrc/activation_kernels.cu

@@ -1,5 +1,5 @@
 #include <ATen/cuda/CUDAContext.h>
-#include <torch/extension.h>
+#include <torch/all.h>
 #include <c10/cuda/CUDAGuard.h>
 
 #include <cmath>
@@ -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/attention/attention_kernels.cu

@@ -17,7 +17,7 @@
  * limitations under the License.
  */
 
-#include <torch/extension.h>
+#include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 #include <algorithm>
@@ -105,9 +105,9 @@ __device__ void paged_attention_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float kv_scale, const int tp_rank, const int blocksparse_local_blocks,
+    const float k_scale, const float v_scale, const int tp_rank,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_head_sliding_step) {
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   const int seq_idx = blockIdx.y;
   const int partition_idx = blockIdx.z;
   const int max_num_partitions = gridDim.z;
@@ -285,7 +285,7 @@ __device__ void paged_attention_kernel(
           Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
               k_ptr + offset1 * BLOCK_SIZE * x + offset2);
           k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
-              k_vec_quant, kv_scale);
+              k_vec_quant, k_scale);
         }
       }
 
@@ -415,7 +415,7 @@ __device__ void paged_attention_kernel(
               *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
           // Vector conversion from V_quant_vec to V_vec.
           v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
-                                                                    kv_scale);
+                                                                    v_scale);
         }
         if (block_idx == num_seq_blocks - 1) {
           // NOTE(woosuk): When v_vec contains the tokens that are out of the
@@ -513,15 +513,15 @@ __global__ void paged_attention_v1_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float kv_scale, const int tp_rank, const int blocksparse_local_blocks,
+    const float k_scale, const float v_scale, const int tp_rank,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_head_sliding_step) {
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
                          KV_DTYPE, IS_BLOCK_SPARSE>(
       /* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
       v_cache, num_kv_heads, scale, block_tables, seq_lens,
       max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
-      kv_head_stride, kv_scale, tp_rank, blocksparse_local_blocks,
+      kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
       blocksparse_vert_stride, blocksparse_block_size,
       blocksparse_head_sliding_step);
 }
@@ -549,14 +549,14 @@ __global__ void paged_attention_v2_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float kv_scale, const int tp_rank, const int blocksparse_local_blocks,
+    const float k_scale, const float v_scale, const int tp_rank,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_head_sliding_step) {
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
                          KV_DTYPE, IS_BLOCK_SPARSE, PARTITION_SIZE>(
       exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
       block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
-      kv_block_stride, kv_head_stride, kv_scale, tp_rank,
+      kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
       blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
       blocksparse_head_sliding_step);
 }
@@ -682,7 +682,7 @@ __global__ void paged_attention_v2_reduce_kernel(
           out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, \
           scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq,    \
           alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,      \
-          kv_scale, tp_rank, blocksparse_local_blocks,                      \
+          k_scale, v_scale, tp_rank, blocksparse_local_blocks,              \
           blocksparse_vert_stride, blocksparse_block_size,                  \
           blocksparse_head_sliding_step);
 
@@ -694,8 +694,8 @@ void paged_attention_v1_launcher(
     torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const c10::optional<torch::Tensor>& alibi_slopes, float kv_scale,
+    const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
-    const int tp_rank, const int blocksparse_local_blocks,
+    float v_scale, const int tp_rank, const int blocksparse_local_blocks,
     const int blocksparse_vert_stride, const int blocksparse_block_size,
     const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
@@ -770,7 +770,7 @@ void paged_attention_v1_launcher(
   paged_attention_v1_launcher<T, CACHE_T, BLOCK_SIZE, KV_DTYPE,              \
                               IS_BLOCK_SPARSE>(                              \
       out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
-      seq_lens, max_seq_len, alibi_slopes, kv_scale, tp_rank,                \
+      seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank,        \
       blocksparse_local_blocks, blocksparse_vert_stride,                     \
       blocksparse_block_size, blocksparse_head_sliding_step);
 
@@ -808,16 +808,17 @@ void paged_attention_v1(
     torch::Tensor&
         key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
     torch::Tensor&
-        value_cache,   // [num_blocks, num_heads, head_size, block_size]
+        value_cache,       // [num_blocks, num_heads, head_size, block_size]
-    int num_kv_heads,  // [num_heads]
+    int64_t num_kv_heads,  // [num_heads]
-    float scale,
+    double scale,
     torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& seq_lens,      // [num_seqs]
-    int block_size, int max_seq_len,
+    int64_t block_size, int64_t max_seq_len,
     const c10::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);
 
   DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
@@ -832,7 +833,7 @@ void paged_attention_v1(
           exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr, key_cache_ptr, \
           value_cache_ptr, num_kv_heads, scale, block_tables_ptr,              \
           seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride,    \
-          kv_block_stride, kv_head_stride, kv_scale, tp_rank,                  \
+          kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,          \
           blocksparse_local_blocks, blocksparse_vert_stride,                   \
           blocksparse_block_size, blocksparse_head_sliding_step);              \
   vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS,            \
@@ -849,8 +850,8 @@ void paged_attention_v2_launcher(
     torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const c10::optional<torch::Tensor>& alibi_slopes, float kv_scale,
+    const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
-    const int tp_rank, const int blocksparse_local_blocks,
+    float v_scale, const int tp_rank, const int blocksparse_local_blocks,
     const int blocksparse_vert_stride, const int blocksparse_block_size,
     const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
@@ -931,8 +932,9 @@ void paged_attention_v2_launcher(
                               IS_BLOCK_SPARSE>(                               \
       out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache,      \
       num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
-      kv_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,   \
+      k_scale, v_scale, tp_rank, blocksparse_local_blocks,                    \
-      blocksparse_block_size, blocksparse_head_sliding_step);
+      blocksparse_vert_stride, blocksparse_block_size,                        \
+      blocksparse_head_sliding_step);
 
 #define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
   switch (is_block_sparse) {                                               \
@@ -972,16 +974,17 @@ void paged_attention_v2(
     torch::Tensor&
         key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
     torch::Tensor&
-        value_cache,   // [num_blocks, num_heads, head_size, block_size]
+        value_cache,       // [num_blocks, num_heads, head_size, block_size]
-    int num_kv_heads,  // [num_heads]
+    int64_t num_kv_heads,  // [num_heads]
-    float scale,
+    double scale,
     torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& seq_lens,      // [num_seqs]
-    int block_size, int max_seq_len,
+    int64_t block_size, int64_t max_seq_len,
     const c10::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);
   DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
                              CALL_V2_LAUNCHER_BLOCK_SIZE)
@@ -990,4 +993,4 @@ void paged_attention_v2(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/cache.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include <torch/extension.h>
+#include <torch/all.h>
 
 #include <map>
 #include <vector>
@@ -8,14 +8,18 @@
 void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
                  const torch::Tensor& block_mapping);
 
-void copy_blocks(std::vector<torch::Tensor>& key_caches,
+// Note: the key_caches and value_caches vectors are constant but
-                 std::vector<torch::Tensor>& value_caches,
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
                  const torch::Tensor& block_mapping);
 
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
                        torch::Tensor& key_cache, torch::Tensor& value_cache,
                        torch::Tensor& slot_mapping,
-                       const std::string& kv_cache_dtype, const float kv_scale);
+                       const std::string& kv_cache_dtype, const double k_scale,
+                       const double v_scale);
 
 void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
                              torch::Tensor& key_cache,
@@ -25,4 +29,4 @@ void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
 
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
-                 const float scale, const std::string& kv_cache_dtype);
+                 const double scale, const std::string& kv_cache_dtype);

csrc/cache_kernels.cu

@@ -1,4 +1,4 @@
-#include <torch/extension.h>
+#include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 
@@ -95,8 +95,11 @@ __global__ void copy_blocks_kernel(int64_t* key_cache_ptrs,
 
 }  // namespace vllm
 
-void copy_blocks(std::vector<torch::Tensor>& key_caches,
+// Note: the key_caches and value_caches vectors are constant but
-                 std::vector<torch::Tensor>& value_caches,
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
                  const torch::Tensor& block_mapping) {
   int num_layers = key_caches.size();
   TORCH_CHECK(num_layers == value_caches.size());
@@ -156,8 +159,8 @@ __global__ void reshape_and_cache_kernel(
                                          // block_size]
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int key_stride, const int value_stride, const int num_heads,
-    const int head_size, const int block_size, const int x,
+    const int head_size, const int block_size, const int x, const float k_scale,
-    const float kv_scale) {
+    const float v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   if (slot_idx < 0) {
@@ -193,9 +196,9 @@ __global__ void reshape_and_cache_kernel(
       value_cache[tgt_value_idx] = tgt_value;
     } else {
       key_cache[tgt_key_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, kv_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, k_scale);
       value_cache[tgt_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, kv_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, v_scale);
     }
   }
 }
@@ -245,7 +248,7 @@ __global__ void reshape_and_cache_flash_kernel(
           reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
           reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
           slot_mapping.data_ptr<int64_t>(), key_stride, value_stride, \
-          num_heads, head_size, block_size, x, kv_scale);
+          num_heads, head_size, block_size, x, k_scale, v_scale);
 
 void reshape_and_cache(
     torch::Tensor& key,    // [num_tokens, num_heads, head_size]
@@ -255,7 +258,8 @@ void reshape_and_cache(
     torch::Tensor&
         value_cache,  // [num_blocks, num_heads, head_size, block_size]
     torch::Tensor& slot_mapping,  // [num_tokens]
-    const std::string& kv_cache_dtype, const float kv_scale) {
+    const std::string& kv_cache_dtype, const double k_scale,
+    const double v_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
@@ -315,13 +319,13 @@ namespace vllm {
 template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
 __global__ void convert_fp8_kernel(const Tin* __restrict__ src_cache,
                                    Tout* __restrict__ dst_cache,
-                                   const float kv_scale,
+                                   const float scale,
                                    const int64_t block_stride) {
   const int64_t block_idx = blockIdx.x;
   for (int i = threadIdx.x; i < block_stride; i += blockDim.x) {
     int64_t idx = block_idx * block_stride + i;
     dst_cache[idx] =
-        fp8::scaled_convert<Tout, Tin, kv_dt>(src_cache[idx], kv_scale);
+        fp8::scaled_convert<Tout, Tin, kv_dt>(src_cache[idx], scale);
   }
 }
 
@@ -330,11 +334,11 @@ __global__ void convert_fp8_kernel(const Tin* __restrict__ src_cache,
 #define CALL_CONVERT_FP8(Tout, Tin, KV_DTYPE)                                \
   vllm::convert_fp8_kernel<Tout, Tin, KV_DTYPE><<<grid, block, 0, stream>>>( \
       reinterpret_cast<Tin*>(src_cache.data_ptr()),                          \
-      reinterpret_cast<Tout*>(dst_cache.data_ptr()), kv_scale, block_stride);
+      reinterpret_cast<Tout*>(dst_cache.data_ptr()), scale, block_stride);
 
 // Only for testing.
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
-                 const float kv_scale, const std::string& kv_cache_dtype) {
+                 const double scale, const std::string& kv_cache_dtype) {
   torch::Device src_device = src_cache.device();
   torch::Device dst_device = dst_cache.device();
   TORCH_CHECK(src_device.is_cuda(), "src must be on a GPU")

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/attention.cpp

@@ -420,13 +420,14 @@ void paged_attention_v1_impl_launcher(
 
 void paged_attention_v1(
     torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
-    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
-    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
-    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
+    int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
-  TORCH_CHECK(kv_scale == 1.0f);
+    const int64_t blocksparse_head_sliding_step) {
+  TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f);
   TORCH_CHECK(blocksparse_vert_stride <= 1,
               "CPU backend does not support blocksparse attention yet.");
   VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v1_impl",
@@ -738,13 +739,14 @@ void paged_attention_v2_impl_launcher(
 void paged_attention_v2(
     torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
     torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
-    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
-    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
-    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
+    int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
-  TORCH_CHECK(kv_scale == 1.0f);
+    const int64_t blocksparse_head_sliding_step) {
+  TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f);
   TORCH_CHECK(blocksparse_vert_stride <= 1,
               "CPU backend does not support blocksparse attention yet.");
   VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v2_impl",

csrc/cpu/cache.cpp

@@ -5,8 +5,8 @@
 
 namespace {
 template <typename scalar_t>
-void copy_blocks_cpu_impl(std::vector<torch::Tensor>& key_caches,
+void copy_blocks_cpu_impl(std::vector<torch::Tensor> const& key_caches,
-                          std::vector<torch::Tensor>& value_caches,
+                          std::vector<torch::Tensor> const& value_caches,
                           const torch::Tensor& mapping_pairs,
                           const int element_num_per_block,
                           const int layer_num) {
@@ -82,8 +82,11 @@ void reshape_and_cache_cpu_impl(
 }
 };  // namespace
 
-void copy_blocks(std::vector<torch::Tensor>& key_caches,
+// Note: the key_caches and value_caches vectors are constant but
-                 std::vector<torch::Tensor>& value_caches,
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
                  const torch::Tensor& block_mapping) {
   unsigned num_layers = key_caches.size();
   TORCH_CHECK(num_layers == value_caches.size());
@@ -104,8 +107,9 @@ void copy_blocks(std::vector<torch::Tensor>& key_caches,
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
                        torch::Tensor& key_cache, torch::Tensor& value_cache,
                        torch::Tensor& slot_mapping,
-                       const std::string& kv_cache_dtype, float kv_scale) {
+                       const std::string& kv_cache_dtype, double k_scale,
-  TORCH_CHECK(kv_scale == 1.0f);
+                       double v_scale) {
+  TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f);
 
   int num_tokens = key.size(0);
   int num_heads = key.size(1);

csrc/cpu/cpu_types.hpp

@@ -2,351 +2,14 @@
 #ifndef CPU_TYPES_HPP
 #define CPU_TYPES_HPP
 
-#include <immintrin.h>
+#if defined(__x86_64__)
-#include <torch/extension.h>
+  //x86 implementation
-
+  #include "cpu_types_x86.hpp"
-namespace vec_op {
+#elif defined(__POWER9_VECTOR__)
-
+  //ppc implementation
-// FIXME: FP16 is not fully supported in Torch-CPU
+  #include "cpu_types_vsx.hpp"
-#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)                                              \
+  #warning "unsupported vLLM cpu implementation"
-  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; }
-};
-
-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