Add benchmarking scripts and utilities for vLLM performance evaluation

benchmarks/auto_tune.sh

+#!/bin/bash
+
+# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
+# The current server parameter combination is  max_num_seqs and max_num_batched_tokens
+# It also supports additional requirement: e2e latency and prefix cache. 
+
+# Pre-requisite:
+# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. 
+# 2. If the model is customized, replace the MODEL's config with the customized config.
+# 3. Set variables (ALL REQUIRED)
+#   BASE: your directory for vllm repo
+#   MODEL: the model served by vllm
+#   SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support.
+#   TP: ways of tensor parallelism
+#   DOWNLOAD_DIR: directory to download and load model weights.
+#   INPUT_LEN: request input len
+#   OUTPUT_LEN: request output len
+#   MIN_CACHE_HIT_PCT: prefix cache rate
+#   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
+#   NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with.
+#   NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with.
+#   Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST.
+# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
+# 5. The final result will be saved in RESULT file. 
+
+
+# Example use cases 
+# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000
+# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500
+# 3. If we want to reach 60% prefix cache, what's the best server parameter? 
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500
+
+TAG=$(date +"%Y_%m_%d_%H_%M")
+BASE=""
+MODEL="meta-llama/Llama-3.1-8B-Instruct"
+SYSTEM="TPU"
+TP=1
+DOWNLOAD_DIR=""
+INPUT_LEN=4000
+OUTPUT_LEN=16
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=100000000000
+NUM_SEQS_LIST="128 256"
+NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
+
+LOG_FOLDER="$BASE/auto-benchmark/$TAG"
+RESULT="$LOG_FOLDER/result.txt"
+PROFILE_PATH="$LOG_FOLDER/profile"
+
+echo "result file: $RESULT"
+echo "model: $MODEL"
+
+rm -rf $LOG_FOLDER
+rm -rf $PROFILE_PATH
+mkdir -p $LOG_FOLDER
+mkdir -p $PROFILE_PATH
+
+cd "$BASE/vllm"
+
+pip install -q datasets
+
+current_hash=$(git rev-parse HEAD)
+echo "hash:$current_hash" >> "$RESULT"
+echo "current_hash: $current_hash"
+
+best_throughput=0
+best_max_num_seqs=0
+best_num_batched_tokens=0
+best_goodput=0
+
+start_server() {
+    local gpu_memory_utilization=$1
+    local max_num_seqs=$2
+    local max_num_batched_tokens=$3
+    local vllm_log=$4
+    local profile_dir=$5
+    
+    pkill -f vllm
+
+    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
+        --disable-log-requests \
+        --port 8004 \
+        --gpu-memory-utilization $gpu_memory_utilization \
+        --max-num-seqs $max_num_seqs \
+        --max-num-batched-tokens $max_num_batched_tokens \
+        --tensor-parallel-size $TP \
+        --enable-prefix-caching \
+        --load-format dummy \
+        --download-dir "$DOWNLOAD_DIR" \
+        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
+
+    # wait for 10 minutes...
+    server_started=0
+    for i in {1..60}; do  
+        RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
+        STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) 
+        if [[ "$STATUS_CODE" -eq 200 ]]; then
+            server_started=1
+            break
+        else
+            sleep 10
+        fi
+    done
+    if (( ! server_started )); then
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
+        return 1
+    else
+        return 0
+    fi
+}
+
+update_best_profile() {
+    local profile_dir=$1
+    local profile_index=$2
+    sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort))
+    selected_profile_file=
+    if [[ "$SYSTEM" == "TPU" ]]; then
+        selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
+    fi 
+    if [[ "$SYSTEM" == "GPU" ]]; then
+        selected_profile_file="${sorted_paths[$profile_index]}"
+    fi 
+    rm -f $PROFILE_PATH/*
+    cp $selected_profile_file $PROFILE_PATH
+}
+
+run_benchmark() {
+    local max_num_seqs=$1
+    local max_num_batched_tokens=$2
+    local gpu_memory_utilization=$3
+    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
+    local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}"
+    echo "vllm_log: $vllm_log"
+    echo
+    rm -f $vllm_log
+    mkdir -p $profile_dir
+    pkill -f vllm
+    local profile_index=0
+
+    echo "starting server..."
+    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir
+    result=$?
+    if [[ "$result" -eq 1 ]]; then
+        echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    else
+        echo "server started."
+    fi
+    echo
+    
+    echo "run benchmark test..."
+    meet_latency_requirement=0
+    # get a basic qps by using request-rate inf
+    bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
+    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
+    python benchmarks/benchmark_serving.py \
+        --backend vllm \
+        --model $MODEL  \
+        --dataset-name random \
+        --random-input-len $INPUT_LEN \
+        --random-output-len $OUTPUT_LEN \
+        --ignore-eos \
+        --disable-tqdm \
+        --request-rate inf \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --num-prompts 1000 \
+        --random-prefix-len $prefix_len \
+        --port 8004 \
+        --profile &> "$bm_log"
+    throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+    e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+    goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+
+    if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+        meet_latency_requirement=1
+        request_rate=inf
+    fi
+
+    if (( ! meet_latency_requirement )); then
+    # start from request-rate as int(throughput) + 1
+        request_rate=$((${throughput%.*} + 1))
+        while ((request_rate > 0)); do
+            profile_index=$((profile_index+1))
+            # clear prefix cache
+            curl -X POST http://0.0.0.0:8004/reset_prefix_cache
+            sleep 5
+            bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
+            python benchmarks/benchmark_serving.py \
+                --backend vllm \
+                --model $MODEL  \
+                --dataset-name random \
+                --random-input-len $INPUT_LEN \
+                --random-output-len $OUTPUT_LEN \
+                --ignore-eos \
+                --disable-tqdm \
+                --request-rate $request_rate \
+                --percentile-metrics ttft,tpot,itl,e2el \
+                --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+                --num-prompts 100 \
+                --random-prefix-len $prefix_len \
+                --port 8004 &> "$bm_log"
+            throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+            goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+                meet_latency_requirement=1
+                break
+            fi
+            request_rate=$((request_rate-1))
+        done
+    fi
+    # write the results and update the best result.
+    if ((meet_latency_requirement)); then
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT"
+        if (( $(echo "$throughput > $best_throughput" | bc -l) )); then
+            best_throughput=$throughput
+            best_max_num_seqs=$max_num_seqs
+            best_num_batched_tokens=$max_num_batched_tokens
+            best_goodput=$goodput
+            if [[ "$SYSTEM" == "TPU" ]]; then
+                update_best_profile "$profile_dir/plugins/profile" $profile_index
+            fi
+            if [[ "$SYSTEM" == "GPU" ]]; then
+                update_best_profile "$profile_dir" $profile_index
+            fi
+        fi
+    else
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT"
+    fi
+
+    echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+
+    pkill vllm
+    sleep 10
+    printf '=%.0s' $(seq 1 20)
+    return 0
+}
+
+read -r -a num_seqs_list <<< "$NUM_SEQS_LIST"
+read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
+
+# first find out the max gpu-memory-utilization without HBM OOM.
+gpu_memory_utilization=0.98
+find_gpu_memory_utilization=0
+while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
+    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
+    result=$?
+    if [[ "$result" -eq 0 ]]; then
+        find_gpu_memory_utilization=1
+        break
+    else
+        gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc)
+    fi
+done
+
+if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then
+    echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model."
+else
+    echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER."
+    exit 1
+fi
+
+for num_seqs in "${num_seqs_list[@]}"; do
+    for num_batched_tokens in "${num_batched_tokens_list[@]}"; do
+        run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization
+    done
+done
+echo "finish permutations"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"
+

benchmarks/benchmark_guided.py

+# SPDX-License-Identifier: Apache-2.0
+"""Benchmark guided decoding throughput."""
+import argparse
+import dataclasses
+import json
+import os
+import random
+import time
+from typing import List
+
+import datasets
+import pandas as pd
+import uvloop
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
+
+from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
+from vllm.entrypoints.openai.api_server import (
+    build_async_engine_client_from_engine_args)
+from vllm.sampling_params import GuidedDecodingParams
+from vllm.utils import FlexibleArgumentParser, merge_async_iterators
+
+
+@dataclasses.dataclass
+class SampleRequest:
+    """A class representing a single inference request for benchmarking.
+
+    Attributes:
+        prompt: The input text prompt for the model.
+        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
+            images).
+        prompt_len: The length of the prompt in tokens.
+        expected_output_len: The expected length of the output in tokens.
+    """
+    prompt: str
+    prompt_len: int
+    expected_output_len: int
+    schema: dict
+    structure_type: str = 'json'
+    completion: str = None
+
+
+def run_vllm(requests: List[SampleRequest],
+             engine_args: EngineArgs,
+             n: int,
+             guided_decoding_rate: float = 1.0,
+             warmup: bool = False) -> float:
+    from vllm import LLM, SamplingParams
+    llm = LLM(**vars(engine_args))
+
+    # Add the requests to the engine.
+    prompts: List[str] = []
+    sampling_params: List[SamplingParams] = []
+    # create a list containing random selected true or false
+    guided_decoding_req_idx = random.sample(
+        range(len(requests)), int(len(requests) * guided_decoding_rate))
+
+    if warmup:
+        print(">>>>> Running warmup prompt, for the first 5")
+        # We setup the first 5 requests to warmup FSM
+        # if using xgrammar dataset, we will skip warmup
+        warmup_requests = requests[:5]
+        for i, request in enumerate(warmup_requests):
+            prompts.append(request.prompt)
+            sampling_params.append(
+                SamplingParams(
+                    n=n,
+                    temperature=1.0,
+                    top_p=1.0,
+                    ignore_eos=True,
+                    max_tokens=request.expected_output_len,
+                    guided_decoding=GuidedDecodingParams(json=request.schema)
+                    if guided_decoding_rate > 0 else None,
+                ))
+        llm.generate(prompts, sampling_params, use_tqdm=False)
+
+    print(">>>>> Benchmark started...")
+    prompts = []
+    sampling_params = []
+    for i, request in enumerate(requests):
+        prompts.append(request.prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                guided_decoding=GuidedDecodingParams(
+                    **{request.structure_type: request.schema})
+                if i in guided_decoding_req_idx else None,
+            ))
+
+    start = time.perf_counter()
+    outputs = llm.generate(prompts, sampling_params, use_tqdm=False)
+    ret = []
+    for output, request in zip(outputs, requests):
+        generated_text = output.outputs[0].text
+        ret.append({
+            "generated": generated_text,
+            "expected": request.completion
+        })
+    end = time.perf_counter()
+    return end - start, ret
+
+
+async def run_vllm_async(
+        requests: List[SampleRequest],
+        engine_args: AsyncEngineArgs,
+        n: int,
+        guided_decoding_rate: float = 1.0,
+        warmup: bool = False,
+        disable_frontend_multiprocessing: bool = False) -> float:
+    from vllm import SamplingParams
+
+    async with build_async_engine_client_from_engine_args(
+            engine_args, disable_frontend_multiprocessing) as llm:
+
+        # Add the requests to the engine.
+        prompts: List[str] = []
+        sampling_params: List[SamplingParams] = []
+        guided_decoding_req_idx = random.sample(
+            range(len(requests)), int(len(requests) * guided_decoding_rate))
+
+        if warmup:
+            print(">>>>>> Running warmup prompt, for the first 5")
+            # We setup the first 5 requests to warmup FSM
+            # if using xgrammar dataset, we will skip warmup
+            warmup_requests = requests[:5]
+            for i, request in enumerate(warmup_requests):
+                prompts.append(request.prompt)
+                sampling_params.append(
+                    SamplingParams(
+                        n=n,
+                        temperature=1.0,
+                        top_p=1.0,
+                        ignore_eos=True,
+                        max_tokens=request.expected_output_len,
+                        guided_decoding=GuidedDecodingParams(
+                            json=request.schema)
+                        if guided_decoding_rate > 0 else None,
+                    ))
+            generators = []
+            for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
+                generator = llm.generate(prompt, sp, request_id=f"test{i}")
+                generators.append(generator)
+            all_gens = merge_async_iterators(*generators)
+            async for i, res in all_gens:
+                pass
+
+        print(">>>>> Benchmark started...")
+        prompts = []
+        sampling_params = []
+        for i, request in enumerate(requests):
+            prompts.append(request.prompt)
+            sampling_params.append(
+                SamplingParams(
+                    n=n,
+                    temperature=1.0,
+                    top_p=1.0,
+                    ignore_eos=True,
+                    max_tokens=request.expected_output_len,
+                    guided_decoding=GuidedDecodingParams(json=request.schema)
+                    if i in guided_decoding_req_idx else None,
+                ))
+
+        generators = []
+        start_time = []
+        latencies = []
+        start = time.perf_counter()
+        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
+            generator = llm.generate(prompt, sp, request_id=f"test{i}")
+            generators.append(generator)
+            start_time.append(time.perf_counter())
+            latencies.append([])
+        all_gens = merge_async_iterators(*generators)
+        generated_texts = [''] * len(requests)
+        async for i, res in all_gens:
+            generated_texts[i] = res.outputs[0].text
+            lat = time.perf_counter() - start_time[i]
+            latencies[i].append(lat)
+        ret = [{
+            'generated': gt,
+            'expected': req.completion
+        } for gt, req in zip(generated_texts, requests)]
+        end = time.perf_counter()
+        first_latency = pd.Series([lat[0] * 1000 for lat in latencies])
+        next_latency = pd.Series([(lat[-1] - lat[0]) / len(lat[1:]) * 1000
+                                  for lat in latencies])
+        return end - start, ret, (first_latency, next_latency)
+
+
+def sample_requests(tokenizer: PreTrainedTokenizerBase,
+                    args: argparse.Namespace) -> List[SampleRequest]:
+    if args.dataset == 'json':
+        if args.json_schema_path is None:
+            dir_path = os.path.dirname(os.path.realpath(__file__))
+            args.json_schema_path = os.path.join(dir_path,
+                                                 "structured_schemas",
+                                                 "structured_schema_1.json")
+        with open(args.json_schema_path) as f:
+            schema = json.load(f)
+        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
+        input_len = len(tokenizer(prompt).input_ids)
+        print(f"Input length of the prompt: {input_len} tokens")
+        requests = [
+            SampleRequest(prompt=prompt,
+                          prompt_len=input_len,
+                          expected_output_len=args.output_len,
+                          schema=schema,
+                          structure_type=args.structure_type)
+            for _ in range(args.num_prompts)
+        ]
+
+    elif args.dataset == "grammar":
+        schema = """
+            ?start: select_statement
+
+            ?select_statement: "SELECT " column_list " FROM " table_name
+
+            ?column_list: column_name ("," column_name)*
+
+            ?table_name: identifier
+
+            ?column_name: identifier
+
+            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
+        """
+        prompt = "Generate an SQL query to show the 'username' \
+            and 'email' from the 'users' table."
+
+        input_len = len(tokenizer(prompt).input_ids)
+        print(f"Input length of the prompt: {input_len} tokens")
+        requests = [
+            SampleRequest(prompt=prompt,
+                          prompt_len=input_len,
+                          expected_output_len=args.output_len,
+                          schema=schema,
+                          structure_type=args.structure_type)
+            for _ in range(args.num_prompts)
+        ]
+
+    elif args.dataset == "regex":
+        regex = r"\w+@\w+\.com\n"
+        args.regex = regex
+        prompt = "Generate an email address for Alan Turing, \
+            who works in Enigma. End in .com and new line. \
+                Example result: alan.turing@enigma.com\n"
+
+        input_len = len(tokenizer(prompt).input_ids)
+        print(f"Input length of the prompt: {input_len} tokens")
+        requests = [
+            SampleRequest(prompt=prompt,
+                          prompt_len=input_len,
+                          expected_output_len=args.output_len,
+                          schema=regex,
+                          structure_type=args.structure_type)
+            for _ in range(args.num_prompts)
+        ]
+
+    elif args.dataset == "choice":
+        choice = ["Positive", "Negative"]
+        args.choice = choice
+        prompt = "Classify this sentiment: vLLM is wonderful!"
+        input_len = len(tokenizer(prompt).input_ids)
+        print(f"Input length of the prompt: {input_len} tokens")
+        requests = [
+            SampleRequest(prompt=prompt,
+                          prompt_len=input_len,
+                          expected_output_len=args.output_len,
+                          schema=choice,
+                          structure_type=args.structure_type)
+            for _ in range(args.num_prompts)
+        ]
+
+    elif args.dataset == "xgrammar_bench":
+        args.warmup = False
+        requests: List[SampleRequest] = []
+        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
+                                        split="train")
+        print(f"dataset has {len(dataset)} entries")
+        len_dataset = len(dataset)
+        for data_point_idx in range(args.num_prompts):
+            idx = data_point_idx
+            while idx >= len_dataset:
+                idx -= len_dataset
+            schema = dataset["schema"][idx]
+            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
+                                                   tokenize=False)
+            input_len = len(tokenizer(prompt).input_ids)
+            completion = dataset["completion"][idx]
+
+            requests.append(
+                SampleRequest(prompt=prompt,
+                              prompt_len=input_len,
+                              expected_output_len=args.output_len,
+                              schema=schema,
+                              completion=completion))
+
+    return requests
+
+
+def evaluate(ret, args):
+
+    def _eval_correctness_json(expected, actual):
+        # extract json string from string using regex
+        import re
+        actual = actual.replace('\n', '').replace(' ', '').strip()
+        try:
+            actual = re.search(r'\{.*\}', actual).group()
+            actual = json.loads(actual)
+        except Exception:
+            return False
+
+        return True
+
+    def _eval_correctness_choice(expected, actual):
+        return actual in args.choice
+
+    def _eval_correctness_regex(expected, actual):
+        import re
+        return re.match(args.regex, actual) is not None
+
+    def _eval_correctness(expected, actual):
+        if args.structure_type == 'json':
+            return _eval_correctness_json(expected, actual)
+        elif args.structure_type == 'regex':
+            return _eval_correctness_regex(expected, actual)
+        elif args.structure_type == 'choice':
+            return _eval_correctness_choice(expected, actual)
+        else:
+            return None
+
+    scores = []
+    for res in ret:
+        score = _eval_correctness(res['expected'], res['generated'])
+        res['correctness'] = score
+        scores.append(score)
+
+    not_none_scores = [score for score in scores if score is not None]
+
+    return (sum(not_none_scores) / len(not_none_scores) *
+            100) if len(not_none_scores) > 0 else None
+
+
+def main(args: argparse.Namespace):
+    print(args)
+    random.seed(args.seed)
+
+    # async engine is working for 'regex', 'choice' and 'grammar'
+    if args.dataset == 'grammar':
+        args.structure_type = 'grammar'
+        args.async_engine = False
+    elif args.dataset == 'regex':
+        args.structure_type = 'regex'
+        args.async_engine = False
+    elif args.dataset == 'choice':
+        args.structure_type = 'choice'
+        args.async_engine = False
+    else:
+        args.structure_type = 'json'
+
+    if args.no_guided_decoding:
+        args.guided_decoding_ratio = 0
+    if args.save_results:
+        result_file_name = f'{args.guided_decoding_ratio}guided'
+        result_file_name += f"_{args.model.split('/')[-1]}"
+        result_file_name += f"_{args.dataset}"
+        result_file_name += f"_{args.num_prompts}"
+        result_file_name += f"_out{args.output_len}"
+        result_file_name += f"_async{args.async_engine}"
+        result_file_name += f"_warmup{args.warmup}"
+        result_file_name += f"_chunkedprefill{args.enable_chunked_prefill}"
+        result_file_name += ".txt"
+    else:
+        result_file_name = None
+
+    # Synthesize a prompt with the given input length.
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.tokenizer, trust_remote_code=args.trust_remote_code)
+    requests = sample_requests(tokenizer, args)
+
+    if args.async_engine:
+        engine_args = AsyncEngineArgs.from_cli_args(args)
+        elapsed_time, ret, (first_latency, next_latency) = uvloop.run(
+            run_vllm_async(requests, engine_args, args.n,
+                           args.guided_decoding_ratio, args.warmup,
+                           args.disable_frontend_multiprocessing))
+    else:
+        engine_args = EngineArgs.from_cli_args(args)
+        elapsed_time, ret = run_vllm(requests, engine_args, args.n,
+                                     args.guided_decoding_ratio, args.warmup)
+        first_latency, next_latency = None, None
+
+    score = evaluate(ret, args)
+    total_num_tokens = sum(request.prompt_len + request.expected_output_len
+                           for request in requests)
+    total_output_tokens = sum(request.expected_output_len
+                              for request in requests)
+    if first_latency is not None:
+        latency_breakdown = "\nFirst token latency(msecs):\n"
+        latency_breakdown += f"{first_latency.describe()}"
+        latency_breakdown += "\nNext token latency(msecs):\n"
+        latency_breakdown += f"{next_latency.describe()}"
+    print(
+        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
+        f"{total_output_tokens / elapsed_time:.2f} output tokens/s",
+        f"Correct rate is {score} %",
+        f"{latency_breakdown if first_latency is not None else ''}")
+
+    # Output JSON results if specified
+    if args.output_json or result_file_name:
+        results = {
+            "elapsed_time": elapsed_time,
+            "num_requests": len(requests),
+            "total_num_tokens": total_num_tokens,
+            "total_output_tokens": total_output_tokens,
+            "requests_per_second": len(requests) / elapsed_time,
+            "tokens_per_second": f"{total_num_tokens / elapsed_time:.2f}",
+            "output_tokens_per_second":
+            f"{total_output_tokens / elapsed_time:.2f}",
+            "correct_rate(%)": score
+        }
+        results = {"outputs": ret, **results}
+        if first_latency is not None:
+            results["first_token_latency(msecs)"] = first_latency.describe(
+            ).to_dict()
+            results["next_token_latency(msecs)"] = next_latency.describe(
+            ).to_dict()
+        if args.output_json:
+            with open(args.output_json, "w") as f:
+                json.dump(results, f, indent=4)
+        elif result_file_name:
+            with open(result_file_name, "w") as f:
+                json.dump(results, f, indent=4)
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark guided decoding.")
+    parser = AsyncEngineArgs.add_cli_args(parser)
+
+    parser.add_argument("--output-len",
+                        type=int,
+                        default=512,
+                        help="Output length for each request. Overrides the "
+                        "output length from the dataset.")
+    parser.add_argument(
+        "--dataset",
+        default='json',
+        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
+    parser.add_argument("--json_schema_path",
+                        type=str,
+                        default=None,
+                        help="Path to json schema.")
+    parser.add_argument("--n",
+                        type=int,
+                        default=1,
+                        help="Number of generated sequences per prompt.")
+    parser.add_argument("--num-prompts",
+                        type=int,
+                        default=10,
+                        help="Number of prompts to process.")
+    parser.add_argument(
+        '--output-json',
+        type=str,
+        default=None,
+        help='Path to save the throughput results in JSON format.')
+    parser.add_argument("--async-engine",
+                        action='store_true',
+                        default=False,
+                        help="Use vLLM async engine rather than LLM class.")
+    parser.add_argument("--no-guided-decoding",
+                        action='store_true',
+                        default=False,
+                        help="Whether to disable JSON decoding or not.")
+    parser.add_argument("--guided-decoding-ratio",
+                        type=float,
+                        default=1.0,
+                        help="Ratio of Guided Decoding requests")
+    parser.add_argument("--disable-frontend-multiprocessing",
+                        action='store_true',
+                        default=False,
+                        help="Disable decoupled async engine frontend.")
+    parser.add_argument("--warmup",
+                        action="store_true",
+                        default=False,
+                        help="Run warmup prompts before benchmark.")
+    parser.add_argument("--save-results",
+                        action="store_true",
+                        default=False,
+                        help="save output results.")
+    args = parser.parse_args()
+    if args.tokenizer is None:
+        args.tokenizer = args.model
+    main(args)
\ No newline at end of file

benchmarks/benchmark_long_document_qa_throughput.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Offline benchmark to test the long document QA throughput.
+
+Example usage:
+    # This workload samples 8 different prompts with a default input
+    # length of 20000 tokens, then replicates each prompt 2 times 
+    # in random order.
+    python benchmark_long_document_qa_throughput.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --enable-prefix-caching \
+        --num-documents 8 \
+        --repeat-count 2 
+
+Commandline arguments:
+    --num-documents: The number of documents to sample prompts from.
+
+    --document-length: The length of each document in tokens. 
+                       (Optional, default: 20000)
+
+    --output-len: The number of tokens to generate for each prompt.
+                  (Optional, default: 10)
+
+    --repeat-count: The number of times to repeat each prompt.
+                    (Optional, default: 2)
+
+    --repeat-mode: The mode to repeat prompts. The supported modes are:
+        - 'random': shuffle the prompts randomly. (Default)
+        - 'tile': the entire prompt list is repeated in sequence. (Potentially
+                  lowest cache hit)
+        - 'interleave': each prompt is repeated consecutively before 
+                        moving to the next element. (Highest cache hit)
+    
+    --shuffle-seed: Random seed when the repeat mode is "random".
+                    (Optional, default: 0)
+
+In the meantime, it also supports all the vLLM engine args to initialize the 
+LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more
+details.
+"""
+
+import dataclasses
+import random
+import time
+
+from vllm import LLM, SamplingParams
+from vllm.engine.arg_utils import EngineArgs
+from vllm.utils import FlexibleArgumentParser
+
+
+def test_long_document_qa(llm=None, sampling_params=None, prompts=None):
+    """
+    Test long document QA with the given prompts and sampling parameters.
+    Print the time spent in processing all the prompts.
+
+    Args:
+        llm: The language model used for generating responses.
+        sampling_params: Sampling parameter used to generate the response.
+        prompts: A list of prompt strings to be processed by the LLM.
+    """
+    start_time = time.time()
+    llm.generate(prompts, sampling_params=sampling_params)
+    end_time = time.time()
+    print(f"Time to execute all requests: {end_time - start_time:.4f} secs")
+
+
+def repeat_prompts(prompts, repeat_count, mode: str):
+    """
+    Repeat each prompt in the list for a specified number of times.
+    The order of prompts in the output list depends on the mode.
+
+    Args:
+        prompts: A list of prompts to be repeated.
+        repeat_count: The number of times each prompt is repeated.
+        mode: The mode of repetition. Supported modes are:
+            - 'random': Shuffle the prompts randomly after repetition.
+            - 'tile': Repeat the entire prompt list in sequence.
+              Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3].
+            - 'interleave': Repeat each prompt consecutively before moving to
+              the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3].
+
+    Returns:
+        A list of repeated prompts in the specified order.
+
+    Raises:
+        ValueError: If an invalid mode is provided.
+    """
+    print("Repeat mode: ", mode)
+    if mode == "random":
+        repeated_prompts = prompts * repeat_count
+        random.shuffle(repeated_prompts)
+        return repeated_prompts
+    elif mode == "tile":
+        return prompts * repeat_count
+    elif mode == "interleave":
+        repeated_prompts = []
+        for prompt in prompts:
+            repeated_prompts.extend([prompt] * repeat_count)
+        return repeated_prompts
+    else:
+        raise ValueError(
+            f"Invalid mode: {mode}, only support 'random', 'tile', 'interleave'"
+        )
+
+
+def main(args):
+    random.seed(args.shuffle_seed)
+
+    # Prepare the prompts:
+    # we append the document id at the beginning to avoid any of the document
+    # being the prefix of other documents
+    prompts = [
+        str(i) + " ".join(["hi"] * args.document_length)
+        for i in range(args.num_documents)
+    ]
+
+    prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode)
+
+    warmup_prompts = [
+        "This is warm up request " + str(i) + " ".join(["hi"] * args.document_length)
+        for i in range(args.num_documents)
+    ]
+
+    # Create the LLM engine
+    engine_args = EngineArgs.from_cli_args(args)
+    llm = LLM(**dataclasses.asdict(engine_args))
+    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+
+    print("------warm up------")
+    test_long_document_qa(
+        llm=llm,
+        prompts=warmup_prompts,
+        sampling_params=sampling_params,
+    )
+
+    print("------start generating------")
+    test_long_document_qa(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+
+def create_argument_parser():
+    parser = FlexibleArgumentParser(
+        description="Benchmark the performance with or "
+        "without automatic prefix caching."
+    )
+
+    parser.add_argument(
+        "--document-length",
+        type=int,
+        # Roughly the number of tokens for a system paper,
+        # excluding images
+        default=20000,
+        help="Range of input lengths for sampling prompts, "
+        'specified as "min:max" (e.g., "128:256").',
+    )
+
+    parser.add_argument(
+        "--num-documents",
+        type=int,
+        default=8,
+        help="Range of input lengths for sampling prompts, "
+        'specified as "min:max" (e.g., "128:256").',
+    )
+
+    parser.add_argument("--output-len", type=int, default=10)
+
+    parser.add_argument(
+        "--repeat-count",
+        type=int,
+        default=2,
+        help="Number of times to repeat each prompt",
+    )
+
+    parser.add_argument(
+        "--repeat-mode",
+        type=str,
+        default="random",
+        help="The mode to repeat prompts. The supported "
+        'modes are "random", "tile", and "interleave". '
+        "See repeat_prompts() in the source code for details.",
+    )
+
+    parser.add_argument(
+        "--shuffle-seed",
+        type=int,
+        default=0,
+        help='Random seed when the repeat mode is "random"',
+    )
+
+    parser = EngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_utils.py

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

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import copy
+import itertools
+import pickle as pkl
+import time
+from collections.abc import Iterable
+from typing import Callable
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+from utils import make_rand_sparse_tensors
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+
+# bench
+def bench_fn(
+    label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs
+) -> TMeasurement:
+    min_run_time = 1
+
+    globals = {
+        "args": args,
+        "kwargs": kwargs,
+        "fn": fn,
+    }
+    return TBenchmark.Timer(
+        stmt="fn(*args, **kwargs)",
+        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
+    b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16)
+
+    out = ops.cutlass_scaled_sparse_mm(
+        a, b_compressed, e, scale_a, scale_b, torch.bfloat16
+    )
+    out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
+
+    if not torch.allclose(out, out_ref):
+        print("Incorrect results")
+        print(out)
+        print(out_ref)
+    else:
+        print("Correct results")
+
+    timers = []
+    # pytorch impl - bfloat16
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_bf16_bf16_bf16_matmul-no-scales",
+            torch.mm,
+            a.to(dtype=torch.bfloat16),
+            b.to(dtype=torch.bfloat16),
+        )
+    )
+
+    # pytorch impl - float16
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp16_fp16_fp16_matmul-no-scales",
+            torch.mm,
+            a.to(dtype=torch.float16),
+            b.to(dtype=torch.float16),
+        )
+    )
+
+    # cutlass impl
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_mm",
+            ops.cutlass_scaled_mm,
+            a,
+            b,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
+
+    # cutlass with bias
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_mm_bias",
+            ops.cutlass_scaled_mm,
+            a,
+            b,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+            bias,
+        )
+    )
+
+    # cutlass sparse impl
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_sparse_mm",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
+
+    # cutlass sparse with bias
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_sparse_mm_bias",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+            bias,
+        )
+    )
+
+    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
+    b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16)
+
+    out = ops.cutlass_scaled_sparse_mm(
+        a, b_compressed, e, scale_a, scale_b, torch.bfloat16
+    )
+    out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
+
+    if not torch.allclose(out, out_ref):
+        print("Incorrect results")
+        print(out)
+        print(out_ref)
+    else:
+        print("Correct results")
+
+    timers = []
+
+    # pytorch impl w. bf16
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_bf16_bf16_bf16_matmul-no-scales",
+            torch.mm,
+            a.to(dtype=torch.bfloat16, device="cuda"),
+            b.to(dtype=torch.bfloat16, device="cuda"),
+        )
+    )
+
+    # pytorch impl: bf16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_bf16_scaled_mm",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.bfloat16,
+        )
+    )
+
+    # pytorch impl: bf16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.bfloat16,
+            use_fast_accum=True,
+        )
+    )
+
+    # pytorch impl: fp16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_fp16_scaled_mm",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.float16,
+        )
+    )
+
+    # pytorch impl: fp16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.float16,
+            use_fast_accum=True,
+        )
+    )
+
+    # cutlass impl: bf16 output
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_bf16_scaled_mm",
+            ops.cutlass_scaled_mm,
+            a,
+            b,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
+
+    # cutlass impl: bf16 output
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_bf16_scaled_sparse_mm",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
+
+    # cutlass impl: fp16 output
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_fp16_scaled_sparse_mm",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.float16,
+        )
+    )
+
+    # cutlass impl: bf16 output, with bias
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+            bias,
+        )
+    )
+
+    # cutlass impl: fp16 output, with bias
+    timers.append(
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.float16,
+            bias.to(dtype=torch.float16),
+        )
+    )
+
+    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/sparse_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/sparse_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/sparse_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/utils.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Cutlass bench utils
+from collections.abc import Iterable
+
+import torch
+
+import vllm._custom_ops as ops
+
+
+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 to_bf16(tensor: torch.Tensor) -> torch.Tensor:
+    return tensor.to(dtype=torch.bfloat16)
+
+
+def to_fp16(tensor: torch.Tensor) -> torch.Tensor:
+    return tensor.to(dtype=torch.float16)
+
+
+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")
+
+
+def prune_to_2_4(tensor):
+    # Reshape tensor to [N, 4] where N is number of groups of 4
+    original_shape = tensor.shape
+    reshaped = tensor.reshape(-1, 4)
+
+    # Get indices of top 2 absolute values in each group of 4
+    _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1)
+
+    # Create binary mask
+    mask = torch.zeros_like(reshaped)
+    mask.scatter_(dim=1, index=indices, src=torch.ones_like(indices, dtype=mask.dtype))
+
+    # Apply mask and reshape back
+    pruned = reshaped * mask
+
+    # Turn all -0.0 to 0.0
+    pruned[pruned == -0.0] = 0.0
+
+    return pruned.reshape(original_shape)
+
+
+def make_rand_sparse_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
+
+    b = prune_to_2_4(b.t()).t()
+
+    if dtype == torch.int8:
+        a, b = to_int8(a), to_int8(b)
+    elif dtype == torch.float8_e4m3fn:
+        a, b = to_fp8(a), to_fp8(b)
+    elif dtype == torch.float16:
+        a, b = to_fp16(a), to_fp16(b)
+    elif dtype == torch.bfloat16:
+        a, b = to_bf16(a), to_bf16(b)
+    else:
+        raise ValueError("unsupported dtype")
+
+    b_compressed, e = ops.cutlass_sparse_compress(b.t())
+
+    # Compressed B, Metadata, Original A, B
+    return b_compressed, e, a, b
+
+
+def make_n_rand_sparse_tensors(
+    num_tensors: int, dtype: torch.dtype, m: int, n: int, k: int
+) -> tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
+    ABs = []
+    for _ in range(num_tensors):
+        b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k)
+        if b_comp is not None:
+            ABs.append(make_rand_sparse_tensors(dtype, m, n, k))
+    BComps, Es, As, Bs = zip(*ABs)
+    return list(BComps), list(Es), list(As), list(Bs)

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

+#!/bin/bash
+
+# benchmark the overhead of disaggregated prefill.
+# methodology:
+# - send all request to prefill vLLM instance. It will buffer KV cache.
+# - then send all request to decode instance. 
+# - The TTFT of decode instance is the overhead.
+
+set -ex
+
+kill_gpu_processes() {
+  # kill all processes on GPU.
+  pgrep pt_main_thread | xargs -r kill -9
+  pgrep python3 | xargs -r kill -9
+  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"
+}
+
+wait_for_server() {
+  # wait for vllm server to start
+  # return 1 if vllm server crashes
+  local port=$1
+  timeout 1200 bash -c "
+    until curl -s localhost:${port}/v1/completions > /dev/null; do
+      sleep 1
+    done" && return 0 || return 1
+}
+
+
+benchmark() {
+
+  export VLLM_LOGGING_LEVEL=DEBUG
+  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
+
+  # compare chunked prefill with disaggregated prefill
+
+  results_folder="./results"
+  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
+  dataset_name="sonnet"
+  dataset_path="../sonnet_4x.txt"
+  num_prompts=10
+  qps=$1
+  prefix_len=50
+  input_len=2048
+  output_len=$2
+
+
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
+    --port 8100 \
+    --max-model-len 10000 \
+    --gpu-memory-utilization 0.6 \
+    --kv-transfer-config \
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    
+
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
+    --port 8200 \
+    --max-model-len 10000 \
+    --gpu-memory-utilization 0.6 \
+    --kv-transfer-config \
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+
+  wait_for_server 8100
+  wait_for_server 8200
+
+  # let the prefill instance finish prefill
+  python3 ../benchmark_serving.py \
+          --backend vllm \
+          --model $model \
+          --dataset-name $dataset_name \
+          --dataset-path $dataset_path \
+          --sonnet-input-len $input_len \
+          --sonnet-output-len "$output_len" \
+          --sonnet-prefix-len $prefix_len \
+          --num-prompts $num_prompts \
+          --port 8100 \
+          --save-result \
+          --result-dir $results_folder \
+          --result-filename disagg_prefill_tp1.json \
+          --request-rate "inf"
+
+
+  # send the request to decode.
+  # The TTFT of this command will be the overhead of disagg prefill impl.
+  python3 ../benchmark_serving.py \
+          --backend vllm \
+          --model $model \
+          --dataset-name $dataset_name \
+          --dataset-path $dataset_path \
+          --sonnet-input-len $input_len \
+          --sonnet-output-len "$output_len" \
+          --sonnet-prefix-len $prefix_len \
+          --num-prompts $num_prompts \
+          --port 8200 \
+          --save-result \
+          --result-dir $results_folder \
+          --result-filename disagg_prefill_tp1_overhead.json \
+          --request-rate "$qps"
+  kill_gpu_processes
+
+}
+
+
+main() {
+
+  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+  (which jq) || (apt-get -y install jq)
+  (which socat) || (apt-get -y install socat)
+
+  pip install quart httpx datasets
+
+  cd "$(dirname "$0")"
+
+  cd ..
+  # create sonnet-4x.txt
+  echo "" > sonnet_4x.txt
+  for _ in {1..4}
+  do
+    cat sonnet.txt >> sonnet_4x.txt
+  done
+  cd disagg_benchmarks
+
+  rm -rf results
+  mkdir results
+
+  default_qps=1
+  default_output_len=1
+  benchmark $default_qps $default_output_len
+
+}
+
+
+main "$@"

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

+#!/bin/bash
+
+# Requirement: 2x GPUs.
+
+
+# Model: meta-llama/Meta-Llama-3.1-8B-Instruct
+# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests
+# Resource: 2x GPU
+# Approaches:
+# 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4
+# 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance
+# Prefilling instance: max_output_token=1
+# Decoding instance: force the input tokens be the same across requests to bypass prefilling
+
+set -ex
+
+kill_gpu_processes() {
+  # kill all processes on GPU.
+  pgrep pt_main_thread | xargs -r kill -9
+  pgrep python3 | xargs -r kill -9
+  for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done
+  sleep 1
+}
+
+wait_for_server() {
+  # wait for vllm server to start
+  # return 1 if vllm server crashes
+  local port=$1
+  timeout 1200 bash -c "
+    until curl -s localhost:${port}/v1/completions > /dev/null; do
+      sleep 1
+    done" && return 0 || return 1
+}
+
+
+launch_chunked_prefill() {
+  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
+  # disagg prefill
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
+    --port 8100 \
+    --max-model-len 10000 \
+    --enable-chunked-prefill \
+    --gpu-memory-utilization 0.6 &
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
+    --port 8200 \
+    --max-model-len 10000 \
+    --enable-chunked-prefill \
+    --gpu-memory-utilization 0.6 &
+  wait_for_server 8100
+  wait_for_server 8200
+  python3 round_robin_proxy.py &
+  sleep 1
+}
+
+
+launch_disagg_prefill() {
+  model="meta-llama/Meta-Llama-3.1-8B-Instruct" 
+  # disagg prefill
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
+    --port 8100 \
+    --max-model-len 10000 \
+    --gpu-memory-utilization 0.6 \
+    --kv-transfer-config \
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
+    --port 8200 \
+    --max-model-len 10000 \
+    --gpu-memory-utilization 0.6 \
+    --kv-transfer-config \
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+
+  wait_for_server 8100
+  wait_for_server 8200
+  python3 disagg_prefill_proxy_server.py &
+  sleep 1
+}
+
+
+benchmark() {
+  results_folder="./results"
+  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
+  dataset_name="sonnet"
+  dataset_path="../sonnet_4x.txt"
+  num_prompts=100
+  qps=$1
+  prefix_len=50
+  input_len=1024
+  output_len=$2
+  tag=$3
+
+  python3 ../benchmark_serving.py \
+          --backend vllm \
+          --model $model \
+          --dataset-name $dataset_name \
+          --dataset-path $dataset_path \
+          --sonnet-input-len $input_len \
+          --sonnet-output-len "$output_len" \
+          --sonnet-prefix-len $prefix_len \
+          --num-prompts $num_prompts \
+          --port 8000 \
+          --save-result \
+          --result-dir $results_folder \
+          --result-filename "$tag"-qps-"$qps".json \
+          --request-rate "$qps"
+
+  sleep 2
+}
+
+
+main() {
+
+  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+  (which jq) || (apt-get -y install jq)
+  (which socat) || (apt-get -y install socat)
+  (which lsof) || (apt-get -y install lsof)
+
+  pip install quart httpx matplotlib aiohttp datasets
+
+  cd "$(dirname "$0")"
+
+  cd ..
+  # create sonnet-4x.txt so that we can sample 2048 tokens for input
+  echo "" > sonnet_4x.txt
+  for _ in {1..4}
+  do
+    cat sonnet.txt >> sonnet_4x.txt
+  done
+  cd disagg_benchmarks
+
+  rm -rf results
+  mkdir results
+
+  default_output_len=6
+
+  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
+
+  launch_chunked_prefill
+  for qps in 2 4 6 8; do
+  benchmark $qps $default_output_len chunked_prefill
+  done
+  kill_gpu_processes
+
+  launch_disagg_prefill
+  for qps in 2 4 6 8; do
+  benchmark $qps $default_output_len disagg_prefill
+  done
+  kill_gpu_processes
+
+  python3 visualize_benchmark_results.py
+
+}
+
+
+main "$@"

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+
+import aiohttp
+from quart import Quart, make_response, request
+
+AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
+
+app = Quart(__name__)
+
+
+async def forward_request(url, data):
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
+        async with session.post(url=url, json=data, headers=headers) as response:
+            if response.status == 200:
+                # if response.headers.get('Transfer-Encoding') == 'chunked':
+                if True:
+                    async for chunk_bytes in response.content.iter_chunked(1024):
+                        yield chunk_bytes
+                else:
+                    content = await response.read()
+                    yield content
+
+
+@app.route("/v1/completions", methods=["POST"])
+async def handle_request():
+    try:
+        original_request_data = await request.get_json()
+
+        prefill_request = original_request_data.copy()
+        # change max_tokens = 1 to let it only do prefill
+        prefill_request["max_tokens"] = 1
+
+        # finish prefill
+        async for _ in forward_request(
+            "http://localhost:8100/v1/completions", prefill_request
+        ):
+            continue
+
+        # return decode
+        generator = forward_request(
+            "http://localhost:8200/v1/completions", original_request_data
+        )
+        response = await make_response(generator)
+        response.timeout = None
+
+        return response
+
+    except Exception as e:
+        import sys
+        import traceback
+
+        exc_info = sys.exc_info()
+        print("Error occurred in disagg prefill proxy server")
+        print(e)
+        print("".join(traceback.format_exception(*exc_info)))
+
+
+if __name__ == "__main__":
+    app.run(port=8000)

benchmarks/disagg_benchmarks/round_robin_proxy.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import asyncio
+import itertools
+
+import aiohttp
+from aiohttp import web
+
+
+class RoundRobinProxy:
+    def __init__(self, target_ports):
+        self.target_ports = target_ports
+        self.port_cycle = itertools.cycle(self.target_ports)
+
+    async def handle_request(self, request):
+        target_port = next(self.port_cycle)
+        target_url = f"http://localhost:{target_port}{request.path_qs}"
+
+        async with aiohttp.ClientSession() as session:
+            try:
+                # Forward the request
+                async with session.request(
+                    method=request.method,
+                    url=target_url,
+                    headers=request.headers,
+                    data=request.content,
+                ) as response:
+                    # Start sending the response
+                    resp = web.StreamResponse(
+                        status=response.status, headers=response.headers
+                    )
+                    await resp.prepare(request)
+
+                    # Stream the response content
+                    async for chunk in response.content.iter_any():
+                        await resp.write(chunk)
+
+                    await resp.write_eof()
+                    return resp
+
+            except Exception as e:
+                return web.Response(text=f"Error: {str(e)}", status=500)
+
+
+async def main():
+    proxy = RoundRobinProxy([8100, 8200])
+    app = web.Application()
+    app.router.add_route("*", "/{path:.*}", proxy.handle_request)
+
+    runner = web.AppRunner(app)
+    await runner.setup()
+    site = web.TCPSite(runner, "localhost", 8000)
+    await site.start()
+
+    print("Proxy server started on http://localhost:8000")
+
+    # Keep the server running
+    await asyncio.Event().wait()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import matplotlib.pyplot as plt
+import pandas as pd
+
+if __name__ == "__main__":
+    data = []
+    for name in ["disagg_prefill", "chunked_prefill"]:
+        for qps in [2, 4, 6, 8]:
+            with open(f"results/{name}-qps-{qps}.json") as f:
+                x = json.load(f)
+                x["name"] = name
+                x["qps"] = qps
+                data.append(x)
+
+    df = pd.DataFrame.from_dict(data)
+    dis_df = df[df["name"] == "disagg_prefill"]
+    chu_df = df[df["name"] == "chunked_prefill"]
+
+    plt.style.use("bmh")
+    plt.rcParams["font.size"] = 20
+
+    for key in [
+        "mean_ttft_ms",
+        "median_ttft_ms",
+        "p99_ttft_ms",
+        "mean_itl_ms",
+        "median_itl_ms",
+        "p99_itl_ms",
+    ]:
+        fig, ax = plt.subplots(figsize=(11, 7))
+        plt.plot(
+            dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4
+        )
+        plt.plot(
+            chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4
+        )
+        ax.legend()
+
+        ax.set_xlabel("QPS")
+        ax.set_ylabel(key)
+        ax.set_ylim(bottom=0)
+        fig.savefig(f"results/{key}.png")
+        plt.close(fig)

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pickle as pkl
+import time
+from collections.abc import Iterable
+from dataclasses import dataclass
+from itertools import product
+from typing import Callable, Optional
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+from tqdm import tqdm
+
+import vllm._custom_ops as ops
+from vllm.model_executor.layers.layernorm import RMSNorm
+
+
+@dataclass
+class bench_params_t:
+    num_tokens: int
+    hidden_size: int
+    add_residual: bool
+    dtype: torch.dtype
+
+    def description(self):
+        return (
+            f"N {self.num_tokens} "
+            f"x D {self.hidden_size} "
+            f"x R {self.add_residual} "
+            f"x DT {self.dtype}"
+        )
+
+
+def get_bench_params() -> list[bench_params_t]:
+    ## Test Fixtures
+    NUM_TOKENS = [2**x for x in range(11)]
+    HIDDEN_SIZES = list(range(1024, 8129, 1024))
+    ADD_RESIDUAL = [True, False]
+    DTYPES = [torch.bfloat16, torch.float]
+
+    combinations = product(NUM_TOKENS, HIDDEN_SIZES, ADD_RESIDUAL, DTYPES)
+    bench_params = list(
+        map(lambda x: bench_params_t(x[0], x[1], x[2], x[3]), combinations)
+    )
+    return bench_params
+
+
+# Reference impls
+def unfused_int8_impl(
+    rms_norm_layer: RMSNorm,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor],
+    quant_dtype: torch.dtype,
+):
+    # Norm
+    torch_out = None
+    if residual is None:
+        torch_out = rms_norm_layer.forward_cuda(x, residual)
+    else:
+        torch_out, _ = rms_norm_layer.forward_cuda(x, residual)
+
+    # Quant
+    torch_out, _, _ = ops.scaled_int8_quant(torch_out)
+
+
+def unfused_fp8_impl(
+    rms_norm_layer: RMSNorm,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor],
+    quant_dtype: torch.dtype,
+):
+    # Norm
+    torch_out = None
+    if residual is None:
+        torch_out = rms_norm_layer.forward_cuda(x, residual)
+    else:
+        torch_out, _ = rms_norm_layer.forward_cuda(x, residual)
+
+    # Quant
+    torch_out, _ = ops.scaled_fp8_quant(torch_out)
+
+
+def fused_impl(
+    rms_norm_layer: RMSNorm,  # this stores the weights
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor],
+    quant_dtype: torch.dtype,
+):
+    out, _ = ops.rms_norm_dynamic_per_token_quant(
+        x, rms_norm_layer.weight, 1e-6, quant_dtype, residual=residual
+    )
+
+
+# Bench functions
+def bench_fn(
+    rms_norm_layer: RMSNorm,
+    x: torch.Tensor,
+    residual: torch.Tensor,
+    quant_dtype: torch.dtype,
+    label: str,
+    sub_label: str,
+    fn: Callable,
+    description: str,
+) -> TMeasurement:
+    min_run_time = 1
+
+    globals = {
+        "rms_norm_layer": rms_norm_layer,
+        "x": x,
+        "residual": residual,
+        "quant_dtype": quant_dtype,
+        "fn": fn,
+    }
+    return TBenchmark.Timer(
+        stmt="fn(rms_norm_layer, x, residual, quant_dtype)",
+        globals=globals,
+        label=label,
+        sub_label=sub_label,
+        description=description,
+    ).blocked_autorange(min_run_time=min_run_time)
+
+
+def bench(params: bench_params_t, label: str, sub_label: str) -> Iterable[TMeasurement]:
+    # Make inputs
+    layer = RMSNorm(params.hidden_size, 1e-6).to(dtype=params.dtype)
+    # Make weights
+    layer.weight.data.normal_(mean=1.0, std=0.1)
+    # Make inputs
+    scale = 1 / params.hidden_size
+    x = (
+        torch.randn(
+            params.num_tokens, params.hidden_size, dtype=params.dtype, device="cuda"
+        )
+        * scale
+    )
+    residual = (
+        (torch.randn_like(x) * scale).to(device="cuda") if params.add_residual else None
+    )
+
+    timers = []
+
+    # unfused int8 impl.
+    timers.append(
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.int8,
+            label,
+            sub_label,
+            unfused_int8_impl,
+            "unfused_int8_impl",
+        )
+    )
+
+    # unfused fp8 impl.
+    timers.append(
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.float8_e4m3fn,
+            label,
+            sub_label,
+            unfused_fp8_impl,
+            "unfused_fp8_impl",
+        )
+    )
+
+    # fused int8 impl.
+    timers.append(
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.int8,
+            label,
+            sub_label,
+            fused_impl,
+            "fused_int8_impl",
+        )
+    )
+
+    # fused fp8 impl.
+    timers.append(
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.float8_e4m3fn,
+            label,
+            sub_label,
+            fused_impl,
+            "fused_fp8_impl",
+        )
+    )
+
+    print_timers(timers)
+
+    return timers
+
+
+# launch bench
+# runner
+def print_timers(timers: Iterable[TMeasurement]):
+    compare = TBenchmark.Compare(timers)
+    compare.print()
+
+
+def main():
+    torch.set_default_device("cuda")
+    bench_params = get_bench_params()
+
+    timers = []
+    for bp in tqdm(bench_params):
+        timers.extend(bench(bp, "rms-norm-dynamic-per-token-quant", bp.description()))
+    print_timers(timers)
+
+    # pickle all the results
+    timestamp = int(time.time())
+    with open(f"rms_norm_dpt_quant-{timestamp}.pkl", "wb") as f:
+        pkl.dump(timers, f)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/bench_fp8_gemm.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
+from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "fp8-tensor-w-token-a": dict(
+        w="tensor", a="token", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a": dict(
+        w="tensor", a="tensor", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-token-a": dict(
+        w="channel", a="token", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-tensor-a": dict(
+        w="channel", a="tensor", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-token-a-noquant": dict(
+        w="tensor", a="token", no_a_quant=True, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a-noquant": dict(
+        w="tensor", a="tensor", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-token-a-noquant": dict(
+        w="channel", a="token", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-tensor-a-noquant": dict(
+        w="channel", a="tensor", no_a_quant=True, enabled=False
+    ),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str):
+    if w_type == "tensor":
+        scale_b = torch.ones(1, device=device, dtype=torch.float32)
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
+    else:
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True)
+    return b_fp8.t(), scale_b_fp8
+
+
+def build_fp8_runner(cfg, a, b, dtype, device):
+    b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device)
+
+    scale_a_const = (
+        torch.ones(1, device=device, dtype=torch.float32)
+        if cfg["a"] == "tensor"
+        else None
+    )
+
+    if cfg["no_a_quant"]:
+        if cfg["a"] == "tensor":
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+        else:
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+
+        def run():
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+        return run
+
+    if cfg["a"] == "tensor":
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    else:
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs FP8 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_fp8_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_fp8_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/bench_int8_gemm.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
+from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "int8-tensor-w-token-a": dict(
+        w="tensor", a="token", no_a_quant=False, enabled=False
+    ),
+    "int8-tensor-w-tensor-a": dict(
+        w="tensor", a="tensor", no_a_quant=False, enabled=True
+    ),
+    "int8-channel-w-token-a": dict(
+        w="channel", a="token", no_a_quant=False, enabled=True
+    ),
+    "int8-channel-w-tensor-a": dict(
+        w="channel", a="tensor", no_a_quant=False, enabled=False
+    ),
+    "int8-tensor-w-token-a-noquant": dict(
+        w="tensor", a="token", no_a_quant=True, enabled=False
+    ),
+    "int8-tensor-w-tensor-a-noquant": dict(
+        w="tensor", a="tensor", no_a_quant=True, enabled=True
+    ),
+    "int8-channel-w-token-a-noquant": dict(
+        w="channel", a="token", no_a_quant=True, enabled=True
+    ),
+    "int8-channel-w-tensor-a-noquant": dict(
+        w="channel", a="tensor", no_a_quant=True, enabled=False
+    ),
+}
+
+
+def _quant_weight(b, w_type, device):
+    if w_type == "tensor":
+        scale_b = torch.ones(1, device=device, dtype=torch.float32)
+        b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b, scale_b)
+        assert scale_b_int8.numel() == 1
+    else:  # channel
+        b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b)
+        assert scale_b_int8.numel() == b.shape[0]
+    return b_int8.t(), scale_b_int8
+
+
+def build_int8_runner(cfg, a, b, dtype, device):
+    # quant before running the kernel
+    b_int8, scale_b_int8 = _quant_weight(b, cfg["w"], device)
+
+    scale_a_const = None
+    if cfg["a"] == "tensor":
+        scale_a_const = torch.ones(1, device=device, dtype=torch.float32)
+
+    # no quant, create activation ahead
+    if cfg["no_a_quant"]:
+        if cfg["a"] == "tensor":
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const)
+        else:  # token
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a)
+
+        def run_quant():
+            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
+
+        return run_quant
+
+    # dynamic quant, create activation inside
+    if cfg["a"] == "tensor":
+
+        def run_quant():
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const)
+            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
+
+    else:  # token
+
+        def run_quant():
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a)
+            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
+
+    return run_quant
+
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v.get("enabled")]
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=[k for k in _enabled],
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs INT8 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_int8_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    KN_model_names = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            KN_model_names.append(KN)
+    return KN_model_names
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+        help="List of models to benchmark",
+    )
+    parser.add_argument(
+        "--tp-sizes",
+        nargs="+",
+        type=int,
+        default=[1],
+        help="List of tensor parallel sizes",
+    )
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs INT8 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_int8_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")