No commit message

No commit message

benchmarks/README.md

+# Benchmarking vLLM
+
+## Downloading the ShareGPT dataset
+
+You can download the dataset by running:
+```bash
+wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+```

benchmarks/backend_request_func.py

+import json
+import os
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+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)
+
+
+@dataclass
+class RequestFuncInput:
+    prompt: str
+    api_url: str
+    prompt_len: int
+    output_len: int
+    model: str
+    best_of: int = 1
+    use_beam_search: bool = False
+    logprobs: Optional[int] = None
+    multi_modal_content: Optional[dict] = None
+
+
+@dataclass
+class RequestFuncOutput:
+    generated_text: str = ""
+    success: bool = False
+    latency: float = 0.0
+    ttft: float = 0.0  # Time to first token
+    itl: List[float] = field(
+        default_factory=list)  # List of inter-token latencies
+    prompt_len: int = 0
+    error: str = ""
+
+
+async def async_request_tgi(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate_stream")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        params = {
+            "best_of": request_func_input.best_of,
+            "max_new_tokens": request_func_input.output_len,
+            "do_sample": True,
+            "temperature": 0.01,  # TGI does not accept 0.0 temperature.
+            "top_p": 0.99,  # TGI does not accept 1.0 top_p.
+        }
+        payload = {
+            "inputs": request_func_input.prompt,
+            "parameters": params,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        ttft = 0.0
+        st = time.perf_counter()
+        most_recent_timestamp = st
+        try:
+            async with session.post(url=api_url, json=payload) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+                        chunk_bytes = chunk_bytes.decode("utf-8")
+
+                        #NOTE: Sometimes TGI returns a ping response without
+                        # 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()
+                        # First token
+                        if ttft == 0.0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+
+                        # Decoding phase
+                        else:
+                            output.itl.append(timestamp -
+                                              most_recent_timestamp)
+
+                        most_recent_timestamp = timestamp
+
+                    output.latency = most_recent_timestamp - st
+                    output.success = True
+                    output.generated_text = data["generated_text"]
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_trt_llm(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate_stream")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        assert request_func_input.best_of == 1
+        payload = {
+            "accumulate_tokens": True,
+            "text_input": request_func_input.prompt,
+            "temperature": 0.0,
+            "top_p": 1.0,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        ttft = 0.0
+        st = time.perf_counter()
+        most_recent_timestamp = st
+        try:
+            async with session.post(url=api_url, json=payload) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data:")
+
+                        data = json.loads(chunk)
+                        output.generated_text += data["text_output"]
+                        timestamp = time.perf_counter()
+                        # First token
+                        if ttft == 0.0:
+                            ttft = time.perf_counter() - st
+                            output.ttft = ttft
+
+                        # Decoding phase
+                        else:
+                            output.itl.append(timestamp -
+                                              most_recent_timestamp)
+
+                        most_recent_timestamp = timestamp
+
+                    output.latency = most_recent_timestamp - st
+                    output.success = True
+
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_deepspeed_mii(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert request_func_input.best_of == 1
+        assert not request_func_input.use_beam_search
+
+        payload = {
+            "prompt": request_func_input.prompt,
+            "max_tokens": request_func_input.output_len,
+            "temperature": 0.01,  # deepspeed-mii does not accept 0.0 temp.
+            "top_p": 1.0,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        # NOTE: DeepSpeed-MII doesn't support streaming as of Jan 28 2024,
+        # will use 0 as placeholder.
+        # See https://github.com/microsoft/DeepSpeed-MII/pull/311
+        output.ttft = 0
+
+        st = time.perf_counter()
+        try:
+            async with session.post(url=request_func_input.api_url,
+                                    json=payload) as response:
+                if response.status == 200:
+                    parsed_resp = await response.json()
+                    output.latency = time.perf_counter() - st
+                    output.generated_text = parsed_resp["text"][0]
+                    output.success = True
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
+async def async_request_openai_completions(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith(
+        ("completions", "profile")
+    ), "OpenAI Completions API URL must end with 'completions' or 'profile'."
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        payload = {
+            "model": request_func_input.model,
+            "prompt": request_func_input.prompt,
+            "temperature": 0.0,
+            "best_of": request_func_input.best_of,
+            "max_tokens": request_func_input.output_len,
+            "logprobs": request_func_input.logprobs,
+            "stream": True,
+        }
+        headers = {
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
+        }
+
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        generated_text = ""
+        ttft = 0.0
+        st = time.perf_counter()
+        most_recent_timestamp = st
+        try:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data: ")
+                        if chunk == "[DONE]":
+                            latency = time.perf_counter() - st
+                        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
+                                if ttft == 0.0:
+                                    ttft = time.perf_counter() - st
+                                    output.ttft = ttft
+
+                                # Decoding phase
+                                else:
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
+
+                                most_recent_timestamp = timestamp
+                                generated_text += data["choices"][0]["text"]
+
+                    output.generated_text = generated_text
+                    output.success = True
+                    output.latency = latency
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
+async def async_request_openai_chat_completions(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith(
+        "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
+        content = [{"type": "text", "text": request_func_input.prompt}]
+        if request_func_input.multi_modal_content:
+            content.append(request_func_input.multi_modal_content)
+        payload = {
+            "model": request_func_input.model,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": content
+                },
+            ],
+            "temperature": 0.0,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
+        }
+
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        generated_text = ""
+        ttft = 0.0
+        st = time.perf_counter()
+        most_recent_timestamp = st
+        try:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data: ")
+                        if chunk == "[DONE]":
+                            latency = time.perf_counter() - st
+                        else:
+                            timestamp = time.perf_counter()
+                            data = json.loads(chunk)
+
+                            delta = data["choices"][0]["delta"]
+                            if delta.get("content", None):
+                                # First token
+                                if ttft == 0.0:
+                                    ttft = time.perf_counter() - st
+                                    output.ttft = ttft
+
+                                # Decoding phase
+                                else:
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
+
+                                generated_text += delta["content"]
+
+                            most_recent_timestamp = timestamp
+
+                    output.generated_text = generated_text
+                    output.success = True
+                    output.latency = latency
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
+# Since vllm must support Python 3.8, we can't use str.removeprefix(prefix)
+# introduced in Python 3.9
+def remove_prefix(text: str, prefix: str) -> str:
+    if text.startswith(prefix):
+        return text[len(prefix):]
+    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,
+    "lmdeploy": async_request_openai_completions,
+    "deepspeed-mii": async_request_deepspeed_mii,
+    "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

+"""Benchmark the latency of processing a single batch of requests."""
+import argparse
+import json
+import time
+from pathlib import Path
+from typing import List, Optional
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from vllm import LLM, SamplingParams
+from vllm.engine.arg_utils import DEVICE_OPTIONS, EngineArgs
+from vllm.inputs import PromptInputs
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.utils import FlexibleArgumentParser
+
+
+def main(args: argparse.Namespace):
+    print(args)
+
+    # NOTE(woosuk): If the request cannot be processed in a single batch,
+    # the engine will automatically process the request in multiple batches.
+    llm = LLM(
+        model=args.model,
+        speculative_model=args.speculative_model,
+        num_speculative_tokens=args.num_speculative_tokens,
+        speculative_draft_tensor_parallel_size=\
+            args.speculative_draft_tensor_parallel_size,
+        tokenizer=args.tokenizer,
+        quantization=args.quantization,
+        tensor_parallel_size=args.tensor_parallel_size,
+        trust_remote_code=args.trust_remote_code,
+        dtype=args.dtype,
+        max_model_len=args.max_model_len,
+        enforce_eager=args.enforce_eager,
+        kv_cache_dtype=args.kv_cache_dtype,
+        quantization_param_path=args.quantization_param_path,
+        device=args.device,
+        ray_workers_use_nsight=args.ray_workers_use_nsight,
+        use_v2_block_manager=args.use_v2_block_manager,
+        enable_chunked_prefill=args.enable_chunked_prefill,
+        download_dir=args.download_dir,
+        block_size=args.block_size,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        load_format=args.load_format,
+        distributed_executor_backend=args.distributed_executor_backend,
+        otlp_traces_endpoint=args.otlp_traces_endpoint,
+        enable_prefix_caching=args.enable_prefix_caching,
+    )
+
+    sampling_params = SamplingParams(
+        n=args.n,
+        temperature=0.0 if args.use_beam_search else 1.0,
+        top_p=1.0,
+        use_beam_search=args.use_beam_search,
+        ignore_eos=True,
+        max_tokens=args.output_len,
+    )
+    print(sampling_params)
+    dummy_prompt_token_ids = np.random.randint(10000,
+                                               size=(args.batch_size,
+                                                     args.input_len))
+    dummy_inputs: List[PromptInputs] = [{
+        "prompt_token_ids": batch
+    } for batch in dummy_prompt_token_ids.tolist()]
+
+    def run_to_completion(profile_dir: Optional[str] = None):
+        if profile_dir:
+            with torch.profiler.profile(
+                    activities=[
+                        torch.profiler.ProfilerActivity.CPU,
+                        torch.profiler.ProfilerActivity.CUDA,
+                    ],
+                    on_trace_ready=torch.profiler.tensorboard_trace_handler(
+                        str(profile_dir))) as p:
+                llm.generate(dummy_inputs,
+                             sampling_params=sampling_params,
+                             use_tqdm=False)
+            print(p.key_averages())
+        else:
+            start_time = time.perf_counter()
+            llm.generate(dummy_inputs,
+                         sampling_params=sampling_params,
+                         use_tqdm=False)
+            end_time = time.perf_counter()
+            latency = end_time - start_time
+            return latency
+
+    print("Warming up...")
+    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
+        run_to_completion(profile_dir=None)
+
+    if args.profile:
+        profile_dir = args.profile_result_dir
+        if not profile_dir:
+            profile_dir = Path(
+                "."
+            ) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
+        print(f"Profiling (results will be saved to '{profile_dir}')...")
+        run_to_completion(profile_dir=profile_dir)
+        return
+
+    # Benchmark.
+    latencies = []
+    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, 99]
+    percentiles = np.percentile(latencies, percentages)
+    print(f'Avg latency: {np.mean(latencies)} seconds')
+    for percentage, percentile in zip(percentages, percentiles):
+        print(f'{percentage}% percentile latency: {percentile} seconds')
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "avg_latency": np.mean(latencies),
+            "latencies": latencies.tolist(),
+            "percentiles": dict(zip(percentages, percentiles.tolist())),
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+
+
+if __name__ == '__main__':
+    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',
+                        choices=[*QUANTIZATION_METHODS, None],
+                        default=None)
+    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
+    parser.add_argument('--input-len', type=int, default=32)
+    parser.add_argument('--output-len', type=int, default=128)
+    parser.add_argument('--batch-size', type=int, default=8)
+    parser.add_argument('--n',
+                        type=int,
+                        default=1,
+                        help='Number of generated sequences per prompt.')
+    parser.add_argument('--use-beam-search', action='store_true')
+    parser.add_argument('--num-iters-warmup',
+                        type=int,
+                        default=10,
+                        help='Number of iterations to run for warmup.')
+    parser.add_argument('--num-iters',
+                        type=int,
+                        default=30,
+                        help='Number of iterations to run.')
+    parser.add_argument('--trust-remote-code',
+                        action='store_true',
+                        help='trust remote code from huggingface')
+    parser.add_argument(
+        '--max-model-len',
+        type=int,
+        default=None,
+        help='Maximum length of a sequence (including prompt and output). '
+        'If None, will be derived from the model.')
+    parser.add_argument(
+        '--dtype',
+        type=str,
+        default='auto',
+        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
+        help='data type for model weights and activations. '
+        'The "auto" option will use FP16 precision '
+        'for FP32 and FP16 models, and BF16 precision '
+        'for BF16 models.')
+    parser.add_argument('--enforce-eager',
+                        action='store_true',
+                        help='enforce eager mode and disable CUDA graph')
+    parser.add_argument(
+        '--kv-cache-dtype',
+        type=str,
+        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
+        default="auto",
+        help='Data type for kv cache storage. If "auto", will use model '
+        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
+        'ROCm (hcu) supports fp8 (=fp8_e4m3)')
+    parser.add_argument(
+        '--quantization-param-path',
+        type=str,
+        default=None,
+        help='Path to the JSON file containing the KV cache scaling factors. '
+        'This should generally be supplied, when KV cache dtype is FP8. '
+        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
+        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
+        'cuda version greater than 11.8. On ROCm (hcu), FP8_E4M3 is '
+        'instead supported for common inference criteria.')
+    parser.add_argument(
+        '--profile',
+        action='store_true',
+        help='profile the generation process of a single batch')
+    parser.add_argument(
+        '--profile-result-dir',
+        type=str,
+        default=None,
+        help=('path to save the pytorch profiler output. Can be visualized '
+              'with ui.perfetto.dev or Tensorboard.'))
+    parser.add_argument("--device",
+                        type=str,
+                        default="auto",
+                        choices=DEVICE_OPTIONS,
+                        help='device type for vLLM execution')
+    parser.add_argument('--block-size',
+                        type=int,
+                        default=16,
+                        help='block size of key/value cache')
+    parser.add_argument(
+        '--enable-chunked-prefill',
+        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",
+        action='store_true',
+        help="If specified, use nsight to profile ray workers",
+    )
+    parser.add_argument('--download-dir',
+                        type=str,
+                        default=None,
+                        help='directory to download and load the weights, '
+                        'default to the default cache dir of huggingface')
+    parser.add_argument(
+        '--output-json',
+        type=str,
+        default=None,
+        help='Path to save the latency results in JSON format.')
+    parser.add_argument('--gpu-memory-utilization',
+                        type=float,
+                        default=0.9,
+                        help='the fraction of GPU memory to be used for '
+                        'the model executor, which can range from 0 to 1.'
+                        'If unspecified, will use the default value of 0.9.')
+    parser.add_argument(
+        '--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

+"""
+Benchmark the efficiency of prefix caching.
+
+This script allows you to benchmark the performance of
+a model with and without prefix caching using either fixed prompts
+or prompts sampled from the ShareGPT dataset.
+
+Fixed example usage:
+    python benchmark_prefix_caching.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --enable-prefix-caching \
+        --num-prompts 1 \
+        --repeat-count 100
+
+ShareGPT example usage:
+    # This command samples 20 prompts with input lengths
+    # between 128 and 256 tokens from the ShareGPT dataset,
+    # then replicates each prompt 5 times.
+    python benchmark_prefix_caching.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json \
+        --enable-prefix-caching \
+        --num-prompts 20 \
+        --repeat-count 5 \
+        --input-length-range 128:256
+"""
+
+import json
+import random
+import time
+from typing import List, Optional, Tuple
+
+from transformers import PreTrainedTokenizerBase
+
+from vllm import LLM, SamplingParams
+from vllm.utils import FlexibleArgumentParser
+import triton
+
+
+try:
+    from vllm.transformers_utils.tokenizer import get_tokenizer
+except ImportError:
+    from backend_request_func import get_tokenizer
+
+triton_version = triton.__version__
+if triton_version.startswith("2.1"):
+    from triton.common.backend import compute_core_version_key
+elif triton_version.startswith("3.0"):
+    from triton.compiler.compiler import triton_key
+else:
+    print(f"TRITON version {triton_version} is not specifically handled.")
+
+
+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
+
+
+def test_prefix(llm=None, sampling_params=None, prompts=None):
+    if triton_version.startswith("2.1"):
+        version_key = compute_core_version_key()
+    if triton_version.startswith("3.0"):
+        version_key = triton_key()
+    start_time = time.time()
+
+    llm.generate(prompts, sampling_params=sampling_params)
+
+    end_time = time.time()
+    print(f"cost time {end_time - start_time}")
+
+
+def sample_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    input_length_range: Tuple[int, int],
+    fixed_output_len: Optional[int],
+) -> List[Tuple[str, int, int]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # Shuffle the dataset.
+    random.shuffle(dataset)
+
+    min_len, max_len = input_length_range
+
+    # Filter out sequences that are too long or too short
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for i in range(len(dataset)):
+        if len(filtered_dataset) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = dataset[i][1]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if min_len <= prompt_len <= max_len:
+            filtered_dataset.append((prompt, prompt_len, output_len))
+
+    return filtered_dataset
+
+
+def repeat_and_sort_requests(requests: List[Tuple[str, int, int]],
+                             repeat_count: int,
+                             sort: bool = False) -> List[str]:
+    repeated_requests = requests * repeat_count
+    if sort:
+        repeated_requests.sort(key=lambda x: x[1])
+    else:
+        random.shuffle(repeated_requests)
+    return [req[0] for req in repeated_requests]
+
+
+def main(args):
+    tokenizer = get_tokenizer(args.model, trust_remote_code=True)
+    input_length_range = tuple(map(int, args.input_length_range.split(':')))
+
+    if args.dataset_path is not None:
+        print(f"Start to sample {args.num_prompts} prompts"
+              "from {args.dataset_path}")
+        filtered_datasets = sample_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            input_length_range=input_length_range,
+            fixed_output_len=args.output_len,
+        )
+    else:
+        prompt_len = len(tokenizer(PROMPT).input_ids)
+        filtered_datasets = [(PROMPT, prompt_len, args.output_len)
+                             ] * args.num_prompts
+
+    llm = LLM(model=args.model,
+              tokenizer_mode='auto',
+              trust_remote_code=True,
+              enforce_eager=True,
+              use_v2_block_manager=args.use_v2_block_manager,
+              tensor_parallel_size=args.tensor_parallel_size,
+              enable_prefix_caching=args.enable_prefix_caching)
+
+    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+
+    print("Testing filtered datasets")
+    prompts = repeat_and_sort_requests(filtered_datasets,
+                                       repeat_count=args.repeat_count,
+                                       sort=args.sort)
+
+    print("------warm up------")
+    test_prefix(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+    print("------start generating------")
+    test_prefix(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description=
+        'Benchmark the performance with or without automatic prefix caching.')
+    parser.add_argument('--model',
+                        type=str,
+                        default='baichuan-inc/Baichuan2-13B-Chat')
+    parser.add_argument("--dataset-path",
+                        type=str,
+                        default=None,
+                        help="Path to the dataset.")
+    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
+    parser.add_argument('--output-len', type=int, default=10)
+    parser.add_argument('--enable-prefix-caching',
+                        action='store_true',
+                        help='enable prefix caching')
+    parser.add_argument('--use-v2-block-manager',
+                        action='store_true',
+                        help='Use BlockSpaceMangerV2')
+    parser.add_argument('--num-prompts',
+                        type=int,
+                        default=1,
+                        help="Number of the prompts sampled from dataset")
+    parser.add_argument('--repeat-count',
+                        type=int,
+                        default=100,
+                        help='Number of times to repeat each prompt')
+    parser.add_argument('--sort',
+                        action='store_true',
+                        help='Sort prompts by input length')
+    parser.add_argument('--input-length-range',
+                        type=str,
+                        default='128:256',
+                        help='Range of input lengths for sampling prompts,'
+                        'specified as "min:max" (e.g., "128:256").')
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_prioritization.py

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

benchmarks/benchmark_serving.py

+"""Benchmark online serving throughput.
+
+On the server side, run one of the following commands:
+    vLLM OpenAI API server
+    vllm serve <your_model> \
+        --swap-space 16 \
+        --disable-log-requests
+
+    (TGI backend)
+    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
+
+On the client side, run:
+    python benchmarks/benchmark_serving.py \
+        --backend <backend> \
+        --model <your_model> \
+        --dataset-name sharegpt \
+        --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.
+"""
+import argparse
+import asyncio
+import base64
+import io
+import json
+import os
+import random
+import time
+import warnings
+from dataclasses import dataclass
+from datetime import datetime
+from typing import Any, AsyncGenerator, Collection, Dict, List, Optional, Tuple
+
+import numpy as np
+from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
+                                  RequestFuncOutput)
+from datasets import load_dataset
+from PIL.Image import Image
+from tqdm.asyncio import tqdm
+from transformers import PreTrainedTokenizerBase
+
+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
+class BenchmarkMetrics:
+    completed: int
+    total_input: int
+    total_output: int
+    request_throughput: float
+    output_throughput: float
+    total_token_throughput: float
+    mean_ttft_ms: float
+    median_ttft_ms: float
+    std_ttft_ms: float
+    percentiles_ttft_ms: List[Tuple[float, float]]
+    mean_tpot_ms: float
+    median_tpot_ms: float
+    std_tpot_ms: float
+    percentiles_tpot_ms: List[Tuple[float, float]]
+    mean_itl_ms: float
+    median_itl_ms: float
+    std_itl_ms: float
+    percentiles_itl_ms: List[Tuple[float, float]]
+    # E2EL stands for end-to-end latency per request.
+    # It is the time taken on the client side from sending
+    # a request to receiving a complete response.
+    mean_e2el_ms: float
+    median_e2el_ms: float
+    std_e2el_ms: float
+    percentiles_e2el_ms: List[Tuple[float, float]]
+
+
+def sample_sharegpt_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int] = None,
+) -> List[Tuple[str, int, int, None]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # Shuffle the dataset.
+    random.shuffle(dataset)
+
+    # Filter out sequences that are too long or too short
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for i in range(len(dataset)):
+        if len(filtered_dataset) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = dataset[i][1]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+        filtered_dataset.append((prompt, prompt_len, output_len, None))
+
+    return filtered_dataset
+
+
+def sample_sonnet_requests(
+    dataset_path: str,
+    num_requests: int,
+    input_len: int,
+    output_len: int,
+    prefix_len: int,
+    tokenizer: PreTrainedTokenizerBase,
+) -> List[Tuple[str, str, int, int, None]]:
+    assert (
+        input_len > prefix_len
+    ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
+
+    # Load the dataset.
+    with open(dataset_path) as f:
+        poem_lines = f.readlines()
+
+    # Tokenize the poem lines.
+    poem_token_ids = tokenizer(poem_lines).input_ids
+    average_poem_len = sum(
+        len(token_ids) for token_ids in poem_token_ids) / len(poem_token_ids)
+
+    # Base prefix for all requests.
+    base_prompt = "Pick as many lines as you can from these poem lines:\n"
+    base_message = [{
+        "role": "user",
+        "content": base_prompt,
+    }]
+    base_prompt_formatted = tokenizer.apply_chat_template(
+        base_message, add_generation_prompt=True, tokenize=False)
+    base_prompt_offset = len(tokenizer(base_prompt_formatted).input_ids)
+
+    assert (
+        input_len > base_prompt_offset
+    ), f"Please set 'args.sonnet-input-len' higher than {base_prompt_offset}."
+    num_input_lines = round(
+        (input_len - base_prompt_offset) / average_poem_len)
+
+    # First approximately `prefix_len` number of tokens in the
+    # prompt are fixed poem lines.
+    assert (
+        prefix_len > base_prompt_offset
+    ), f"Please set 'args.sonnet-prefix-len' higher than {base_prompt_offset}."
+
+    num_prefix_lines = round(
+        (prefix_len - base_prompt_offset) / average_poem_len)
+    prefix_lines = poem_lines[:num_prefix_lines]
+
+    # Sample the rest of lines per request.
+    sampled_requests: List[Tuple[str, int, int]] = []
+    for _ in range(num_requests):
+        sampled_lines = "".join(
+            prefix_lines +
+            random.sample(poem_lines, num_input_lines - num_prefix_lines))
+
+        prompt = f"{base_prompt}{sampled_lines}"
+        message = [
+            {
+                "role": "user",
+                "content": prompt,
+            },
+        ]
+        prompt_formatted = tokenizer.apply_chat_template(
+            message, add_generation_prompt=True, tokenize=False)
+        prompt_len = len(tokenizer(prompt_formatted).input_ids)
+        sampled_requests.append(
+            (prompt, prompt_formatted, prompt_len, output_len, None))
+
+    return sampled_requests
+
+
+def sample_hf_requests(
+    dataset_path: str,
+    dataset_subset: str,
+    dataset_split: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int] = None,
+) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
+    dataset = load_dataset(dataset_path,
+                           name=dataset_subset,
+                           split=dataset_split,
+                           streaming=True)
+    assert "conversations" in dataset.features, (
+        "HF Dataset must have 'conversations' column.")
+    filtered_dataset = dataset.shuffle().filter(
+        lambda x: len(x["conversations"]) >= 2)
+    sampled_requests: List[Tuple[str, int, int, Dict[str,
+                                                     Collection[str]]]] = []
+    for data in filtered_dataset:
+        if len(sampled_requests) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = data["conversations"][0]["value"]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = data["conversations"][1]["value"]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+
+        if "image" in data and isinstance(data["image"], Image):
+            image: Image = data["image"]
+            image = image.convert("RGB")
+            image_data = io.BytesIO()
+            image.save(image_data, format='JPEG')
+            image_base64 = base64.b64encode(
+                image_data.getvalue()).decode("utf-8")
+            mm_content = {
+                "type": "image_url",
+                "image_url": {
+                    "url": f"data:image/jpeg;base64,{image_base64}"
+                },
+            }
+        else:
+            mm_content = None
+
+        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
+
+    return sampled_requests
+
+
+def sample_random_requests(
+    prefix_len: int,
+    input_len: int,
+    output_len: int,
+    num_prompts: int,
+    range_ratio: float,
+    tokenizer: PreTrainedTokenizerBase,
+) -> List[Tuple[str, int, int]]:
+    prefix_token_ids = np.random.randint(0,
+                                         tokenizer.vocab_size,
+                                         size=prefix_len).tolist()
+
+    input_lens = np.random.randint(
+        int(input_len * range_ratio),
+        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(prefix_token_ids +
+                                  [(offsets[i] + i + j) % tokenizer.vocab_size
+                                   for j in range(input_lens[i])])
+
+        input_requests.append((prompt, int(prefix_len + input_lens[i]),
+                               int(output_lens[i]), None))
+
+    return input_requests
+
+
+async def get_request(
+    input_requests: List[Tuple[str, int, int]],
+    request_rate: float,
+) -> AsyncGenerator[Tuple[str, int, int], None]:
+    input_requests = iter(input_requests)
+    for request in input_requests:
+        yield 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.
+        await asyncio.sleep(interval)
+
+
+def calculate_metrics(
+    input_requests: List[Tuple[str, int, int]],
+    outputs: List[RequestFuncOutput],
+    dur_s: float,
+    tokenizer: PreTrainedTokenizerBase,
+    selected_percentile_metrics: List[str],
+    selected_percentiles: List[float],
+) -> Tuple[BenchmarkMetrics, List[int]]:
+    actual_output_lens: List[int] = []
+    total_input = 0
+    completed = 0
+    itls: List[float] = []
+    tpots: List[float] = []
+    ttfts: List[float] = []
+    e2els: List[float] = []
+    for i in range(len(outputs)):
+        if outputs[i].success:
+            # We use the tokenizer to count the number of output tokens for all
+            # 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)
+            e2els.append(outputs[i].latency)
+            completed += 1
+        else:
+            actual_output_lens.append(0)
+
+    if completed == 0:
+        warnings.warn(
+            "All requests failed. This is likely due to a misconfiguration "
+            "on the benchmark arguments.",
+            stacklevel=2)
+    metrics = BenchmarkMetrics(
+        completed=completed,
+        total_input=total_input,
+        total_output=sum(actual_output_lens),
+        request_throughput=completed / dur_s,
+        output_throughput=sum(actual_output_lens) / dur_s,
+        total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s,
+        mean_ttft_ms=np.mean(ttfts or 0) *
+        1000,  # ttfts is empty if streaming is not supported by backend
+        std_ttft_ms=np.std(ttfts or 0) * 1000,
+        median_ttft_ms=np.median(ttfts or 0) * 1000,
+        percentiles_ttft_ms=[(p, np.percentile(ttfts or 0, p) * 1000)
+                             for p in selected_percentiles],
+        mean_tpot_ms=np.mean(tpots or 0) * 1000,
+        std_tpot_ms=np.std(tpots or 0) * 1000,
+        median_tpot_ms=np.median(tpots or 0) * 1000,
+        percentiles_tpot_ms=[(p, np.percentile(tpots or 0, p) * 1000)
+                             for p in selected_percentiles],
+        mean_itl_ms=np.mean(itls or 0) * 1000,
+        std_itl_ms=np.std(itls or 0) * 1000,
+        median_itl_ms=np.median(itls or 0) * 1000,
+        percentiles_itl_ms=[(p, np.percentile(itls or 0, p) * 1000)
+                            for p in selected_percentiles],
+        mean_e2el_ms=np.median(e2els or 0) * 1000,
+        std_e2el_ms=np.std(e2els or 0) * 1000,
+        median_e2el_ms=np.mean(e2els or 0) * 1000,
+        percentiles_e2el_ms=[(p, np.percentile(e2els or 0, p) * 1000)
+                             for p in selected_percentiles],
+    )
+
+    return metrics, actual_output_lens
+
+
+async def benchmark(
+    backend: str,
+    api_url: str,
+    base_url: str,
+    model_id: str,
+    tokenizer: PreTrainedTokenizerBase,
+    input_requests: List[Tuple[str, int, int]],
+    logprobs: Optional[int],
+    best_of: int,
+    use_beam_search: bool,
+    request_rate: float,
+    disable_tqdm: bool,
+    profile: bool,
+    selected_percentile_metrics: List[str],
+    selected_percentiles: List[str],
+):
+    if backend in ASYNC_REQUEST_FUNCS:
+        request_func = ASYNC_REQUEST_FUNCS[backend]
+    else:
+        raise ValueError(f"Unknown backend: {backend}")
+
+    print("Starting initial single prompt test run...")
+    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
+        input_requests[0])
+    if backend != "openai-chat" and test_mm_content is not None:
+        # multi-modal benchmark is only available on OpenAI Chat backend.
+        raise ValueError(
+            "Multi-modal content is only supported on 'openai-chat' backend.")
+    test_input = RequestFuncInput(
+        model=model_id,
+        prompt=test_prompt,
+        api_url=api_url,
+        prompt_len=test_prompt_len,
+        output_len=test_output_len,
+        logprobs=logprobs,
+        best_of=best_of,
+        use_beam_search=use_beam_search,
+        multi_modal_content=test_mm_content,
+    )
+    test_output = await request_func(request_func_input=test_input)
+    if not test_output.success:
+        raise ValueError(
+            "Initial test run failed - Please make sure benchmark arguments "
+            f"are correctly specified. Error: {test_output.error}")
+    else:
+        print("Initial test run completed. Starting main benchmark run...")
+
+    if profile:
+        print("Starting profiler...")
+        profile_input = RequestFuncInput(
+            model=model_id,
+            prompt=test_prompt,
+            api_url=base_url + "/start_profile",
+            prompt_len=test_prompt_len,
+            output_len=test_output_len,
+            logprobs=logprobs,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+            multi_modal_content=test_mm_content,
+        )
+        profile_output = await request_func(request_func_input=profile_input)
+        if profile_output.success:
+            print("Profiler started")
+
+    print(f"Traffic request rate: {request_rate}")
+
+    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
+
+    benchmark_start_time = time.perf_counter()
+    tasks: List[asyncio.Task] = []
+    async for request in get_request(input_requests, request_rate):
+        prompt, prompt_len, output_len, mm_content = request
+        request_func_input = RequestFuncInput(
+            model=model_id,
+            prompt=prompt,
+            api_url=api_url,
+            prompt_len=prompt_len,
+            output_len=output_len,
+            logprobs=logprobs,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+            multi_modal_content=mm_content,
+        )
+        tasks.append(
+            asyncio.create_task(
+                request_func(request_func_input=request_func_input,
+                             pbar=pbar)))
+    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
+
+    if profile:
+        print("Stopping profiler...")
+        profile_input = RequestFuncInput(
+            model=model_id,
+            prompt=test_prompt,
+            api_url=base_url + "/stop_profile",
+            prompt_len=test_prompt_len,
+            output_len=test_output_len,
+            logprobs=logprobs,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+        )
+        profile_output = await request_func(request_func_input=profile_input)
+        if profile_output.success:
+            print("Profiler stopped")
+
+    if pbar is not None:
+        pbar.close()
+
+    benchmark_duration = time.perf_counter() - benchmark_start_time
+
+    metrics, actual_output_lens = calculate_metrics(
+        input_requests=input_requests,
+        outputs=outputs,
+        dur_s=benchmark_duration,
+        tokenizer=tokenizer,
+        selected_percentile_metrics=selected_percentile_metrics,
+        selected_percentiles=selected_percentiles,
+    )
+
+    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
+    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
+    print("{:<40} {:<10.2f}".format("Benchmark duration (s):",
+                                    benchmark_duration))
+    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
+    print("{:<40} {:<10}".format("Total generated tokens:",
+                                 metrics.total_output))
+    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
+                                    metrics.request_throughput))
+    print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
+                                    metrics.output_throughput))
+    print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
+                                    metrics.total_token_throughput))
+
+    result = {
+        "duration": benchmark_duration,
+        "completed": metrics.completed,
+        "total_input_tokens": metrics.total_input,
+        "total_output_tokens": metrics.total_output,
+        "request_throughput": metrics.request_throughput,
+        "output_throughput": metrics.output_throughput,
+        "total_token_throughput": metrics.total_token_throughput,
+        "input_lens": [output.prompt_len for output in outputs],
+        "output_lens": actual_output_lens,
+        "ttfts": [output.ttft for output in outputs],
+        "itls": [output.itl for output in outputs],
+        "generated_texts": [output.generated_text for output in outputs],
+        "errors": [output.error for output in outputs],
+    }
+
+    def process_one_metric(
+        # E.g., "ttft"
+        metric_attribute_name: str,
+        # E.g., "TTFT"
+        metric_name: str,
+        # E.g., "Time to First Token"
+        metric_header: str,
+    ):
+        # This function print and add statistics of the specified
+        # metric.
+        if metric_attribute_name not in selected_percentile_metrics:
+            return
+        print("{s:{c}^{n}}".format(s=metric_header, n=50, c='-'))
+        print("{:<40} {:<10.2f}".format(
+            f"Mean {metric_name} (ms):",
+            getattr(metrics, f"mean_{metric_attribute_name}_ms")))
+        print("{:<40} {:<10.2f}".format(
+            f"Median {metric_name} (ms):",
+            getattr(metrics, f"median_{metric_attribute_name}_ms")))
+        result[f"mean_{metric_attribute_name}_ms"] = getattr(
+            metrics, f"mean_{metric_attribute_name}_ms")
+        result[f"median_{metric_attribute_name}_ms"] = getattr(
+            metrics, f"median_{metric_attribute_name}_ms")
+        result[f"std_{metric_attribute_name}_ms"] = getattr(
+            metrics, f"std_{metric_attribute_name}_ms")
+        for p, value in getattr(metrics,
+                                f"percentiles_{metric_attribute_name}_ms"):
+            p_word = str(int(p)) if int(p) == p else str(p)
+            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):",
+                                            value))
+            result[f"p{p_word}_{metric_attribute_name}_ms"] = value
+
+    process_one_metric("ttft", "TTFT", "Time to First Token")
+    process_one_metric("tpot", "TPOT",
+                       "Time per Output Token (excl. 1st token)")
+    process_one_metric("itl", "ITL", "Inter-token Latency")
+    process_one_metric("e2el", "E2EL", "End-to-end Latency")
+
+    print("=" * 50)
+
+    return result
+
+
+def main(args: argparse.Namespace):
+    print(args)
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+
+    backend = args.backend
+    model_id = args.model
+    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
+
+    if args.base_url is not None:
+        api_url = f"{args.base_url}{args.endpoint}"
+        base_url = f"{args.base_url}"
+    else:
+        api_url = f"http://{args.host}:{args.port}{args.endpoint}"
+        base_url = f"http://{args.host}:{args.port}"
+
+    tokenizer = get_tokenizer(tokenizer_id,
+                              trust_remote_code=args.trust_remote_code)
+
+    if args.dataset is not None:
+        warnings.warn(
+            "The '--dataset' argument will be deprecated in the next "
+            "release. Please use '--dataset-name' and "
+            "'--dataset-path' in the future runs.",
+            stacklevel=2)
+        input_requests = sample_sharegpt_requests(
+            dataset_path=args.dataset,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            fixed_output_len=args.sharegpt_output_len,
+        )
+
+    elif args.dataset_name == "sharegpt":
+        input_requests = sample_sharegpt_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            fixed_output_len=args.sharegpt_output_len,
+        )
+
+    elif args.dataset_name == "sonnet":
+        # Do not format the prompt, pass to message directly
+        if args.backend == "openai-chat":
+            input_requests = sample_sonnet_requests(
+                dataset_path=args.dataset_path,
+                num_requests=args.num_prompts,
+                input_len=args.sonnet_input_len,
+                output_len=args.sonnet_output_len,
+                prefix_len=args.sonnet_prefix_len,
+                tokenizer=tokenizer,
+            )
+            input_requests = [(prompt, prompt_len, output_len, None)
+                              for prompt, prompt_formatted, prompt_len,
+                              output_len, _ in input_requests]
+        else:
+            assert (
+                tokenizer.chat_template or tokenizer.default_chat_template
+            ), "Tokenizer/model must have chat template for sonnet dataset."
+            input_requests = sample_sonnet_requests(
+                dataset_path=args.dataset_path,
+                num_requests=args.num_prompts,
+                input_len=args.sonnet_input_len,
+                output_len=args.sonnet_output_len,
+                prefix_len=args.sonnet_prefix_len,
+                tokenizer=tokenizer,
+            )
+            input_requests = [(prompt_formatted, prompt_len, output_len, None)
+                              for prompt, prompt_formatted, prompt_len,
+                              output_len, _ in input_requests]
+
+    elif args.dataset_name == "hf":
+        input_requests = sample_hf_requests(
+            dataset_path=args.dataset_path,
+            dataset_subset=args.hf_subset,
+            dataset_split=args.hf_split,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            fixed_output_len=args.hf_output_len,
+        )
+
+    elif args.dataset_name == "random":
+        input_requests = sample_random_requests(
+            prefix_len=args.random_prefix_len,
+            input_len=args.random_input_len,
+            output_len=args.random_output_len,
+            num_prompts=args.num_prompts,
+            range_ratio=args.random_range_ratio,
+            tokenizer=tokenizer,
+        )
+
+    else:
+        raise ValueError(f"Unknown dataset: {args.dataset_name}")
+
+    benchmark_result = asyncio.run(
+        benchmark(
+            backend=backend,
+            api_url=api_url,
+            base_url=base_url,
+            model_id=model_id,
+            tokenizer=tokenizer,
+            input_requests=input_requests,
+            logprobs=args.logprobs,
+            best_of=args.best_of,
+            use_beam_search=args.use_beam_search,
+            request_rate=args.request_rate,
+            disable_tqdm=args.disable_tqdm,
+            profile=args.profile,
+            selected_percentile_metrics=args.percentile_metrics.split(","),
+            selected_percentiles=[
+                float(p) for p in args.metric_percentiles.split(",")
+            ],
+        ))
+
+    # Save config and results to json
+    if args.save_result:
+        result_json: Dict[str, Any] = {}
+
+        # Setup
+        current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
+        result_json["date"] = current_dt
+        result_json["backend"] = backend
+        result_json["model_id"] = model_id
+        result_json["tokenizer_id"] = tokenizer_id
+        result_json["best_of"] = args.best_of
+        result_json["use_beam_search"] = args.use_beam_search
+        result_json["num_prompts"] = args.num_prompts
+
+        # Metadata
+        if args.metadata:
+            for item in args.metadata:
+                if "=" in item:
+                    kvstring = item.split("=")
+                    result_json[kvstring[0].strip()] = kvstring[1].strip()
+                else:
+                    raise ValueError(
+                        "Invalid metadata format. Please use KEY=VALUE format."
+                    )
+
+        # Traffic
+        result_json["request_rate"] = (
+            args.request_rate if args.request_rate < float("inf") else "inf")
+
+        # Merge with benchmark result
+        result_json = {**result_json, **benchmark_result}
+
+        # 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:
+            json.dump(result_json, outfile)
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark the online serving throughput.")
+    parser.add_argument(
+        "--backend",
+        type=str,
+        default="vllm",
+        choices=list(ASYNC_REQUEST_FUNCS.keys()),
+    )
+    parser.add_argument(
+        "--base-url",
+        type=str,
+        default=None,
+        help="Server or API base url if not using http host and port.",
+    )
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument(
+        "--endpoint",
+        type=str,
+        default="/v1/completions",
+        help="API endpoint.",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help="Path to the ShareGPT dataset, will be deprecated in the "
+        "next release.",
+    )
+    parser.add_argument(
+        "--dataset-name",
+        type=str,
+        default="sharegpt",
+        choices=["sharegpt", "sonnet", "random", "hf"],
+        help="Name of the dataset to benchmark on.",
+    )
+    parser.add_argument("--dataset-path",
+                        type=str,
+                        default=None,
+                        help="Path to the sharegpt/sonnet dataset. "
+                        "Or the huggingface dataset ID if using HF dataset.")
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Name of the model.",
+    )
+    parser.add_argument(
+        "--tokenizer",
+        type=str,
+        help=
+        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+    )
+    parser.add_argument(
+        "--best-of",
+        type=int,
+        default=1,
+        help="Generates `best_of` sequences per prompt and "
+        "returns the best one.",
+    )
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        default=1000,
+        help="Number of prompts to process.",
+    )
+    parser.add_argument(
+        "--logprobs",
+        type=int,
+        default=None,
+        help=("Number of logprobs-per-token to compute & return as part of "
+              "the request. If unspecified, then either (1) if beam search "
+              "is disabled, no logprobs are computed & a single dummy "
+              "logprob is returned for each token; or (2) if beam search "
+              "is enabled 1 logprob per token is computed"),
+    )
+    parser.add_argument(
+        "--request-rate",
+        type=float,
+        default=float("inf"),
+        help="Number of requests per second. If this is inf, "
+        "then all the requests are sent at time 0. "
+        "Otherwise, we use Poisson process to synthesize "
+        "the request arrival times.",
+    )
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Trust remote code from huggingface",
+    )
+    parser.add_argument(
+        "--disable-tqdm",
+        action="store_true",
+        help="Specify to disable tqdm progress bar.",
+    )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="Use Torch Profiler. The endpoint must be launched with "
+        "VLLM_TORCH_PROFILER_DIR to enable profiler.",
+    )
+    parser.add_argument(
+        "--save-result",
+        action="store_true",
+        help="Specify to save benchmark results to a json file",
+    )
+    parser.add_argument(
+        "--metadata",
+        metavar="KEY=VALUE",
+        nargs="*",
+        help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) "
+        "for metadata of this run to be saved in the result JSON file "
+        "for record keeping purposes.",
+    )
+    parser.add_argument(
+        "--result-dir",
+        type=str,
+        default=None,
+        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.",
+    )
+    parser.add_argument(
+        "--percentile-metrics",
+        type=str,
+        default="ttft,tpot,itl",
+        help="Comma-seperated list of selected metrics to report percentils. "
+        "This argument specifies the metrics to report percentiles. "
+        "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
+        "Default value is \"ttft,tpot,itl\".")
+    parser.add_argument(
+        "--metric-percentiles",
+        type=str,
+        default="99",
+        help="Comma-seperated list of percentiles for selected metrics. "
+        "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
+        "Default value is \"99\". "
+        "Use \"--percentile-metrics\" to select metrics.",
+    )
+
+    # group for dataset specific arguments
+    sonnet_group = parser.add_argument_group("sonnet dataset options")
+    sonnet_group.add_argument(
+        "--sonnet-input-len",
+        type=int,
+        default=550,
+        help=
+        "Number of input tokens per request, used only for sonnet dataset.",
+    )
+    sonnet_group.add_argument(
+        "--sonnet-output-len",
+        type=int,
+        default=150,
+        help=
+        "Number of output tokens per request, used only for sonnet dataset.",
+    )
+    sonnet_group.add_argument(
+        "--sonnet-prefix-len",
+        type=int,
+        default=200,
+        help=
+        "Number of prefix tokens per request, used only for sonnet dataset.",
+    )
+
+    sharegpt_group = parser.add_argument_group("sharegpt dataset options")
+    sharegpt_group.add_argument(
+        "--sharegpt-output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the output length "
+        "from the ShareGPT dataset.")
+
+    random_group = parser.add_argument_group("random dataset options")
+    random_group.add_argument(
+        "--random-input-len",
+        type=int,
+        default=1024,
+        help=
+        "Number of input tokens per request, used only for random sampling.",
+    )
+    random_group.add_argument(
+        "--random-output-len",
+        type=int,
+        default=128,
+        help=
+        "Number of output tokens per request, used only for random sampling.",
+    )
+    random_group.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=1.0,
+        help="Range of sampled ratio of input/output length, "
+        "used only for random sampling.",
+    )
+    random_group.add_argument(
+        "--random-prefix-len",
+        type=int,
+        default=0,
+        help="Number of fixed prefix tokens before random "
+        " context. The length range of context in a random "
+        " request is [random-prefix-len, "
+        " random-prefix-len + random-prefix-len * random-range-ratio).")
+
+    hf_group = parser.add_argument_group("hf dataset options")
+    hf_group.add_argument("--hf-subset",
+                          type=str,
+                          default=None,
+                          help="Subset of the HF dataset.")
+    hf_group.add_argument("--hf-split",
+                          type=str,
+                          default=None,
+                          help="Split of the HF dataset.")
+    hf_group.add_argument(
+        "--hf-output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the output lengths "
+        "from the sampled HF dataset.",
+    )
+
+    args = parser.parse_args()
+    main(args)
\ No newline at end of file

benchmarks/benchmark_throughput.py

+"""Benchmark offline inference throughput."""
+import argparse
+import json
+import random
+import time
+from typing import List, Optional, Tuple
+
+import numpy as np
+import torch
+import uvloop
+from tqdm import tqdm
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          PreTrainedTokenizerBase)
+
+from vllm.inputs import PromptInputs
+from vllm.engine.arg_utils import DEVICE_OPTIONS, AsyncEngineArgs, EngineArgs
+from vllm.entrypoints.openai.api_server import (
+    build_async_engine_client_from_engine_args)
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.utils import FlexibleArgumentParser, merge_async_iterators
+
+
+def sample_requests(
+    dataset_path: str,
+    num_requests: int,
+    tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int],
+) -> List[Tuple[str, int, int]]:
+    if fixed_output_len is not None and fixed_output_len < 4:
+        raise ValueError("output_len too small")
+
+    # Load the dataset.
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+    # Filter out the conversations with less than 2 turns.
+    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
+    # Only keep the first two turns of each conversation.
+    dataset = [(data["conversations"][0]["value"],
+                data["conversations"][1]["value"]) for data in dataset]
+
+    # Shuffle the dataset.
+    random.shuffle(dataset)
+
+    # Filter out sequences that are too long or too short
+    filtered_dataset: List[Tuple[str, int, int]] = []
+    for i in range(len(dataset)):
+        if len(filtered_dataset) == num_requests:
+            break
+
+        # Tokenize the prompts and completions.
+        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(prompt).input_ids
+        completion = dataset[i][1]
+        completion_token_ids = tokenizer(completion).input_ids
+        prompt_len = len(prompt_token_ids)
+        output_len = len(completion_token_ids
+                         ) if fixed_output_len is None else fixed_output_len
+        if prompt_len < 4 or output_len < 4:
+            # Prune too short sequences.
+            continue
+        if prompt_len > 1024 or prompt_len + output_len > 2048:
+            # Prune too long sequences.
+            continue
+        filtered_dataset.append((prompt, prompt_len, output_len))
+
+    return filtered_dataset
+
+
+def run_vllm(
+    warmup_requests: List[Tuple[str, int, int]],
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: str,
+    quantization: Optional[str],
+    tensor_parallel_size: int,
+    seed: int,
+    n: int,
+    use_beam_search: bool,
+    trust_remote_code: bool,
+    dtype: str,
+    max_model_len: Optional[int],
+    enforce_eager: bool,
+    kv_cache_dtype: str,
+    quantization_param_path: Optional[str],
+    device: str,
+    enable_prefix_caching: bool,
+    enable_chunked_prefill: bool,
+    max_num_batched_tokens: int,
+    distributed_executor_backend: Optional[str],
+    gpu_memory_utilization: float = 0.9,
+    num_scheduler_steps: int = 1,
+    use_v2_block_manager: bool = False,
+    download_dir: Optional[str] = None,
+    load_format: str = EngineArgs.load_format,
+    disable_async_output_proc: bool = False,
+    use_new_beam_search_impl: bool = False,
+) -> float:
+    from vllm import LLM, SamplingParams
+    llm = LLM(
+        model=model,
+        tokenizer=tokenizer,
+        quantization=quantization,
+        tensor_parallel_size=tensor_parallel_size,
+        seed=seed,
+        trust_remote_code=trust_remote_code,
+        dtype=dtype,
+        max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_memory_utilization,
+        enforce_eager=enforce_eager,
+        kv_cache_dtype=kv_cache_dtype,
+        quantization_param_path=quantization_param_path,
+        device=device,
+        enable_prefix_caching=enable_prefix_caching,
+        download_dir=download_dir,
+        enable_chunked_prefill=enable_chunked_prefill,
+        max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
+        load_format=load_format,
+        num_scheduler_steps=num_scheduler_steps,
+        use_v2_block_manager=use_v2_block_manager,
+        disable_async_output_proc=disable_async_output_proc,
+    )
+
+    # Add the requests to the engine.
+    prompts: List[str] = []
+    sampling_params: List[SamplingParams] = []
+    for prompt, _, output_len in requests:
+        prompts.append(prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=0.0 if use_beam_search else 1.0,
+                top_p=1.0,
+                use_beam_search=use_beam_search,
+                ignore_eos=True,
+                max_tokens=output_len,
+            ))
+
+    # warmup
+    warmup_prompts = []
+    warmup_sampling_params = []
+    for prompt, _, output_len in warmup_requests:
+        warmup_prompts.append(prompt)
+        warmup_sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=0.0 if use_beam_search else 1.0,
+                top_p=1.0,
+                use_beam_search=use_beam_search,
+                ignore_eos=True,
+                max_tokens=output_len,
+            ))
+        
+    print("Warming up...")
+    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
+        llm.generate(warmup_prompts, warmup_sampling_params, use_tqdm=True)
+    
+    # dummy_prompt_token_ids = np.random.randint(10000,
+    #                                            size=(args.num_prompts,
+    #                                                  args.input_len))
+    # dummy_inputs: List[PromptInputs] = [{
+    #     "prompt_token_ids": batch
+    # } for batch in dummy_prompt_token_ids.tolist()]
+
+    # def run_to_completion():
+    #     llm.generate(dummy_inputs,
+    #                     sampling_params=sampling_params,
+    #                     use_tqdm=False)
+
+    # print("Warming up...")
+    # for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
+    #     run_to_completion()
+
+    if not use_new_beam_search_impl:
+        start = time.perf_counter()
+        llm.generate(prompts, sampling_params, use_tqdm=True)
+        end = time.perf_counter()
+    else:
+        assert use_beam_search
+        prompts = [prompt for prompt, _, _ in requests]
+        # output_len should be the same for all requests.
+        output_len = requests[0][2]
+        for prompt, input_len, _output_len in requests:
+            assert _output_len == output_len
+        start = time.perf_counter()
+        llm.beam_search(prompts,
+                        beam_width=n,
+                        max_tokens=output_len,
+                        ignore_eos=True)
+        end = time.perf_counter()
+    return end - start
+
+
+async def run_vllm_async(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: str,
+    quantization: Optional[str],
+    tensor_parallel_size: int,
+    seed: int,
+    n: int,
+    use_beam_search: bool,
+    trust_remote_code: bool,
+    dtype: str,
+    max_model_len: Optional[int],
+    enforce_eager: bool,
+    kv_cache_dtype: str,
+    quantization_param_path: Optional[str],
+    device: str,
+    enable_prefix_caching: bool,
+    enable_chunked_prefill: bool,
+    max_num_batched_tokens: int,
+    distributed_executor_backend: Optional[str],
+    gpu_memory_utilization: float = 0.9,
+    num_scheduler_steps: int = 1,
+    use_v2_block_manager: bool = False,
+    download_dir: Optional[str] = None,
+    load_format: str = EngineArgs.load_format,
+    disable_async_output_proc: bool = False,
+    disable_frontend_multiprocessing: bool = False,
+) -> float:
+    from vllm import SamplingParams
+    engine_args = AsyncEngineArgs(
+        model=model,
+        tokenizer=tokenizer,
+        quantization=quantization,
+        tensor_parallel_size=tensor_parallel_size,
+        seed=seed,
+        trust_remote_code=trust_remote_code,
+        dtype=dtype,
+        max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_memory_utilization,
+        enforce_eager=enforce_eager,
+        kv_cache_dtype=kv_cache_dtype,
+        quantization_param_path=quantization_param_path,
+        device=device,
+        enable_prefix_caching=enable_prefix_caching,
+        download_dir=download_dir,
+        enable_chunked_prefill=enable_chunked_prefill,
+        max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
+        load_format=load_format,
+        num_scheduler_steps=num_scheduler_steps,
+        use_v2_block_manager=use_v2_block_manager,
+        disable_async_output_proc=disable_async_output_proc,
+        worker_use_ray=False,
+        disable_log_requests=True,
+    )
+
+    async with build_async_engine_client_from_engine_args(
+            engine_args, disable_frontend_multiprocessing) as llm:
+
+        # Add the requests to the engine.
+        prompts: List[str] = []
+        sampling_params: List[SamplingParams] = []
+        for prompt, _, output_len in requests:
+            prompts.append(prompt)
+            sampling_params.append(
+                SamplingParams(
+                    n=n,
+                    temperature=0.0 if use_beam_search else 1.0,
+                    top_p=1.0,
+                    use_beam_search=use_beam_search,
+                    ignore_eos=True,
+                    max_tokens=output_len,
+                ))
+
+        generators = []
+        start = time.perf_counter()
+        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
+            generator = llm.generate(prompt, sp, request_id=f"test{i}")
+            generators.append(generator)
+        all_gens = merge_async_iterators(*generators)
+        async for i, res in all_gens:
+            pass
+        end = time.perf_counter()
+        return end - start
+
+
+def run_hf(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tokenizer: PreTrainedTokenizerBase,
+    n: int,
+    use_beam_search: bool,
+    max_batch_size: int,
+    trust_remote_code: bool,
+) -> float:
+    assert not use_beam_search
+    llm = AutoModelForCausalLM.from_pretrained(
+        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
+    if llm.config.model_type == "llama":
+        # To enable padding in the HF backend.
+        tokenizer.pad_token = tokenizer.eos_token
+    llm = llm.cuda()
+
+    pbar = tqdm(total=len(requests))
+    start = time.perf_counter()
+    batch: List[str] = []
+    max_prompt_len = 0
+    max_output_len = 0
+    for i in range(len(requests)):
+        prompt, prompt_len, output_len = requests[i]
+        # Add the prompt to the batch.
+        batch.append(prompt)
+        max_prompt_len = max(max_prompt_len, prompt_len)
+        max_output_len = max(max_output_len, output_len)
+        if len(batch) < max_batch_size and i != len(requests) - 1:
+            # Check if we can add more requests to the batch.
+            _, next_prompt_len, next_output_len = requests[i + 1]
+            if (max(max_prompt_len, next_prompt_len) +
+                    max(max_output_len, next_output_len)) <= 2048:
+                # We can add more requests to the batch.
+                continue
+
+        # Generate the sequences.
+        input_ids = tokenizer(batch, return_tensors="pt",
+                              padding=True).input_ids
+        llm_outputs = llm.generate(
+            input_ids=input_ids.cuda(),
+            do_sample=not use_beam_search,
+            num_return_sequences=n,
+            temperature=1.0,
+            top_p=1.0,
+            use_cache=True,
+            max_new_tokens=max_output_len,
+        )
+        # Include the decoding time.
+        tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        pbar.update(len(batch))
+
+        # Clear the batch.
+        batch = []
+        max_prompt_len = 0
+        max_output_len = 0
+    end = time.perf_counter()
+    return end - start
+
+
+def run_mii(
+    requests: List[Tuple[str, int, int]],
+    model: str,
+    tensor_parallel_size: int,
+    output_len: int,
+) -> float:
+    from mii import client, serve
+    llm = serve(model, tensor_parallel=tensor_parallel_size)
+    prompts = [prompt for prompt, _, _ in requests]
+
+    start = time.perf_counter()
+    llm.generate(prompts, max_new_tokens=output_len)
+    end = time.perf_counter()
+    client = client(model)
+    client.terminate_server()
+    return end - start
+
+
+def main(args: argparse.Namespace):
+    print(args)
+    random.seed(args.seed)
+
+    # Sample the requests.
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.tokenizer, trust_remote_code=args.trust_remote_code)
+    warmup_prompt = "hi" * 10
+    warmup_requests = [(warmup_prompt, 10, 10)
+                for _ in range(1)]
+    if args.dataset is None:
+        # Synthesize a prompt with the given input length.
+        prompt = "hi" * (args.input_len - 1)
+        requests = [(prompt, args.input_len, args.output_len)
+                    for _ in range(args.num_prompts)]
+    else:
+        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
+                                   args.output_len)
+
+    if args.backend == "vllm":
+        if args.async_engine:
+            run_args = [
+                requests, args.model, args.tokenizer, args.quantization,
+                args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
+                args.trust_remote_code, args.dtype, args.max_model_len,
+                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.distributed_executor_backend,
+                args.gpu_memory_utilization, args.num_scheduler_steps,
+                args.use_v2_block_manager, args.download_dir, args.load_format,
+                args.disable_async_output_proc
+            ]
+        else:
+            run_args = [
+                warmup_requests, requests, args.model, args.tokenizer, args.quantization,
+                args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
+                args.trust_remote_code, args.dtype, args.max_model_len,
+                args.enforce_eager, args.kv_cache_dtype,
+                args.quantization_param_path, args.device,
+                args.enable_prefix_caching, args.enable_chunked_prefill,
+                args.max_num_batched_tokens, args.distributed_executor_backend,
+                args.gpu_memory_utilization, args.num_scheduler_steps,
+                args.use_v2_block_manager, args.download_dir, args.load_format,
+                args.disable_async_output_proc
+            ]
+
+        if args.async_engine:
+            run_args.append(args.disable_frontend_multiprocessing)
+            elapsed_time = uvloop.run(run_vllm_async(*run_args))
+        else:
+            elapsed_time = run_vllm(*run_args, args.use_new_beam_search_impl)
+    elif args.backend == "hf":
+        assert args.tensor_parallel_size == 1
+        elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
+                              args.use_beam_search, args.hf_max_batch_size,
+                              args.trust_remote_code)
+    elif args.backend == "mii":
+        elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
+                               args.output_len)
+    else:
+        raise ValueError(f"Unknown backend: {args.backend}")
+    total_num_tokens = sum(prompt_len + output_len
+                           for _, prompt_len, output_len in requests)
+    
+    if args.dataset is None:
+        total_out_tokens = args.output_len * args.num_prompts
+    else:
+        total_out_tokens = sum(output_len for _, _, output_len in requests) 
+    print(f"Latency: {elapsed_time:.2f} s")
+    print(f"All Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+          f"{total_num_tokens / elapsed_time:.2f} tokens/s")
+    print(f"Generate Throughput: {total_out_tokens / elapsed_time:.2f} tokens/s")
+
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "elapsed_time": elapsed_time,
+            "num_requests": len(requests),
+            "total_num_tokens": total_num_tokens,
+            "requests_per_second": len(requests) / elapsed_time,
+            "tokens_per_second": total_num_tokens / elapsed_time,
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
+    parser.add_argument("--backend",
+                        type=str,
+                        choices=["vllm", "hf", "mii"],
+                        default="vllm")
+    parser.add_argument("--dataset",
+                        type=str,
+                        default=None,
+                        help="Path to the dataset.")
+    parser.add_argument("--input-len",
+                        type=int,
+                        default=None,
+                        help="Input prompt length for each request")
+    parser.add_argument("--output-len",
+                        type=int,
+                        default=None,
+                        help="Output length for each request. Overrides the "
+                        "output length from the dataset.")
+    parser.add_argument("--model", type=str, default="facebook/opt-125m")
+    parser.add_argument("--tokenizer", type=str, default=None)
+    parser.add_argument('--quantization',
+                        '-q',
+                        choices=[*QUANTIZATION_METHODS, None],
+                        default=None)
+    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
+    parser.add_argument("--n",
+                        type=int,
+                        default=1,
+                        help="Number of generated sequences per prompt.")
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument('--num-iters-warmup',
+                        type=int,
+                        default=1,
+                        help='Number of iterations to run for warmup.')
+    parser.add_argument("--use-new-beam-search-impl", action="store_true")
+    parser.add_argument("--num-prompts",
+                        type=int,
+                        default=1000,
+                        help="Number of prompts to process.")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--hf-max-batch-size",
+                        type=int,
+                        default=None,
+                        help="Maximum batch size for HF backend.")
+    parser.add_argument('--trust-remote-code',
+                        action='store_true',
+                        help='trust remote code from huggingface')
+    parser.add_argument(
+        '--max-model-len',
+        type=int,
+        default=None,
+        help='Maximum length of a sequence (including prompt and output). '
+        'If None, will be derived from the model.')
+    parser.add_argument(
+        '--dtype',
+        type=str,
+        default='auto',
+        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
+        help='data type for model weights and activations. '
+        'The "auto" option will use FP16 precision '
+        'for FP32 and FP16 models, and BF16 precision '
+        'for BF16 models.')
+    parser.add_argument('--gpu-memory-utilization',
+                        type=float,
+                        default=0.9,
+                        help='the fraction of GPU memory to be used for '
+                        'the model executor, which can range from 0 to 1.'
+                        'If unspecified, will use the default value of 0.9.')
+    parser.add_argument("--enforce-eager",
+                        action="store_true",
+                        help="enforce eager execution")
+    parser.add_argument(
+        '--kv-cache-dtype',
+        type=str,
+        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
+        default="auto",
+        help='Data type for kv cache storage. If "auto", will use model '
+        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
+        'ROCm (hcu) supports fp8 (=fp8_e4m3)')
+    parser.add_argument(
+        '--quantization-param-path',
+        type=str,
+        default=None,
+        help='Path to the JSON file containing the KV cache scaling factors. '
+        'This should generally be supplied, when KV cache dtype is FP8. '
+        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
+        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
+        'cuda version greater than 11.8. On ROCm (hcu), FP8_E4M3 is '
+        'instead supported for common inference criteria.')
+    parser.add_argument("--device",
+                        type=str,
+                        default="auto",
+                        choices=DEVICE_OPTIONS,
+                        help='device type for vLLM execution')
+    parser.add_argument(
+        "--num-scheduler-steps",
+        type=int,
+        default=1,
+        help="Maximum number of forward steps per scheduler call.")
+    parser.add_argument("--use-v2-block-manager",
+                        action='store_true',
+                        help="Enable block manager v2.")
+    parser.add_argument(
+        "--enable-prefix-caching",
+        action='store_true',
+        help="Enable automatic prefix caching for vLLM backend.")
+    parser.add_argument("--enable-chunked-prefill",
+                        action='store_true',
+                        help="enable chunked prefill for vLLM backend.")
+    parser.add_argument('--max-num-batched-tokens',
+                        type=int,
+                        default=None,
+                        help='maximum number of batched tokens per '
+                        'iteration')
+    parser.add_argument('--download-dir',
+                        type=str,
+                        default=None,
+                        help='directory to download and load the weights, '
+                        'default to the default cache dir of huggingface')
+    parser.add_argument(
+        '--output-json',
+        type=str,
+        default=None,
+        help='Path to save the throughput results in JSON format.')
+    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')
+    parser.add_argument(
+        "--disable-async-output-proc",
+        action='store_true',
+        default=False,
+        help="Disable async output processor for vLLM backend.")
+    parser.add_argument("--async-engine",
+                        action='store_true',
+                        default=False,
+                        help="Use vLLM async engine rather than LLM class.")
+    parser.add_argument("--disable-frontend-multiprocessing",
+                        action='store_true',
+                        default=False,
+                        help="Disable decoupled async engine frontend.")
+    args = parser.parse_args()
+    if args.tokenizer is None:
+        args.tokenizer = args.model
+    if args.dataset is None:
+        assert args.input_len is not None
+        assert args.output_len is not None
+    else:
+        assert args.input_len is None
+
+    if args.backend == "vllm":
+        if args.hf_max_batch_size is not None:
+            raise ValueError("HF max batch size is only for HF backend.")
+    elif args.backend == "hf":
+        if args.hf_max_batch_size is None:
+            raise ValueError("HF max batch size is required for HF backend.")
+        if args.quantization is not None:
+            raise ValueError("Quantization is only for vLLM backend.")
+    elif args.backend == "mii":
+        if args.dtype != "auto":
+            raise ValueError("dtype must be auto for MII backend.")
+        if args.n != 1:
+            raise ValueError("n must be 1 for MII backend.")
+        if args.use_beam_search:
+            raise ValueError("Beam search is not supported for MII backend.")
+        if args.quantization is not None:
+            raise ValueError("Quantization is only for vLLM backend.")
+        if args.hf_max_batch_size is not None:
+            raise ValueError("HF max batch size is only for HF backend.")
+        if args.tokenizer != args.model:
+            raise ValueError("Tokenizer must be the same as the model for MII "
+                             "backend.")
+    main(args)
\ No newline at end of file

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

+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())
+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")
+
+
+# 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
+    a, b = make_rand_tensors(torch.int8, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+    azp = torch.zeros((m, ), device="cuda", dtype=torch.int32)
+    azp_adj = torch.zeros((n, ), device="cuda", dtype=torch.int32)
+
+    timers = []
+    # pytorch impl - 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 with azp per-tensor
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj))
+
+    # cutlass with azp per-tensor + bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_bias",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj, None, bias))
+
+    # cutlass with azp per-token
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj, azp))
+
+    # cutlass with azp per-token + bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias",
+                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
+                 torch.bfloat16, azp_adj, azp, bias))
+
+    return timers
+
+
+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)
+    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+
+    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: fp16 output
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16))
+
+    # cutlass impl: bf16 output, with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm_bias",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
+                 bias))
+
+    # cutlass impl: fp16 output, with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm_bias",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16,
+                 bias.to(dtype=torch.float16)))
+
+    return timers
+
+
+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

+# 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

+import os
+import sys
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+
+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'
+
+
+def torch_mult(
+        input: torch.Tensor,  #  [..., in_features]
+        weights: torch.Tensor,
+        scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+) -> torch.Tensor:
+    output = F.linear(input, weights)
+    return output
+
+
+def dequant_out_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    if bias is None:
+        output = F.linear(input, weights, bias)
+        orig_shape = output.shape
+        flattened_output = output.view(-1, output.size(-1))
+        f_scales = scales.view(-1, scales.shape[0])
+        b_scales = f_scales.expand(flattened_output.shape[0], -1)
+        flattened_output *= b_scales
+        return flattened_output.view(orig_shape)
+    else:
+        b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+            -1, weights.shape[1])
+        weights *= b_scales
+        return F.linear(input, weights, bias)
+
+
+def dequant_weight_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+        -1, weights.shape[1])
+    weights *= b_scales
+    return F.linear(input, weights, bias)
+
+
+def dequant_no_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    return F.linear(input, weights, bias)
+
+
+# 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:
+
+    n = int(parts.sum().item())
+
+    device = torch.device('cuda:0')
+
+    code_range = (1 << bits) // 2
+    ingroups = 8
+
+    codes = torch.randint(-code_range,
+                          code_range,
+                          size=(n, k // ingroups, nbooks),
+                          dtype=get_int_dtype(bits),
+                          device=device)
+
+    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+                            dtype=torch.float16,
+                            device=device)
+
+    count = 0
+    for index in range(16):
+        for i in range(8):
+            for book in range(nbooks):
+                codebooks[book, index, 0, i] = count * (10**book)
+            count += 1
+
+    print("codes shape", codes.shape)
+
+    for i in range(16):
+        for book in range(nbooks):
+            codes[0, i, book] = i
+            codes[0, -i, book] = i
+
+    weights = dequantize_weight(codes, codebooks, None)
+    weights2 = ops.aqlm_dequant(codes, codebooks, parts)
+
+    print("weights shape:", weights.shape)
+    print("weights2 shape:", weights2.shape)
+
+    print("weights are:", weights)
+    print("weights2 are:", weights2)
+
+    print("first 128 weights are", weights[0, 0:128].to(torch.int32))
+    print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32))
+
+    print("last 128 weights are", weights[0, -128:])
+    print("last 128 weights2 are:", weights2[0, -128:])
+
+
+def main():
+
+    parser = FlexibleArgumentParser(description="Benchmark aqlm performance.")
+
+    # Add arguments
+    parser.add_argument("--nbooks",
+                        type=int,
+                        default=1,
+                        help="Number of codebooks (default: 1)")
+    parser.add_argument("--bits",
+                        type=int,
+                        default=16,
+                        help="Number of bits per code element (default: 16)")
+    parser.add_argument(
+        "--test",
+        type=bool,
+        default=False,
+        help="Run the decompression/dequant tester rather than benchmarking "
+        "(default: False)")
+
+    # Parse the arguments
+    args = parser.parse_args()
+
+    # Extract values
+    nbooks = args.nbooks
+    bits = args.bits
+
+    if args.test:
+        dequant_test(4096, torch.tensor((4096, )), nbooks, bits)
+        return
+
+    # Otherwise, benchmark.
+    methods = [
+        ops.aqlm_gemm,
+        dequant_out_scale,
+        generic_dequantize_gemm,
+        optimized_dequantize_gemm,
+        dequant_weight_scale,
+        torch_mult,
+        dequant_no_scale,
+    ]
+
+    filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv"
+    print(f"writing benchmarks to file {filename}")
+    with open(filename, "w") as f:
+        sys.stdout = f
+
+        print('m | k | n | n parts', end='')
+        for method in methods:
+            print(f" | {method.__name__.replace('_', ' ')} (µs)", end='')
+        print('')
+
+        # These are reasonable prefill sizes.
+        ksandpartions = ((4096, (4096, 4096, 4096)), (4096, (4096, )),
+                         (4096, (11008, 11008)), (11008, (4096, )))
+
+        # reasonable ranges for m.
+        for m in [
+                1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 48, 52, 56, 64, 96, 112,
+                128, 256, 512, 1024, 1536, 2048, 3072, 4096
+        ]:
+            print(f'{m}', file=sys.__stdout__)
+            for ksp in ksandpartions:
+                run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits,
+                         methods)
+
+        sys.stdout = sys.__stdout__
+
+
+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 :)
+    num_warmup_trials = 1
+    num_trials = 1
+
+    num_calls = 100
+
+    # warmup.
+    for method in methods:
+        for _ in range(num_warmup_trials):
+            run_timing(
+                num_calls=num_calls,
+                m=m,
+                k=k,
+                parts=parts,
+                nbooks=nbooks,
+                bits=bits,
+                method=method,
+            )
+
+    n = parts.sum().item()
+    print(f'{m} | {k} | {n} | {parts.tolist()}', end='')
+
+    for method in methods:
+        best_time_us = 1e20
+        for _ in range(num_trials):
+            kernel_dur_ms = run_timing(
+                num_calls=num_calls,
+                m=m,
+                k=k,
+                parts=parts,
+                nbooks=nbooks,
+                bits=bits,
+                method=method,
+            )
+
+            kernel_dur_us = 1000 * kernel_dur_ms
+
+            if kernel_dur_us < best_time_us:
+                best_time_us = kernel_dur_us
+
+        print(f' | {kernel_dur_us:.0f}', end='')
+
+    print('')
+
+
+def run_timing(num_calls: int, m: int, k: int, parts: torch.Tensor,
+               nbooks: int, bits: int, method) -> float:
+
+    n = int(parts.sum().item())
+
+    device = torch.device('cuda:0')
+
+    input = torch.randn((1, m, k), dtype=torch.float16, device=device)
+
+    code_range = (1 << bits) // 2
+    ingroups = 8
+
+    codes = torch.randint(-code_range,
+                          code_range,
+                          size=(n, k // ingroups, nbooks),
+                          dtype=get_int_dtype(bits),
+                          device=device)
+
+    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+                            dtype=torch.float16,
+                            device=device)
+
+    scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device)
+
+    # for comparison to just a pytorch mult.
+    weights = torch.randn((n, k), dtype=torch.float16, device=device)
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    start_event.record()
+
+    if method is torch_mult:
+        for i in range(num_calls):
+            torch_mult(input, weights, scales)
+    else:
+        for i in range(num_calls):
+            method(input, codes, codebooks, scales, parts, None)
+
+    end_event.record()
+    end_event.synchronize()
+
+    dur_ms = start_event.elapsed_time(end_event) / num_calls
+    return dur_ms
+
+
+if __name__ == "__main__":
+    sys.exit(main())

benchmarks/kernels/benchmark_layernorm.py

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

benchmarks/kernels/benchmark_machete.py

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

benchmarks/kernels/benchmark_marlin.py

+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_24 import (
+    GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
+    GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES)
+from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
+    MARLIN_SUPPORTED_GROUP_SIZES, query_marlin_supported_quant_types)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
+    MarlinWorkspace, marlin_quantize)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
+    marlin_24_quantize)
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    gptq_pack, gptq_quantize_weights, sort_weights)
+from vllm.scalar_type import ScalarType
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+
+ACT_ORDER_OPTS = [False, True]
+K_FULL_OPTS = [False, True]
+
+
+def bench_run(results: List[benchmark.Measurement], model: str,
+              act_order: bool, is_k_full: bool, quant_type: ScalarType,
+              group_size: int, size_m: int, size_k: int, size_n: int):
+    label = "Quant Matmul"
+
+    sub_label = ("{}, act={} k_full={}, q={}, g={}, "
+                 "MKN=({}x{}x{})".format(model, act_order, is_k_full,
+                                         str(quant_type), group_size, size_m,
+                                         size_k, size_n))
+
+    print(f"Testing: {sub_label}")
+
+    a = torch.randn(size_m, size_k).to(torch.half).cuda()
+    b = torch.rand(size_k, size_n).to(torch.half).cuda()
+
+    a_tmp = (torch.zeros(size_m, size_k).to(torch.half).cuda())
+
+    # Marlin quant
+    (
+        marlin_w_ref,
+        marlin_q_w,
+        marlin_s,
+        marlin_g_idx,
+        marlin_sort_indices,
+        marlin_rand_perm,
+    ) = marlin_quantize(b, quant_type, group_size, act_order)
+
+    # Marlin_24 quant
+    (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta,
+     marlin_24_s) = marlin_24_quantize(b, quant_type, group_size)
+
+    marlin_zp = torch.empty(0, dtype=torch.int, device=b.device)
+
+    # GPTQ quant
+    (w_ref, q_w, s, g_idx,
+     rand_perm) = gptq_quantize_weights(b, quant_type, group_size, act_order)
+    q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
+
+    # For act_order, sort the "weights" and "g_idx"
+    # so that group ids are increasing
+    repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
+    if act_order:
+        (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
+
+    # Prepare
+    marlin_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_MIN_THREAD_N,
+                                       GPTQ_MARLIN_MAX_PARALLEL)
+
+    marlin_24_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_24_MIN_THREAD_N,
+                                          GPTQ_MARLIN_24_MAX_PARALLEL)
+    marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)
+
+    globals = {
+        # Gen params
+        "quant_type": quant_type,
+        "group_size": group_size,
+        "size_m": size_m,
+        "size_n": size_n,
+        "size_k": size_k,
+        "a": a,
+        "a_tmp": a_tmp,
+        # Marlin params
+        "marlin_w_ref": marlin_w_ref,
+        "marlin_q_w": marlin_q_w,
+        "marlin_s": marlin_s,
+        "marlin_zp": marlin_zp,
+        "marlin_g_idx": marlin_g_idx,
+        "marlin_sort_indices": marlin_sort_indices,
+        "marlin_rand_perm": marlin_rand_perm,
+        "marlin_workspace": marlin_workspace,
+        "is_k_full": is_k_full,
+        # Marlin_24 params
+        "marlin_24_w_ref": marlin_24_w_ref,
+        "marlin_24_q_w_comp": marlin_24_q_w_comp,
+        "marlin_24_meta": marlin_24_meta,
+        "marlin_24_s": marlin_24_s,
+        "marlin_24_workspace": marlin_24_workspace,
+        # GPTQ params
+        "q_w_gptq": q_w_gptq,
+        "repack_sort_indices": repack_sort_indices,
+        # Kernels
+        "gptq_marlin_gemm": ops.gptq_marlin_gemm,
+        "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
+        "gptq_marlin_repack": ops.gptq_marlin_repack,
+    }
+
+    min_run_time = 1
+
+    # Warmup pytorch
+    for i in range(5):
+        torch.matmul(a, marlin_w_ref)
+
+    results.append(
+        benchmark.Timer(
+            stmt="torch.matmul(a, marlin_w_ref)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="pytorch_gemm",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_gemm_fp16",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_gemm_fp32",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    if (quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
+            and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES):
+        results.append(
+            benchmark.Timer(
+                stmt=
+                "output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)",  # noqa: E501
+                globals=globals,
+                label=label,
+                sub_label=sub_label,
+                description="gptq_marlin_24_gemm",
+            ).blocked_autorange(min_run_time=min_run_time))
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_repack",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+
+def main(args):
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    results: List[benchmark.Measurement] = []
+
+    for model in args.models:
+        for layer in WEIGHT_SHAPES[model]:
+            size_k = layer[0]
+            size_n = layer[1]
+
+            if len(args.limit_k) > 0 and size_k not in args.limit_k:
+                continue
+
+            if len(args.limit_n) > 0 and size_n not in args.limit_n:
+                continue
+
+            for act_order in ACT_ORDER_OPTS:
+                if len(args.limit_act_order
+                       ) > 0 and act_order not in args.limit_act_order:
+                    continue
+
+                for is_k_full in K_FULL_OPTS:
+                    if len(args.limit_k_full
+                           ) > 0 and is_k_full not in args.limit_k_full:
+                        continue
+
+                    for quant_type in query_marlin_supported_quant_types(
+                            False):
+                        if len(args.limit_num_bits) > 0 and \
+                            quant_type.size_bits not in args.limit_num_bits:
+                            continue
+
+                        for group_size in MARLIN_SUPPORTED_GROUP_SIZES:
+                            if len(
+                                    args.limit_group_size
+                            ) > 0 and group_size not in args.limit_group_size:
+                                continue
+
+                            # For act_order, the group_size must be less than
+                            # size_k
+                            if act_order and (group_size == size_k
+                                              or group_size == -1):
+                                continue
+
+                            for size_m in args.batch_sizes:
+                                bench_run(results, model, act_order, is_k_full,
+                                          quant_type, group_size, size_m,
+                                          size_k, size_n)
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+# For quick benchmarking use:
+#   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 = FlexibleArgumentParser(
+        description="Benchmark Marlin across specified models/shapes/batches")
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES.keys(),
+    )
+    parser.add_argument("--batch-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-group-size", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-bits", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-act-order", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-k-full", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_moe.py

+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, seed_everything
+
+
+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_w8a8: bool,
+    use_int8_w8a16: bool,
+    num_iters: int = 100,
+) -> float:
+    init_dtype = torch.float16 if use_fp8_w8a8 else dtype
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    if use_int8_w8a16:
+        w1 = torch.randint(-127,
+                           127, (
+                               num_experts,
+                               shard_intermediate_size,
+                               hidden_size,
+                           ),
+                           dtype=torch.int8)
+        w2 = torch.randint(-127,
+                           127, (
+                               num_experts,
+                               hidden_size,
+                               shard_intermediate_size // 2,
+                           ),
+                           dtype=torch.int8)
+    else:
+        w1 = torch.randn(num_experts,
+                         shard_intermediate_size,
+                         hidden_size,
+                         dtype=init_dtype)
+        w2 = torch.randn(num_experts,
+                         hidden_size,
+                         shard_intermediate_size // 2,
+                         dtype=init_dtype)
+    gating_output = torch.randn(num_iters,
+                                num_tokens,
+                                num_experts,
+                                dtype=torch.float32)
+
+    w1_scale = None
+    w2_scale = None
+    a1_scale = None
+    a2_scale = None
+    if use_int8_w8a16:
+        w1_scale = torch.randn((num_experts, 2 * shard_intermediate_size),
+                               dtype=torch.float32)
+        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
+    if use_fp8_w8a8:
+        w1_scale = torch.randn(num_experts, dtype=torch.float32)
+        w2_scale = torch.randn(num_experts, dtype=torch.float32)
+        a1_scale = torch.randn(1, dtype=torch.float32)
+        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_w8a8=use_fp8_w8a8,
+            use_int8_w8a16=use_int8_w8a16,
+            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")
+        seed_everything(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_w8a8: bool,
+        use_int8_w8a16: bool,
+    ) -> Tuple[Dict[str, int], float]:
+        seed_everything(self.seed)
+        dtype_str = get_config_dtype_str(dtype,
+                                         use_int8_w8a16=use_int8_w8a16,
+                                         use_fp8_w8a8=use_fp8_w8a8)
+        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
+        # is the intermediate size after silu_and_mul.
+        op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
+                                    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_w8a8,
+                                       use_int8_w8a16)
+        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_w8a8: bool,
+        use_int8_w8a16: bool,
+        search_space: List[Dict[str, int]],
+    ) -> Dict[str, int]:
+        best_config = None
+        best_time = float("inf")
+        for config in tqdm(search_space):
+            try:
+                kernel_time = benchmark_config(config,
+                                               num_tokens,
+                                               num_experts,
+                                               shard_intermediate_size,
+                                               hidden_size,
+                                               topk,
+                                               dtype,
+                                               use_fp8_w8a8,
+                                               use_int8_w8a16,
+                                               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_w8a8: bool,
+                 use_int8_w8a16: bool) -> None:
+    dtype_str = get_config_dtype_str(dtype,
+                                     use_int8_w8a16=use_int8_w8a16,
+                                     use_fp8_w8a8=use_fp8_w8a8)
+
+    # NOTE(woosuk): The current naming convention uses w2.shape[2], which
+    # is the intermediate size after silu_and_mul.
+    filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
+                                    dtype_str)
+
+    print(f"Writing best config to {filename}...")
+    with open(filename, "w") as f:
+        json.dump(configs, f, indent=4)
+        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
+    elif config.architectures[0] == "JambaForCausalLM":
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    else:
+        # Default: Mixtral.
+        E = config.num_local_experts
+        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_w8a8 = args.dtype == "fp8_w8a8"
+    use_int8_w8a16 = args.dtype == "int8_w8a16"
+
+    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(address=None,
+                 ignore_reinit_error=True,
+                 num_gpus=args.tp_size)
+    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_w8a8, use_int8_w8a16, search_space)
+                     for batch_size in batch_sizes])
+        best_configs = {
+            M: sort_config(config)
+            for M, config in zip(batch_sizes, configs)
+        }
+        save_configs(best_configs, E, shard_intermediate_size, hidden_size,
+                     topk, dtype, use_fp8_w8a8, use_int8_w8a16)
+        end = time.time()
+        print(f"Tuning took {end - start:.2f} seconds")
+    else:
+        outputs = _distribute(
+            "benchmark", [(batch_size, E, shard_intermediate_size, hidden_size,
+                           topk, dtype, use_fp8_w8a8, use_int8_w8a16)
+                          for batch_size in batch_sizes])
+
+        for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
+            print(f"Batch size: {batch_size}, config: {config}")
+            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_w8a8", "int8_w8a16"],
+                        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

+import random
+import time
+from typing import List, Optional
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
+                        create_kv_caches_with_random, seed_everything)
+import vllm.envs as envs
+
+NUM_BLOCKS = 1024
+PARTITION_SIZE = 512
+
+
+@torch.inference_mode()
+def main(
+    version: str,
+    num_seqs: int,
+    seq_len: int,
+    num_query_heads: int,
+    num_kv_heads: int,
+    head_size: int,
+    use_alibi: bool,
+    block_size: int,
+    dtype: torch.dtype,
+    seed: int,
+    do_profile: bool,
+    device: str = "cuda",
+    kv_cache_dtype: Optional[str] = None,
+) -> None:
+    seed_everything(seed)
+
+    scale = float(1.0 / (head_size**0.5))
+    query = torch.empty(num_seqs,
+                        num_query_heads,
+                        head_size,
+                        dtype=dtype,
+                        device=device)
+    query.uniform_(-scale, scale)
+
+    assert num_query_heads % num_kv_heads == 0
+    alibi_slopes = None
+    if use_alibi:
+        alibi_slopes = torch.randn(num_query_heads,
+                                   dtype=torch.float,
+                                   device=device)
+
+    seq_lens = [seq_len for _ in range(num_seqs)]
+    max_seq_len = max(seq_lens)
+    seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device)
+
+    # Create the block tables.
+    max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
+    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_lst.append(block_table)
+
+    block_tables = torch.tensor(block_tables_lst,
+                                dtype=torch.int,
+                                device=device)
+
+    # Create the KV cache.
+    key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
+                                                            block_size,
+                                                            1,
+                                                            num_kv_heads,
+                                                            head_size,
+                                                            kv_cache_dtype,
+                                                            dtype,
+                                                            device=device)
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # Prepare for the paged attention kernel.
+    output = torch.empty_like(query)
+    if version == "v2":
+        num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
+        tmp_output = torch.empty(
+            size=(num_seqs, num_query_heads, num_partitions, head_size),
+            dtype=output.dtype,
+            device=output.device,
+        )
+        exp_sums = torch.empty(
+            size=(num_seqs, num_query_heads, num_partitions),
+            dtype=torch.float32,
+            device=output.device,
+        )
+        max_logits = torch.empty_like(exp_sums)
+
+    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
+        torch.cuda.synchronize()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        start_time = time.perf_counter()
+
+        # Using default kv_scale
+        k_scale = v_scale = 1.0
+
+        for _ in range(num_iters):
+            if version == "v1":
+                if envs.VLLM_USE_OPT_OP:
+                    if envs.VLLM_USE_TC_PAGED_ATTN:
+                        ops.paged_attention_v1_opt_tc(
+                            output,
+                            query,
+                            key_cache,
+                            value_cache,
+                            num_kv_heads,
+                            scale,
+                            block_tables,
+                            seq_lens,
+                            block_size,
+                            max_seq_len,
+                            alibi_slopes,
+                            kv_cache_dtype,
+                            k_scale,
+                            v_scale,
+                        )
+                    else:
+                        ops.paged_attention_v1_opt(
+                            output,
+                            query,
+                            key_cache,
+                            value_cache,
+                            num_kv_heads,
+                            scale,
+                            block_tables,
+                            seq_lens,
+                            block_size,
+                            max_seq_len,
+                            alibi_slopes,
+                            kv_cache_dtype,
+                            k_scale,
+                            v_scale,
+                        )
+                else:
+                    ops.paged_attention_v1(
+                    output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    num_kv_heads,
+                    scale,
+                    block_tables,
+                    seq_lens,
+                    block_size,
+                    max_seq_len,
+                    alibi_slopes,
+                    kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
+            elif version == "v2":
+                if envs.VLLM_USE_OPT_OP:
+                    if envs.VLLM_USE_TC_PAGED_ATTN:
+                        ops.paged_attention_v2_opt_tc(
+                            output,
+                            exp_sums,
+                            max_logits,
+                            tmp_output,
+                            query,
+                            key_cache,
+                            value_cache,
+                            num_kv_heads,
+                            scale,
+                            block_tables,
+                            seq_lens,
+                            block_size,
+                            max_seq_len,
+                            alibi_slopes,
+                            kv_cache_dtype,
+                            k_scale,
+                            v_scale,
+                        )
+                    else:
+                        ops.paged_attention_v2_opt(
+                            output,
+                            exp_sums,
+                            max_logits,
+                            tmp_output,
+                            query,
+                            key_cache,
+                            value_cache,
+                            num_kv_heads,
+                            scale,
+                            block_tables,
+                            seq_lens,
+                            block_size,
+                            max_seq_len,
+                            alibi_slopes,
+                            kv_cache_dtype,
+                            k_scale,
+                            v_scale,
+                        )
+                else:
+                    ops.paged_attention_v2(
+                    output,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    num_kv_heads,
+                    scale,
+                    block_tables,
+                    seq_lens,
+                    block_size,
+                    max_seq_len,
+                    alibi_slopes,
+                    kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
+            else:
+                raise ValueError(f"Invalid version: {version}")
+        torch.cuda.synchronize()
+
+        end_time = time.perf_counter()
+        if profile:
+            torch.cuda.cudart().cudaProfilerStart()
+        return (end_time - start_time) / num_iters
+
+    # Warmup.
+    print("Warming up...")
+    run_benchmark = run_cuda_benchmark
+    run_benchmark(num_iters=3, profile=False)
+
+    # Benchmark.
+    if do_profile:
+        latency = run_benchmark(num_iters=1, profile=True)
+    else:
+        latency = run_benchmark(num_iters=100, profile=False)
+    print(f"Kernel running time: {latency * 1000000:.3f} us")
+
+
+if __name__ == '__main__':
+    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("--num-query-heads", type=int, default=64)
+    parser.add_argument("--num-kv-heads", type=int, default=8)
+    parser.add_argument("--head-size",
+                        type=int,
+                        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+                        default=128)
+    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
+    parser.add_argument("--use-alibi", action="store_true")
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["half", "bfloat16", "float"],
+                        default="half")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--profile", action="store_true")
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"],
+        default="auto",
+        help="Data type for kv cache storage. If 'auto', will use model "
+        "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
+        "ROCm (hcu) supports fp8 (=fp8_e4m3)")
+    args = parser.parse_args()
+    print(args)
+
+    if args.num_query_heads % args.num_kv_heads != 0:
+        raise ValueError("num_query_heads must be divisible by num_kv_heads")
+    main(
+        version=args.version,
+        num_seqs=args.batch_size,
+        seq_len=args.seq_len,
+        num_query_heads=args.num_query_heads,
+        num_kv_heads=args.num_kv_heads,
+        head_size=args.head_size,
+        block_size=args.block_size,
+        use_alibi=args.use_alibi,
+        dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+        seed=args.seed,
+        do_profile=args.profile,
+        kv_cache_dtype=args.kv_cache_dtype,
+    )

benchmarks/kernels/benchmark_quant.py

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

benchmarks/kernels/benchmark_rope.py

+from itertools import accumulate
+from typing import List, Optional
+
+import nvtx
+import torch
+
+from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
+                                                         get_rope)
+from vllm.utils import FlexibleArgumentParser, seed_everything
+
+
+def benchmark_rope_kernels_multi_lora(
+    is_neox_style: bool,
+    batch_size: int,
+    seq_len: int,
+    num_heads: int,
+    head_size: int,
+    rotary_dim: Optional[int],
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+    max_position: int = 8192,
+    base: int = 10000,
+) -> None:
+    seed_everything(seed)
+    torch.set_default_device(device)
+    if rotary_dim is None:
+        rotary_dim = head_size
+    # silulating serving 4 LoRAs
+    scaling_factors = [1, 2, 4, 8]
+    # batched RoPE can take multiple scaling factors
+    batched_rope = get_rope(head_size, rotary_dim, max_position, base,
+                            is_neox_style, {
+                                "type": "linear",
+                                "factor": tuple(scaling_factors)
+                            })
+    # non-batched RoPE takes only one scaling factor, we create multiple
+    # instances to simulate the same behavior
+    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,
+                     {
+                         "type": "linear",
+                         "factor": (scaling_factor, )
+                     }))
+
+    positions = torch.randint(0, max_position, (batch_size, seq_len))
+    query = torch.randn(batch_size,
+                        seq_len,
+                        num_heads * head_size,
+                        dtype=dtype)
+    key = torch.randn_like(query)
+
+    # create query offsets for batched RoPE, we concat multiple kv cache
+    # together and each query needs to find the right kv cache of its type
+    offset_map = torch.tensor(
+        list(
+            accumulate([0] + [
+                max_position * scaling_factor * 2
+                for scaling_factor in scaling_factors[:-1]
+            ])))
+    query_types = torch.randint(0,
+                                len(scaling_factors), (batch_size, seq_len),
+                                device=device)
+    # map query types to offsets
+    query_offsets = offset_map[query_types]
+    # the kernel takes flattened offsets
+    flatten_offsets = query_offsets.flatten()
+
+    # batched queries of the same type together for non-batched RoPE
+    queries = [query[query_types == i] for i in range(len(scaling_factors))]
+    keys = [key[query_types == i] for i in range(len(scaling_factors))]
+    packed_qkr = zip(queries, keys, non_batched_ropes)
+    # synchronize before start timing
+    torch.cuda.synchronize()
+    with nvtx.annotate("non-batched", color="yellow"):
+        for q, k, r in packed_qkr:
+            r.forward(positions, q, k)
+    torch.cuda.synchronize()
+    with nvtx.annotate("batched", color="green"):
+        batched_rope.forward(positions, query, key, flatten_offsets)
+    torch.cuda.synchronize()
+
+
+if __name__ == '__main__':
+    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)
+    parser.add_argument("--seq-len", type=int, default=512)
+    parser.add_argument("--num-heads", type=int, default=8)
+    parser.add_argument("--head-size",
+                        type=int,
+                        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+                        default=128)
+    parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32)
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["bfloat16", "float"],
+                        default="float")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--device",
+                        type=str,
+                        choices=["cuda:0", "cuda:1"],
+                        default="cuda:0")
+    args = parser.parse_args()
+    print(args)
+
+    benchmark_rope_kernels_multi_lora(
+        is_neox_style=args.is_neox_style,
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        num_heads=args.num_heads,
+        head_size=args.head_size,
+        rotary_dim=args.rotary_dim,
+        dtype=getattr(torch, args.dtype),
+        seed=args.seed,
+        device=args.device,
+    )

benchmarks/kernels/benchmark_shapes.py

+WEIGHT_SHAPES = {
+    "ideal": [[4 * 256 * 32, 256 * 32]],
+    "mistralai/Mistral-7B-v0.1/TP1": [
+        [4096, 6144],
+        [4096, 4096],
+        [4096, 28672],
+        [14336, 4096],
+    ],
+    "mistralai/Mistral-7B-v0.1/TP2": [
+        [4096, 3072],
+        [2048, 4096],
+        [4096, 14336],
+        [7168, 4096],
+    ],
+    "mistralai/Mistral-7B-v0.1/TP4": [
+        [4096, 1536],
+        [1024, 4096],
+        [4096, 7168],
+        [3584, 4096],
+    ],
+    "meta-llama/Llama-2-7b-hf/TP1": [
+        [4096, 12288],
+        [4096, 4096],
+        [4096, 22016],
+        [11008, 4096],
+    ],
+    "meta-llama/Llama-2-7b-hf/TP2": [
+        [4096, 6144],
+        [2048, 4096],
+        [4096, 11008],
+        [5504, 4096],
+    ],
+    "meta-llama/Llama-2-7b-hf/TP4": [
+        [4096, 3072],
+        [1024, 4096],
+        [4096, 5504],
+        [2752, 4096],
+    ],
+    "meta-llama/Llama-2-13b-hf/TP1": [
+        [5120, 15360],
+        [5120, 5120],
+        [5120, 27648],
+        [13824, 5120],
+    ],
+    "meta-llama/Llama-2-13b-hf/TP2": [
+        [5120, 7680],
+        [2560, 5120],
+        [5120, 13824],
+        [6912, 5120],
+    ],
+    "meta-llama/Llama-2-13b-hf/TP4": [
+        [5120, 3840],
+        [1280, 5120],
+        [5120, 6912],
+        [3456, 5120],
+    ],
+    "meta-llama/Llama-2-70b-hf/TP1": [
+        [8192, 10240],
+        [8192, 8192],
+        [8192, 57344],
+        [28672, 8192],
+    ],
+    "meta-llama/Llama-2-70b-hf/TP2": [
+        [8192, 5120],
+        [4096, 8192],
+        [8192, 28672],
+        [14336, 8192],
+    ],
+    "meta-llama/Llama-2-70b-hf/TP4": [
+        [8192, 2560],
+        [2048, 8192],
+        [8192, 14336],
+        [7168, 8192],
+    ],
+}

benchmarks/kernels/graph_machete_bench.py

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

benchmarks/kernels/requirements.txt

+pandas
\ No newline at end of file