# Copyright 2020-2026 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import argparse import base64 import json import logging import math import os from collections.abc import Sequence from contextlib import asynccontextmanager from dataclasses import dataclass, field from io import BytesIO from itertools import chain from multiprocessing import Pipe, Process from multiprocessing.connection import Connection # We use CUDA with multiprocessing, so we must use the 'spawn' start method. Otherwise, we will get the following # error: RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use # the 'spawn' start method os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" class WeightSyncWorkerExtension: """ A vLLM worker extension that enables weight synchronization between a client and multiple server workers. This worker uses a `StatelessProcessGroup` to establish communication and a `PyNcclCommunicator` or `ProcessGroupXCCL` to handle efficient GPU-based communication using NCCL. The primary purpose of this class is to receive updated model weights from a client process and distribute them to all worker processes participating in model inference. """ # The following attributes are initialized when `init_communicator` method is called. communicator = None # Communicator for weight updates client_rank = None # Source rank for broadcasting updated weights def init_communicator(self, host: str, port: int, world_size: int, client_device_uuid: str) -> None: """ Initializes the weight update communicator using a stateless process group. This method creates a `StatelessProcessGroup` that allows external training processes to communicate with vLLM workers without interfering with the global torch distributed group. Args: host (`str`): Hostname or IP address of the master node. port (`int`): Port number to be used for communication. world_size (`int`): Total number of participating processes in the update group. client_device_uuid (`str`): UUID of the device of client main process. Used to assert that devices are different from vllm workers devices. """ import torch import torch.distributed.distributed_c10d as c10d from transformers import is_torch_xpu_available from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator from vllm.distributed.parallel_state import get_world_group from vllm.distributed.utils import StatelessProcessGroup from trl.import_utils import is_vllm_ascend_available if is_vllm_ascend_available(): from vllm_ascend.distributed.device_communicators.pyhccl import PyHcclCommunicator as PyNcclCommunicator if self.communicator is not None: raise RuntimeError("Weight update group already initialized. Call close_communicator first.") # TODO: will remove after torch xpu 2.9 support uuid in get_device_properties if torch.cuda.is_available() or ( is_torch_xpu_available() and hasattr(torch.xpu.get_device_properties(self.device), "uuid") ): accelerator_module = torch.xpu if is_torch_xpu_available() else torch.cuda if client_device_uuid == str(accelerator_module.get_device_properties(self.device).uuid): raise RuntimeError( f"Attempting to use the same CUDA device (UUID: {client_device_uuid}) for multiple distinct " "roles/ranks within the same communicator. This setup is unsupported and will likely lead to program " "hangs or incorrect behavior. Ensure that trainer is using different devices than vLLM server." ) # Get the rank of the current worker in the global world group. rank = get_world_group().rank if is_torch_xpu_available(): store = torch.distributed.TCPStore(host_name=host, port=port, world_size=world_size, is_master=(rank == 0)) prefixed_store = c10d.PrefixStore("client2server", store) xccl_options = c10d.ProcessGroupXCCL.Options() pg = c10d.ProcessGroupXCCL( store=prefixed_store, rank=rank, size=world_size, options=xccl_options, ) self.communicator = pg else: # Create a stateless process group to manage communication between training processes and vLLM workers. # Initialize the NCCL-based communicator for weight synchronization. pg = StatelessProcessGroup.create(host=host, port=port, rank=rank, world_size=world_size) self.communicator = PyNcclCommunicator(pg, device=self.device) # The client process that sends updated weights has the highest rank (world_size - 1). self.client_rank = world_size - 1 def update_named_param(self, name: str, dtype: str, shape: Sequence[int]) -> None: """ Receives updated weights from the client process and updates the named parameter in the model. Args: name (`str`): Name of the weight tensor being updated. dtype (`str`): Data type of the weight tensor as a string (e.g., `"torch.float32"`). shape (`Sequence[int]`): Shape of the weight tensor. """ import torch from transformers import is_torch_xpu_available if self.communicator is None: raise RuntimeError("Communicator not initialized. Call `init_communicator` first.") dtype = getattr(torch, dtype.split(".")[-1]) # Allocate memory for the incoming weight tensor on the correct device. weight = torch.empty(shape, dtype=dtype, device=self.device) if is_torch_xpu_available(): # Use XCCL to broadcast the updated weights from the client (src) to all workers. self.communicator.broadcast(weight, root=self.client_rank) self.communicator.barrier() else: # Use NCCL to broadcast the updated weights from the client (src) to all workers. self.communicator.broadcast(weight, src=self.client_rank) self.communicator.group.barrier() # Load the received weights into the model. self.model_runner.model.load_weights(weights=[(name, weight)]) def close_communicator(self) -> None: """ Closes the communicator when weight synchronization is no longer needed. This method deletes the NCCL communicator to release associated resources. """ if self.communicator is not None: del self.communicator self.communicator = None # Ensure attribute is reset to None self.client_rank = None # Ensure attribute is reset to None @dataclass class ScriptArguments: r""" Arguments for the script. Args: model (`str`): Model name or path to load the model from. revision (`str`, *optional*): Revision to use for the model. If not specified, the default branch will be used. tensor_parallel_size (`int`, *optional*, defaults to `1`): Number of tensor parallel workers to use. data_parallel_size (`int`, *optional*, defaults to `1`): Number of data parallel workers to use. For dense models, keep this at 1. Starting from vLLM `0.14.0`, setting this above `1` for dense models is no longer supported/useful and will error out (see vLLM PR #30739). host (`str`, *optional*, defaults to `"0.0.0.0"`): Host address to run the server on. port (`int`, *optional*, defaults to `8000`): Port to run the server on. gpu_memory_utilization (`float`, *optional*, defaults to `0.9`): Ratio (between 0 and 1) of GPU memory to reserve for the model weights, activations, and KV cache on the device dedicated to generation powered by vLLM. Higher values will increase the KV cache size and thus improve the model's throughput. However, if the value is too high, it may cause out-of-memory (OOM) errors during initialization. dtype (`str`, *optional*, defaults to `"auto"`): Data type to use for vLLM generation. If set to `"auto"`, the data type will be automatically determined based on the model configuration. Find the supported values in the vLLM documentation. max_model_len (`int`, *optional*): If set, the `max_model_len` to use for vLLM. This can be useful when running with reduced `vllm_gpu_memory_utilization`, leading to a reduced KV cache size. If not set, vLLM will use the model context size, which might be much larger than the KV cache, leading to inefficiencies. enable_prefix_caching (`bool`, *optional*): Whether to enable prefix caching in vLLM. If set to `True`, ensure that the model and the hardware support this feature. enforce_eager (`bool`, *optional*, defaults to `False`): Whether to enforce eager execution. If set to `True`, we will disable CUDA graph and always execute the model in eager mode. If `False` (default behavior), we will use CUDA graph and eager execution in hybrid. vllm_model_impl (`str`, *optional*, defaults to `"vllm"`): Model implementation to use for vLLM. Must be one of `"transformers"` or `"vllm"`. `"transformers"`: Use the `transformers` backend for model implementation. `"vllm"`: Use the `vllm` library for model implementation. kv_cache_dtype (`str`, *optional*, defaults to `"auto"`): Data type to use for KV cache. If set to `"auto"`, the dtype will default to the model data type. trust_remote_code (`bool`, *optional*, defaults to `False`): Whether to trust remote code when loading models. Set to `True` to allow executing code from model repositories. This is required for some custom models but introduces security risks. log_level (`str`, *optional*, defaults to `"info"`): Log level for uvicorn. Possible choices: `"critical"`, `"error"`, `"warning"`, `"info"`, `"debug"`, `"trace"`. distributed_executor_backend (`str` or `None`, *optional*): Distributed executor backend for vLLM. Set to `"ray"` to distribute tensor parallel workers across multiple nodes via a Ray cluster. Required when `tensor_parallel_size` exceeds the number of local GPUs. If not set, vLLM defaults to the multiproc backend (single-node only). speculative_config (`str`, *optional*): JSON string for vLLM speculative decoding config, forwarded to `LLM(speculative_config=...)`. When unset, speculative decoding is disabled. Example: `'{"method": "qwen3_next_mtp", "num_speculative_tokens": 5}'`. """ model: str = field( metadata={"help": "Model name or path to load the model from."}, ) revision: str | None = field( default=None, metadata={"help": "Revision to use for the model. If not specified, the default branch will be used."}, ) tensor_parallel_size: int = field( default=1, metadata={"help": "Number of tensor parallel workers to use."}, ) data_parallel_size: int = field( default=1, metadata={ "help": "Number of data parallel workers to use. For dense models, keep this at 1. Starting from vLLM " "`0.14.0`, setting this above `1` for dense models is no longer supported/useful and will error out (see " "vLLM PR #30739)." }, ) host: str = field( default="0.0.0.0", metadata={"help": "Host address to run the server on."}, ) port: int = field( default=8000, metadata={"help": "Port to run the server on."}, ) gpu_memory_utilization: float = field( default=0.9, metadata={ "help": "Ratio (between 0 and 1) of GPU memory to reserve for the model weights, activations, and KV " "cache on the device dedicated to generation powered by vLLM. Higher values will increase the KV cache " "size and thus improve the model's throughput. However, if the value is too high, it may cause " "out-of-memory (OOM) errors during initialization." }, ) dtype: str = field( default="auto", metadata={ "help": "Data type to use for vLLM generation. If set to 'auto', the data type will be automatically " "determined based on the model configuration. Find the supported values in the vLLM documentation." }, ) max_model_len: int | None = field( default=None, metadata={ "help": "If set, the `max_model_len` to use for vLLM. This can be useful when running with reduced " "`vllm_gpu_memory_utilization`, leading to a reduced KV cache size. If not set, vLLM will use the model " "context size, which might be much larger than the KV cache, leading to inefficiencies." }, ) enable_prefix_caching: bool | None = field( default=None, metadata={ "help": "Whether to enable prefix caching in vLLM. If set to `True`, ensure that the model and the " "hardware support this feature." }, ) enforce_eager: bool | None = field( default=False, metadata={ "help": "Whether to enforce eager execution. If set to `True`, we will disable CUDA graph and always " "execute the model in eager mode. If `False` (default behavior), we will use CUDA graph and eager " "execution in hybrid." }, ) kv_cache_dtype: str = field( default="auto", metadata={ "help": "Data type to use for KV cache. If set to 'auto', the dtype will default to the model data type." }, ) trust_remote_code: bool = field( default=False, metadata={ "help": "Whether to trust remote code when loading models. Set to True to allow executing code from model " "repositories. This is required for some custom models but introduces security risks." }, ) log_level: str = field( default="info", metadata={ "help": "Log level for uvicorn. Possible choices: 'critical', 'error', 'warning', 'info', 'debug', " "'trace'." }, ) vllm_model_impl: str = field( default="vllm", metadata={ "help": "Model implementation to use for vLLM. Must be one of `transformers` or `vllm`. `transformers`: " "Use the `transformers` backend for model implementation. `vllm`: Use the `vllm` library for " "model implementation." }, ) distributed_executor_backend: str | None = field( default=None, metadata={ "help": "Distributed executor backend for vLLM. When set to 'ray', vLLM uses Ray to distribute tensor " "parallel workers across multiple nodes. Required when tensor_parallel_size exceeds the number of local " "GPUs. If not set, vLLM defaults to the multiproc backend (single-node only)." }, ) speculative_config: str | None = field( default=None, metadata={ "help": "JSON string for vLLM speculative decoding config. " 'Example: \'{"method": "qwen3_next_mtp", "num_speculative_tokens": 5}\'' }, ) def llm_worker( script_args: ScriptArguments, data_parallel_rank: int, master_port: int, connection: Connection ) -> None: from vllm import LLM # Set required environment variables for DP to work with vLLM os.environ["VLLM_DP_RANK"] = str(data_parallel_rank) os.environ["VLLM_DP_RANK_LOCAL"] = str(data_parallel_rank) os.environ["VLLM_DP_SIZE"] = str(script_args.data_parallel_size) os.environ["VLLM_DP_MASTER_PORT"] = str(master_port) llm = LLM( model=script_args.model, revision=script_args.revision, tensor_parallel_size=script_args.tensor_parallel_size, gpu_memory_utilization=script_args.gpu_memory_utilization, enforce_eager=script_args.enforce_eager, dtype=script_args.dtype, # Automatic Prefix Caching caches the KV cache of existing queries, so that a new query can # directly reuse the KV cache if it shares the same prefix with one of the existing queries. # This is particularly useful here because we generate completions from the same prompts. enable_prefix_caching=script_args.enable_prefix_caching, kv_cache_dtype=script_args.kv_cache_dtype, max_model_len=script_args.max_model_len, worker_extension_cls="trl.scripts.vllm_serve.WeightSyncWorkerExtension", trust_remote_code=script_args.trust_remote_code, model_impl=script_args.vllm_model_impl, distributed_executor_backend=script_args.distributed_executor_backend, # Important so temperature scaling/logit tweaking affects the TIS log probs logprobs_mode="processed_logprobs", speculative_config=json.loads(script_args.speculative_config) if script_args.speculative_config else None, ) # Send ready signal to parent process connection.send({"status": "ready"}) while True: # Wait for commands from the parent process try: command = connection.recv() except KeyboardInterrupt: llm.collective_rpc(method="close_communicator") break # Handle commands if command["type"] in ["call", "fire_and_forget"]: method_name = command["method"] args, kwargs = command.get("args", ()), command.get("kwargs", {}) method = getattr(llm, method_name) result = method(*args, **kwargs) if command["type"] == "call": connection.send(result) elif command["type"] == "shutdown": break def chunk_list(lst: list, n: int) -> list[list]: """ Split list `lst` into `n` evenly distributed sublists. Example: ```python >>> chunk_list([1, 2, 3, 4, 5, 6], 2) [[1, 2, 3], [4, 5, 6]] >>> chunk_list([1, 2, 3, 4, 5, 6], 4) [[1, 2], [3, 4], [5], [6]] >>> chunk_list([1, 2, 3, 4, 5, 6], 8) [[1], [2], [3], [4], [5], [6], [], []] ``` """ k, r = divmod(len(lst), n) return [lst[i * k + min(i, r) : (i + 1) * k + min(i + 1, r)] for i in range(n)] def main(script_args: ScriptArguments): import asyncio from transformers import is_vision_available from trl.generation.vllm_generation import extract_logprobs from trl.import_utils import ( is_fastapi_available, is_pydantic_available, is_uvicorn_available, is_vllm_available, ) if not is_fastapi_available(): raise ImportError( "FastAPI is required to run the vLLM serve script. Please install it using `pip install fastapi`." ) if not is_pydantic_available(): raise ImportError( "Pydantic is required to run the vLLM serve script. Please install it using `pip install pydantic`." ) if not is_uvicorn_available(): raise ImportError( "Uvicorn is required to run the vLLM serve script. Please install it using `pip install uvicorn`." ) if not is_vllm_available(): raise ImportError("vLLM is required to run the vLLM serve script. Please install it using `pip install vllm`.") import uvicorn from fastapi import FastAPI from pydantic import BaseModel from vllm import SamplingParams from vllm.sampling_params import StructuredOutputsParams from vllm.utils.network_utils import get_open_port if is_vision_available(): from PIL import Image logger = logging.getLogger(__name__) # Spawn dp workers, and setup pipes for communication master_port = get_open_port() connections = [] processes = [] for data_parallel_rank in range(script_args.data_parallel_size): parent_connection, child_connection = Pipe() process = Process(target=llm_worker, args=(script_args, data_parallel_rank, master_port, child_connection)) process.start() connections.append(parent_connection) processes.append(process) @asynccontextmanager async def lifespan(app: FastAPI): # Wait for all workers to send "ready" ready_connections = set() while len(ready_connections) < script_args.data_parallel_size: for connection in connections: msg = connection.recv() if isinstance(msg, dict) and msg.get("status") == "ready": ready_connections.add(connection) # Start the logprob request batcher background task batcher_task = asyncio.create_task(_logprob_batcher()) yield batcher_task.cancel() # Wait for processes to terminate for process in processes: process.join(timeout=10) # Wait for 10 seconds for the process to terminate if process.is_alive(): logger.warning(f"Process {process} is still alive after 10 seconds, attempting to terminate...") process.terminate() process.join() # ensure process termination after calling terminate() app = FastAPI(lifespan=lifespan) # Define the endpoints for the model server @app.get("/health/") async def health(): """ Health check endpoint to verify that the server is running. """ return {"status": "ok"} @app.get("/get_world_size/") async def get_world_size(): """ Retrieves the world size of the LLM engine, which is `tensor_parallel_size * data_parallel_size`. Returns: `dict`: A dictionary containing the world size. Example response: ```json {"world_size": 8} ``` """ return {"world_size": script_args.tensor_parallel_size * script_args.data_parallel_size} class GenerateRequest(BaseModel): prompts: list[str] | list[list[int]] images: list[list[str] | None] | None = None n: int = 1 repetition_penalty: float = 1.0 temperature: float = 1.0 top_p: float = 1.0 top_k: int = -1 min_p: float = 0.0 max_tokens: int = 16 logprobs: int | None = 0 structured_outputs_regex: str | None = None generation_kwargs: dict = field(default_factory=dict) class GenerateResponse(BaseModel): prompt_ids: list[list[int]] completion_ids: list[list[int]] logprobs: list[list[list[float | None]]] | None logprob_token_ids: list[list[list[int]]] | None @app.post("/generate/", response_model=GenerateResponse) async def generate(request: GenerateRequest): """ Generates completions for the provided prompts. Args: request (`GenerateRequest`): - `prompts` (list of `str` or list of list of `int`): A list of prompts. It accepts either text strings or pre-tokenized token ID lists. When text strings are provided, `images` can optionally be included. - `images` (list of list of `str` or `None`, *optional*): A list of image lists. Each element is a list of base64-encoded images for the corresponding prompt, or `None` if no images for that prompt. - `n` (`int`, *optional*, defaults to `1`): Number of completions to generate for each prompt. - `repetition_penalty` (`float`, *optional*, defaults to `1.0`): Repetition penalty to apply during generation. - `temperature` (`float`, *optional*, defaults to `1.0`): Temperature for sampling. Higher values lead to more random outputs. - `top_p` (`float`, *optional*, defaults to `1.0`): Top-p (nucleus) sampling parameter. It controls the diversity of the generated text. - `top_k` (`int`, *optional*, defaults to `-1`): Top-k sampling parameter. If set to `-1`, it disables top-k sampling. - `min_p` (`float`, *optional*, defaults to `0.0`): Minimum probability threshold for sampling. - `max_tokens` (`int`, *optional*, defaults to `16`): Maximum number of tokens to generate for each completion. - `logprobs` (`int`, *optional*, defaults to `0`): Number of top logprobs to return per token. When 0, only the sampled token's logprob is returned. When N>0, returns up to N+1 logprobs sorted by descending probability, because vLLM always includes the sampled token's logprob (which may fall outside the top-N). - `structured_outputs_regex` (`str`, *optional*): A regex pattern for structured outputs. If provided, the model will only generate tokens that match this regex pattern. - `generation_kwargs` (`dict`, *optional*): Additional generation parameters to pass to the vLLM `SamplingParams`. This can include parameters like `seed`, `frequency_penalty`, etc. If it contains keys that conflict with the other parameters, they will override them. Returns: `GenerateResponse`: - `prompt_ids` (list of list of `int`): A list of lists of token IDs for each input prompt. - `completion_ids` (list of list of `int`): A list of lists of token IDs for each generated completion. - `logprobs` (list of list of list of `float`): Per-token logprobs of shape (num_sequences, seq_len, num_logprobs), sorted by descending probability. - `logprob_token_ids` (list of list of list of `int`): Token IDs corresponding to each logprob, same shape as `logprobs`. Example request (text prompts): ```json {"prompts": ["Hello world", "What is AI?"]} ``` Example request (token IDs): ```json {"prompts": [[101, 102], [201, 202]]} ``` Example response: ```json { "prompt_ids": [[101, 102], [201, 202]], "completion_ids": [[103, 104, 105], [203, 204, 205]], "logprobs": [[[-0.1], [-0.2], [-0.3]], [[-0.4], [-0.5], [-0.6]]], "logprob_token_ids": [[[103], [104], [105]], [[203], [204], [205]]] } ``` """ # Build vLLM-compatible prompt inputs is_token_ids = request.prompts and isinstance(request.prompts[0], list) request.images = request.images or [None] * len(request.prompts) prompts = [] for prompt, image_list in zip(request.prompts, request.images, strict=True): row = {"prompt_token_ids": prompt} if is_token_ids else {"prompt": prompt} if image_list is not None: row["multi_modal_data"] = {"image": [Image.open(BytesIO(base64.b64decode(img))) for img in image_list]} prompts.append(row) generation_kwargs = { "n": request.n, "repetition_penalty": request.repetition_penalty, "temperature": request.temperature, "top_p": request.top_p, "top_k": request.top_k, "min_p": request.min_p, "max_tokens": request.max_tokens, "logprobs": request.logprobs, } generation_kwargs.update(request.generation_kwargs) # Structured outputs, if enabled if request.structured_outputs_regex is not None: if generation_kwargs.get("structured_outputs") is not None: logger.warning( "Both `structured_outputs_regex` and `generation_kwargs['structured_outputs']` are set; " "`structured_outputs_regex` takes precedence." ) generation_kwargs["structured_outputs"] = StructuredOutputsParams(regex=request.structured_outputs_regex) elif isinstance(structured_outputs_kwargs := generation_kwargs.get("structured_outputs"), dict): generation_kwargs["structured_outputs"] = StructuredOutputsParams(**structured_outputs_kwargs) sampling_params = SamplingParams(**generation_kwargs) # Evenly distribute prompts across DP ranks chunked_prompts = chunk_list(prompts, script_args.data_parallel_size) # Send the prompts to each worker for connection, prompts in zip(connections, chunked_prompts, strict=True): # When the number of prompts is less than data_parallel_size, some workers will receive empty prompts. # However, vLLM requires that we always send at least one prompt. So we send a placeholder prompt to comply # with vLLM's requirement, and we later ignore the result. if not prompts: prompts = [""] kwargs = {"prompts": prompts, "sampling_params": sampling_params} connection.send({"type": "call", "method": "generate", "kwargs": kwargs}) # Receive results all_outputs = [connection.recv() for connection in connections] # Handle empty prompts (see above) all_outputs = [output for output, prompts in zip(all_outputs, chunked_prompts, strict=True) if prompts] # Flatten and combine all results all_outputs = list(chain.from_iterable(all_outputs)) # from list of list to single list prompt_ids = [output.prompt_token_ids for output in all_outputs] completion_ids = [list(output.token_ids) for outputs in all_outputs for output in outputs.outputs] logprobs, logprob_token_ids = extract_logprobs(all_outputs) return { "prompt_ids": prompt_ids, "completion_ids": completion_ids, "logprobs": logprobs, "logprob_token_ids": logprob_token_ids, } class SequenceLogprobsRequest(BaseModel): sequences: list[list[int]] prompt_lengths: list[int] top_logprobs: int = 100 temperature: float = 1.0 response_format: str = "json" # "json" (legacy) or "binary" (base64 numpy arrays) class SequenceLogprobsResponse(BaseModel): logprobs: list[list[list[float | None]]] | None = None logprob_token_ids: list[list[list[int]]] | None = None # Binary format fields (base64-encoded numpy arrays) logprobs_b64: str | None = None token_ids_b64: str | None = None actual_logprobs_b64: str | None = None actual_token_ids_b64: str | None = None shape: list[int] | None = None # [batch_size, max_completion_len, top_logprobs] completion_lengths: list[int] | None = None # actual completion length per sample def _run_prompt_logprobs(prompts, sampling_params): """Send prompts to DP workers and collect outputs.""" chunked_prompts = chunk_list(prompts, script_args.data_parallel_size) for connection, chunk in zip(connections, chunked_prompts, strict=True): if not chunk: chunk = [{"prompt_token_ids": [0]}] kwargs = {"prompts": chunk, "sampling_params": sampling_params} connection.send({"type": "call", "method": "generate", "kwargs": kwargs}) all_outputs = [connection.recv() for connection in connections] all_outputs = [output for output, chunk in zip(all_outputs, chunked_prompts, strict=True) if chunk] return list(chain.from_iterable(all_outputs)) # ── Request batching for get_sequence_logprobs ── # Collects concurrent requests into batches and dispatches them together so that # all DP workers stay busy. Without this, async endpoint handlers block the event # loop during pipe I/O, serializing requests and leaving DP workers idle. _logprob_queue: asyncio.Queue = asyncio.Queue() # Maximum time (seconds) to wait for more requests before dispatching a batch. _BATCH_WAIT_S = 0.005 # 5ms - short enough to not add much latency when lightly loaded # Maximum number of HTTP requests to collect per batcher cycle _MAX_BATCH_REQUESTS = max(script_args.data_parallel_size * 4, 16) # Maximum total tokens per batch. prompt_logprobs materializes full-vocab logits # during the forward pass, so each worker can safely handle ~1 max-length sequence. # Budget = max_model_len * dp_size gives ~1 sequence per worker at max length. _max_model_len = script_args.max_model_len or 8192 _MAX_BATCH_TOKENS = _max_model_len * script_args.data_parallel_size async def _logprob_batcher(): """Background task that continuously drains the queue, batches requests, and dispatches.""" loop = asyncio.get_running_loop() while True: batch = [] try: # Wait for the first request batch_tokens = 0 item = await _logprob_queue.get() batch.append(item) # Count tokens in this item's sequences for prompt in item[0]: batch_tokens += len(prompt.get("prompt_token_ids", [])) # Collect more requests up to batch limit, timeout, or token budget deadline = loop.time() + _BATCH_WAIT_S while len(batch) < _MAX_BATCH_REQUESTS and batch_tokens < _MAX_BATCH_TOKENS: remaining = deadline - loop.time() if remaining <= 0: break try: item = await asyncio.wait_for(_logprob_queue.get(), timeout=remaining) # Check if adding this item would exceed the token budget item_tokens = sum(len(p.get("prompt_token_ids", [])) for p in item[0]) if batch_tokens + item_tokens > _MAX_BATCH_TOKENS and len(batch) > 0: # Put it back and dispatch current batch await _logprob_queue.put(item) break batch.append(item) batch_tokens += item_tokens except asyncio.TimeoutError: break # batch is a list of (prompts, prompt_lengths, top_logprobs, temperature, response_format, future) # All items in a batch must share the same (top_logprobs, temperature) pair. # Group by those execution parameters to handle mixed requests. groups = {} for prompts, prompt_lengths, top_logprobs, temperature, response_format, future in batch: key = (top_logprobs, temperature) if key not in groups: groups[key] = [] groups[key].append((prompts, prompt_lengths, response_format, future)) for (top_logprobs, temperature), items in groups.items(): # Merge all sequences into a single batch all_prompts = [] all_prompt_lengths = [] offsets = [] # (start_idx, count) per original request for prompts, prompt_lengths, _response_format, _future in items: start = len(all_prompts) all_prompts.extend(prompts) all_prompt_lengths.extend(prompt_lengths) offsets.append((start, len(prompts))) sampling_params = SamplingParams( max_tokens=1, temperature=temperature, prompt_logprobs=top_logprobs, ) # Dispatch to workers in a thread to avoid blocking the event loop try: all_outputs = await loop.run_in_executor( None, _run_prompt_logprobs, all_prompts, sampling_params ) # Split results back to individual requests for (start, count), (_, prompt_lengths, response_format, future) in zip( offsets, items, strict=True ): outputs_slice = all_outputs[start : start + count] if not future.done(): future.set_result((outputs_slice, prompt_lengths, top_logprobs, response_format)) except Exception as e: # Signal error to all waiting requests in this execution-parameter group for *_, future in items: if not future.done(): future.set_exception(e) except Exception as e: # Prevent killing the batcher task — signal error to all unfulfilled futures for *_, future in batch: if not future.done(): future.set_exception(e) def _format_logprob_response(all_outputs, prompt_lengths, top_k, response_format): """Format vLLM outputs into the response dict (runs in any thread).""" import numpy as np batch_size = len(all_outputs) use_binary = response_format == "binary" if use_binary: from starlette.responses import Response comp_lengths = [] for output, prompt_length in zip(all_outputs, prompt_lengths, strict=True): prompt_lps = output.prompt_logprobs if prompt_lps is None: raise ValueError("prompt_logprobs is None.") comp_lengths.append(len(prompt_lps) - prompt_length) max_comp_len = max(comp_lengths) if comp_lengths else 0 # logprobs_arr / token_ids_arr: teacher's sorted top-k logprobs + token ids (for forward KL). # actual_logprobs_arr / actual_token_ids_arr: actual token's teacher logprob (for reverse KL). logprobs_arr = np.full((batch_size, max_comp_len, top_k), float("-inf"), dtype=np.float32) token_ids_arr = np.zeros((batch_size, max_comp_len, top_k), dtype=np.int32) actual_logprobs_arr = np.full((batch_size, max_comp_len, 1), float("-inf"), dtype=np.float32) actual_token_ids_arr = np.zeros((batch_size, max_comp_len, 1), dtype=np.int32) for i, (output, prompt_length) in enumerate(zip(all_outputs, prompt_lengths, strict=True)): prompt_lps = output.prompt_logprobs seq_tokens = output.prompt_token_ids if comp_lengths[i] == 0: continue for pos in range(prompt_length, len(prompt_lps)): lp = prompt_lps[pos] if lp is None: continue t = pos - prompt_length actual_token = seq_tokens[pos] # Actual token's logprob (for reverse KL) if actual_token in lp: val = lp[actual_token].logprob if not math.isnan(val): actual_logprobs_arr[i, t, 0] = val actual_token_ids_arr[i, t, 0] = actual_token # Teacher's top-k logprobs (for forward KL) if top_k == 1: # Fast path: find rank-1 directly instead of sorting for token_id, logprob_obj in lp.items(): if logprob_obj.rank == 1: val = logprob_obj.logprob if not math.isnan(val): logprobs_arr[i, t, 0] = val token_ids_arr[i, t, 0] = token_id break else: sorted_items = sorted(lp.items(), key=lambda x: x[1].rank) for k_idx, (token_id, logprob_obj) in enumerate(sorted_items[:top_k]): val = logprob_obj.logprob if not math.isnan(val): logprobs_arr[i, t, k_idx] = val token_ids_arr[i, t, k_idx] = token_id payload = { "logprobs_b64": base64.b64encode(logprobs_arr.tobytes()).decode("ascii"), "token_ids_b64": base64.b64encode(token_ids_arr.tobytes()).decode("ascii"), "actual_logprobs_b64": base64.b64encode(actual_logprobs_arr.tobytes()).decode("ascii"), "actual_token_ids_b64": base64.b64encode(actual_token_ids_arr.tobytes()).decode("ascii"), "shape": [batch_size, max_comp_len, top_k], "completion_lengths": comp_lengths, } try: import orjson return Response(content=orjson.dumps(payload), media_type="application/json") except ImportError: return payload else: all_logprobs = [] all_token_ids = [] for output, prompt_length in zip(all_outputs, prompt_lengths, strict=True): prompt_lps = output.prompt_logprobs if prompt_lps is None: raise ValueError("prompt_logprobs is None.") seq_logprobs = [] seq_token_ids = [] for pos in range(prompt_length, len(prompt_lps)): lp = prompt_lps[pos] if lp is None: seq_logprobs.append([]) seq_token_ids.append([]) continue sorted_items = sorted(lp.items(), key=lambda x: x[1].rank) seq_token_ids.append([token_id for token_id, _ in sorted_items]) seq_logprobs.append( [None if math.isnan(item.logprob) else item.logprob for _, item in sorted_items] ) all_logprobs.append(seq_logprobs) all_token_ids.append(seq_token_ids) return {"logprobs": all_logprobs, "logprob_token_ids": all_token_ids} @app.post("/get_sequence_logprobs/", response_model=SequenceLogprobsResponse) async def get_sequence_logprobs(request: SequenceLogprobsRequest): """ Computes teacher logprobs for existing token sequences without generating new tokens. Concurrent requests are automatically batched and dispatched together to maximize GPU utilization across DP workers. This avoids the event-loop-blocking problem where synchronous pipe I/O serializes requests despite having multiple DP workers. Args: request (`SequenceLogprobsRequest`): - `sequences` (list of list of `int`): Full token sequences (prompt + completion) per sample. - `prompt_lengths` (list of `int`): Number of prompt tokens per sequence; completion logprobs start after each prompt. - `top_logprobs` (`int`, *optional*, defaults to `100`): Number of top teacher logprobs to return per completion position (sorted by vLLM rank). - `temperature` (`float`, *optional*, defaults to `1.0`): Sampling temperature passed to vLLM for logprob computation. - `response_format` (`str`, *optional*, defaults to `"json"`): Either `"json"` (nested lists, backward-compatible) or `"binary"` (base64-encoded numpy arrays for fast serialization). Returns: `SequenceLogprobsResponse` or Starlette `Response`: When `response_format` is `"json"`, a JSON object with: - `logprobs` (list of list of list of `float` or `None`): Top-k teacher logprobs per completion token. - `logprob_token_ids` (list of list of list of `int`): Token IDs aligned with `logprobs`. When `response_format` is `"binary"`, a JSON response (Starlette `Response` if `orjson` is installed) whose body is a JSON object with base64-encoded float32/int32 arrays: `logprobs_b64`, `token_ids_b64`, `actual_logprobs_b64`, `actual_token_ids_b64`, plus `shape` (`list[int]`, `[batch_size, max_completion_len, top_k]`) and `completion_lengths` (`list[int]`). """ if len(request.sequences) != len(request.prompt_lengths): raise ValueError("sequences and prompt_lengths must have the same length.") for i, (seq, pl) in enumerate(zip(request.sequences, request.prompt_lengths, strict=True)): if pl < 0 or pl > len(seq): raise ValueError( f"Sequence {i} has prompt_length={pl} which is out of range [0, {len(seq)}]. " f"prompt_length must be between 0 and the sequence length inclusive." ) # Validate sequence lengths against max_model_len to prevent worker OOM crashes if _max_model_len: for i, seq in enumerate(request.sequences): if len(seq) > _max_model_len: raise ValueError( f"Sequence {i} has length {len(seq)} which exceeds max_model_len={_max_model_len}. " f"Truncate sequences or increase --max-model-len." ) prompts = [{"prompt_token_ids": seq} for seq in request.sequences] # Submit to the batching queue and await result loop = asyncio.get_running_loop() future = loop.create_future() await _logprob_queue.put( ( prompts, list(request.prompt_lengths), request.top_logprobs, request.temperature, request.response_format, future, ) ) # Wait for the batcher to process our request all_outputs, prompt_lengths, top_k, response_format = await future return await loop.run_in_executor( None, _format_logprob_response, all_outputs, prompt_lengths, top_k, response_format ) class ChatRequest(BaseModel): messages: list[list[dict]] n: int = 1 repetition_penalty: float = 1.0 temperature: float = 1.0 top_p: float = 1.0 top_k: int = -1 min_p: float = 0.0 max_tokens: int = 16 logprobs: int | None = 0 structured_outputs_regex: str | None = None generation_kwargs: dict = field(default_factory=dict) chat_template_kwargs: dict = field(default_factory=dict) tools: list | None = None class ChatResponse(BaseModel): prompt_ids: list[list[int]] completion_ids: list[list[int]] logprobs: list[list[list[float | None]]] | None logprob_token_ids: list[list[list[int]]] | None @app.post("/chat/", response_model=ChatResponse) async def chat(request: ChatRequest): """ Generates completions for the provided chat messages. Args: request (`ChatRequest`): - `messages` (list of `dict`): A list of messages (dicts with "role" and "content" keys) for the model to generate completions. - `n` (`int`, *optional*, defaults to `1`): Number of completions to generate for each prompt. - `repetition_penalty` (`float`, *optional*, defaults to `1.0`): Repetition penalty to apply during generation. - `temperature` (`float`, *optional*, defaults to `1.0`): Temperature for sampling. Higher values lead to more random outputs. - `top_p` (`float`, *optional*, defaults to `1.0`): Top-p (nucleus) sampling parameter. It controls the diversity of the generated text. - `top_k` (`int`, *optional*, defaults to `-1`): Top-k sampling parameter. If set to `-1`, it disables top-k sampling. - `min_p` (`float`, *optional*, defaults to `0.0`): Minimum probability threshold for sampling. - `max_tokens` (`int`, *optional*, defaults to `16`): Maximum number of tokens to generate for each completion. - `logprobs` (`int`, *optional*, defaults to `0`): Number of top logprobs to return per token. When 0, only the sampled token's logprob is returned. When N>0, returns up to N+1 logprobs sorted by descending probability, because vLLM always includes the sampled token's logprob (which may fall outside the top-N). - `structured_outputs_regex` (`str`, *optional*): A regex pattern for structured outputs. If provided, the model will only generate tokens that match this regex pattern. - `generation_kwargs` (`dict`, *optional*): Additional generation parameters to pass to the vLLM `SamplingParams`. This can include parameters like `seed`, `frequency_penalty`, etc. If it contains keys that conflict with the other parameters, they will override them. - `chat_template_kwargs` (`dict`, *optional*): Additional keyword arguments to pass to the chat template. Returns: `ChatResponse`: - `prompt_ids` (list of list of `int`): A list of lists of token IDs for each input prompt. - `completion_ids` (list of list of `int`): A list of lists of token IDs for each generated completion. - `logprobs` (list of list of list of `float`): Per-token logprobs of shape (num_sequences, seq_len, num_logprobs), sorted by descending probability. - `logprob_token_ids` (list of list of list of `int`): Token IDs corresponding to each logprob, same shape as `logprobs`. Example request: ```bash curl -X POST 'http://0.0.0.0:8000/chat/' \ -H 'Content-Type: application/json' \ -d '{"messages": [[{ "role": "user", "content": "Hello!" }]]}' ``` Example response: ```json { "prompt_ids": [[151644, 872, 198, 9707, 0, 151645, 198, 151644, 77091, 198]], "completion_ids": [[151667, 198, 32313, 11, 279]], "logprobs": [[[-0.0003], [-3.58e-07], [-0.0902], [-6.39e-05], [-0.0387]]], "logprob_token_ids": [[[151667], [198], [32313], [11], [279]]] } ``` """ # Convert PIL images to base64 strings for message_list in request.messages: for message in message_list: if isinstance(message["content"], list): for part in message["content"]: if part["type"] == "image_pil": part["image_pil"] = Image.open(BytesIO(base64.b64decode(part["image_pil"]))) generation_kwargs = { "n": request.n, "repetition_penalty": request.repetition_penalty, "temperature": request.temperature, "top_p": request.top_p, "top_k": request.top_k, "min_p": request.min_p, "max_tokens": request.max_tokens, "logprobs": request.logprobs, } generation_kwargs.update(request.generation_kwargs) # Structured outputs, if enabled if request.structured_outputs_regex is not None: if generation_kwargs.get("structured_outputs") is not None: logger.warning( "Both `structured_outputs_regex` and `generation_kwargs['structured_outputs']` are set; " "`structured_outputs_regex` takes precedence." ) generation_kwargs["structured_outputs"] = StructuredOutputsParams(regex=request.structured_outputs_regex) elif isinstance(structured_outputs_kwargs := generation_kwargs.get("structured_outputs"), dict): generation_kwargs["structured_outputs"] = StructuredOutputsParams(**structured_outputs_kwargs) sampling_params = SamplingParams(**generation_kwargs) # Evenly distribute prompts across DP ranks chunked_messages = chunk_list(request.messages, script_args.data_parallel_size) # Send the messages to each worker for connection, messages in zip(connections, chunked_messages, strict=True): # When the number of messages is less than data_parallel_size, some workers will receive empty messages. # However, vLLM requires that we always send at least one prompt. So we send a placeholder prompt to comply # with vLLM's requirement, and we later ignore the result. if not messages: messages = [[{"role": "user", "content": ""}]] kwargs = { "messages": messages, "sampling_params": sampling_params, "chat_template_kwargs": request.chat_template_kwargs, "tools": request.tools, } connection.send({"type": "call", "method": "chat", "kwargs": kwargs}) # Receive results all_outputs = [connection.recv() for connection in connections] # Handle empty prompts (see above) all_outputs = [output for output, prompts in zip(all_outputs, chunked_messages, strict=True) if prompts] # Flatten and combine all results all_outputs = list(chain.from_iterable(all_outputs)) # from list of list to single list prompt_ids = [output.prompt_token_ids for output in all_outputs] completion_ids = [list(output.token_ids) for outputs in all_outputs for output in outputs.outputs] logprobs, logprob_token_ids = extract_logprobs(all_outputs) return { "prompt_ids": prompt_ids, "completion_ids": completion_ids, "logprobs": logprobs, "logprob_token_ids": logprob_token_ids, } class InitCommunicatorRequest(BaseModel): host: str port: int world_size: int client_device_uuid: str @app.post("/init_communicator/") async def init_communicator(request: InitCommunicatorRequest): """ Initializes the communicator for synchronizing model weights between a client and multiple server workers. Args: request (`InitCommunicatorRequest`): - `host` (`str`): Hostname or IP address of the master node. - `port` (`int`): Port number to be used for communication. - `world_size` (`int`): Total number of participating processes in the group. - `client_device_uuid` (`str`): UUID of the device of client main process. Used to assert that devices are different from vLLM workers devices. """ world_size = script_args.tensor_parallel_size * script_args.data_parallel_size + 1 # The function init_communicator is called this way: init_communicator(host, port, world_size) # So with collective_rpc we need to call it this way: # llm.collective_rpc(method="init_communicator", args=(host, port, world_size)) kwargs = { "method": "init_communicator", "args": (request.host, request.port, world_size, request.client_device_uuid), } for connection in connections: connection.send({"type": "fire_and_forget", "method": "collective_rpc", "kwargs": kwargs}) return {"message": "Request received, initializing communicator"} class UpdateWeightsRequest(BaseModel): name: str dtype: str shape: list[int] @app.post("/update_named_param/") async def update_named_param(request: UpdateWeightsRequest): """ Updates the model weights with the provided tensor. Once this endpoint is called, the client process should broadcast the updated weights to all server workers. Args: request (`UpdateWeightsRequest`): - `name` (`str`): Name of the weight tensor being updated. - `dtype` (`str`): Data type of the weight tensor (e.g., `"torch.float32"`). - `shape` (list of `int`): Shape of the weight """ # The function update_named_param is called this way: update_named_param("name", "torch.float32", (10, 10)) # So with collective_rpc we need to call it this way: # llm.collective_rpc("update_named_param", args=("name", "torch.float32", (10, 10))) kwargs = {"method": "update_named_param", "args": (request.name, request.dtype, tuple(request.shape))} for connection in connections: connection.send({"type": "fire_and_forget", "method": "collective_rpc", "kwargs": kwargs}) return {"message": "Request received, updating named parameter"} @app.post("/reset_prefix_cache/") async def reset_prefix_cache(): """ Resets the prefix cache for the model. """ for connection in connections: connection.send({"type": "call", "method": "reset_prefix_cache"}) # Wait for and collect all results all_outputs = [connection.recv() for connection in connections] success = all(output for output in all_outputs) return {"message": "Request received, resetting prefix cache status: " + str(success)} @app.post("/close_communicator/") async def close_communicator(): """ Closes the weight update group and cleans up associated resources. """ kwargs = {"method": "close_communicator"} for connection in connections: connection.send({"type": "fire_and_forget", "method": "collective_rpc", "kwargs": kwargs}) return {"message": "Request received, closing communicator"} # Start the server uvicorn.run(app, host=script_args.host, port=script_args.port, log_level=script_args.log_level) def make_parser(subparsers: argparse._SubParsersAction | None = None, prog: str | None = None): from trl import TrlParser if subparsers is not None: parser = subparsers.add_parser("vllm-serve", help="Run the vLLM serve script", dataclass_types=ScriptArguments) else: parser = TrlParser(ScriptArguments, prog=prog) return parser if __name__ == "__main__": parser = make_parser() (script_args,) = parser.parse_args_and_config() main(script_args)