|
| 1 | +# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 2 | +# SPDX-License-Identifier: Apache-2.0 |
| 3 | + |
| 4 | +# Usage: `python -m dynamo.llama_cpp --model-path /data/models/Qwen3-0.6B-Q8_0.gguf [args]` |
| 5 | + |
| 6 | +import argparse |
| 7 | +import logging |
| 8 | +import sys |
| 9 | +from typing import Optional |
| 10 | + |
| 11 | +import uvloop |
| 12 | +from llama_cpp import Llama |
| 13 | + |
| 14 | +from dynamo.llm import ModelType, register_llm |
| 15 | +from dynamo.runtime import DistributedRuntime, dynamo_worker |
| 16 | +from dynamo.runtime.logging import configure_dynamo_logging |
| 17 | + |
| 18 | +DEFAULT_ENDPOINT = "dyn://dynamo.backend.generate" |
| 19 | + |
| 20 | +configure_dynamo_logging() |
| 21 | + |
| 22 | + |
| 23 | +class Config: |
| 24 | + """Command line parameters or defaults""" |
| 25 | + |
| 26 | + namespace: str |
| 27 | + component: str |
| 28 | + endpoint: str |
| 29 | + model_path: str |
| 30 | + model_name: Optional[str] |
| 31 | + context_length: int |
| 32 | + |
| 33 | + |
| 34 | +@dynamo_worker(static=False) |
| 35 | +async def worker(runtime: DistributedRuntime): |
| 36 | + config = cmd_line_args() |
| 37 | + |
| 38 | + component = runtime.namespace(config.namespace).component(config.component) |
| 39 | + await component.create_service() |
| 40 | + |
| 41 | + model_type = ModelType.Chat # llama.cpp does the pre-processing |
| 42 | + endpoint = component.endpoint(config.endpoint) |
| 43 | + await register_llm(model_type, endpoint, config.model_path, config.model_name) |
| 44 | + |
| 45 | + # Initialize the engine |
| 46 | + # For more parameters see: |
| 47 | + # https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#high-level-api |
| 48 | + kwargs = { |
| 49 | + "model_path": config.model_path, |
| 50 | + "n_gpu_layers": -1, # GPU if we can |
| 51 | + "n_threads": 16, # Otherwise give it some CPU |
| 52 | + } |
| 53 | + if config.context_length: |
| 54 | + kwargs["n_ctx"] = config.context_length |
| 55 | + engine = Llama(**kwargs) |
| 56 | + |
| 57 | + await endpoint.serve_endpoint(RequestHandler(engine).generate) |
| 58 | + |
| 59 | + |
| 60 | +class RequestHandler: |
| 61 | + def __init__(self, engine): |
| 62 | + self.engine_client = engine |
| 63 | + |
| 64 | + async def generate(self, request): |
| 65 | + gen = self.engine_client.create_chat_completion( |
| 66 | + request["messages"], stream=True |
| 67 | + ) |
| 68 | + # TODO this is a synchronous generator in an async method. |
| 69 | + # Move it to a thread so it doesn't block the event loop. |
| 70 | + for res in gen: |
| 71 | + logging.debug(f"res: {res}") |
| 72 | + yield res |
| 73 | + |
| 74 | + |
| 75 | +def cmd_line_args(): |
| 76 | + parser = argparse.ArgumentParser( |
| 77 | + description="llama.cpp server integrated with Dynamo LLM." |
| 78 | + ) |
| 79 | + parser.add_argument( |
| 80 | + "--model-path", |
| 81 | + type=str, |
| 82 | + required=True, |
| 83 | + help="Path to a local GGUF file.", |
| 84 | + ) |
| 85 | + parser.add_argument( |
| 86 | + "--endpoint", |
| 87 | + type=str, |
| 88 | + default=DEFAULT_ENDPOINT, |
| 89 | + help=f"Dynamo endpoint string in 'dyn://namespace.component.endpoint' format. Default: {DEFAULT_ENDPOINT}", |
| 90 | + ) |
| 91 | + parser.add_argument( |
| 92 | + "--model-name", |
| 93 | + type=str, |
| 94 | + default="", |
| 95 | + help="Name to serve the model under. Defaults to deriving it from model path.", |
| 96 | + ) |
| 97 | + parser.add_argument( |
| 98 | + "--context-length", |
| 99 | + type=int, |
| 100 | + default=None, |
| 101 | + help="Max model context length. Defaults to models max, usually model_max_length from tokenizer_config.json. Reducing this reduces VRAM requirements.", |
| 102 | + ) |
| 103 | + args = parser.parse_args() |
| 104 | + |
| 105 | + config = Config() |
| 106 | + config.model_path = args.model_path |
| 107 | + if args.model_name: |
| 108 | + config.model_name = args.model_name |
| 109 | + else: |
| 110 | + # This becomes an `Option` on the Rust side |
| 111 | + config.model_name = None |
| 112 | + |
| 113 | + endpoint_str = args.endpoint.replace("dyn://", "", 1) |
| 114 | + endpoint_parts = endpoint_str.split(".") |
| 115 | + if len(endpoint_parts) != 3: |
| 116 | + logging.error( |
| 117 | + f"Invalid endpoint format: '{args.endpoint}'. Expected 'dyn://namespace.component.endpoint' or 'namespace.component.endpoint'." |
| 118 | + ) |
| 119 | + sys.exit(1) |
| 120 | + |
| 121 | + parsed_namespace, parsed_component_name, parsed_endpoint_name = endpoint_parts |
| 122 | + |
| 123 | + config.namespace = parsed_namespace |
| 124 | + config.component = parsed_component_name |
| 125 | + config.endpoint = parsed_endpoint_name |
| 126 | + config.context_length = args.context_length |
| 127 | + return config |
| 128 | + |
| 129 | + |
| 130 | +def main(): |
| 131 | + uvloop.run(worker()) |
| 132 | + |
| 133 | + |
| 134 | +if __name__ == "__main__": |
| 135 | + main() |
0 commit comments