File size: 9,531 Bytes
ec0c335
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
"""
A model worker that executes the model based on vLLM.

See documentations at docs/vllm_integration.md
"""

import argparse
import asyncio
import json
from typing import List

from fastapi import FastAPI, Request, BackgroundTasks
from fastapi.responses import StreamingResponse, JSONResponse
import uvicorn
from vllm import AsyncLLMEngine
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.sampling_params import SamplingParams
from vllm.utils import random_uuid

from fastchat.serve.base_model_worker import BaseModelWorker
from fastchat.serve.model_worker import (
    logger,
    worker_id,
)
from fastchat.utils import get_context_length, is_partial_stop


app = FastAPI()


class VLLMWorker(BaseModelWorker):
    def __init__(
        self,
        controller_addr: str,
        worker_addr: str,
        worker_id: str,
        model_path: str,
        model_names: List[str],
        limit_worker_concurrency: int,
        no_register: bool,
        llm_engine: AsyncLLMEngine,
        conv_template: str,
    ):
        super().__init__(
            controller_addr,
            worker_addr,
            worker_id,
            model_path,
            model_names,
            limit_worker_concurrency,
            conv_template,
        )

        logger.info(
            f"Loading the model {self.model_names} on worker {worker_id}, worker type: vLLM worker..."
        )
        self.tokenizer = llm_engine.engine.tokenizer
        self.context_len = get_context_length(llm_engine.engine.model_config.hf_config)

        if not no_register:
            self.init_heart_beat()

    async def generate_stream(self, params):
        self.call_ct += 1

        context = params.pop("prompt")
        request_id = params.pop("request_id")
        temperature = float(params.get("temperature", 1.0))
        top_p = float(params.get("top_p", 1.0))
        top_k = params.get("top_k", -1.0)
        presence_penalty = float(params.get("presence_penalty", 0.0))
        frequency_penalty = float(params.get("frequency_penalty", 0.0))
        max_new_tokens = params.get("max_new_tokens", 256)
        stop_str = params.get("stop", None)
        stop_token_ids = params.get("stop_token_ids", None) or []
        if self.tokenizer.eos_token_id is not None:
            stop_token_ids.append(self.tokenizer.eos_token_id)
        echo = params.get("echo", True)
        use_beam_search = params.get("use_beam_search", False)
        best_of = params.get("best_of", None)

        # Handle stop_str
        stop = set()
        if isinstance(stop_str, str) and stop_str != "":
            stop.add(stop_str)
        elif isinstance(stop_str, list) and stop_str != []:
            stop.update(stop_str)

        for tid in stop_token_ids:
            if tid is not None:
                stop.add(self.tokenizer.decode(tid))

        # make sampling params in vllm
        top_p = max(top_p, 1e-5)
        if temperature <= 1e-5:
            top_p = 1.0

        sampling_params = SamplingParams(
            n=1,
            temperature=temperature,
            top_p=top_p,
            use_beam_search=use_beam_search,
            stop=list(stop),
            stop_token_ids=stop_token_ids,
            max_tokens=max_new_tokens,
            top_k=top_k,
            presence_penalty=presence_penalty,
            frequency_penalty=frequency_penalty,
            best_of=best_of,
        )
        results_generator = engine.generate(context, sampling_params, request_id)

        async for request_output in results_generator:
            prompt = request_output.prompt
            if echo:
                text_outputs = [
                    prompt + output.text for output in request_output.outputs
                ]
            else:
                text_outputs = [output.text for output in request_output.outputs]
            text_outputs = " ".join(text_outputs)

            partial_stop = any(is_partial_stop(text_outputs, i) for i in stop)
            # prevent yielding partial stop sequence
            if partial_stop:
                continue

            prompt_tokens = len(request_output.prompt_token_ids)
            completion_tokens = sum(
                len(output.token_ids) for output in request_output.outputs
            )
            ret = {
                "text": text_outputs,
                "error_code": 0,
                "usage": {
                    "prompt_tokens": prompt_tokens,
                    "completion_tokens": completion_tokens,
                    "total_tokens": prompt_tokens + completion_tokens,
                },
                "cumulative_logprob": [
                    output.cumulative_logprob for output in request_output.outputs
                ],
                "finish_reason": request_output.outputs[0].finish_reason
                if len(request_output.outputs) == 1
                else [output.finish_reason for output in request_output.outputs],
            }
            # Emit twice here to ensure a 'finish_reason' with empty content in the OpenAI API response.
            # This aligns with the behavior of model_worker.
            if request_output.finished:
                yield (json.dumps(ret | {"finish_reason": None}) + "\0").encode()
            yield (json.dumps(ret) + "\0").encode()

    async def generate(self, params):
        async for x in self.generate_stream(params):
            pass
        return json.loads(x[:-1].decode())


def release_worker_semaphore():
    worker.semaphore.release()


def acquire_worker_semaphore():
    if worker.semaphore is None:
        worker.semaphore = asyncio.Semaphore(worker.limit_worker_concurrency)
    return worker.semaphore.acquire()


def create_background_tasks(request_id):
    async def abort_request() -> None:
        await engine.abort(request_id)

    background_tasks = BackgroundTasks()
    background_tasks.add_task(release_worker_semaphore)
    background_tasks.add_task(abort_request)
    return background_tasks


@app.post("/worker_generate_stream")
async def api_generate_stream(request: Request):
    logger.info("vllm_worker")
    params = await request.json()
    await acquire_worker_semaphore()
    request_id = random_uuid()
    params["request_id"] = request_id
    generator = worker.generate_stream(params)
    background_tasks = create_background_tasks(request_id)
    return StreamingResponse(generator, background=background_tasks)


@app.post("/worker_generate")
async def api_generate(request: Request):
    params = await request.json()
    await acquire_worker_semaphore()
    request_id = random_uuid()
    params["request_id"] = request_id
    output = await worker.generate(params)
    release_worker_semaphore()
    await engine.abort(request_id)
    return JSONResponse(output)


@app.post("/worker_get_status")
async def api_get_status(request: Request):
    return worker.get_status()


@app.post("/count_token")
async def api_count_token(request: Request):
    params = await request.json()
    return worker.count_token(params)


@app.post("/worker_get_conv_template")
async def api_get_conv(request: Request):
    return worker.get_conv_template()


@app.post("/model_details")
async def api_model_details(request: Request):
    return {"context_length": worker.context_len}


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", type=str, default="localhost")
    parser.add_argument("--port", type=int, default=21002)
    parser.add_argument("--worker-address", type=str, default="http://localhost:21002")
    parser.add_argument(
        "--controller-address", type=str, default="http://localhost:21001"
    )
    parser.add_argument("--model-path", type=str, default="lmsys/vicuna-7b-v1.5")
    parser.add_argument(
        "--model-names",
        type=lambda s: s.split(","),
        help="Optional display comma separated names",
    )
    parser.add_argument("--limit-worker-concurrency", type=int, default=1024)
    parser.add_argument("--no-register", action="store_true")
    parser.add_argument("--num-gpus", type=int, default=1)
    parser.add_argument(
        "--conv-template", type=str, default=None, help="Conversation prompt template."
    )
    parser.add_argument(
        "--trust_remote_code",
        action="store_false",
        default=True,
        help="Trust remote code (e.g., from HuggingFace) when"
        "downloading the model and tokenizer.",
    )
    parser.add_argument(
        "--gpu_memory_utilization",
        type=float,
        default=0.9,
        help="The ratio (between 0 and 1) of GPU memory to"
        "reserve for the model weights, activations, and KV cache. 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.",
    )

    parser = AsyncEngineArgs.add_cli_args(parser)
    args = parser.parse_args()
    if args.model_path:
        args.model = args.model_path
    if args.num_gpus > 1:
        args.tensor_parallel_size = args.num_gpus

    engine_args = AsyncEngineArgs.from_cli_args(args)
    engine = AsyncLLMEngine.from_engine_args(engine_args)
    worker = VLLMWorker(
        args.controller_address,
        args.worker_address,
        worker_id,
        args.model_path,
        args.model_names,
        args.limit_worker_concurrency,
        args.no_register,
        engine,
        args.conv_template,
    )
    uvicorn.run(app, host=args.host, port=args.port, log_level="info")