|
4 | 4 |
|
5 | 5 | ```bash
|
6 | 6 | pip install fastapi uvicorn sse-starlette
|
7 |
| -export MODEL=../models/7B/ggml-model.bin |
8 |
| -uvicorn fastapi_server_chat:app --reload |
| 7 | +export MODEL=../models/7B/... |
9 | 8 | ```
|
10 | 9 |
|
11 |
| -Then visit http://localhost:8000/docs to see the interactive API docs. |
12 |
| -
|
13 |
| -""" |
14 |
| -import os |
15 |
| -import json |
16 |
| -from typing import List, Optional, Literal, Union, Iterator, Dict |
17 |
| -from typing_extensions import TypedDict |
18 |
| - |
19 |
| -import llama_cpp |
20 |
| - |
21 |
| -from fastapi import FastAPI |
22 |
| -from fastapi.middleware.cors import CORSMiddleware |
23 |
| -from pydantic import BaseModel, BaseSettings, Field, create_model_from_typeddict |
24 |
| -from sse_starlette.sse import EventSourceResponse |
25 |
| - |
26 |
| - |
27 |
| -class Settings(BaseSettings): |
28 |
| - model: str |
29 |
| - n_ctx: int = 2048 |
30 |
| - n_batch: int = 8 |
31 |
| - n_threads: int = int(os.cpu_count() / 2) or 1 |
32 |
| - f16_kv: bool = True |
33 |
| - use_mlock: bool = False # This causes a silent failure on platforms that don't support mlock (e.g. Windows) took forever to figure out... |
34 |
| - embedding: bool = True |
35 |
| - last_n_tokens_size: int = 64 |
36 |
| - |
37 |
| - |
38 |
| -app = FastAPI( |
39 |
| - title="🦙 llama.cpp Python API", |
40 |
| - version="0.0.1", |
41 |
| -) |
42 |
| -app.add_middleware( |
43 |
| - CORSMiddleware, |
44 |
| - allow_origins=["*"], |
45 |
| - allow_credentials=True, |
46 |
| - allow_methods=["*"], |
47 |
| - allow_headers=["*"], |
48 |
| -) |
49 |
| -settings = Settings() |
50 |
| -llama = llama_cpp.Llama( |
51 |
| - settings.model, |
52 |
| - f16_kv=settings.f16_kv, |
53 |
| - use_mlock=settings.use_mlock, |
54 |
| - embedding=settings.embedding, |
55 |
| - n_threads=settings.n_threads, |
56 |
| - n_batch=settings.n_batch, |
57 |
| - n_ctx=settings.n_ctx, |
58 |
| - last_n_tokens_size=settings.last_n_tokens_size, |
59 |
| -) |
60 |
| - |
61 |
| - |
62 |
| -class CreateCompletionRequest(BaseModel): |
63 |
| - prompt: str |
64 |
| - suffix: Optional[str] = Field(None) |
65 |
| - max_tokens: int = 16 |
66 |
| - temperature: float = 0.8 |
67 |
| - top_p: float = 0.95 |
68 |
| - echo: bool = False |
69 |
| - stop: List[str] = [] |
70 |
| - stream: bool = False |
71 |
| - |
72 |
| - # ignored or currently unsupported |
73 |
| - model: Optional[str] = Field(None) |
74 |
| - n: Optional[int] = 1 |
75 |
| - logprobs: Optional[int] = Field(None) |
76 |
| - presence_penalty: Optional[float] = 0 |
77 |
| - frequency_penalty: Optional[float] = 0 |
78 |
| - best_of: Optional[int] = 1 |
79 |
| - logit_bias: Optional[Dict[str, float]] = Field(None) |
80 |
| - user: Optional[str] = Field(None) |
81 |
| - |
82 |
| - # llama.cpp specific parameters |
83 |
| - top_k: int = 40 |
84 |
| - repeat_penalty: float = 1.1 |
85 |
| - |
86 |
| - class Config: |
87 |
| - schema_extra = { |
88 |
| - "example": { |
89 |
| - "prompt": "\n\n### Instructions:\nWhat is the capital of France?\n\n### Response:\n", |
90 |
| - "stop": ["\n", "###"], |
91 |
| - } |
92 |
| - } |
93 |
| - |
94 |
| - |
95 |
| -CreateCompletionResponse = create_model_from_typeddict(llama_cpp.Completion) |
96 |
| - |
97 |
| - |
98 |
| -@app.post( |
99 |
| - "/v1/completions", |
100 |
| - response_model=CreateCompletionResponse, |
101 |
| -) |
102 |
| -def create_completion(request: CreateCompletionRequest): |
103 |
| - if request.stream: |
104 |
| - chunks: Iterator[llama_cpp.CompletionChunk] = llama(**request.dict()) # type: ignore |
105 |
| - return EventSourceResponse(dict(data=json.dumps(chunk)) for chunk in chunks) |
106 |
| - return llama( |
107 |
| - **request.dict( |
108 |
| - exclude={ |
109 |
| - "model", |
110 |
| - "n", |
111 |
| - "logprobs", |
112 |
| - "frequency_penalty", |
113 |
| - "presence_penalty", |
114 |
| - "best_of", |
115 |
| - "logit_bias", |
116 |
| - "user", |
117 |
| - } |
118 |
| - ) |
119 |
| - ) |
120 |
| - |
121 |
| - |
122 |
| -class CreateEmbeddingRequest(BaseModel): |
123 |
| - model: Optional[str] |
124 |
| - input: str |
125 |
| - user: Optional[str] |
126 |
| - |
127 |
| - class Config: |
128 |
| - schema_extra = { |
129 |
| - "example": { |
130 |
| - "input": "The food was delicious and the waiter...", |
131 |
| - } |
132 |
| - } |
133 |
| - |
134 |
| - |
135 |
| -CreateEmbeddingResponse = create_model_from_typeddict(llama_cpp.Embedding) |
136 |
| - |
137 |
| - |
138 |
| -@app.post( |
139 |
| - "/v1/embeddings", |
140 |
| - response_model=CreateEmbeddingResponse, |
141 |
| -) |
142 |
| -def create_embedding(request: CreateEmbeddingRequest): |
143 |
| - return llama.create_embedding(**request.dict(exclude={"model", "user"})) |
144 |
| - |
145 |
| - |
146 |
| -class ChatCompletionRequestMessage(BaseModel): |
147 |
| - role: Union[Literal["system"], Literal["user"], Literal["assistant"]] |
148 |
| - content: str |
149 |
| - user: Optional[str] = None |
150 |
| - |
151 |
| - |
152 |
| -class CreateChatCompletionRequest(BaseModel): |
153 |
| - model: Optional[str] |
154 |
| - messages: List[ChatCompletionRequestMessage] |
155 |
| - temperature: float = 0.8 |
156 |
| - top_p: float = 0.95 |
157 |
| - stream: bool = False |
158 |
| - stop: List[str] = [] |
159 |
| - max_tokens: int = 128 |
160 |
| - |
161 |
| - # ignored or currently unsupported |
162 |
| - model: Optional[str] = Field(None) |
163 |
| - n: Optional[int] = 1 |
164 |
| - presence_penalty: Optional[float] = 0 |
165 |
| - frequency_penalty: Optional[float] = 0 |
166 |
| - logit_bias: Optional[Dict[str, float]] = Field(None) |
167 |
| - user: Optional[str] = Field(None) |
168 |
| - |
169 |
| - # llama.cpp specific parameters |
170 |
| - repeat_penalty: float = 1.1 |
171 |
| - |
172 |
| - class Config: |
173 |
| - schema_extra = { |
174 |
| - "example": { |
175 |
| - "messages": [ |
176 |
| - ChatCompletionRequestMessage( |
177 |
| - role="system", content="You are a helpful assistant." |
178 |
| - ), |
179 |
| - ChatCompletionRequestMessage( |
180 |
| - role="user", content="What is the capital of France?" |
181 |
| - ), |
182 |
| - ] |
183 |
| - } |
184 |
| - } |
185 |
| - |
186 |
| - |
187 |
| -CreateChatCompletionResponse = create_model_from_typeddict(llama_cpp.ChatCompletion) |
188 |
| - |
189 |
| - |
190 |
| -@app.post( |
191 |
| - "/v1/chat/completions", |
192 |
| - response_model=CreateChatCompletionResponse, |
193 |
| -) |
194 |
| -async def create_chat_completion( |
195 |
| - request: CreateChatCompletionRequest, |
196 |
| -) -> Union[llama_cpp.ChatCompletion, EventSourceResponse]: |
197 |
| - completion_or_chunks = llama.create_chat_completion( |
198 |
| - **request.dict( |
199 |
| - exclude={ |
200 |
| - "model", |
201 |
| - "n", |
202 |
| - "presence_penalty", |
203 |
| - "frequency_penalty", |
204 |
| - "logit_bias", |
205 |
| - "user", |
206 |
| - } |
207 |
| - ), |
208 |
| - ) |
209 |
| - |
210 |
| - if request.stream: |
211 |
| - |
212 |
| - async def server_sent_events( |
213 |
| - chat_chunks: Iterator[llama_cpp.ChatCompletionChunk], |
214 |
| - ): |
215 |
| - for chat_chunk in chat_chunks: |
216 |
| - yield dict(data=json.dumps(chat_chunk)) |
217 |
| - yield dict(data="[DONE]") |
218 |
| - |
219 |
| - chunks: Iterator[llama_cpp.ChatCompletionChunk] = completion_or_chunks # type: ignore |
220 |
| - |
221 |
| - return EventSourceResponse( |
222 |
| - server_sent_events(chunks), |
223 |
| - ) |
224 |
| - completion: llama_cpp.ChatCompletion = completion_or_chunks # type: ignore |
225 |
| - return completion |
226 |
| - |
227 |
| - |
228 |
| -class ModelData(TypedDict): |
229 |
| - id: str |
230 |
| - object: Literal["model"] |
231 |
| - owned_by: str |
232 |
| - permissions: List[str] |
| 10 | +Then run: |
| 11 | +``` |
| 12 | +uvicorn llama_cpp.server.app:app --reload |
| 13 | +``` |
233 | 14 |
|
| 15 | +or |
234 | 16 |
|
235 |
| -class ModelList(TypedDict): |
236 |
| - object: Literal["list"] |
237 |
| - data: List[ModelData] |
| 17 | +``` |
| 18 | +python3 -m llama_cpp.server |
| 19 | +``` |
238 | 20 |
|
| 21 | +Then visit http://localhost:8000/docs to see the interactive API docs. |
239 | 22 |
|
240 |
| -GetModelResponse = create_model_from_typeddict(ModelList) |
241 | 23 |
|
| 24 | +To actually see the implementation of the server, see llama_cpp/server/app.py |
242 | 25 |
|
243 |
| -@app.get("/v1/models", response_model=GetModelResponse) |
244 |
| -def get_models() -> ModelList: |
245 |
| - return { |
246 |
| - "object": "list", |
247 |
| - "data": [ |
248 |
| - { |
249 |
| - "id": llama.model_path, |
250 |
| - "object": "model", |
251 |
| - "owned_by": "me", |
252 |
| - "permissions": [], |
253 |
| - } |
254 |
| - ], |
255 |
| - } |
| 26 | +""" |
| 27 | +import os |
| 28 | +import uvicorn |
256 | 29 |
|
| 30 | +from llama_cpp.server.app import create_app |
257 | 31 |
|
258 | 32 | if __name__ == "__main__":
|
259 |
| - import os |
260 |
| - import uvicorn |
| 33 | + app = create_app() |
261 | 34 |
|
262 |
| - uvicorn.run(app, host=os.getenv("HOST", "localhost"), port=os.getenv("PORT", 8000)) |
| 35 | + uvicorn.run( |
| 36 | + app, host=os.getenv("HOST", "localhost"), port=int(os.getenv("PORT", 8000)) |
| 37 | + ) |
0 commit comments