Upload 3 files
Browse files- Dockerfile +42 -0
- app.py +445 -0
- start.sh +19 -0
Dockerfile
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM ubuntu:22.04
|
| 2 |
+
|
| 3 |
+
RUN apt-get update && \
|
| 4 |
+
apt-get install -y \
|
| 5 |
+
build-essential \
|
| 6 |
+
libssl-dev \
|
| 7 |
+
zlib1g-dev \
|
| 8 |
+
libopenblas-dev \
|
| 9 |
+
libomp-dev \
|
| 10 |
+
cmake \
|
| 11 |
+
pkg-config \
|
| 12 |
+
git \
|
| 13 |
+
python3-pip \
|
| 14 |
+
curl \
|
| 15 |
+
libcurl4-openssl-dev \
|
| 16 |
+
wget && \
|
| 17 |
+
rm -rf /var/lib/apt/lists/*
|
| 18 |
+
|
| 19 |
+
RUN pip3 install --upgrade pip && \
|
| 20 |
+
pip3 install openai fastapi uvicorn pydantic orjson httptools
|
| 21 |
+
|
| 22 |
+
RUN pip install httpx[http2]
|
| 23 |
+
|
| 24 |
+
RUN git clone https://github.com/ggerganov/llama.cpp && \
|
| 25 |
+
cd llama.cpp && \
|
| 26 |
+
cmake -B build -S . \
|
| 27 |
+
-DLLAMA_BUILD_SERVER=ON \
|
| 28 |
+
-DGGML_BLAS=ON \
|
| 29 |
+
-DGGML_BLAS_VENDOR=OpenBLAS \
|
| 30 |
+
-DCMAKE_BUILD_TYPE=Release && \
|
| 31 |
+
cmake --build build --config Release --target llama-server -j $(nproc)
|
| 32 |
+
|
| 33 |
+
RUN mkdir -p /models && \
|
| 34 |
+
wget -O /models/model.gguf https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-UD-Q8_K_XL.gguf
|
| 35 |
+
|
| 36 |
+
COPY app.py /app.py
|
| 37 |
+
COPY start.sh /start.sh
|
| 38 |
+
RUN chmod +x /start.sh
|
| 39 |
+
|
| 40 |
+
EXPOSE 7860
|
| 41 |
+
|
| 42 |
+
CMD ["/start.sh"]
|
app.py
ADDED
|
@@ -0,0 +1,445 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import time
|
| 2 |
+
import orjson
|
| 3 |
+
import asyncio
|
| 4 |
+
from typing import List, AsyncGenerator
|
| 5 |
+
from fastapi import FastAPI, HTTPException
|
| 6 |
+
import os
|
| 7 |
+
from fastapi.responses import StreamingResponse, ORJSONResponse
|
| 8 |
+
from fastapi.middleware.cors import CORSMiddleware
|
| 9 |
+
from pydantic import BaseModel, Field
|
| 10 |
+
import httpx
|
| 11 |
+
import uvicorn
|
| 12 |
+
|
| 13 |
+
# Use orjson for faster JSON serialization
|
| 14 |
+
app = FastAPI(
|
| 15 |
+
title="Qwen3 API",
|
| 16 |
+
description="Streaming API for Qwen3-0.6B model",
|
| 17 |
+
version="2.0.0",
|
| 18 |
+
default_response_class=ORJSONResponse
|
| 19 |
+
)
|
| 20 |
+
|
| 21 |
+
app.add_middleware(
|
| 22 |
+
CORSMiddleware,
|
| 23 |
+
allow_origins=["*"],
|
| 24 |
+
allow_credentials=True,
|
| 25 |
+
allow_methods=["*"],
|
| 26 |
+
allow_headers=["*"],
|
| 27 |
+
)
|
| 28 |
+
|
| 29 |
+
# Async HTTP client with connection pooling
|
| 30 |
+
BASE_URL = "http://localhost:8080/v1"
|
| 31 |
+
http_client: httpx.AsyncClient = None
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
@app.on_event("startup")
|
| 35 |
+
async def startup():
|
| 36 |
+
global http_client
|
| 37 |
+
http_client = httpx.AsyncClient(
|
| 38 |
+
base_url=BASE_URL,
|
| 39 |
+
timeout=httpx.Timeout(300.0, connect=10.0),
|
| 40 |
+
limits=httpx.Limits(max_keepalive_connections=10, max_connections=20),
|
| 41 |
+
http2=True
|
| 42 |
+
)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
@app.on_event("shutdown")
|
| 46 |
+
async def shutdown():
|
| 47 |
+
global http_client
|
| 48 |
+
if http_client:
|
| 49 |
+
await http_client.aclose()
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
# ===== Models =====
|
| 53 |
+
|
| 54 |
+
class Message(BaseModel):
|
| 55 |
+
role: str
|
| 56 |
+
content: str
|
| 57 |
+
|
| 58 |
+
class Config:
|
| 59 |
+
extra = "ignore"
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
class ChatRequest(BaseModel):
|
| 63 |
+
messages: List[Message]
|
| 64 |
+
temperature: float = Field(default=0.6, ge=0.0, le=2.0)
|
| 65 |
+
top_p: float = Field(default=0.95, ge=0.0, le=1.0)
|
| 66 |
+
max_tokens: int = Field(default=4096, ge=1, le=32768)
|
| 67 |
+
stream: bool = Field(default=True)
|
| 68 |
+
|
| 69 |
+
class Config:
|
| 70 |
+
extra = "ignore"
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
class SimpleChatRequest(BaseModel):
|
| 74 |
+
prompt: str
|
| 75 |
+
temperature: float = Field(default=0.6, ge=0.0, le=2.0)
|
| 76 |
+
top_p: float = Field(default=0.95, ge=0.0, le=1.0)
|
| 77 |
+
max_tokens: int = Field(default=4096, ge=1, le=32768)
|
| 78 |
+
stream: bool = Field(default=True)
|
| 79 |
+
|
| 80 |
+
class Config:
|
| 81 |
+
extra = "ignore"
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
# ===== Optimized Think Tag Parser =====
|
| 85 |
+
|
| 86 |
+
__slots_parser__ = ['answer', 'thought', 'in_think', 'start_time', 'total_think_time', 'buffer']
|
| 87 |
+
|
| 88 |
+
class ParserState:
|
| 89 |
+
__slots__ = ['answer', 'thought', 'in_think', 'start_time', 'total_think_time']
|
| 90 |
+
|
| 91 |
+
def __init__(self):
|
| 92 |
+
self.answer = []
|
| 93 |
+
self.thought = []
|
| 94 |
+
self.in_think = False
|
| 95 |
+
self.start_time = 0.0
|
| 96 |
+
self.total_think_time = 0.0
|
| 97 |
+
|
| 98 |
+
def get_answer(self) -> str:
|
| 99 |
+
return ''.join(self.answer)
|
| 100 |
+
|
| 101 |
+
def get_thought(self) -> str:
|
| 102 |
+
return ''.join(self.thought)
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def parse_chunk(content: str, state: ParserState) -> float:
|
| 106 |
+
buffer = content
|
| 107 |
+
|
| 108 |
+
while buffer:
|
| 109 |
+
if not state.in_think:
|
| 110 |
+
idx = buffer.find('<think>')
|
| 111 |
+
if idx != -1:
|
| 112 |
+
if idx > 0:
|
| 113 |
+
state.answer.append(buffer[:idx])
|
| 114 |
+
state.in_think = True
|
| 115 |
+
state.start_time = time.perf_counter()
|
| 116 |
+
buffer = buffer[idx + 7:]
|
| 117 |
+
else:
|
| 118 |
+
for i in range(min(6, len(buffer)), 0, -1):
|
| 119 |
+
if '<think>'[:i] == buffer[-i:]:
|
| 120 |
+
state.answer.append(buffer[:-i])
|
| 121 |
+
return 0.0
|
| 122 |
+
state.answer.append(buffer)
|
| 123 |
+
return 0.0
|
| 124 |
+
else:
|
| 125 |
+
idx = buffer.find('</think>')
|
| 126 |
+
if idx != -1:
|
| 127 |
+
if idx > 0:
|
| 128 |
+
state.thought.append(buffer[:idx])
|
| 129 |
+
state.total_think_time += time.perf_counter() - state.start_time
|
| 130 |
+
state.in_think = False
|
| 131 |
+
buffer = buffer[idx + 8:]
|
| 132 |
+
else:
|
| 133 |
+
for i in range(min(7, len(buffer)), 0, -1):
|
| 134 |
+
if '</think>'[:i] == buffer[-i:]:
|
| 135 |
+
state.thought.append(buffer[:-i])
|
| 136 |
+
return time.perf_counter() - state.start_time
|
| 137 |
+
state.thought.append(buffer)
|
| 138 |
+
return time.perf_counter() - state.start_time
|
| 139 |
+
|
| 140 |
+
return time.perf_counter() - state.start_time if state.in_think else 0.0
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
# ===== Async Streaming Functions =====
|
| 144 |
+
|
| 145 |
+
async def stream_from_backend(messages: list, temperature: float, top_p: float, max_tokens: int) -> AsyncGenerator[str, None]:
|
| 146 |
+
payload = {
|
| 147 |
+
"model": "",
|
| 148 |
+
"messages": messages,
|
| 149 |
+
"temperature": temperature,
|
| 150 |
+
"top_p": top_p,
|
| 151 |
+
"max_tokens": max_tokens,
|
| 152 |
+
"stream": True
|
| 153 |
+
}
|
| 154 |
+
|
| 155 |
+
async with http_client.stream(
|
| 156 |
+
"POST",
|
| 157 |
+
"/chat/completions",
|
| 158 |
+
json=payload,
|
| 159 |
+
headers={"Accept": "text/event-stream"}
|
| 160 |
+
) as response:
|
| 161 |
+
async for line in response.aiter_lines():
|
| 162 |
+
if line.startswith("data: "):
|
| 163 |
+
data = line[6:]
|
| 164 |
+
if data == "[DONE]":
|
| 165 |
+
break
|
| 166 |
+
try:
|
| 167 |
+
chunk = orjson.loads(data)
|
| 168 |
+
if chunk.get("choices") and chunk["choices"][0].get("delta", {}).get("content"):
|
| 169 |
+
yield chunk["choices"][0]["delta"]["content"]
|
| 170 |
+
except orjson.JSONDecodeError:
|
| 171 |
+
continue
|
| 172 |
+
|
| 173 |
+
|
| 174 |
+
async def generate_stream_fast(request: ChatRequest) -> AsyncGenerator[bytes, None]:
|
| 175 |
+
messages = [{"role": m.role, "content": m.content} for m in request.messages]
|
| 176 |
+
state = ParserState()
|
| 177 |
+
chunk_id = f"chatcmpl-{int(time.time() * 1000)}"
|
| 178 |
+
created = int(time.time())
|
| 179 |
+
|
| 180 |
+
try:
|
| 181 |
+
async for content in stream_from_backend(
|
| 182 |
+
messages, request.temperature, request.top_p, request.max_tokens
|
| 183 |
+
):
|
| 184 |
+
elapsed = parse_chunk(content, state)
|
| 185 |
+
|
| 186 |
+
sse_chunk = {
|
| 187 |
+
"id": chunk_id,
|
| 188 |
+
"object": "chat.completion.chunk",
|
| 189 |
+
"created": created,
|
| 190 |
+
"model": "qwen3-0.6b",
|
| 191 |
+
"choices": [{
|
| 192 |
+
"index": 0,
|
| 193 |
+
"delta": {"content": content},
|
| 194 |
+
"finish_reason": None
|
| 195 |
+
}],
|
| 196 |
+
"thinking": {
|
| 197 |
+
"in_progress": state.in_think,
|
| 198 |
+
"elapsed": elapsed if state.in_think else state.total_think_time
|
| 199 |
+
}
|
| 200 |
+
}
|
| 201 |
+
yield b"data: " + orjson.dumps(sse_chunk) + b"\n\n"
|
| 202 |
+
|
| 203 |
+
final_chunk = {
|
| 204 |
+
"id": chunk_id,
|
| 205 |
+
"object": "chat.completion.chunk",
|
| 206 |
+
"created": created,
|
| 207 |
+
"model": "qwen3-0.6b",
|
| 208 |
+
"choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}],
|
| 209 |
+
"thinking": {
|
| 210 |
+
"in_progress": False,
|
| 211 |
+
"total_think_time": state.total_think_time,
|
| 212 |
+
"thought_content": state.get_thought(),
|
| 213 |
+
"answer_content": state.get_answer()
|
| 214 |
+
}
|
| 215 |
+
}
|
| 216 |
+
yield b"data: " + orjson.dumps(final_chunk) + b"\n\n"
|
| 217 |
+
yield b"data: [DONE]\n\n"
|
| 218 |
+
|
| 219 |
+
except Exception as e:
|
| 220 |
+
yield b"data: " + orjson.dumps({"error": {"message": str(e)}}) + b"\n\n"
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
async def generate_complete_fast(request: ChatRequest) -> dict:
|
| 224 |
+
messages = [{"role": m.role, "content": m.content} for m in request.messages]
|
| 225 |
+
state = ParserState()
|
| 226 |
+
response_parts = []
|
| 227 |
+
|
| 228 |
+
try:
|
| 229 |
+
async for content in stream_from_backend(
|
| 230 |
+
messages, request.temperature, request.top_p, request.max_tokens
|
| 231 |
+
):
|
| 232 |
+
response_parts.append(content)
|
| 233 |
+
parse_chunk(content, state)
|
| 234 |
+
|
| 235 |
+
full_response = ''.join(response_parts)
|
| 236 |
+
|
| 237 |
+
return {
|
| 238 |
+
"id": f"chatcmpl-{int(time.time() * 1000)}",
|
| 239 |
+
"object": "chat.completion",
|
| 240 |
+
"created": int(time.time()),
|
| 241 |
+
"model": "qwen3-0.6b",
|
| 242 |
+
"choices": [{
|
| 243 |
+
"index": 0,
|
| 244 |
+
"message": {
|
| 245 |
+
"role": "assistant",
|
| 246 |
+
"content": full_response,
|
| 247 |
+
"thinking": {
|
| 248 |
+
"thought_content": state.get_thought(),
|
| 249 |
+
"answer_content": state.get_answer(),
|
| 250 |
+
"total_think_time": state.total_think_time
|
| 251 |
+
}
|
| 252 |
+
},
|
| 253 |
+
"finish_reason": "stop"
|
| 254 |
+
}]
|
| 255 |
+
}
|
| 256 |
+
except Exception as e:
|
| 257 |
+
raise HTTPException(status_code=500, detail=str(e))
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
# ===== Endpoints =====
|
| 261 |
+
|
| 262 |
+
@app.get("/")
|
| 263 |
+
async def root():
|
| 264 |
+
return {"status": "ok", "message": "Qwen3 API is running"}
|
| 265 |
+
|
| 266 |
+
|
| 267 |
+
@app.get("/health")
|
| 268 |
+
async def health():
|
| 269 |
+
try:
|
| 270 |
+
response = await http_client.get("/models")
|
| 271 |
+
return {"status": "healthy" if response.status_code == 200 else "unhealthy"}
|
| 272 |
+
except Exception as e:
|
| 273 |
+
return {"status": "unhealthy", "error": str(e)}
|
| 274 |
+
|
| 275 |
+
|
| 276 |
+
@app.get("/v1/models")
|
| 277 |
+
async def list_models():
|
| 278 |
+
return {
|
| 279 |
+
"object": "list",
|
| 280 |
+
"data": [{
|
| 281 |
+
"id": "qwen3-0.6b",
|
| 282 |
+
"object": "model",
|
| 283 |
+
"created": int(time.time()),
|
| 284 |
+
"owned_by": "local"
|
| 285 |
+
}]
|
| 286 |
+
}
|
| 287 |
+
|
| 288 |
+
|
| 289 |
+
@app.post("/v1/chat/completions")
|
| 290 |
+
async def chat_completions(request: ChatRequest):
|
| 291 |
+
if request.stream:
|
| 292 |
+
return StreamingResponse(
|
| 293 |
+
generate_stream_fast(request),
|
| 294 |
+
media_type="text/event-stream",
|
| 295 |
+
headers={
|
| 296 |
+
"Cache-Control": "no-cache",
|
| 297 |
+
"Connection": "keep-alive",
|
| 298 |
+
"X-Accel-Buffering": "no",
|
| 299 |
+
"Transfer-Encoding": "chunked"
|
| 300 |
+
}
|
| 301 |
+
)
|
| 302 |
+
return await generate_complete_fast(request)
|
| 303 |
+
|
| 304 |
+
|
| 305 |
+
@app.post("/chat")
|
| 306 |
+
async def simple_chat(request: SimpleChatRequest):
|
| 307 |
+
chat_request = ChatRequest(
|
| 308 |
+
messages=[Message(role="user", content=request.prompt)],
|
| 309 |
+
temperature=request.temperature,
|
| 310 |
+
top_p=request.top_p,
|
| 311 |
+
max_tokens=request.max_tokens,
|
| 312 |
+
stream=request.stream
|
| 313 |
+
)
|
| 314 |
+
|
| 315 |
+
if request.stream:
|
| 316 |
+
return StreamingResponse(
|
| 317 |
+
generate_stream_fast(chat_request),
|
| 318 |
+
media_type="text/event-stream",
|
| 319 |
+
headers={
|
| 320 |
+
"Cache-Control": "no-cache",
|
| 321 |
+
"Connection": "keep-alive",
|
| 322 |
+
"X-Accel-Buffering": "no"
|
| 323 |
+
}
|
| 324 |
+
)
|
| 325 |
+
return await generate_complete_fast(chat_request)
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
async def raw_stream_fast(request: ChatRequest) -> AsyncGenerator[bytes, None]:
|
| 329 |
+
messages = [{"role": m.role, "content": m.content} for m in request.messages]
|
| 330 |
+
|
| 331 |
+
try:
|
| 332 |
+
async for content in stream_from_backend(
|
| 333 |
+
messages, request.temperature, request.top_p, request.max_tokens
|
| 334 |
+
):
|
| 335 |
+
yield content.encode()
|
| 336 |
+
except Exception as e:
|
| 337 |
+
yield f"\n\nError: {str(e)}".encode()
|
| 338 |
+
|
| 339 |
+
|
| 340 |
+
@app.post("/chat/raw")
|
| 341 |
+
async def raw_chat(request: SimpleChatRequest):
|
| 342 |
+
chat_request = ChatRequest(
|
| 343 |
+
messages=[Message(role="user", content=request.prompt)],
|
| 344 |
+
temperature=request.temperature,
|
| 345 |
+
top_p=request.top_p,
|
| 346 |
+
max_tokens=request.max_tokens,
|
| 347 |
+
stream=True
|
| 348 |
+
)
|
| 349 |
+
|
| 350 |
+
return StreamingResponse(
|
| 351 |
+
raw_stream_fast(chat_request),
|
| 352 |
+
media_type="text/plain",
|
| 353 |
+
headers={
|
| 354 |
+
"Cache-Control": "no-cache",
|
| 355 |
+
"Connection": "keep-alive",
|
| 356 |
+
"X-Accel-Buffering": "no"
|
| 357 |
+
}
|
| 358 |
+
)
|
| 359 |
+
|
| 360 |
+
|
| 361 |
+
@app.post("/fast")
|
| 362 |
+
async def fast_chat(prompt: str = "", max_tokens: int = 512):
|
| 363 |
+
messages = [{"role": "user", "content": prompt}]
|
| 364 |
+
response_parts = []
|
| 365 |
+
|
| 366 |
+
async for content in stream_from_backend(messages, 0.6, 0.95, max_tokens):
|
| 367 |
+
response_parts.append(content)
|
| 368 |
+
|
| 369 |
+
return {"response": ''.join(response_parts)}
|
| 370 |
+
|
| 371 |
+
# ===== Mini-server load tracking & coordination endpoints =====
|
| 372 |
+
|
| 373 |
+
# How many concurrent requests this mini should handle
|
| 374 |
+
MAX_CONCURRENT_REQUESTS = int(os.environ.get("MAX_CONCURRENT_REQUESTS", "1"))
|
| 375 |
+
|
| 376 |
+
# In-memory tracking per process
|
| 377 |
+
current_requests = 0
|
| 378 |
+
|
| 379 |
+
# For identification / debugging
|
| 380 |
+
MINI_SERVER_ID = os.environ.get("MINI_SERVER_ID", "mini-1")
|
| 381 |
+
|
| 382 |
+
|
| 383 |
+
class MiniStatus(BaseModel):
|
| 384 |
+
server_id: str
|
| 385 |
+
max_concurrent: int
|
| 386 |
+
current_requests: int
|
| 387 |
+
status: str
|
| 388 |
+
|
| 389 |
+
|
| 390 |
+
@app.get("/status")
|
| 391 |
+
async def mini_status():
|
| 392 |
+
"""
|
| 393 |
+
Used by the main server to know if this mini is idle/busy.
|
| 394 |
+
"""
|
| 395 |
+
status = "busy" if current_requests >= MAX_CONCURRENT_REQUESTS else "idle"
|
| 396 |
+
return MiniStatus(
|
| 397 |
+
server_id=MINI_SERVER_ID,
|
| 398 |
+
max_concurrent=MAX_CONCURRENT_REQUESTS,
|
| 399 |
+
current_requests=current_requests,
|
| 400 |
+
status=status,
|
| 401 |
+
)
|
| 402 |
+
|
| 403 |
+
|
| 404 |
+
@app.post("/reserve")
|
| 405 |
+
async def reserve_slot():
|
| 406 |
+
"""
|
| 407 |
+
Called by the main server BEFORE it forwards a chat request.
|
| 408 |
+
If this mini is full, returns 429 so main server can try another mini.
|
| 409 |
+
"""
|
| 410 |
+
global current_requests
|
| 411 |
+
if current_requests >= MAX_CONCURRENT_REQUESTS:
|
| 412 |
+
raise HTTPException(status_code=429, detail="Mini server busy")
|
| 413 |
+
current_requests += 1
|
| 414 |
+
return {
|
| 415 |
+
"server_id": MINI_SERVER_ID,
|
| 416 |
+
"current_requests": current_requests,
|
| 417 |
+
"max_concurrent": MAX_CONCURRENT_REQUESTS,
|
| 418 |
+
}
|
| 419 |
+
|
| 420 |
+
|
| 421 |
+
@app.post("/release")
|
| 422 |
+
async def release_slot():
|
| 423 |
+
"""
|
| 424 |
+
Called by the main server after request is finished (stream closed/response sent).
|
| 425 |
+
"""
|
| 426 |
+
global current_requests
|
| 427 |
+
if current_requests > 0:
|
| 428 |
+
current_requests -= 1
|
| 429 |
+
return {
|
| 430 |
+
"server_id": MINI_SERVER_ID,
|
| 431 |
+
"current_requests": current_requests,
|
| 432 |
+
"max_concurrent": MAX_CONCURRENT_REQUESTS,
|
| 433 |
+
}
|
| 434 |
+
|
| 435 |
+
|
| 436 |
+
if __name__ == "__main__":
|
| 437 |
+
uvicorn.run(
|
| 438 |
+
app,
|
| 439 |
+
host="0.0.0.0",
|
| 440 |
+
port=7860,
|
| 441 |
+
loop="uvloop",
|
| 442 |
+
http="httptools",
|
| 443 |
+
access_log=False,
|
| 444 |
+
workers=1
|
| 445 |
+
)
|
start.sh
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
cd /llama.cpp/build
|
| 2 |
+
./bin/llama-server \
|
| 3 |
+
--host 0.0.0.0 \
|
| 4 |
+
--port 8080 \
|
| 5 |
+
--model /models/model.gguf \
|
| 6 |
+
--ctx-size 32768 \
|
| 7 |
+
--threads 2 &
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
echo "Waiting for llama.cpp server..."
|
| 11 |
+
until curl -s "http://localhost:8080/v1/models" >/dev/null 2>&1; do
|
| 12 |
+
sleep 1
|
| 13 |
+
done
|
| 14 |
+
echo "llama.cpp server is ready."
|
| 15 |
+
|
| 16 |
+
# Start FastAPI
|
| 17 |
+
echo "Starting FastAPI server on port 7860..."
|
| 18 |
+
cd /
|
| 19 |
+
python3 -m uvicorn app:app --host 0.0.0.0 --port 7860
|