File size: 2,978 Bytes
4625807 ec97b47 4625807 d6ff847 4625807 76f1775 4625807 76f1775 4625807 76f1775 4625807 7e2c46b 4625807 76f1775 4625807 76f1775 4625807 76f1775 4625807 76f1775 |
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 |
# vllm_backend.py
import time, logging
from typing import Any, Dict, AsyncIterable
from vllm.sampling_params import SamplingParams
from backends_base import ChatBackend, ImagesBackend
logger = logging.getLogger(__name__)
try:
import spaces
except ImportError:
spaces = None
class VLLMChatBackend(ChatBackend):
"""
On ZeroGPU: build vLLM engine per request (no persistent state).
Returns a single ChatCompletionChunk with the full text.
"""
async def stream(self, request: Dict[str, Any]) -> AsyncIterable[Dict[str, Any]]:
messages = request.get("messages", [])
prompt = messages[-1]["content"] if messages else "(empty)"
params = SamplingParams(
temperature=float(request.get("temperature", 0.7)),
max_tokens=int(request.get("max_tokens", 512))
)
rid = f"chatcmpl-local-{int(time.time())}"
now = int(time.time())
model_name = request.get("model", "local-vllm")
# GPU wrapper for ZeroGPU
if spaces:
@spaces.GPU(duration=60)
def run_once(prompt: str) -> str:
from vllm.engine.async_llm_engine import AsyncLLMEngine
from vllm.engine.arg_utils import AsyncEngineArgs
args = AsyncEngineArgs(model=model_name, trust_remote_code=True)
engine = AsyncLLMEngine.from_engine_args(args)
# synchronous generate
outputs = list(engine.generate(prompt, params, request_id=rid))
return outputs[-1].outputs[0].text if outputs else ""
else:
def run_once(prompt: str) -> str:
from vllm.engine.async_llm_engine import AsyncLLMEngine
from vllm.engine.arg_utils import AsyncEngineArgs
args = AsyncEngineArgs(model=model_name, trust_remote_code=True)
engine = AsyncLLMEngine.from_engine_args(args)
outputs = list(engine.generate(prompt, params, request_id=rid))
return outputs[-1].outputs[0].text if outputs else ""
try:
text = run_once(prompt)
yield {
"id": rid,
"object": "chat.completion.chunk",
"created": now,
"model": model_name,
"choices": [
{"index": 0, "delta": {"content": text}, "finish_reason": "stop"}
],
}
except Exception:
logger.exception("vLLM inference failed")
raise
class StubImagesBackend(ImagesBackend):
"""
vLLM does not support image generation.
For now, return a transparent PNG placeholder.
"""
async def generate_b64(self, request: Dict[str, Any]) -> str:
logger.warning("Image generation not supported in local vLLM backend.")
return (
"iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR4nGP4BwQACfsD/etCJH0AAAAASUVORK5CYII="
)
|