Upload mirothinker server.py
Browse files- mirothinker server.py +119 -0
mirothinker server.py
ADDED
|
@@ -0,0 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Minimal OpenAI-compatible server for MiroThinker using plain transformers.
|
| 2 |
+
# No vLLM/SGLang needed. Works wherever PyTorch CUDA works (incl. Windows).
|
| 3 |
+
#
|
| 4 |
+
# pip install transformers accelerate fastapi uvicorn torch
|
| 5 |
+
# python mirothinker_server.py
|
| 6 |
+
#
|
| 7 |
+
# Then set in miroflow-agent .env / config:
|
| 8 |
+
# base_url: http://localhost:61005/v1 (provider "qwen", any api_key)
|
| 9 |
+
#
|
| 10 |
+
# Handles the two non-standard params miroflow-agent sends via extra_body:
|
| 11 |
+
# - repetition_penalty
|
| 12 |
+
# - continue_final_message / add_generation_prompt (resume truncated output)
|
| 13 |
+
|
| 14 |
+
import time
|
| 15 |
+
import uuid
|
| 16 |
+
|
| 17 |
+
import torch
|
| 18 |
+
import uvicorn
|
| 19 |
+
from fastapi import FastAPI
|
| 20 |
+
from pydantic import BaseModel
|
| 21 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 22 |
+
|
| 23 |
+
MODEL_ID = "miromind-ai/MiroThinker-v1.5-30B" # change if using 1.7 etc.
|
| 24 |
+
PORT = 61005
|
| 25 |
+
|
| 26 |
+
print(f"Loading {MODEL_ID} ... (this takes a while)")
|
| 27 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 28 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 29 |
+
MODEL_ID,
|
| 30 |
+
torch_dtype=torch.bfloat16,
|
| 31 |
+
device_map="auto", # splits layers across both A100s
|
| 32 |
+
trust_remote_code=True,
|
| 33 |
+
)
|
| 34 |
+
model.eval()
|
| 35 |
+
print("Model loaded. Device map:", model.hf_device_map if hasattr(model, "hf_device_map") else "n/a")
|
| 36 |
+
|
| 37 |
+
app = FastAPI()
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
class ChatRequest(BaseModel):
|
| 41 |
+
model: str = MODEL_ID
|
| 42 |
+
messages: list
|
| 43 |
+
max_tokens: int | None = 16384
|
| 44 |
+
max_completion_tokens: int | None = None
|
| 45 |
+
temperature: float = 1.0
|
| 46 |
+
top_p: float = 0.95
|
| 47 |
+
stream: bool = False
|
| 48 |
+
# extra_body params arrive as top-level fields with the openai SDK
|
| 49 |
+
repetition_penalty: float | None = None
|
| 50 |
+
continue_final_message: bool | None = None
|
| 51 |
+
add_generation_prompt: bool | None = None
|
| 52 |
+
|
| 53 |
+
class Config:
|
| 54 |
+
extra = "allow"
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
@app.get("/health")
|
| 58 |
+
def health():
|
| 59 |
+
return {"status": "ok"}
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
@app.get("/v1/models")
|
| 63 |
+
def models():
|
| 64 |
+
return {"object": "list", "data": [{"id": MODEL_ID, "object": "model"}]}
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
@app.post("/v1/chat/completions")
|
| 68 |
+
def chat(req: ChatRequest):
|
| 69 |
+
continue_final = bool(req.continue_final_message)
|
| 70 |
+
|
| 71 |
+
# Build the prompt with the model's own chat template.
|
| 72 |
+
input_ids = tokenizer.apply_chat_template(
|
| 73 |
+
req.messages,
|
| 74 |
+
add_generation_prompt=not continue_final,
|
| 75 |
+
continue_final_message=continue_final,
|
| 76 |
+
return_tensors="pt",
|
| 77 |
+
).to(model.device)
|
| 78 |
+
|
| 79 |
+
prompt_tokens = input_ids.shape[-1]
|
| 80 |
+
max_new = req.max_completion_tokens or req.max_tokens or 16384
|
| 81 |
+
|
| 82 |
+
gen_kwargs = dict(
|
| 83 |
+
max_new_tokens=max_new,
|
| 84 |
+
do_sample=req.temperature > 0,
|
| 85 |
+
temperature=max(req.temperature, 1e-5),
|
| 86 |
+
top_p=req.top_p,
|
| 87 |
+
pad_token_id=tokenizer.eos_token_id,
|
| 88 |
+
)
|
| 89 |
+
if req.repetition_penalty and req.repetition_penalty != 1.0:
|
| 90 |
+
gen_kwargs["repetition_penalty"] = req.repetition_penalty
|
| 91 |
+
|
| 92 |
+
with torch.inference_mode():
|
| 93 |
+
out = model.generate(input_ids, **gen_kwargs)
|
| 94 |
+
|
| 95 |
+
new_tokens = out[0][prompt_tokens:]
|
| 96 |
+
text = tokenizer.decode(new_tokens, skip_special_tokens=True)
|
| 97 |
+
completion_tokens = new_tokens.shape[-1]
|
| 98 |
+
finish = "length" if completion_tokens >= max_new else "stop"
|
| 99 |
+
|
| 100 |
+
return {
|
| 101 |
+
"id": f"chatcmpl-{uuid.uuid4().hex[:12]}",
|
| 102 |
+
"object": "chat.completion",
|
| 103 |
+
"created": int(time.time()),
|
| 104 |
+
"model": req.model,
|
| 105 |
+
"choices": [{
|
| 106 |
+
"index": 0,
|
| 107 |
+
"message": {"role": "assistant", "content": text},
|
| 108 |
+
"finish_reason": finish,
|
| 109 |
+
}],
|
| 110 |
+
"usage": {
|
| 111 |
+
"prompt_tokens": prompt_tokens,
|
| 112 |
+
"completion_tokens": completion_tokens,
|
| 113 |
+
"total_tokens": prompt_tokens + completion_tokens,
|
| 114 |
+
},
|
| 115 |
+
}
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
if __name__ == "__main__":
|
| 119 |
+
uvicorn.run(app, host="0.0.0.0", port=PORT)
|