Create app.py

app.py

@@ -0,0 +1,105 @@
+from fastapi import FastAPI
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    TextIteratorStreamer
+)
+
+import torch
+from threading import Thread
+
+# ============================================
+# MODEL
+# ============================================
+
+MODEL_NAME = "junaid17/qwen-0.5b-16bit_merged"
+
+tokenizer = AutoTokenizer.from_pretrained(
+    MODEL_NAME,
+    trust_remote_code=True
+)
+
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_NAME,
+    trust_remote_code=True,
+    torch_dtype=torch.float16,
+    device_map="auto"
+)
+
+# ============================================
+# FASTAPI
+# ============================================
+
+app = FastAPI()
+
+# ============================================
+# REQUEST SCHEMA
+# ============================================
+
+class ChatRequest(BaseModel):
+    query: str
+    max_new_tokens: int = 256
+    temperature: float = 0.7
+
+# ============================================
+# STREAM CHAT
+# ============================================
+
+@app.post("/chat")
+async def chat(request: ChatRequest):
+
+    messages = [
+        {
+            "role": "system",
+            "content": "You are a helpful AI assistant."
+        },
+        {
+            "role": "user",
+            "content": request.query
+        }
+    ]
+
+    prompt = tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=True
+    )
+
+    inputs = tokenizer(
+        prompt,
+        return_tensors="pt"
+    ).to(model.device)
+
+    streamer = TextIteratorStreamer(
+        tokenizer,
+        skip_prompt=True,
+        skip_special_tokens=True
+    )
+
+    generation_kwargs = dict(
+        **inputs,
+        streamer=streamer,
+        max_new_tokens=request.max_new_tokens,
+        temperature=request.temperature,
+        do_sample=True,
+        pad_token_id=tokenizer.eos_token_id
+    )
+
+    thread = Thread(
+        target=model.generate,
+        kwargs=generation_kwargs
+    )
+
+    thread.start()
+
+    def generate_tokens():
+        for token in streamer:
+            yield token
+
+    return StreamingResponse(
+        generate_tokens(),
+        media_type="text/plain"
+    )