Upload serve_ministral.py with huggingface_hub

serve_ministral.py

@@ -5,7 +5,7 @@ Usage: python serve_ministral.py
 """
 
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoProcessor, AutoModelForImageTextToText
 from fastapi import FastAPI
 from pydantic import BaseModel
 from typing import List, Optional
@@ -16,7 +16,7 @@ app = FastAPI()
 
 # Global model and tokenizer
 model = None
-tokenizer = None
+processor = None
 
 class Message(BaseModel):
     role: str
@@ -39,15 +39,14 @@ class ChatResponse(BaseModel):
 
 @app.on_event("startup")
 async def load_model():
-    global model, tokenizer
+    global model, processor
     print("Loading Ministral 14B...")
 
     model_id = "RoleModel/ministral-14b-merged-official"
 
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+    processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
 
-    # Load just the text model weights
-    model = AutoModelForCausalLM.from_pretrained(
+    model = AutoModelForImageTextToText.from_pretrained(
         model_id,
         torch_dtype=torch.bfloat16,
         device_map="auto",
@@ -58,30 +57,33 @@ async def load_model():
 
 @app.post("/v1/chat/completions")
 async def chat_completions(request: ChatRequest):
-    global model, tokenizer
+    global model, processor
 
     # Format messages using chat template
-    chat_text = tokenizer.apply_chat_template(
-        [{"role": m.role, "content": m.content} for m in request.messages],
+    messages = [{"role": m.role, "content": m.content} for m in request.messages]
+
+    chat_text = processor.apply_chat_template(
+        messages,
         tokenize=False,
         add_generation_prompt=True
     )
 
-    inputs = tokenizer(chat_text, return_tensors="pt").to(model.device)
+    inputs = processor(text=chat_text, return_tensors="pt").to(model.device)
 
     with torch.no_grad():
         outputs = model.generate(
             **inputs,
             max_new_tokens=request.max_tokens,
-            temperature=request.temperature,
+            temperature=request.temperature if request.temperature > 0 else None,
             top_p=request.top_p,
-            do_sample=True,
-            pad_token_id=tokenizer.eos_token_id,
+            do_sample=request.temperature > 0,
+            pad_token_id=processor.tokenizer.eos_token_id,
         )
 
     # Decode only the new tokens
-    new_tokens = outputs[0][inputs["input_ids"].shape[1]:]
-    response_text = tokenizer.decode(new_tokens, skip_special_tokens=True)
+    input_len = inputs["input_ids"].shape[1]
+    new_tokens = outputs[0][input_len:]
+    response_text = processor.decode(new_tokens, skip_special_tokens=True)
 
     return ChatResponse(
         id=f"chatcmpl-{int(time.time())}",
@@ -93,9 +95,9 @@ async def chat_completions(request: ChatRequest):
             "finish_reason": "stop"
         }],
         usage={
-            "prompt_tokens": inputs["input_ids"].shape[1],
+            "prompt_tokens": input_len,
             "completion_tokens": len(new_tokens),
-            "total_tokens": inputs["input_ids"].shape[1] + len(new_tokens)
+            "total_tokens": input_len + len(new_tokens)
         }
     )