Update app.py

app.py

@@ -12,6 +12,7 @@ import torch
 from fastapi.responses import StreamingResponse
 from tqdm import tqdm
 import logging
+import json
 
 load_dotenv()
 
@@ -49,6 +50,7 @@ class S3DirectStream:
         )
         self.bucket_name = bucket_name
 
+
     def stream_from_s3(self, key):
         try:
             logger.info(f"Downloading {key} from S3...")
@@ -71,6 +73,7 @@ class S3DirectStream:
     def load_model_from_stream(self, model_prefix, revision):
         try:
             logger.info(f"Loading model {model_prefix} (revision {revision})...")
+
             if self.file_exists_in_s3(f"{model_prefix}/config.json") and \
                (self.file_exists_in_s3(f"{model_prefix}/pytorch_model.bin") or self.file_exists_in_s3(f"{model_prefix}/model.safetensors")):
                 logger.info(f"Model {model_prefix} found in S3. Loading...")
@@ -84,18 +87,22 @@ class S3DirectStream:
             logger.error(f"Error loading model: {e}")
             return None
 
+
     def load_model_from_existing_s3(self, model_prefix):
         logger.info(f"Loading config for {model_prefix} from S3...")
         config_stream = self.stream_from_s3(f"{model_prefix}/config.json")
-        config = AutoConfig.from_pretrained(config_stream)
+        config_dict = json.load(config_stream)
+        config = AutoConfig.from_pretrained(config_dict)
         logger.info(f"Config loaded for {model_prefix}.")
 
+
         if self.file_exists_in_s3(f"{model_prefix}/model.safetensors"):
             logger.info(f"Loading safetensors model for {model_prefix} from S3...")
             model_stream = self.stream_from_s3(f"{model_prefix}/model.safetensors")
             model = AutoModelForCausalLM.from_config(config)
             model.load_state_dict(safetensors.torch.load_stream(model_stream))
             logger.info(f"Safetensors model loaded for {model_prefix}.")
+
         elif self.file_exists_in_s3(f"{model_prefix}/pytorch_model.bin"):
             logger.info(f"Loading PyTorch model for {model_prefix} from S3...")
             model_stream = self.stream_from_s3(f"{model_prefix}/pytorch_model.bin")
@@ -106,15 +113,18 @@ class S3DirectStream:
         else:
             logger.error(f"No model file found for {model_prefix} in S3")
             raise EnvironmentError(f"No model file found for {model_prefix} in S3")
+
         return model
 
     def load_tokenizer_from_stream(self, model_prefix):
         try:
             logger.info(f"Loading tokenizer for {model_prefix}...")
+
             if self.file_exists_in_s3(f"{model_prefix}/tokenizer.json"):
                 logger.info(f"Tokenizer for {model_prefix} found in S3. Loading...")
                 return self.load_tokenizer_from_existing_s3(model_prefix)
 
+
             logger.info(f"Tokenizer for {model_prefix} not found in S3. Downloading and uploading...")
             self.download_and_upload_to_s3(model_prefix)
             logger.info(f"Downloaded and uploaded tokenizer for {model_prefix}. Loading from S3...")
@@ -123,20 +133,24 @@ class S3DirectStream:
             logger.error(f"Error loading tokenizer: {e}")
             return None
 
+
     def load_tokenizer_from_existing_s3(self, model_prefix):
         logger.info(f"Loading tokenizer from S3 for {model_prefix}...")
         tokenizer_stream = self.stream_from_s3(f"{model_prefix}/tokenizer.json")
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_stream)
+        tokenizer_config = json.load(tokenizer_stream) # Corrected this line
+        tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=None, config=tokenizer_config) # Corrected this line
         logger.info(f"Tokenizer loaded for {model_prefix}.")
         return tokenizer
 
 
+
     def download_and_upload_to_s3(self, model_prefix, revision="main"):
         model_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/pytorch_model.bin"
         safetensors_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/model.safetensors"
         tokenizer_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/tokenizer.json"
         config_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/config.json"
 
+
         logger.info(f"Downloading and uploading model files for {model_prefix} to S3...")
         self.download_and_upload_to_s3_url(model_url, f"{model_prefix}/pytorch_model.bin")
         self.download_and_upload_to_s3_url(safetensors_url, f"{model_prefix}/model.safetensors")
@@ -146,15 +160,18 @@ class S3DirectStream:
 
     def download_and_upload_to_s3_url(self, url, s3_key):
         logger.info(f"Downloading from {url}...")
+
         with requests.get(url, stream=True) as response:
             if response.status_code == 200:
                 total_size_in_bytes= int(response.headers.get('content-length', 0))
                 block_size = 1024
                 progress_bar = tqdm(total=total_size_in_bytes, unit='iB', unit_scale=True)
                 logger.info(f"Uploading to S3: {s3_key}...")
+
                 self.s3_client.upload_fileobj(response.raw, self.bucket_name, s3_key, Callback=lambda bytes_transferred: progress_bar.update(bytes_transferred))
                 progress_bar.close()
                 logger.info(f"Uploaded {s3_key} to S3 successfully.")
+
             elif response.status_code == 404:
                 logger.error(f"File not found at {url}")
                 raise HTTPException(status_code=404, detail=f"Error downloading file from {url}. File not found.")
@@ -167,26 +184,31 @@ class S3DirectStream:
 async def predict(model_request: DownloadModelRequest):
     try:
         logger.info(f"Received request: Model={model_request.model_name}, Task={model_request.pipeline_task}, Input={model_request.input_text}")
+
         model_name = model_request.model_name
         revision = model_request.revision
 
         streamer = S3DirectStream(S3_BUCKET_NAME)
-        logger.info(f"Loading model and tokenizer...")
+        logger.info("Loading model and tokenizer...")
         model = streamer.load_model_from_stream(model_name, revision)
+
         tokenizer = streamer.load_tokenizer_from_stream(model_name)
-        logger.info(f"Model and tokenizer loaded.")
+        logger.info("Model and tokenizer loaded.")
+
 
         task = model_request.pipeline_task
         if task not in ["text-generation", "sentiment-analysis", "translation", "fill-mask", "question-answering",  "summarization", "zero-shot-classification"]:
             raise HTTPException(status_code=400, detail="Unsupported pipeline task")
 
+
         if task == "text-generation":
-            logger.info(f"Starting text generation...")
+            logger.info("Starting text generation...")
             text_streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
             inputs = tokenizer(model_request.input_text, return_tensors="pt").to(model.device)
             generation_kwargs = dict(inputs, streamer=text_streamer)
             model.generate(**generation_kwargs)
-            logger.info(f"Text generation finished.")
+            logger.info("Text generation finished.")
+
             return StreamingResponse(iter([tokenizer.decode(token) for token in text_streamer]), media_type="text/event-stream")
 
         else:
@@ -196,10 +218,13 @@ async def predict(model_request: DownloadModelRequest):
             logger.info(f"Pipeline task {task} finished.")
             return {"result": outputs}
 
+
+
     except Exception as e:
         logger.exception(f"Error processing request: {e}")
         raise HTTPException(status_code=500, detail=f"Error processing request: {str(e)}")
 
 
+
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=7860)