Added download of model extras into weights folder, separate download of tokenizer, making engine.py utilize downloaded tokenizer, model and tokenizer revision

Dockerfile

@@ -15,10 +15,16 @@ COPY src /src
 
 # Setup for Option 2: Building the Image with the Model included
 ARG MODEL_NAME=""
+ARG TOKENIZER_NAME=""
 ARG BASE_PATH="/runpod-volume"
 ARG QUANTIZATION=""
+ARG MODEL_REVISION="main"
+ARG TOKENIZER_REVISION="main"
 
 ENV MODEL_NAME=$MODEL_NAME \
+    MODEL_REVISION=$REVISION \
+    TOKENIZER_NAME=$TOKENIZER_NAME \
+    TOKENIZER_REVISION=$TOKENIZER_REVISION \
     BASE_PATH=$BASE_PATH \
     QUANTIZATION=$QUANTIZATION \
     HF_DATASETS_CACHE="${BASE_PATH}/huggingface-cache/datasets" \

README.md

@@ -58,8 +58,7 @@ Development Image: ```runpod/worker-vllm:dev```
 
 **Optional**:
 - LLM Settings:
-  - `TOKENIZER_NAME`: Tokenizer repository if you would like to use a different tokenizer than the one that comes with the model. (default: `None`)
+  - `MODEL_REVISION`: Model revision to load (default: `main`).
-  - `CUSTOM_CHAT_TEMPLATE`: Custom chat jinja template, read more about Hugging Face chat templates [here](https://huggingface.co/docs/transformers/chat_templating). (default: `None`) 
   - `MAX_MODEL_LENGTH`: Maximum number of tokens for the engine to be able to handle. (default: maximum supported by the model)
   - `BASE_PATH`: Storage directory where huggingface cache and model will be located. (default: `/runpod-volume`, which will utilize network storage if you attach it or create a local directory within the image if you don't)
   - `LOAD_FORMAT`: Format to load model in (default: `auto`).
@@ -67,8 +66,12 @@ Development Image: ```runpod/worker-vllm:dev```
   - `QUANTIZATION`: AWQ (`awq`), SqueezeLLM (`squeezellm`) or GPTQ (`gptq`) Quantization. The specified Model Repo must be of a quantized model. (default: `None`)
   - `TRUST_REMOTE_CODE`: Trust remote code for Hugging Face (default: `0`)
   
-- Tensor Parallelism:
+- Tokenizer Settings:
+  - `TOKENIZER_NAME`: Tokenizer repository if you would like to use a different tokenizer than the one that comes with the model. (default: `None`, which uses the model's tokenizer)
+  - `TOKENIZER_REVISION`: Tokenizer revision to load (default: `main`).
+  - `CUSTOM_CHAT_TEMPLATE`: Custom chat jinja template, read more about Hugging Face chat templates [here](https://huggingface.co/docs/transformers/chat_templating). (default: `None`) 
 
+- Tensor Parallelism:
   Note that the more GPUs you split a model's weights accross, the slower it will be due to inter-GPU communication overhead. If you can fit the model on a single GPU, it is recommended to do so. 
   - `TENSOR_PARALLEL_SIZE`: Number of GPUs to shard the model across (default: `1`).
   - If you are having issues loading your model with Tensor Parallelism, try decreasing `VLLM_CPU_FRACTION` (default: `1`).
@@ -96,9 +99,12 @@ To build an image with the model baked in, you must specify the following docker
 - **Required**
   - `MODEL_NAME`
 - **Optional**
+  - `MODEL_REVISION`: Model revision to load (default: `main`).
   - `BASE_PATH`: Storage directory where huggingface cache and model will be located. (default: `/runpod-volume`, which will utilize network storage if you attach it or create a local directory within the image if you don't. If your intention is to bake the model into the image, you should set this to something like `/models` to make sure there are no issues if you were to accidentally attach network storage.)
   - `QUANTIZATION`
   - `WORKER_CUDA_VERSION`: `11.8.0` or `12.1.0` (default: `11.8.0` due to a small amount of workers not having CUDA 12.1 support yet. `12.1.0` is recommended for optimal performance).
+  - `TOKENIZER_NAME`: Tokenizer repository if you would like to use a different tokenizer than the one that comes with the model. (default: `None`, which uses the model's tokenizer)
+  - `TOKENIZER_REVISION`: Tokenizer revision to load (default: `main`).
 
 For the remaining settings, you may apply them as environment variables when running the container. Supported environment variables are listed in the [Environment Variables](#environment-variables) section.
 

src/download_model.py

@@ -1,11 +1,16 @@
 import os
+import shutil
 import logging
 from huggingface_hub import snapshot_download
-from vllm.model_executor.weight_utils import prepare_hf_model_weights
+from vllm.model_executor.weight_utils import Disabledtqdm, prepare_hf_model_weights
 
 if __name__ == "__main__":
     model = os.getenv("MODEL_NAME")
     download_dir = os.getenv("HF_HOME")
+    tokenizer = os.getenv("TOKENIZER_NAME", model)
+    model_revision = os.getenv("MODEL_REVISION", "main")
+    tokenizer_revision = os.getenv("TOKENIZER_REVISION", "main")
+    
     if not model or not download_dir:
         raise ValueError(f"Must specify model and download_dir. Model: {model}, download_dir: {download_dir}")
 
@@ -14,22 +19,41 @@ if __name__ == "__main__":
     
     logging.info(f"Downloading model {model} to {download_dir}")
         
-    hf_folder, hf_weights_files, use_safetensors = prepare_hf_model_weights(
+    model_folder, hf_weights_files, use_safetensors = prepare_hf_model_weights(
         model_name_or_path=model,
         cache_dir=download_dir,
-    )
+        revision=model_revision,
-
-    snapshot_download(
-        model,
-        cache_dir=download_dir,
-        allow_patterns=[
-            "*token*",
-            "config.json",
-        ]
     )
     
-    logging.info(f"Finished downloading model {model} to {download_dir}")
+    model_extras_folder = snapshot_download(model,
+                      allow_patterns=[
+                          "*token*",
+                          "*.json"
+                          ],
+                      cache_dir=download_dir + "/extras",
+                      tqdm_class=Disabledtqdm,
+                      revision=model_revision)
+    
+    # Move extras to hf_folder
+    for f in os.listdir(model_extras_folder):
+        shutil.move(model_extras_folder + "/" + f, model_folder + "/" + f)
     
     # Wrie hf_folder to file
     with open("/local_model_path.txt", "w") as f:
-        f.write(hf_folder) 
+        f.write(model_folder) 
+        
+    logging.info(f"Finished downloading model {model} to {download_dir}")
+    
+    if tokenizer != model:
+        logging.info(f"Downloading tokenizer {tokenizer} to {download_dir}")
+        tokenizer_folder = snapshot_download(tokenizer,
+                          cache_dir=download_dir,
+                          tqdm_class=Disabledtqdm,
+                          revision=tokenizer_revision)
+        with open("/local_tokenizer_path.txt", "w") as f:
+            f.write(tokenizer_folder) 
+        
+        logging.info(f"Finished downloading tokenizer {tokenizer} to {download_dir}")
+    
+  
+    

src/engine.py

@@ -41,7 +41,7 @@ class vLLMEngine:
         load_dotenv() # For local development
         self.config = self._initialize_config()
         logging.info("vLLM config: %s", self.config)
-        self.tokenizer = Tokenizer(os.getenv("TOKENIZER_NAME", os.getenv("MODEL_NAME")))
+        self.tokenizer = Tokenizer(self.config["tokenizer_name_or_path"])
         self.llm = self._initialize_llm() if engine is None else engine
         self.openai_engine = self._initialize_openai()
         self.max_concurrency = int(os.getenv("MAX_CONCURRENCY", DEFAULT_MAX_CONCURRENCY))
@@ -166,6 +166,9 @@ class vLLMEngine:
     def _initialize_config(self):
         quantization = self._get_quantization()
         model, download_dir = self._get_model_name_and_path()
+        tokenizer_name_or_path = self._get_tokenizer_name_or_path()
+        if not tokenizer_name_or_path:
+            tokenizer_name_or_path = model
         
         return {
             "model": model,
@@ -173,7 +176,7 @@ class vLLMEngine:
             "quantization": quantization,
             "load_format": os.getenv("LOAD_FORMAT", "auto"),
             "dtype": "half" if quantization else "auto",
-            "tokenizer": os.getenv("TOKENIZER_NAME"),
+            "tokenizer": tokenizer_name_or_path,
             "disable_log_stats": bool(int(os.getenv("DISABLE_LOG_STATS", 1))),
             "disable_log_requests": bool(int(os.getenv("DISABLE_LOG_REQUESTS", 1))),
             "trust_remote_code": bool(int(os.getenv("TRUST_REMOTE_CODE", 0))),
@@ -209,6 +212,14 @@ class vLLMEngine:
         else:
             model, download_dir = os.getenv("MODEL_NAME"), os.getenv("HF_HOME")  
         return model, download_dir
+    
+    def _get_tokenizer_name_or_path(self):
+        if os.path.exists("/local_tokenizer_path.txt"):
+            tokenizer_name_or_path = open("/local_tokenizer_path.txt", "r").read().strip()
+            logging.info("Using local tokenizer at %s", tokenizer_name_or_path)
+        else:
+            tokenizer_name_or_path = os.getenv("TOKENIZER_NAME")
+        return tokenizer_name_or_path
         
     def _get_num_gpu_shard(self):
         num_gpu_shard = int(os.getenv("TENSOR_PARALLEL_SIZE", 1))