[Distributed] Support index + multi-bin loading

Author: kwen2501

torchchat/distributed/checkpoint_utils.py

@@ -19,7 +19,8 @@
 from torchchat.cli.builder import BuilderArgs, _load_checkpoint
 
 
-_DEFAULT_SAFETENSOR_FILE_NAME = "model.safetensors.index.json"
+_DEFAULT_SAFETENSOR_INDEX = "model.safetensors.index.json"
+_DEFAULT_BIN_INDEX = "pytorch_model.bin.index.json"
 _CONFIG_NAME = "config.json"
 
 
@@ -81,31 +82,6 @@ def get_hf_path_from_model_id(model_id: str) -> str:
     return file_location
 
 
-def get_hf_weight_map_and_path(
-    model_id: str,
-) -> Tuple[Dict[str, str], str, Dict[str, str]]:
-    """Get the weight map for a given HF model id and also the cache path for loading the weights"""
-    index_file = cached_file(model_id, _DEFAULT_SAFETENSOR_FILE_NAME)
-    if not os.path.exists(index_file):
-        raise FileNotFoundError(
-            f"Weight index file for {model_id} does not exist in HF cache."
-        )
-    logger.info(
-        f"Loading weight map from: {index_file}"
-    )
-    weight_map = read_weights_from_json(index_file)
-    if weight_map is None:
-        raise ValueError(f"Weight map not found in config file {index_file}")
-    weight_map, new_to_old_keymap = remap_weight_keys(weight_map)
-    weight_path = os.path.dirname(index_file)
-    if not os.path.exists(weight_path):
-        raise FileNotFoundError(f"Weight path {weight_path} does not exist")
-    logger.info(
-        f"Loading weights from: {weight_path}"
-    )
-    return weight_map, weight_path, new_to_old_keymap
-
-
 def remap_weight_keys(dictionary):
     """Remap the keys of a dictionary to match the expected format of the tune model."""
     # hf_key : dist_model_key
@@ -141,12 +117,13 @@ def remap_weight_keys(dictionary):
     return new_dict, key_mapping
 
 
-def load_safetensor_weights(
+def load_weights_per_map(
     stage_module: Module,
     weight_map: Dict[str, str],
     file_location: str,
     new_to_old_keymap: Dict[str, str],
-    device: torch.device = "cuda",
+    device: torch.device,
+    is_safetensor: bool,
     purge_model_prefix: bool = True,
     ignore_cache_layers: bool = True,
     model_config: Optional[Dict] = None,
@@ -160,6 +137,7 @@ def load_safetensor_weights(
         file_location (str): Directory containing the weight files.
         new_to_old_keymap (Dict[str, str]): Mapping of new parameter names to old ones.
         device (torch.device): The device to load tensors onto.
+        is_safetensor (bool): Whether the files are safetensors.
         purge_model_prefix (bool): Whether to remove 'model.' prefix from keys.
         ignore_cache_layers (bool): Whether to ignore cache layers when reporting missing keys.
         model_config (Optional[Dict]): Model configuration.
@@ -178,9 +156,13 @@ def load_safetensor_weights(
     for file in needed_files:
         full_path = os.path.join(file_location, file)
         # logger.info(f"Loading checkpoint file: {full_path}")
-        try:
-            checkpoint = load_safetensor_file(full_path, "cpu")  # device)
+        # TODO: directly load to device
+        if is_safetensor:
+            checkpoint = load_safetensor_file(full_path)
+        else:
+            checkpoint = torch.load(full_path, mmap=True, weights_only=True)
 
+        try:
             update_state_dict(
                 stage_state_dict,
                 checkpoint,
@@ -189,10 +171,9 @@ def load_safetensor_weights(
                 new_to_old_keymap=new_to_old_keymap,
                 updated_states=updated_states,
             )
-        except FileNotFoundError:
-            logger.error(f"File not found: {full_path}")
         except Exception as e:
-            logger.error(f"Error during checkpoint processing of {full_path}: {str(e)}")
+            logger.error(f"Error during checkpoint processing:")
+            raise e
 
     missing_keys = handle_missing_keys(
         stage_state_dict, updated_states, ignore_cache_layers
@@ -244,12 +225,14 @@ def get_needed_files(
     return needed_files
 
 
-def load_safetensor_file(full_path: str, device: torch.device) -> Dict[str, torch.Tensor]:
+def load_safetensor_file(
+    full_path: str,
+    device: str = "cpu",
+) -> Dict[str, torch.Tensor]:
     tensors = {}
     with safe_open(full_path, framework="pt", device=device) as f:
         for k in f.keys():
             tensors[k] = f.get_tensor(k)
-    logger.info(f"Loaded {len(tensors)} tensors from {full_path}")
     return tensors
 
 
@@ -378,15 +361,35 @@ def load_weights_from_hf_format(stage_module, distribution, device, model_config
     files), and fill into `stage_module`.  Model config is needed b/c we permute
     wq and wk weights based on attn heads.
     """
+    # Get the weight map for a given HF model id
+    try:
+        index_file = cached_file(distribution, _DEFAULT_SAFETENSOR_INDEX)
+        is_safetensor = True
+    except:
+        index_file = cached_file(distribution, _DEFAULT_BIN_INDEX)
+        is_safetensor = False
+    logger.info(f"Loading weight map from: {index_file}")
+
+    # Read the weight map from the index file
+    weight_map = read_weights_from_json(index_file)
+    if weight_map is None:
+        raise ValueError(f"Weight map not found in config file {index_file}")
+
+    # Remap the FQNs to the FQNs in HF checkpoints
+    weight_map, new_to_old_keymap = remap_weight_keys(weight_map)
 
-    weight_map, weight_path, key_map = get_hf_weight_map_and_path(distribution)
+    # Get the dir containing the weight files
+    weight_dir = os.path.dirname(index_file)
+    logger.info(f"Loading weights from: {weight_dir}")
 
-    num_loaded_weights, num_missing_weights = load_safetensor_weights(
+    # Load the weights into the stage module
+    num_loaded_weights, num_missing_weights = load_weights_per_map(
         stage_module,
         weight_map,
-        weight_path,
-        key_map,
+        weight_dir,
+        new_to_old_keymap,
         device,
+        is_safetensor,
         model_config=model_config,
     )
     logger.info(