llama export with input vocab pruning (#6421)

Summary:
Pull Request resolved: #6421

D62143905 added llama model export with output vocab pruning. In the similar lines, this diff applies the same for input vocabulary pruning.

The assumption here is: we have trained the model with full vocab and we are pruning out the input vocab after the model training, at export time.

Reviewed By: iseeyuan

Differential Revision: D64723663

Author: navsud

examples/models/llama/TARGETS

@@ -84,7 +84,7 @@ runtime.python_library(
         "source_transformation/apply_spin_quant_r1_r2.py",
         "source_transformation/lora.py",
         "source_transformation/pre_quantization.py",
-        "source_transformation/prune_output.py",
+        "source_transformation/prune_vocab.py",
         "source_transformation/quantize.py",
         "source_transformation/quantized_kv_cache.py",
         "source_transformation/rms_norm.py",

examples/models/llama/export_llama_lib.py

@@ -437,6 +437,12 @@ def build_args_parser() -> argparse.ArgumentParser:
         default=None,
         help="path to the output pruning token mapping file (token_map.json)",
     )
+
+    parser.add_argument(
+        "--input_prune_map",
+        default=None,
+        help="path to the input pruning token mapping file (token_map.json)",
+    )
     return parser
 
 
@@ -525,6 +531,7 @@ def _prepare_for_llama_export(modelname: str, args) -> LLMEdgeManager:
             tokenizer_path=args.tokenizer_path,
             verbose=args.verbose,
             max_seq_len=args.max_seq_length,
+            input_prune_map_path=args.input_prune_map,
             output_prune_map_path=args.output_prune_map,
             metadata_str=args.metadata,
             dtype_override=dtype_override,
@@ -766,6 +773,7 @@ def _load_llama_model(
     tokenizer_path: Optional[str] = None,
     verbose: bool = False,
     max_seq_len: int = 128,
+    input_prune_map_path: Optional[str] = None,
     output_prune_map_path: Optional[str] = None,
     metadata_str: Optional[str] = None,
     dtype_override: Optional[DType] = None,
@@ -795,6 +803,7 @@ def _load_llama_model(
         fairseq2=weight_type == WeightType.FAIRSEQ2,
         max_seq_len=max_seq_len,
         enable_dynamic_shape=enable_dynamic_shape,
+        input_prune_map_path=input_prune_map_path,
         output_prune_map_path=output_prune_map_path,
         args=args,
     )

examples/models/llama/llama_transformer.py

@@ -103,6 +103,8 @@ class ModelArgs:
     generate_full_logits: bool = False
     enable_dynamic_shape: bool = False  # export model with dynamic shape support
     # A dictionary mapping from pruned token-id to original token-id
+    input_prune_map: Optional[Dict[int, int]] = None
+    # A dictionary mapping from pruned token-id to original token-id
     output_prune_map: Optional[Dict[int, int]] = None
     use_hf_rope: bool = False  # Use HuggingFace's RoPE implementation
     rope_theta: Optional[float] = (
@@ -461,6 +463,7 @@ def __init__(self, params: ModelArgs):
         self.use_kv_cache = params.use_kv_cache
         self.generate_full_logits = params.generate_full_logits
         self.max_seq_len = params.max_seq_len
+        self.input_prune_map = params.input_prune_map
         self.output_prune_map = params.output_prune_map
         if params.use_hf_rope:
             self.precompute_freqs_cis = hf_precompute_freqs_cis

examples/models/llama/model.py

@@ -49,6 +49,7 @@ def __init__(self, **kwargs):
         self.use_sdpa_with_kv_cache_op = kwargs.get("use_sdpa_with_kv_cache", False)
         self.generate_full_logits = kwargs.get("generate_full_logits", False)
         self.enable_dynamic_shape = kwargs.get("enable_dynamic_shape", False)
+        self.input_prune_map_path = kwargs.get("input_prune_map_path", None)
         self.output_prune_map_path = kwargs.get("output_prune_map_path", None)
         self.max_seq_len = kwargs.get("max_seq_len", 128)
         self.args = kwargs.get("args", None)
@@ -126,13 +127,20 @@ def __init__(self, **kwargs):
                 output_prune_map = json.load(f)
             # Change keys from string to int (json only supports string keys).
             output_prune_map = {int(k): v for (k, v) in output_prune_map.items()}
+        input_prune_map = None
+        if self.input_prune_map_path is not None:
+            with open(self.input_prune_map_path, "r") as f:
+                input_prune_map = json.load(f)
+            # Change keys from string to int (json only supports string keys).
+            input_prune_map = {int(k): v for (k, v) in input_prune_map.items()}
 
         model_args: ModelArgs = ModelArgs(
             max_seq_len=self.max_seq_len,
             max_batch_size=1,
             use_kv_cache=self.use_kv_cache,
             use_sdpa_with_kv_cache_op=self.use_sdpa_with_kv_cache_op,
             generate_full_logits=self.generate_full_logits,
+            input_prune_map=input_prune_map,
             output_prune_map=output_prune_map,
             enable_dynamic_shape=self.enable_dynamic_shape,
             **params,
@@ -208,10 +216,16 @@ def __init__(self, **kwargs):
             print("============= unexpected keys ================")
             print(unexpected)
             print("============= /unexpected ================")
+        
+        # Prune the input layer if input_prune_map is provided
+        if input_prune_map is not None:
+            from .source_transformation.prune_vocab import prune_input_vocab
+
+            self.model_ = prune_input_vocab(self.model_, input_prune_map)
 
         # Prune the output layer if output_prune_map is provided
         if output_prune_map is not None:
-            from .source_transformation.prune_output import prune_output_vocab
+            from .source_transformation.prune_vocab import prune_output_vocab
 
             self.model_ = prune_output_vocab(self.model_, output_prune_map)
 

examples/models/llama/source_transformation/prune_output.py renamed to examples/models/llama/source_transformation/prune_vocab.py

@@ -69,3 +69,50 @@ def prune_output_vocab(
     setattr(model, output_layer_name, pruned_layer)
 
     return model
+
+def prune_input_vocab(
+    model: torch.nn.Module,
+    token_map: Dict[int, int],
+    imput_layer_name: str = "tok_embeddings",
+) -> torch.nn.Module:
+    """Prune the model input embedding layer while keeping the tokens in the token map.
+    
+    Note: Pruning is performed in-place.
+
+    Args:
+        model: The model to prune.
+        token_map: A dictionary mapping from new token ids to the old token ids to preserve.
+            e.g. {0: 221, 1: 1325, 2: 1542, 3: 1728, 4: 18243}
+        imput_layer_name: name of the input embedding layer to prune
+
+    Returns:
+        The pruned model.
+    """
+    assert hasattr(
+        model, imput_layer_name
+    ), f"Model does not have {imput_layer_name} layer"
+    input_layer = getattr(model, imput_layer_name)
+    assert isinstance(
+        input_layer, torch.nn.Embedding
+    ), "Input layer is not an Embedding layer"
+    original_shape = input_layer.weight.shape
+    num_pruned_tokens = len(token_map)
+    weight_dtype = input_layer.weight.dtype
+    pruned_layer = torch.nn.Embedding(num_pruned_tokens, original_shape[1])
+    pruned_layer.to(dtype=weight_dtype)
+    pruned_layer_weights = np.zeros(pruned_layer.weight.shape, dtype=np.float32)
+    for i, token_id in token_map.items():
+        # Copy the weights from the original layer to the pruned layer
+        pruned_wt = input_layer.weight[token_id].detach()
+        if weight_dtype == torch.bfloat16:
+            pruned_wt = pruned_wt.float()
+        pruned_layer_weights[i] = pruned_wt.numpy()
+    with torch.no_grad():
+        pruned_layer.weight.copy_(
+            torch.tensor(pruned_layer_weights, dtype=weight_dtype)
+        )
+
+    # Replace the original layer with the pruned layer
+    setattr(model, imput_layer_name, pruned_layer)
+
+    return model