Merge pull request #4 from axolotl-ai-cloud/fix-untrained-distributed

handle distributed embeddings

Author: winglian

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "axolotl-contribs-lgpl"
-version = "0.0.4.dev0"
+version = "0.0.4"
 description = "LGPL contributions to the axolotl framework"
 authors = [{ name="Wing Lian" }]
 readme = "README.md"

src/axolotl/contribs/lgpl/__init__.py

@@ -0,0 +1 @@
+from .unsloth import fix_untrained_tokens  # noqa: F401

src/axolotl/contribs/lgpl/unsloth.py

@@ -14,56 +14,68 @@
 import itertools
 import logging
 from collections import Counter
+from contextlib import nullcontext
+from functools import partial
 
 import datasets
+import deepspeed
 import numpy as np
 import torch
 
 LOG = logging.getLogger("axolotl.contribs.lgpl.unsloth")
 
 
-@torch.inference_mode()
-def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
-        model, tokenizer, train_dataset, ignored_tokenizer_names=None, eps=1e-16, token_ids_to_fix=None,
-):
+def get_embedding_mean(input_embeddings, output_embeddings, tokenizer, train_dataset, eps=1e-16, token_ids_to_fix=None):
     """
-    Llama-3 for eg has untrained vectors in the base model.
-    These include <|eot_id|>, <|start_header_id|>, <|end_header_id|>
-    We reset them to the mean of the rest of the tokens
+    Detects tokens with untrained embeddings (values close to zero or duplicated) and
+    prepares scaled mean embedding values for these tokens based on their frequency in the training data.
+
+    Args:
+        input_embeddings: The model's input embedding layer (embed_tokens)
+        output_embeddings: The model's output embedding layer (lm_head)
+        tokenizer: The tokenizer used with the model
+        train_dataset: The training dataset to analyze token frequencies
+        eps: Small epsilon value to detect near-zero embeddings (default: 1e-16)
+        token_ids_to_fix: Additional token IDs to include in the fixing process
+
+    Returns:
+        tuple: If untrained tokens are found in the training data, returns:
+            - mean_embedding_repeated: Scaled mean embeddings for input layer
+            - mean_lm_head_repeated: Scaled mean embeddings for output layer
+            - tokens_to_update: List of token IDs that need updating
+        None: If no untrained tokens are found in the training data
+
+    Raises:
+        ValueError: If embedding matrices have incorrect shapes or if untrained tokens are found but embeddings are not trainable
     """
-    # Code licensed under LGPL
-    if not token_ids_to_fix:
-        token_ids_to_fix = []
-    embedding_matrix = model.get_input_embeddings().weight
-    lm_head_matrix = model.get_output_embeddings().weight
+
+    # Check if we still have an issue with the shapes
+    if input_embeddings.weight.shape[0] == 0 or output_embeddings.weight.shape[0] == 0:
+        raise ValueError(
+            f"Could not gather embedding matrices properly. "
+            f"Shapes: embedding={input_embeddings.weight.shape}, lm_head={output_embeddings.weight.shape}. "
+            f"This might indicate a DeepSpeed configuration issue."
+        )
+
+    # Get the chat template if available
     chat_template = getattr(tokenizer, "chat_template", None)
     tokenizer = tokenizer.tokenizer if hasattr(tokenizer, "tokenizer") else tokenizer
 
-    # Ignore some model checks for now
-    if not ignored_tokenizer_names:
-        ignored_tokenizer_names = []
-    if (
-            model.config._name_or_path  # pylint: disable=protected-access
-            in ignored_tokenizer_names
-    ):
-        return
-
     # Sometimes the sizes can be different like in vision models
     # Ie <image> is in input, but not in output
-    min_size = min(embedding_matrix.shape[1], lm_head_matrix.shape[1])
-    embedding_matrix = embedding_matrix[:, :min_size]
-    lm_head_matrix = lm_head_matrix[:, :min_size]
+    min_size = min(input_embeddings.weight.shape[1], output_embeddings.weight.shape[1])
+    input_embeddings.weight = torch.nn.Parameter(input_embeddings.weight[:, :min_size])
+    output_embeddings.weight = torch.nn.Parameter(output_embeddings.weight[:, :min_size])
 
     # Get untrained tokens
-    indicator_untrained1 = torch.amax(embedding_matrix, axis=1) <= eps
-    # Check lm_head as well
+    indicator_untrained1 = torch.amax(input_embeddings.weight, axis=1) <= eps
 
     # Does NOT work for Llama 3.1!!
-    indicator_untrained2 = torch.amax(lm_head_matrix, axis=1) <= eps
+    indicator_untrained2 = torch.amax(output_embeddings.weight, axis=1) <= eps
 
     # We instead check for repeated vectors
     lm_head_where = torch.where(indicator_untrained1)[0]
-    lm_head_bad = lm_head_matrix[lm_head_where]
+    lm_head_bad = output_embeddings.weight[lm_head_where]
     lm_head_bad = lm_head_bad.cpu().float().numpy().round(3)
     counter = Counter()
     for row in lm_head_bad:
@@ -75,7 +87,8 @@ def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
     for j, row in enumerate(lm_head_bad):
         if hash(row.data.tobytes()) in counter:
             final_bad_lm_head.append(lm_head_where[j])
-    indicator_untrained2 = indicator_untrained2 | torch.zeros_like(indicator_untrained2)
+
+    indicator_untrained2 = indicator_untrained2 | torch.zeros_like(indicator_untrained1, dtype=torch.bool)
     indicator_untrained2[final_bad_lm_head] = True
 
     # Combine both checks
@@ -89,7 +102,7 @@ def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
 
     where_untrained = torch.where(indicator_untrained)[0]
     n_untrained = where_untrained.shape[0]
-    n_trained = embedding_matrix.shape[0] - n_untrained
+    n_trained = input_embeddings.weight.shape[0] - n_untrained
 
     # Get set and actual tokens
     where_untrained = where_untrained.tolist()
@@ -145,9 +158,9 @@ def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
 
     # Check if lm_head / embed_token are trainable!
     bad_not_trainable = False
-    if not embedding_matrix.requires_grad:
+    if not input_embeddings.weight.requires_grad:
         bad_not_trainable = True
-    if not lm_head_matrix.requires_grad:
+    if not output_embeddings.weight.requires_grad:
         bad_not_trainable = True
 
     if bad_not_trainable:  # pylint: disable=too-many-nested-blocks
@@ -176,8 +189,6 @@ def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
 
         # If no bad tokens, possibly chat template itself has issues?
         if len(final_bad_items) == 0:
-            # Recheck 2000 and last 2000 items
-            size_dataset = len(train_dataset)
             size = min(size_dataset, 2000)
             for j in range(size):
                 input_ids = train_dataset[j]
@@ -207,7 +218,7 @@ def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
 
     # Count all the possible bad tokens
     final_counts = np.zeros(
-        max(len(tokenizer), embedding_matrix.shape[0]), dtype=np.int64
+        max(len(tokenizer), input_embeddings.weight.shape[0]), dtype=np.int64
     )
 
     def mapping(examples):
@@ -235,15 +246,15 @@ def mapping(examples):
     )
 
     # Get sum of all items
-    sum_embedding = torch.sum(embedding_matrix, dtype=torch.float32, axis=0)
-    sum_lm_head = torch.sum(lm_head_matrix, dtype=torch.float32, axis=0)
+    sum_embedding = torch.sum(input_embeddings.weight, dtype=torch.float32, axis=0)
+    sum_lm_head = torch.sum(output_embeddings.weight, dtype=torch.float32, axis=0)
 
     # Remove bad tokens
     sum_embedding -= torch.sum(
-        embedding_matrix[where_untrained], dtype=torch.float32, axis=0
+        input_embeddings.weight[where_untrained], dtype=torch.float32, axis=0
     )
     sum_lm_head -= torch.sum(
-        lm_head_matrix[where_untrained], dtype=torch.float32, axis=0
+        output_embeddings.weight[where_untrained], dtype=torch.float32, axis=0
     )
 
     # Find correct average by dividing by sum of trained tokens
@@ -262,14 +273,60 @@ def mapping(examples):
             mean_lm_head.unsqueeze(0).repeat(len(tokens_to_update), 1) * scaling
     )
 
-    # Update embeddings only for tokens seen in train_dataset
-    embedding_matrix[tokens_to_update] = mean_embedding_repeated.to(
-        embedding_matrix.dtype
-    )
-    lm_head_matrix[tokens_to_update] = mean_lm_head_repeated.to(lm_head_matrix.dtype)
+    return mean_embedding_repeated, mean_lm_head_repeated, tokens_to_update
+
+def fix_untrained_tokens(
+        model, tokenizer, train_dataset, ignored_tokenizer_names=None, eps=1e-16, token_ids_to_fix=None, is_ds_zero3=False
+):
+    """
+    Some base models have untrained vectors for embeddings/tokens. Update these embeddings to
+    the mean of the rest of the tokens. This additionally handles distributed settings like DeepSpeed ZeRO-3.
+
+    Args:
+        model: The model to fix embeddings for
+        tokenizer: The tokenizer used with the model
+        train_dataset: The training dataset to analyze token frequencies
+        ignored_tokenizer_names: List of model names to skip processing (default: None)
+        eps: Small epsilon value to detect near-zero embeddings (default: 1e-16)
+        token_ids_to_fix: Additional token IDs to include in the fixing process (default: None)
+        is_ds_zero3: Whether DeepSpeed ZeRO-3 is being used (default: False)
+    """
+
+    if not token_ids_to_fix:
+        token_ids_to_fix = []
+    if not ignored_tokenizer_names:
+        ignored_tokenizer_names = []
+
+    # Check if we should ignore this model
+    if (
+            hasattr(model, "config") and
+            hasattr(model.config, "_name_or_path") and
+            model.config._name_or_path in ignored_tokenizer_names
+    ):
+        return
+
+    with torch.no_grad():
+        # Get the embedding layer and lm_head
+        embedding_layer = model.get_input_embeddings()
+        lm_head_layer = model.get_output_embeddings()
+
+        context = nullcontext
+        if is_ds_zero3:
+            # Get the full parameters if using DeepSpeed
+            context = partial(deepspeed.zero.GatheredParameters, [embedding_layer.weight, lm_head_layer.weight], modifier_rank=0)
+
+        with context():
+            input_embeddings = model.get_input_embeddings()
+            output_embeddings = model.get_output_embeddings()
+            mean_embedding_repeated, mean_lm_head_repeated, tokens_to_update = get_embedding_mean(input_embeddings, output_embeddings, tokenizer, train_dataset, eps=eps, token_ids_to_fix=token_ids_to_fix)
+            input_embeddings.weight[tokens_to_update] = mean_embedding_repeated.to(
+                input_embeddings.weight.dtype
+            )
+            model.set_input_embeddings(input_embeddings)
+            output_embeddings.weight[tokens_to_update] = mean_lm_head_repeated.to(output_embeddings.weight.dtype)
+            model.set_output_embeddings(output_embeddings)
 
     # Clean up
     for _ in range(3):
         gc.collect()
         torch.cuda.empty_cache()
-    return