[Transformers backend] Replace vocab embeddings in recursive_replace (#54760)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-01 15:35:24 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 82b7d49a6e
commit f1e5fdd7f2
3 changed files with 128 additions and 70 deletions
@@ -224,41 +224,22 @@ def _rebase_on_vocab_parallel(cls: type[nn.Embedding]) -> type[VocabParallelEmbe
def replace_embedding_class(
embedding: nn.Module,
embedding: nn.Embedding,
quant_config: "QuantizationConfig | None" = None,
*,
prefix: str = "",
) -> nn.Module:
"""Replace the `nn.Embedding` in `embedding` with `VocabParallelEmbedding`.
"""Replace `nn.Embedding` with `VocabParallelEmbedding`.
Args:
embedding: The module returned by `model.get_input_embeddings()`.
embedding: The `nn.Embedding` holding a vocab table.
quant_config: Quantization config for the new embedding.
prefix: Qualname of `embedding`, used to look up its quantization method.
Returns:
The module to install with `model.set_input_embeddings()`. Composing and
inheriting modules are mutated in place and returned as-is.
Raises:
ValueError: If `embedding` composes anything other than one `nn.Embedding`,
which would leave the input embedding weights ambiguous.
The module to install in `embedding`'s place.
"""
# If `embedding` composes its `nn.Embedding`, recurse into it
if not isinstance(embedding, nn.Embedding):
composed = [
(name, module)
for name, module in embedding.named_modules()
if isinstance(module, nn.Embedding)
]
if len(composed) != 1:
raise ValueError(
f"Expected {type(embedding).__name__} to be an `nn.Embedding` or to "
f"compose exactly one, but found {len(composed)}."
)
name, module = composed[0]
new_embedding = replace_embedding_class(
module, quant_config, prefix=maybe_prefix(prefix, name)
)
attrsetter(name)(embedding, new_embedding)
# Tied tables are reached more than once, so don't rebase an already new embedding
if isinstance(embedding, VocabParallelEmbedding):
return embedding
kwargs = dict(