Fix weight tying (#51665)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-08-20 10:31:55 +00:00
committed by GitHub
parent 6e85feb1b7
commit f0c14b4f77
70 changed files with 526 additions and 238 deletions
+1 -6
View File
@@ -467,10 +467,5 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
# With tie_word_embeddings, we can skip lm_head.weight
# The weight might appear unnecessarily in the files if the model is
# processed with quantization, LoRA, fine-tuning, etc.
skip_prefixes = ["lm_head."] if self.config.tie_word_embeddings else None
loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
loader = AutoWeightsLoader(self)
return loader.load_weights(weights)