[Bugfix] Transformers backend: apply learned lm_head.bias for tied-embedding models (#46835)

Signed-off-by: John Langford <jl@hunch.net>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
John
2026-06-29 08:59:15 +00:00
committed by GitHub
co-authored by Harry Mellor
parent 9e86352c60
commit 0e207dac78
3 changed files with 26 additions and 15 deletions
@@ -542,19 +542,16 @@ class ParallelLMHead(VocabParallelEmbedding):
)
self.quant_config = quant_config
if bias:
self.bias = Parameter(
torch.empty(self.num_embeddings_per_partition, dtype=params_dtype)
)
set_weight_attrs(
self.bias,
{
"output_dim": 0,
"weight_loader": self.weight_loader,
},
)
self._register_bias()
else:
self.register_parameter("bias", None)
def _register_bias(self):
data = torch.empty(self.num_embeddings_per_partition, dtype=self.params_dtype)
self.bias = Parameter(data, requires_grad=False)
weight_attrs = dict(output_dim=0, weight_loader=self.weight_loader)
set_weight_attrs(weight=self.bias, weight_attrs=weight_attrs)
def tie_weights(self, embed_tokens: VocabParallelEmbedding):
"""Tie the weights with word embeddings."""
return self.quant_method.tie_weights(self, embed_tokens)