Standardise weight tying on ParallelLMHead.tie_weights (#52147)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
7553aac77b
commit
80d6d557f3
@@ -324,14 +324,13 @@ class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant):
|
||||
self.transformer = BloomModel(
|
||||
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
|
||||
)
|
||||
self.lm_head = ParallelLMHead(
|
||||
self.config.vocab_size,
|
||||
self.config.hidden_size,
|
||||
prefix=maybe_prefix(prefix, "lm_head"),
|
||||
)
|
||||
if self.config.tie_word_embeddings:
|
||||
self.lm_head = self.transformer.word_embeddings
|
||||
else:
|
||||
self.lm_head = ParallelLMHead(
|
||||
self.config.vocab_size,
|
||||
self.config.hidden_size,
|
||||
prefix=maybe_prefix(prefix, "lm_head"),
|
||||
)
|
||||
self.lm_head = self.lm_head.tie_weights(self.transformer.word_embeddings)
|
||||
|
||||
self.logits_processor = LogitsProcessor(config.vocab_size)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
|
||||
Reference in New Issue
Block a user