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:
Harry Mellor
2026-08-13 15:10:09 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 7553aac77b
commit 80d6d557f3
56 changed files with 191 additions and 214 deletions
+6 -7
View File
@@ -349,14 +349,13 @@ class OPTForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
self.model = OPTModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
self.lm_head = ParallelLMHead(
config.vocab_size,
config.word_embed_proj_dim,
prefix=maybe_prefix(prefix, "lm_head"),
)
if self.config.tie_word_embeddings:
self.lm_head = self.model.decoder.embed_tokens
else:
self.lm_head = ParallelLMHead(
config.vocab_size,
config.word_embed_proj_dim,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.lm_head = self.lm_head.tie_weights(self.model.decoder.embed_tokens)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors