Fully generalise input embedding handling in Transformers modelling backend (#51247)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-08-06 16:21:24 +00:00
committed by GitHub
parent 566c80edf9
commit e7b8d59460
6 changed files with 300 additions and 47 deletions
@@ -20,7 +20,10 @@ from collections.abc import Iterable
from typing import TYPE_CHECKING
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.models.interfaces_base import VllmModelForTextGeneration
from vllm.model_executor.models.utils import PPMissingLayer, maybe_prefix
@@ -51,9 +54,10 @@ class CausalMixin(VllmModelForTextGeneration):
prefix=maybe_prefix(prefix, "lm_head"),
)
if tie_word_embeddings:
self.lm_head = self.lm_head.tie_weights(
self.model.get_input_embeddings()
)
for module in self.model.get_input_embeddings().modules():
if isinstance(module, VocabParallelEmbedding):
self.lm_head = self.lm_head.tie_weights(module)
break
logit_scale = getattr(self.text_config, "logit_scale", 1.0)
self.logits_processor = LogitsProcessor(