huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-06-17 23:02:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: prefer patch to be vlm specific

Browse Source
This commit is contained in:
drbh 2025-05-06 00:02:38 +00:00
parent b32cd97b71
commit 783ca66926
2 changed files with 9 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
server/text_generation_server/models/flash_causal_lm.py
View File

@ -1267,15 +1267,6 @@ class FlashCausalLM(Model):
prefix = None
model = model_class(prefix, config, weights)
if model.config.vocab_size != tokenizer.vocab_size:
logger.warning(
f"Tokenizer vocab size {tokenizer.vocab_size} does not match model vocab size {model.config.vocab_size}. Updating tokenizer vocab size."
)
# TODO: HUGE HACK! This is a workaround for the fact that Qwen2TokenizerFast
# returns the incorrect vocab size for the 2B model.
tokenizer._vocab_size = model.config.vocab_size
torch.distributed.barrier(group=self.process_group)
# VLM models define the config we care about in their text_config

9
server/text_generation_server/models/vlm_causal_lm.py
View File

@ -414,6 +414,15 @@ class VlmCausalLM(FlashCausalLM):
**kwargs,
)
if self.config.vocab_size != self.tokenizer.vocab_size:
logger.warning(
f"Tokenizer vocab size {self.tokenizer.vocab_size} does not match model vocab size {self.config.vocab_size}. Updating tokenizer vocab size."
)
# TODO: HUGE HACK! This is a workaround to update the vocab size
# in the tokenizer. When the tokenizer is updated within the model
# the vocab size is not updated in the tokenizer.
self.tokenizer._vocab_size = self.config.vocab_size
@property
def batch_type(self) -> Type[VlmCausalLMBatch]:
return self.batch_class