huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-25 20:12:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Do not schedule decode if max_new_tokens is equal to 1 (#183)

Browse Source 
Co-authored-by: Bartosz Kowalski <bkowalski@habana.ai>
This commit is contained in:
bkowalskiINTEL 2024-07-16 14:53:24 +02:00 committed by GitHub
parent 15e5df1cc4
commit 0ca54b55f8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
server/text_generation_server/models/causal_lm.py
View File

@ -985,6 +985,10 @@ class CausalLM(Model):
batch.past_key_values,
bypass_hpu_graph=prefill and self.limit_hpu_graph if self.enable_hpu_graph else None,
)
elif all([req.stopping_criteria.max_new_tokens == 1 for req in batch.requests]):
# Don't schedule next forward if max_new_tokens for all requests equals 1
# - we've already generated the first and only needed token in the prefill phase
pass
else:
token_idx = torch.tensor(batch.attention_mask.shape[-1] - batch.right_padding).to(self.device)
input_ids = torch.index_select(batch.input_ids, 1, token_idx - 1)