huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-11 20:34:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: prefer inplace softmax to avoid copy

Browse Source
This commit is contained in:
David Holtz 2024-10-17 02:53:32 +00:00
parent a6a0c97ed9
commit 8d7448de9f
1 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
server/text_generation_server/models/flash_causal_lm.py
View File

@ -1922,8 +1922,9 @@ class FlashCausalLM(Model):
batch.adapter_meta.adapter_indices = next_adapter_indices batch.adapter_meta.adapter_indices = next_adapter_indices
if prefill and prefill_logprobs: if prefill and prefill_logprobs:
# Get prefill logprobs # Get prefill logprobs with inplace softmax (avoid copying the `out` tensor (max_batch_size * vocab_size))
prefill_logprobs_tensor = torch.log_softmax(out, -1) torch.log_softmax(out, -1, out=out)
prefill_logprobs_tensor = out
prefill_logprobs = torch.gather( prefill_logprobs = torch.gather(
prefill_logprobs_tensor, 1, prefill_tokens_indices.view(-1, 1) prefill_logprobs_tensor, 1, prefill_tokens_indices.view(-1, 1)
) )