@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / cfba4def5d
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 19:30 UTC

[Bugfix] Fix logit soft cap in flash-attn backend (#7425)

Woosuk Kwon committed Aug 12, 2024 at 09:58 UTC cfba4def5d422dfbafbc21b8f695bdcb2295aa19
1 file changed +1
vllm/attention/backends/flash_attn.py
+1
@@ -563,6 +563,7 @@ class FlashAttentionImpl(AttentionImpl):
563 softmax_scale=self.scale,
564 causal=True,
565 alibi_slopes=self.alibi_slopes,
566 + softcap=self.logits_soft_cap,
567 ).squeeze(1)
568
569 # Reshape the output tensor.