kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-02-16 15:55:08 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

[https://nvbugs/5624818][fix] Work around accuracy issue by enforcing paged_context_fmha on Hopper for fmha_v2 (#11192)

Browse Source 
Signed-off-by: eopXD <yuehtingc@nvidia.com>
This commit is contained in:
Yueh-Ting (eop) Chen 2026-02-04 19:21:50 +08:00 committed by GitHub
parent 3d8c1a51bd
commit f6fff18142
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 10 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
tensorrt_llm/_torch/attention_backend/trtllm.py
View File

@ -1549,6 +1549,11 @@ class TrtllmAttention(AttentionBackend[TrtllmAttentionMetadata]):
or metadata.runtime_features.has_speculative_draft_tokens
) if metadata.runtime_features else False
# This is a workaround for https://nvbugs/5624818
# Paged context FMHA is forced on SM90 for correctness
if get_sm_version() == 90:
use_paged_context_fmha = True
return self.wrapper.is_nvfp4_output_kernel_available(
tokens_per_block=metadata.tokens_per_block,
attention_mask=attention_mask,
@ -1648,6 +1653,11 @@ class TrtllmAttention(AttentionBackend[TrtllmAttentionMetadata]):
or metadata.runtime_features.has_speculative_draft_tokens
) if metadata.runtime_features else False
# This is a workaround for https://nvbugs/5624818
# Paged context FMHA is forced on SM90 for correctness
if get_sm_version() == 90:
use_paged_context_fmha = True
if self.is_mla_enable:
# Context MLA uses separate qkv instead of paged_context_fmha
use_paged_context_fmha = False