[Bugfix][MLA] Do not use Dense MHA for GLM-5.2 (#52512)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
ad5e71b276
commit
01e56caaf2
@@ -160,6 +160,7 @@ class DeepseekV32Attention(MLAAttention):
|
||||
indexer_cls: "type[DeepseekV32Indexer]" = DeepseekV32Indexer
|
||||
|
||||
require_fp8_kv_cache: bool = True
|
||||
supports_dense_mha_prefill = False
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
@@ -265,6 +266,7 @@ class DeepseekV32Attention(MLAAttention):
|
||||
and not skip_topk
|
||||
and not self.use_pcp
|
||||
and current_platform.is_cuda()
|
||||
and self.supports_dense_mha_prefill
|
||||
)
|
||||
self._dense_mha_metadata_layer_name = (
|
||||
self.layer_name if enable_short_prefill_scoring_skip else ""
|
||||
|
||||
Reference in New Issue
Block a user