[ROCm][Kimi-K3] Don't read W_UK_T before weight loading in the MLA wrapper (#61654)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
4c5ea86f53
commit
9a648b4649
@@ -50,8 +50,9 @@ class KimiK3MultiHeadLatentAttentionWrapper(MultiHeadLatentAttentionWrapper):
|
||||
if attn.hisparse_cache is not None or attn.q_pad_num_heads is not None:
|
||||
return False
|
||||
# Only the W_UK variants reproduced in _fused_decode are supported.
|
||||
return attn.is_aiter_triton_fp4_bmm_enabled or (
|
||||
not attn.is_aiter_triton_fp8_bmm_enabled and attn.W_UK_T is not None
|
||||
return (
|
||||
attn.is_aiter_triton_fp4_bmm_enabled
|
||||
or not attn.is_aiter_triton_fp8_bmm_enabled
|
||||
)
|
||||
|
||||
def get_input_quant_key(self) -> QuantKey | None:
|
||||
|
||||
Reference in New Issue
Block a user