[Bugfix][ROCm][MLA] Pass q/kv dtypes to get_mla_metadata_v1 in FP8 decode (#46997)

Signed-off-by: pei.zhang <pei.zhang@amd.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
peizhang56
2026-06-30 05:31:16 -07:00
committed by GitHub
co-authored by TJian
parent 536047755e
commit aab7af0bcb
3 changed files with 209 additions and 0 deletions
@@ -199,6 +199,10 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
else:
kv_cache_dtype_str = "bf16"
kv_dtype = dtypes.d_dtypes.get(kv_cache_dtype_str, dtypes.bf16)
# Persist for get_mla_metadata_v1 (decode build): omitting these causes
# wrong split/reduce metadata for the gfx950 fp8 nhead=32 fold path.
self._mla_q_dtype = q_dtype
self._mla_kv_dtype = kv_dtype
(
(work_meta_data_size, work_meta_data_type),
(work_indptr_size, work_indptr_type),
@@ -534,6 +538,8 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
max_seqlen_qo=max_qo_len,
uni_seqlen_qo=max_qo_len,
fast_mode=True,
dtype_q=self._mla_q_dtype,
dtype_kv=self._mla_kv_dtype,
)
has_persistent_metadata = True