[Attention][MiniMax-M3] Add MSA speculative decode verification (#50032)
This commit is contained in:
@@ -2655,6 +2655,8 @@ def fused_minimax_m3_qknorm_rope_kv_insert(
|
||||
index_q_out: torch.Tensor | None = None,
|
||||
kv_cache_dtype: str = "auto",
|
||||
skip_index_branch: bool = False,
|
||||
q_fp8_out: torch.Tensor | None = None,
|
||||
q_fp8_scale: float = 1.0,
|
||||
) -> None:
|
||||
"""Fused MiniMax-M3 attention pre-processing (in-place).
|
||||
|
||||
@@ -2677,6 +2679,9 @@ def fused_minimax_m3_qknorm_rope_kv_insert(
|
||||
callers skip a separate ``.contiguous()`` copy before the SM100 sparse
|
||||
attention's flat TMA descriptor.
|
||||
|
||||
If ``q_fp8_out`` is given, the same normalized q is also written in FP8
|
||||
E4M3 using ``q_fp8_scale`` as its dequantization scale.
|
||||
|
||||
When ``skip_index_branch`` is true, sparse rows still keep their packed
|
||||
``[index_q | index_k]`` tail, but the kernel only processes the main q/k/v
|
||||
branches and main KV cache. This is used by MiniMax-M3 index-topk reuse
|
||||
@@ -2704,6 +2709,8 @@ def fused_minimax_m3_qknorm_rope_kv_insert(
|
||||
index_q_out,
|
||||
kv_cache_dtype,
|
||||
skip_index_branch,
|
||||
q_fp8_out,
|
||||
q_fp8_scale,
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user