[ROCm][MiniMax-M3] Add AITER sparse paged attention (#47287)

Signed-off-by: Tan Pin Siang <tanpinsiang@gmail.com>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Co-authored-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: Hongxia Yang <hongxia.yang@amd.com>
Co-authored-by: Jun Kang Chow <junkangchow@gmail.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
Tan Pin Siang
2026-07-12 19:27:29 -07:00
committed by GitHub
co-authored by vllmellm Hongxia Yang Jun Kang Chow tjtanaa
parent e26264f3ef
commit ee5a89f4d7
15 changed files with 1554 additions and 128 deletions
+7
View File
@@ -2516,6 +2516,7 @@ def fused_minimax_m3_qknorm_rope_kv_insert(
q_out: torch.Tensor | None = None,
index_q_out: torch.Tensor | None = None,
kv_cache_dtype: str = "auto",
skip_index_branch: bool = False,
) -> None:
"""Fused MiniMax-M3 attention pre-processing (in-place).
@@ -2537,6 +2538,11 @@ def fused_minimax_m3_qknorm_rope_kv_insert(
instead of in place — folding the de-interleave into this kernel's store so
callers skip a separate ``.contiguous()`` copy before the SM100 sparse
attention's flat TMA descriptor.
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
layers that consume top-k block ids selected by an earlier sparse layer.
"""
torch.ops._C.fused_minimax_m3_qknorm_rope_kv_insert(
qkv,
@@ -2559,6 +2565,7 @@ def fused_minimax_m3_qknorm_rope_kv_insert(
q_out,
index_q_out,
kv_cache_dtype,
skip_index_branch,
)