[Kernel][DSv4.1] Bump FlashMLA to the 2026.09.30 release and its interleaved V4.1 KV cache (#61371)

Signed-off-by: juntianl <juntianl@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kimi Code <noreply@moonshot.ai>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
This commit is contained in:
Juntian Liu
2026-10-15 13:58:25 -07:00
committed by GitHub
co-authored by Claude Opus 5.5 Kimi Code Yifan Qiao
parent 33212db2c2
commit f8460ef236
13 changed files with 262 additions and 165 deletions
@@ -427,7 +427,7 @@ def sparse_mqa_logits_paged_decode(
q: [rows, 1, H, D] packed Q values (MXFP4 viewed as int8).
q_scale: [rows, 1, H] int32 packed UE8M0 Q scales.
kv_cache: 4D [num_pages, page_kv, 1, head_bytes] uint8 fused cache
view; the page stride must be 512B-aligned.
view; the page stride must be 256B-aligned.
weights: [rows, H] bf16 per-head weights.
context_lens: [rows] int32 per-row (compressed) context lengths.
block_table: [rows, P] int32 per-row page table, ``stride(-1) == 1``.
@@ -451,8 +451,8 @@ def sparse_mqa_logits_paged_decode(
rows = q.shape[0]
page_kv = kv_cache.shape[1]
assert weights.dtype == torch.bfloat16, weights.dtype
assert kv_cache.stride(0) % 512 == 0, (
"paged cache page stride must be 512B-aligned: "
assert kv_cache.stride(0) % 256 == 0, (
"paged cache page stride must be 256B-aligned: "
f"shape={tuple(kv_cache.shape)} stride={kv_cache.stride()}"
)
assert page_kv % sparse_block_kv == 0, (page_kv, sparse_block_kv)