[DSA] Wire DeepGEMM sparse MQA logits into the DeepSeek V4.1 indexer (#56254)

Signed-off-by: Jared Wen <w13431838023@gmail.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
Jared Wen
2026-09-15 14:07:14 -07:00
committed by GitHub
co-authored by Claude Fable 5.1
parent f2aad6aa70
commit 1257512305
13 changed files with 1830 additions and 26 deletions
+10 -1
View File
@@ -653,6 +653,12 @@ class MLAAttentionSpec(FullAttentionSpec):
is_index_group_leader: bool = False
storage_block_size: int | None = None
"""Token width used to view storage when it differs from the kernel block."""
block_stride_alignment: int | None = None
"""Required alignment, in bytes, of the distance between consecutive
blocks of this cache. In block-major layouts that distance is the whole
block (all layers' pages), so the allocator rounds the block up to it.
DeepGEMM's paged sparse MQA-logits kernels address pages as
``base + page * stride`` and need it 512B-aligned."""
# Group capability enabled when any member flattens a non-causal query block
# into decode rows. Runtime metadata still selects causal vs. non-causal mode.
non_causal_multi_token_decode: bool = False
@@ -674,6 +680,7 @@ class MLAAttentionSpec(FullAttentionSpec):
cache_role_set = {spec.cache_role for spec in specs}
index_group_leader_set = {spec.is_index_group_leader for spec in specs}
storage_block_size_set = set(spec.storage_block_size for spec in specs)
block_stride_alignment_set = {spec.block_stride_alignment for spec in specs}
assert (
len(cache_dtype_str_set) == 1
and len(tokens_per_state_set) == 1
@@ -681,10 +688,11 @@ class MLAAttentionSpec(FullAttentionSpec):
and len(cache_role_set) == 1
and len(index_group_leader_set) == 1
and len(storage_block_size_set) == 1
and len(block_stride_alignment_set) == 1
), (
"All attention layers in the same KV cache group must use the same "
"quantization method, tokens per state, model version, cache role, "
"index-sharing role, and storage block size."
"index-sharing role, storage block size and block stride alignment."
)
merged_spec = cls(
block_size=specs[0].block_size,
@@ -701,6 +709,7 @@ class MLAAttentionSpec(FullAttentionSpec):
cache_role=cache_role_set.pop(),
is_index_group_leader=index_group_leader_set.pop(),
storage_block_size=storage_block_size_set.pop(),
block_stride_alignment=block_stride_alignment_set.pop(),
non_causal_multi_token_decode=any(
spec.non_causal_multi_token_decode for spec in specs
),