[Kernel][Build] Bump DeepGEMM to the 26/09/30 release and adapt the sparse indexer (#61149)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-10-12 23:24:43 -07:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 86b1c2ef15
commit a6009ea85b
23 changed files with 516 additions and 208 deletions
@@ -14,6 +14,7 @@ from vllm.models.deepseek_v4.sparse_mla import DeepseekV4SparseMLABackend
from vllm.models.deepseek_v41.sparse_mla import (
DeepseekV4SparseMLABackend as DeepseekV41SparseMLABackend,
)
from vllm.utils.deep_gemm import mqa_logits_use_mx_scales
from vllm.v1.attention.backend import CommonAttentionMetadata
from vllm.v1.attention.backends.mla import indexer as indexer_module
from vllm.v1.attention.backends.mla.compressor_utils import (
@@ -113,10 +114,17 @@ def test_indexer_draft_decode_metadata_update_is_capture_safe():
num_reqs, 4
)
num_sms = torch.cuda.get_device_properties(device).multi_processor_count
# SM100's paged MQA logits only take varlen rows.
indices = (
torch.arange(num_reqs, dtype=torch.int32, device=device)
if mqa_logits_use_mx_scales()
else None
)
schedule_metadata = indexer_module.get_paged_mqa_logits_metadata(
torch.div(seq_lens, 16, rounding_mode="floor").unsqueeze(1),
64,
num_sms,
indices=indices,
).clone()
builder = object.__new__(DeepseekV32IndexerMetadataBuilder)
@@ -139,6 +147,7 @@ def test_indexer_draft_decode_metadata_update_is_capture_safe():
decode_lens=torch.zeros(num_reqs, dtype=torch.int32, device=device),
requires_padding=False,
schedule_metadata=schedule_metadata,
indices=indices,
),
block_table=block_table,
)
@@ -157,7 +166,7 @@ def test_indexer_draft_decode_metadata_update_is_capture_safe():
torch.testing.assert_close(
schedule_metadata,
indexer_module.get_paged_mqa_logits_metadata(
expected_decode_seq_lens.unsqueeze(1), 64, num_sms
expected_decode_seq_lens.unsqueeze(1), 64, num_sms, indices=indices
),
)
assert metadata.decode is not None
@@ -36,6 +36,7 @@ def _chunk(block_table, token_start, token_end, row_ends):
cu_seq_lens=torch.zeros(1, dtype=torch.int32),
token_to_seq=torch.zeros(1, dtype=torch.int32),
total_seq_lens=1,
max_seqlen_k=int(ends.max()),
token_start=token_start,
token_end=token_end,
num_reqs=block_table.shape[0],