[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
@@ -467,6 +467,9 @@ class DeepseekV32IndexerPrefillChunkMetadata:
cu_seq_lens: torch.Tensor
token_to_seq: torch.Tensor
total_seq_lens: int
# Host upper bound on cu_seqlen_ke - cu_seqlen_ks, sizing the compressed
# DeepGEMM prefill logits.
max_seqlen_k: int
token_start: int
token_end: int
num_reqs: int
@@ -1957,6 +1960,7 @@ def build_prefill_chunk_metadata(
cu_seq_lens=cu_seq_lens,
token_to_seq=token_to_seq,
total_seq_lens=total_seq_lens,
max_seqlen_k=int(compressed_seq_lens_cpu[start_idx:end_idx].max()),
block_table=block_table[start_idx:end_idx],
token_start=token_start,
token_end=token_end,
+7
View File
@@ -154,6 +154,13 @@ class PackSeqTritonKernel(VllmTritonJitKernel["PackSeqTritonKernel.CompileKey"])
)
@triton.jit
def pack_ue8m0_x4(scale):
"""Encode a power-of-two fp32 scale as four equal packed UE8M0 bytes, the
MX scales of a 128-wide row (see `mqa_logits_use_mx_scales`)."""
return ((scale.to(tl.int32, bitcast=True) >> 23) & 0xFF) * 0x01010101
def pack_seq_triton(
x: torch.Tensor,
lengths: torch.Tensor,