[Bugfix] Fix DeepGEMM context lens contiguity in MLA indexer (#42135)

Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
This commit is contained in:
Mohammad Miadh Angkad
2026-05-15 23:29:58 +08:00
committed by GitHub
parent 491e8d8539
commit ee58665aac
+16 -18
View File
@@ -288,20 +288,14 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
dtype=torch.int32, dtype=torch.int32,
device=self.device, device=self.device,
) )
if not self.use_flattening and next_n > 1: # Shared workspace for decode seq_lens. Native MTP views this as
# Native MTP: 2D buffer for per-token seq_lens. # (B, max_decode_len) at runtime, keeping context_lens contiguous even
self.decode_seq_lens_buffer = torch.zeros( # when max_decode_len is smaller than next_n.
(scheduler_config.max_num_seqs, next_n), self.decode_seq_lens_buffer = torch.zeros(
dtype=torch.int32, (scheduler_config.max_num_batched_tokens,),
device=self.device, dtype=torch.int32,
) device=self.device,
else: )
# Flattening or no MTP: 1D buffer for expanded per-token seq_lens.
self.decode_seq_lens_buffer = torch.zeros(
(scheduler_config.max_num_batched_tokens,),
dtype=torch.int32,
device=self.device,
)
self.arange_buffer = torch.arange( self.arange_buffer = torch.arange(
max( max(
scheduler_config.max_num_seqs * next_n, scheduler_config.max_num_seqs * next_n,
@@ -373,7 +367,8 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
Plain decode or spec-decode with 2D per-token context lengths. Plain decode or spec-decode with 2D per-token context lengths.
Returns (seq_lens, block_table, decode_lens, batch_size, requires_padding). Returns (seq_lens, block_table, decode_lens, batch_size, requires_padding).
seq_lens is 1D (batch_size,) for flatten/plain, 2D (B, next_n) for native MTP. seq_lens is 1D (batch_size,) for flatten/plain, 2D (B, max_decode_len)
for native MTP.
""" """
min_decode_len = int(decode_lens_cpu.min().item()) min_decode_len = int(decode_lens_cpu.min().item())
if not use_native and max_decode_len > 1: if not use_native and max_decode_len > 1:
@@ -454,16 +449,19 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
# (requires_padding) instead. # (requires_padding) instead.
requires_padding = min_decode_len != max_decode_len requires_padding = min_decode_len != max_decode_len
if use_native and next_n > 1: if use_native and next_n > 1:
assert self.decode_seq_lens_buffer.dim() == 2 assert self.decode_seq_lens_buffer.dim() == 1
# (B, max_decode_len): token j attends to # (B, max_decode_len): token j attends to
# L - max_decode_len + j + 1 KV tokens. # L - max_decode_len + j + 1 KV tokens.
self.decode_seq_lens_buffer[:num_decodes, :max_decode_len] = ( seq_lens_buffer = self.decode_seq_lens_buffer[
: num_decodes * max_decode_len
].view(num_decodes, max_decode_len)
seq_lens_buffer[:] = (
seq_lens.unsqueeze(1) seq_lens.unsqueeze(1)
- max_decode_len - max_decode_len
+ 1 + 1
+ self.offsets_buffer[:max_decode_len] + self.offsets_buffer[:max_decode_len]
) )
seq_lens = self.decode_seq_lens_buffer[:num_decodes, :max_decode_len] seq_lens = seq_lens_buffer
return seq_lens, block_table, decode_lens, num_decodes, requires_padding return seq_lens, block_table, decode_lens, num_decodes, requires_padding
def build( def build(