Skip to content

Fix DCP sparse indexer decode metadata #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lukealonso merged 1 commit into from
Jun 7, 2026
Merged

Fix DCP sparse indexer decode metadata #4

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
79 changes: 68 additions & 11 deletions vllm/v1/attention/backends/mla/indexer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -380,15 +380,18 @@ def __init__(self, *args, **kwargs):
if isinstance(self.kv_cache_spec, MLAAttentionSpec):
self.compress_ratio = self.kv_cache_spec.compress_ratio

# Pre-allocate buffers for CUDA graph compatibility when
if self.compress_ratio > 1:
# compress_ratio > 1 (DeepseekV4)
# Compressed slot mapping output buffer
# DCP writes the indexer KV cache through rank-local pages even when
# compress_ratio == 1 (GLM/Kimi). Keep the mapped slots graph-stable.
if self.compress_ratio > 1 or self.dcp_world_size > 1:
self.compressed_slot_mapping_buffer = torch.zeros(
(scheduler_config.max_num_batched_tokens,),
dtype=torch.int64,
device=self.device,
)

# Pre-allocate buffers for CUDA graph compatibility when
if self.compress_ratio > 1:
# compress_ratio > 1 (DeepseekV4)
# Buffer for compressed seq_lens in decode path
self.expanded_seq_lens_buffer = torch.zeros(
(scheduler_config.max_num_batched_tokens,),
Expand Down Expand Up @@ -463,6 +466,7 @@ def _prepare_decode_tensors(
use_native: bool,
next_n: int,
max_decode_len: int,
global_seq_lens: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, bool]:
"""Expand seq_lens/block_table/decode_lens for the decode kernels.

Expand Down Expand Up @@ -494,6 +498,22 @@ def _prepare_decode_tensors(
max_decode_len,
BLOCK_SIZE=1024,
)
if global_seq_lens is not None and self.dcp_world_size > 1:
expanded_global = (
global_seq_lens[:num_decodes].unsqueeze(1)
- max_decode_len
+ 1
+ self.offsets_buffer[:max_decode_len]
).reshape(-1)
self.decode_seq_lens_buffer[:num_decode_tokens].copy_(
get_dcp_local_seq_lens(
expanded_global,
self.dcp_world_size,
self.dcp_rank,
self.cp_kv_cache_interleave_size,
),
non_blocking=True,
)
self.decode_seq_lens_buffer[num_decode_tokens:] = 0
seq_lens = self.decode_seq_lens_buffer[:num_decode_tokens]
block_table = self.expanded_block_table_buffer[:num_decode_tokens]
Expand All @@ -516,16 +536,27 @@ def _prepare_decode_tensors(
# Example: offsets = [7-0, 6-3, 8-4, 0-8] = [7, 3, 4, -8]
# expanded_offsets = [7, 7, 7, 3, 4, 4, 4, 4]
# result = [8, 9, 10, 7, 9, 10, 11, 12]
seq_lens_for_expansion = (
global_seq_lens if global_seq_lens is not None else seq_lens
)
expanded_offsets = torch.repeat_interleave(
seq_lens - decode_lens - query_start_loc,
seq_lens_for_expansion - decode_lens - query_start_loc,
decode_lens,
output_size=actual_expanded,
)

# [8, 9, 10, 7, 9, 10, 11, 12, ...] where ... is unused buffer space
self.decode_seq_lens_buffer[:actual_expanded] = (
expanded_seq_lens = (
expanded_offsets + self.arange_buffer[:actual_expanded] + 1
)
if global_seq_lens is not None and self.dcp_world_size > 1:
expanded_seq_lens = get_dcp_local_seq_lens(
expanded_seq_lens,
self.dcp_world_size,
self.dcp_rank,
self.cp_kv_cache_interleave_size,
)
self.decode_seq_lens_buffer[:actual_expanded] = expanded_seq_lens
self.decode_seq_lens_buffer[actual_expanded:] = 0
seq_lens = self.decode_seq_lens_buffer[:num_decode_tokens]

Expand Down Expand Up @@ -562,12 +593,23 @@ def _prepare_decode_tensors(
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_for_expansion = (
global_seq_lens if global_seq_lens is not None else seq_lens
)
expanded_seq_lens = (
seq_lens_for_expansion.unsqueeze(1)
- max_decode_len
+ 1
+ self.offsets_buffer[:max_decode_len]
)
if global_seq_lens is not None and self.dcp_world_size > 1:
expanded_seq_lens = get_dcp_local_seq_lens(
expanded_seq_lens.reshape(-1),
self.dcp_world_size,
self.dcp_rank,
self.cp_kv_cache_interleave_size,
).view(num_decodes, max_decode_len)
seq_lens_buffer[:] = expanded_seq_lens
seq_lens = seq_lens_buffer
return seq_lens, block_table, decode_lens, num_decodes, requires_padding

Expand Down Expand Up @@ -606,7 +648,7 @@ def build(
compressed_slot_mapping = slot_mapping
logical_compressed_seq_lens = seq_lens
compressed_seq_lens = seq_lens
if self.compress_ratio > 1:
if self.compress_ratio > 1 or self.dcp_world_size > 1:
compressed_slot_mapping = get_compressed_slot_mapping(
num_tokens,
query_start_loc,
Expand Down Expand Up @@ -716,7 +758,19 @@ def build(
decode_lens = self.decode_lens_buffer[:num_decodes]
decode_lens_np = query_lens_np[:num_decodes]

seq_lens = common_attn_metadata.seq_lens[:num_decodes]
global_decode_seq_lens = common_attn_metadata.seq_lens[:num_decodes]
dcp_local_seq_lens = (
common_attn_metadata.dcp_local_seq_lens[:num_decodes]
if self.compress_ratio == 1
and self.dcp_world_size > 1
and common_attn_metadata.dcp_local_seq_lens is not None
else None
)
seq_lens = (
dcp_local_seq_lens
if dcp_local_seq_lens is not None
else global_decode_seq_lens
)
block_table = common_attn_metadata.block_table_tensor[:num_decodes, ...]

max_decode_len = int(decode_lens_np.max())
Expand All @@ -735,6 +789,9 @@ def build(
use_native=use_native,
next_n=next_n,
max_decode_len=max_decode_len,
global_seq_lens=global_decode_seq_lens
if dcp_local_seq_lens is not None
else None,
)
)

Expand Down Expand Up @@ -779,7 +836,7 @@ def build(
)
self.expanded_seq_lens_buffer[num_decodes:num_decode_tokens] = 0
seq_lens = self.expanded_seq_lens_buffer[:num_decode_tokens]
elif self.dcp_world_size > 1:
elif self.dcp_world_size > 1 and dcp_local_seq_lens is None:
if seq_lens.dim() == 1:
seq_lens = get_dcp_local_seq_lens(
seq_lens,
Expand Down
Loading