Skip to content

Use repeat_interleave in Mamba2 seq_idx computation to reduce cpu-gpu syncs #16743

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

Closed
wants to merge 2 commits into from
Closed

Use repeat_interleave in Mamba2 seq_idx computation to reduce cpu-gpu syncs #16743

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
828358c
Use repeat_interleave to reduce cpu-gpu syncs
cyang49 Apr 16, 2025
9827d4d
remove unnecessary computations for cu_seqlens
cyang49 Apr 16, 2025
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
45 changes: 26 additions & 19 deletions vllm/model_executor/layers/mamba/mamba2_metadata.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,21 +24,23 @@ class Mamba2Metadata:
chunk_offsets: torch.Tensor


def _seq_idx_to_chunk_indices_offsets(seq_idx, chunk_size: int):
def _query_start_loc_to_chunk_indices_offsets(query_start_loc: torch.Tensor,
chunk_size: int,
total_seqlens: int):

# convert seq_idx to chunk indices and offsets
# - derive the cu_seqlens
_, cu_seqlens = torch.where(seq_idx.diff())
cu_seqlens += 1
cu_seqlens = query_start_loc[1:] # remove prepended 0

# outputs will have length expansion of chunks that do not divide
# chunk_size
N = math.ceil(seq_idx.shape[-1] / chunk_size) + (cu_seqlens % chunk_size
> 0).sum()
chunk_indices = torch.arange(N, dtype=torch.int, device=seq_idx.device)
chunk_offsets = torch.zeros((N, ), dtype=torch.int, device=seq_idx.device)
N = math.ceil(total_seqlens / chunk_size) + (cu_seqlens[:-1] % chunk_size
> 0).sum()
chunk_indices = torch.arange(N,
dtype=torch.int,
device=query_start_loc.device)
chunk_offsets = torch.zeros((N, ),
dtype=torch.int,
device=query_start_loc.device)

cu_seqlens = cu_seqlens.tolist() + [seq_idx.shape[-1]]
p = 0 # num of insertions
for s, e in zip(cu_seqlens[:-1], cu_seqlens[1:]):

Expand Down Expand Up @@ -80,13 +82,15 @@ def prepare_mamba2_metadata(
seq_idx = None
chunk_indices, chunk_offsets = None, None
if has_prefill:
seq_idx = torch.zeros_like(input_ids, dtype=torch.int32)
for i, (srt, end) in enumerate(
zip(
attn_metadata.query_start_loc,
attn_metadata.query_start_loc[1:],
)):
seq_idx[srt:end] = i
seqlens = attn_metadata.query_start_loc.diff()
total_seqlens = len(input_ids)

seq_idx = torch.repeat_interleave(torch.arange(
len(attn_metadata.query_start_loc) - 1,
dtype=torch.int32,
device=attn_metadata.query_start_loc.device),
seqlens,
output_size=total_seqlens)
seq_idx.unsqueeze_(0)

# compute metadata for chunked prefill.
Expand All @@ -97,8 +101,11 @@ def prepare_mamba2_metadata(
# compute them once at the top level model forward and reuse
# them in mamba layers. If not needed, they will be ignored
# inside mamba kernels.
chunk_indices, chunk_offsets = _seq_idx_to_chunk_indices_offsets(
seq_idx, chunk_size)
chunk_indices, chunk_offsets = \
_query_start_loc_to_chunk_indices_offsets(
attn_metadata.query_start_loc,
chunk_size=chunk_size,
total_seqlens=total_seqlens)

return Mamba2Metadata(has_prefill=has_prefill,
has_initial_states=has_initial_states,
Expand Down