Skip to content

support sparse gqa and fix flashattn #1039

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
helloyongyang merged 2 commits into from
Apr 28, 2026
Merged

support sparse gqa and fix flashattn #1039

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d32c61d
fix as gemini-code-assist suggested
STwangyingrui Apr 24, 2026
79c6d34
sparse way of sla and sparge support GQA
STwangyingrui Apr 24, 2026
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
16 changes: 4 additions & 12 deletions lightx2v/common/ops/attn/flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,8 +5,8 @@
from .utils.sparge_util import block_map_ordinal_lut_triton, get_block_map_meansim

try:
from flash_attn import flash_attn_func_v2
from flash_attn.flash_attn_interface import flash_attn_varlen_func_v2
from flash_attn import flash_attn_func as flash_attn_func_v2
from flash_attn.flash_attn_interface import flash_attn_varlen_func as flash_attn_varlen_func_v2
except ImportError:
logger.info("flash_attn2 not found, please install flash_attn2 first")
flash_attn_func_v2 = None
Expand Down Expand Up @@ -59,16 +59,12 @@ def apply(
q = q.unsqueeze(0)
k = k.unsqueeze(0)
v = v.unsqueeze(0)
x = flash_attn_func_v2(q, k, v).reshape(bs * max_seqlen_q, -1)
x = flash_attn_func_v2(q, k, v).reshape(total_seqlen, -1)
else:
if cu_seqlens_q.is_cpu:
cu_seqlens_q = cu_seqlens_q.to(q.device, non_blocking=True)
if cu_seqlens_kv.is_cpu:
cu_seqlens_kv = cu_seqlens_kv.to(k.device, non_blocking=True)
if max_seqlen_q.is_cpu:
max_seqlen_q = max_seqlen_q.to(q.device, non_blocking=True)
if max_seqlen_kv.is_cpu:
max_seqlen_kv = max_seqlen_kv.to(k.device, non_blocking=True)
if len(q.shape) == 4:
q = q.reshape(-1, q.shape[-2], q.shape[-1])
k = k.reshape(-1, k.shape[-2], k.shape[-1])
Expand Down Expand Up @@ -113,16 +109,12 @@ def apply(
q = q.unsqueeze(0)
k = k.unsqueeze(0)
v = v.unsqueeze(0)
x = flash_attn_func_v3(q, k, v).reshape(bs * max_seqlen_q, -1)
x = flash_attn_func_v3(q, k, v).reshape(total_seqlen, -1)
else:
if cu_seqlens_q.is_cpu:
cu_seqlens_q = cu_seqlens_q.to(q.device, non_blocking=True)
if cu_seqlens_kv.is_cpu:
cu_seqlens_kv = cu_seqlens_kv.to(k.device, non_blocking=True)
if max_seqlen_q.is_cpu:
max_seqlen_q = max_seqlen_q.to(q.device, non_blocking=True)
if max_seqlen_kv.is_cpu:
max_seqlen_kv = max_seqlen_kv.to(k.device, non_blocking=True)
if len(q.shape) == 4:
q = q.reshape(-1, q.shape[-2], q.shape[-1])
k = k.reshape(-1, k.shape[-2], k.shape[-1])
Expand Down
9 changes: 9 additions & 0 deletions lightx2v/common/ops/attn/utils/sla_util.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,15 @@ def get_block_map(q, k, topk_ratio, BLKQ=64, BLKK=64):
arg_k = k - torch.mean(k, dim=-2, keepdim=True) # smooth-k technique in SageAttention
pooled_qblocks = mean_pool(q, BLKQ)
pooled_kblocks = mean_pool(arg_k, BLKK)

# GQA
num_q_heads = q.size(1)
num_kv_heads = k.size(1)
if num_q_heads != num_kv_heads:
assert num_q_heads % num_kv_heads == 0, f"Number of Q heads ({num_q_heads}) must be divisible by number of KV heads ({num_kv_heads})"
repeat_factor = num_q_heads // num_kv_heads
pooled_kblocks = pooled_kblocks.repeat_interleave(repeat_factor, dim=1)

pooled_score = pooled_qblocks @ pooled_kblocks.transpose(-1, -2)

K = pooled_score.shape[-1]
Expand Down
9 changes: 9 additions & 0 deletions lightx2v/common/ops/attn/utils/sparge_util.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,6 +213,15 @@ def get_block_map_meansim(q, k, is_causal=False, BLKQ=128, BLKK=64, simthreshd1=
pooled_qblocks, sim_qblocks = get_pool_sim_triton_simmean(q, BLKQ, simthreshd1)
pooled_kblocks, sim_kblocks = get_pool_sim_triton_simmean(k, BLKK, simthreshd1)

# GQA
num_q_heads = q.size(1)
num_kv_heads = k.size(1)
if num_q_heads != num_kv_heads:
assert num_q_heads % num_kv_heads == 0, f"Number of Q heads ({num_q_heads}) must be divisible by number of KV heads ({num_kv_heads})"
repeat_factor = num_q_heads // num_kv_heads
pooled_kblocks = pooled_kblocks.repeat_interleave(repeat_factor, dim=1)
sim_kblocks = sim_kblocks.repeat_interleave(repeat_factor, dim=1)

sim_kblocks = sim_kblocks.unsqueeze(-2).expand(-1, -1, nq, -1) # faster than repeat
sim_qblocks = sim_qblocks.unsqueeze(-1).expand(-1, -1, -1, nk)
pooled_score = pooled_qblocks @ pooled_kblocks.transpose(-1, -2) * q.shape[-1] ** -0.5
Expand Down
4 changes: 2 additions & 2 deletions lightx2v/models/networks/bagel/infer/transformer_infer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,8 +160,8 @@ def self_attn(
merged_value_states = packed_value_states
key_values_lens = query_lens

cu_seqlens_q = torch.nn.functional.pad(torch.cumsum(query_lens, dim=0), (1, 0))
cu_seqlens_k = torch.nn.functional.pad(torch.cumsum(key_values_lens, dim=0), (1, 0))
cu_seqlens_q = torch.nn.functional.pad(torch.cumsum(query_lens, dim=0), (1, 0)).to(AI_DEVICE)
cu_seqlens_k = torch.nn.functional.pad(torch.cumsum(key_values_lens, dim=0), (1, 0)).to(AI_DEVICE)
Comment on lines +163 to +164
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Apr 24, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The cu_seqlens_q and cu_seqlens_k tensors are recomputed and moved to the device for every layer in the transformer. Since query_lens and the updated key_values_lens are constant across all layers within a single inference step, these cumulative sequence lengths should ideally be computed once outside the layer loop (e.g., in the infer method) and passed down to self_attn to avoid redundant computations and expensive host-to-device transfers.

Additionally, you can combine the device move and type conversion into a single .to(device=AI_DEVICE, dtype=torch.int32) call to improve efficiency, as flash_attn requires int32 tensors.

Suggested change
cu_seqlens_q = torch.nn.functional.pad(torch.cumsum(query_lens, dim=0), (1, 0)).to(AI_DEVICE)
cu_seqlens_k = torch.nn.functional.pad(torch.cumsum(key_values_lens, dim=0), (1, 0)).to(AI_DEVICE)
cu_seqlens_q = torch.nn.functional.pad(torch.cumsum(query_lens, dim=0), (1, 0)).to(device=AI_DEVICE, dtype=torch.int32)
cu_seqlens_k = torch.nn.functional.pad(torch.cumsum(key_values_lens, dim=0), (1, 0)).to(device=AI_DEVICE, dtype=torch.int32)


packed_attn_output = flash_attn_varlen_func(
q=packed_query_states,
Expand Down
Loading