[MM][OOT] Support CPU seq_lens for OOT MMEncoderAttention kernels (#36605)

Signed-off-by: shen-shanshan <467638484@qq.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Shanshan Shen
2026-03-12 03:28:23 -07:00
committed by GitHub
co-authored by Isotr0py
parent 57431d8231
commit f0d3658c0f
5 changed files with 52 additions and 40 deletions
+6 -9
View File
@@ -297,11 +297,10 @@ def test_mha_attn_varlen_forward_flashinfer(
hidden_size = num_heads * head_size
tp_size = 1
sequence_lengths_np = MMEncoderAttention.maybe_compute_sequence_lengths(
AttentionBackendEnum.FLASHINFER, cu_seqlens_np
)
sequence_lengths = torch.from_numpy(sequence_lengths_np).to(
device, dtype=torch.int32, non_blocking=True
sequence_lengths = MMEncoderAttention.maybe_compute_seq_lens(
AttentionBackendEnum.FLASHINFER,
cu_seqlens_np,
device,
)
max_seqlen_val = MMEncoderAttention.compute_max_seqlen(
@@ -309,14 +308,12 @@ def test_mha_attn_varlen_forward_flashinfer(
)
max_seqlen = torch.tensor(max_seqlen_val, device=device, dtype=torch.int32)
cu_seqlens_np = MMEncoderAttention.maybe_recompute_cu_seqlens(
cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
AttentionBackendEnum.FLASHINFER,
cu_seqlens_np,
hidden_size,
tp_size,
)
cu_seqlens = torch.from_numpy(cu_seqlens_np).to(
device, dtype=torch.int32, non_blocking=True
device,
)
scale = 1.0 / head_size**0.5