[Perf][GLM-5.3-Flash] Dense/masked-MHA sparse prefill for the NoPE (256, 0, 256) layout + skip the NoPE K concat (#55738)

Signed-off-by: Jared Wen <jaredwen@inferact.ai>
Signed-off-by: Jared Wen <w13431838023@gmail.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Jared Wen
2026-09-14 09:36:39 +00:00
committed by GitHub
co-authored by Claude Fable 5.1 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 79f0be21ff
commit 238cb2b191
6 changed files with 96 additions and 0 deletions
@@ -2871,6 +2871,10 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
Returns:
Tensor of shape [..., nope_dim + pe_dim]
"""
if k_pe.shape[-1] == 0:
# NoPE MLA: nothing to append, so no copy either.
return k_nope
k = torch.empty(
(*k_nope.shape[:-1], k_nope.shape[-1] + k_pe.shape[-1]),
dtype=k_nope.dtype,
@@ -111,6 +111,8 @@ def _is_masked_mha_available(
if model_dims not in (
(128, 512, 128, 64, 128),
(64, 512, 192, 64, 256),
# GLM-5.3-Flash: NoPE, qk_head_dim 256 == the (192, 64, 256) kernel.
(64, 512, 256, 0, 256),
):
return False
qk_head_dim = qk_nope_head_dim + qk_rope_head_dim