[Spec] DFlash: remove per-step host syncs so the CPU runs a full step ahead (spec-v2 overlap) (#31468)

Co-authored-by: Hao Phan <htphan@nvidia.com>
This commit is contained in:
Thanhhao
2026-07-17 23:22:07 -07:00
committed by GitHub
co-authored by Hao Phan
parent 7fbe91c6ea
commit 72c4ed1a3f
11 changed files with 528 additions and 50 deletions
@@ -19,6 +19,8 @@ register_cpu_ci(est_time=5, suite="base-a-test-cpu")
class _FakeBackend:
def __init__(self, name):
self.name = name
# Real backends always carry this (AttentionBackend class attribute).
self.needs_cpu_seq_lens = True
def test_split_full_attention_applies_model_wrapper_once():