Skip to content

frost(sdpa): SM107 fp8 sibling missing the #585 LPT_L2 scheduler port — all causal fp8 declines on Rubin #653

Description

@vedaanta

Symptom

On cc10.7 at develop tip, every causal/masked per-tensor FP8 SDPA graph fails plan build:

NotImplementedError: sdpa_fwd_prefill_sm100_d128_fp8: SCHED_LPT_L2 decode requires qh_per_kh and seqlen_kv at every call site

test_sdpa_fwd_fp8_sm100.py -m 'L0 or L1': 29 failed / 4 passed on a Rubin board (the 4 survivors are the nomask cases). The same suite is 33/33 on SM100.

Cause

#585 added the LPT_L2 tile scheduler to prefill_d128_fp8_sm100.py (and the adapter's auto causal sched policy now selects it), but the hunk-symmetric port to the SM107 sibling prefill_d128_fp8_sm107.py was not done — the sibling carries only the decode guard, which correctly refuses. 27 vs 7 SCHED_LPT|qh_per_kh references between the two files.

This is the sibling-lockstep hazard (see #579's history): kernel-shared changes must land in both files, and nothing gates it when no Rubin silicon is in the merge loop.

Fix

Port #585's scheduler hunks to the sibling (mechanical, hunk-symmetric), validate causal/BR/SWA fp8 on cc10.7. Interim mitigation if needed: pin sched_policy=SCHED_NATURAL on the Rubin route until the port lands.

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:global_attentioncat-bugReports of incorrect behavior, crashes, regressions, or unexpected results.mod-frostorig-nv-engReported or requested by NVIDIA engineering.

    Type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions