frost(sdpa): SM80 adapters still normalization-copy non-compact operands (`_stage`/`_to_bshd` `.contiguous()`) — should address declared strides natively

[egilliam-nv]

Follow-up from #716 review (requested by @vedaanta): remove the SDPA SM80 adapters' operand staging — most immediately the per-execute .contiguous() allocations inside it — in favor of kernels that address declared operand strides natively.

Current state

Backward — python/cudnn/sdpa/bwd/api.py execute._stage (on develop today):

return _pad_last_dim(view.contiguous() if not view.is_contiguous() else view, fd)
...
return view.contiguous()

Every execute with a dense_flex (non-BSHD-compact) operand or a head-dim pad allocates fresh tensors — an AGENTS.md Rule 1 violation ("no per-execute allocation") that #716's carving pass fixed for the kernel-internal buffers but not for these operand gathers.

Forward — the SdpaFwdDsl base helpers have the same issue for wrapper-path callers:

• _to_bshd: return view if view.is_contiguous() else view.contiguous()
• _to_bshd_writable: torch.empty_like(view, ...) per execute

Two milestones

1. Kill the allocations (Rule 1) — gather into carved workspace instead of .contiguous(). For the backward this lands with frost(sdpa): port the SM80 backward onto SdpaBwdDsl + TemplateParams with sym_int THD extents (issue #604) #765, which deletes bwd/api.py; its replacement _stage carves via the Convert torch-native SM80 SDPA adapters to scratch workspace carving #514 machinery (zero allocation when a workspace is provided). The forward's _to_bshd/_to_bshd_writable residual remains open after frost(sdpa): port the SM80 backward onto SdpaBwdDsl + TemplateParams with sym_int THD extents (issue #604) #765.

2. Kill the staging copies entirely (Rule 2 end-state) — even a carved gather is a copy per execute. The kernels should address non-compact operand layouts natively via declared strides, the way the forward f16 kernels already handle THD declared token strides ("no normalization-copy fallback") and the way frost(sdpa): SM80 backward native strided-LSE reads + THD max_s_kv grid hint #766 makes the backward's LSE reads stride-aware (after frost(sdpa): enable sdpa_fwd engines to write dense LSE directly to non-contiguous, dense-compatible layouts #712 did the forward's LSE writes). Concretely: stride-parameterized global loads for Q/K/V/O/dO (and stores for dQ/dK/dV) in the SM80 kernels, with the plan-time declared strides carried on the template compile() seam like lse_stride. Once operands bind natively, _stage and the gather sizing in scratch_workspace_bytes() can be deleted outright. Declarations the hardware/addressing cannot express should be REJECTED in check_support, not silently copied.

Head-dim padding (flavor envelopes) is a separate concern from layout normalization and may legitimately keep a carved staging path — or move into the kernels as predicated loads; to be decided as part of milestone 2.

Related

• Convert torch-native SM80 SDPA adapters to scratch workspace carving #514 (workspace carving), frost(sdpa): enable sdpa_fwd engines to write dense LSE directly to non-contiguous, dense-compatible layouts #712 (fwd native strided LSE writes), frost(sdpa): port the SM80 backward onto SdpaBwdDsl + TemplateParams with sym_int THD extents (issue #604) #765 (bwd adapter port; fixes the backward's milestone 1), frost(sdpa): SM80 backward native strided-LSE reads + THD max_s_kv grid hint #766 (bwd native strided LSE reads)