able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be

[mkeshavaNV]

[jkiviluoto-nv]

@nv-slang-bot Mukund (mkeshavaNV) won't be returning to this work for a while. Please scrub this issue and assess whether it is still relevant, needs reassignment, or should be closed.

[nv-slang-bot]

Scrub & reassess (requested by @jkiviluoto-nv). Verified against main @ 507b4cf, confirmed identical to the live remote head at time of writing.

• Status: Open, unimplemented, zero commits reference able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be #822 (also zero for take a thread count, in the event the kernel has a dimensionality of 0 (no arguments are vectorized) #819/able to generate a minimal wrapper for an entry point when the only thing missing is the compute shader / cuda kernel tag #821/able to directly call an entry point without wrapping it in a trampoline with a CallData structure if it is already tagged as a compute shader or cuda kernel #820/.dispatch() does not support torch.Tensor arguments #832/Support 2D/3D dispatch with configurable thread group size for torch tensor workflows #844; clone has full 949-commit history, so this absence is conclusive). Sub-issue of epic Support raw dispatch in slangpy #768 — its spec is checklist item 4 there. ⚠️ This issue's body is empty — the entire specification is its title. That's the epic's convention (take a thread count, in the event the kernel has a dimensionality of 0 (no arguments are vectorized) #819/able to directly call an entry point without wrapping it in a trampoline with a CallData structure if it is already tagged as a compute shader or cuda kernel #820/able to generate a minimal wrapper for an entry point when the only thing missing is the compute shader / cuda kernel tag #821 are also empty), but it means a new owner inherits no acceptance criteria, no repro, no design sketch.
• Q1 — Still relevant? YES, both title claims are absent at 507b4cf. (a) Raw entry point for the backwards pass — raw dispatch is structurally forward-only: DispatchData hardcodes CallMode.prim (slangpy/core/dispatchdata.py:50) and never reads call_mode, while backwards is unconditionally trampolined (generator.py:937, with the in-source note "we assume a trampoline is always needed for bwds"). The two paths are disjoint, and raw dispatch additionally rejects out/inout params and return values (dispatchdata.py:72-75) — contract-incompatible with the inout DifferentialPair signature a derivative needs, so this is more than "not yet written". test_raw_dispatch.py has zero bwds/grad references. (b) Detect [CUDAKernel] on forward ⇒ require it on backwards — CUDAKernel appears 32× outside external/ and zero times in library logic (all hits are authored .slang test/benchmark sources and docs prose); no fwd/derivative attribute consistency check exists anywhere. Corroboration that the epic is live: this repo's own tests still list "Support raw dispatch in slangpy #768 (raw dispatch support)" as a prerequisite for true slangtorch parity — slangpy/tests/slangpy_tests/test_torch_autograd_workflows.py:18-21, last touched 2026-07-31. Partial mitigation: _thread_count (Support _threadcount and float<N> for TensorView/DiffTensorview #818, merged 2026-03-05) means differentiable [CUDAKernel] backwards does work today via the CallData/trampoline path — documented at docs/src/basics/firstfunctions.rst:102-114, covered by test_difftensorview.py:289-310. So the user-facing capability is served; what's missing is the raw / no-trampoline path.
• Q2 — Needs reassignment? YES. Assignee field read explicitly: .assignee.login → mkeshavaNV (sole assignee). Since he is departing, that field is no longer coverage. On the sibling precedent — able to directly call an entry point without wrapping it in a trampoline with a CallData structure if it is already tagged as a compute shader or cuda kernel #820/able to generate a minimal wrapper for an entry point when the only thing missing is the compute shader / cuda kernel tag #821 went to @ccummingsNV on 2026-03-13, but the recorded reason was specific context, not generic ownability: "@ccummingsNV - should I move this to you? I believe you've started looking at it?" So that precedent alone doesn't justify routing able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be #822 anywhere. The independent case for @ccummingsNV is stronger: able to directly call an entry point without wrapping it in a trampoline with a CallData structure if it is already tagged as a compute shader or cuda kernel #820 and able to generate a minimal wrapper for an entry point when the only thing missing is the compute shader / cuda kernel tag #821 are prerequisites for able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be #822 (you can't wrap a raw entry point for backwards before the forward raw path exists) and he already owns both, is active on the adjacent surface (kernel-gen/torch-interop merges through 2026-07-31), and has Fix derived function reflection hot reload #1086 open on derivative reflection hot-reload. Sequencing matters more than the name here.
• Q3 — Should be closed? NO — recommend keep open + reassign. Not fixed (verified above), not a duplicate, not obsolete. But there is a real scope question a maintainer should settle rather than me: Support raw dispatch in slangpy #768's own body says slangpy's .dispatch() "is not well maintained and should be retired", and the direction that actually landed is _thread_count on the CallData path. So able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be #822's framing may be superseded even though the capability is genuinely absent. If a raw backwards path is no longer wanted, that is the close rationale — and it would apply to Support raw dispatch in slangpy #768's checklist item 4, not just this issue. Leaving the close decision to a maintainer; not closing unilaterally.
• Next action: (1) reassign off mkeshavaNV — to @ccummingsNV together with prerequisites able to directly call an entry point without wrapping it in a trampoline with a CallData structure if it is already tagged as a compute shader or cuda kernel #820/able to generate a minimal wrapper for an entry point when the only thing missing is the compute shader / cuda kernel tag #821, or explicitly park able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be #822 behind them; (2) restate the acceptance criteria in the body as part of the handoff — a bare reassignment of an empty-body issue hands over a title, not a task; (3) maintainer to answer the supersede question above. One caveat for whoever takes claim (b): confirm [CUDAKernel] is reflectable at all before designing a check — SlangPy's find_user_attribute_by_name (src/sgl/device/reflection.cpp:349) wraps Slang's findUserAttributeByName, i.e. user attributes, and [CUDAKernel] is a built-in modifier. I could not verify reachability locally (Slang ships as a prebuilt binary, SGL_SLANG_VERSION 2026.12); if built-in modifiers aren't reflectable, claim (b) becomes an upstream Slang reflection request.
• Blocker: none technical. Two human decisions: the reassignment itself, and the keep-vs-supersede scope call on Support raw dispatch in slangpy #768 item 4.

_{Also noted while scrubbing, outside this issue's scope: #832 (".dispatch() does not support torch.Tensor arguments", same departing assignee) may already be fixed — test_raw_dispatch.py:264 test_dispatch_torch_tensor was added 2026-07-31 and does exactly that; worth a separate check. And #768's checklist is stale — item 1 (= #819, closed/completed 2026-03-05) is still unchecked.}

_{🤖 Generated by an automated SlangPy coworker — may be inaccurate. A human maintainer should verify.}