level3 conv-family: accept seed to make ref ↔ cand verification possible

[zhshgmail]

Problem

All five level-3 conv ops (6_ConvStandard1d, 7_ConvStandard2d, 8_ConvStandard3d, 9_ConvDepthwise2d, 10_ConvTranspose2d) have Model.forward() that constructs nn.ConvNd(...) inline on every call. The verification harness utils/verification_ascendc.py:306 does torch.manual_seed(0) once before the per-case for-loop. Inside each iteration it calls ref_model.forward(*args) then cand_model.forward(*args) back-to-back.

Because nn.ConvNd.__init__ draws random weights from the global RNG, and the RNG state advances after the ref call, cand draws different weights than ref. Verification then compares two different random outputs → ref ≠ cand by construction, regardless of whether the candidate kernel is correct.

This makes the conv family structurally unverifiable. It hits any worker porting nn.Conv*d to a candidate kernel: even an exact PyTorch re-implementation (e.g. F.conv1d(x, conv.weight, conv.bias, ...)) would diverge from the ref because the ref's conv.weight ≠ the ones the cand sees.

Discovered during an op-gen cold-start of op#6 ConvStandard1d on Ascend950PR (2026-04-28). The downstream workaround would be a workspace-local model.py with torch.manual_seed(...) inside forward — but that has to be reapplied per-op per-cold-start. Fixing it upstream unblocks every future cold-start at no cost.

Fix

Add a new optional seed kwarg (default 0) to each conv forward(). At the start of forward(), call torch.manual_seed(seed) so the inline conv construction draws weights from a fixed RNG state. Both ref and cand pass the same seed → both construct conv with identical weights → comparison is meaningful.

Also adds .to(x.device) after construction so the conv's weights track the input device (this is the existing CPU-vs-NPU device-mismatch error that the inline-construction pattern produces when called with NPU input — was already a separate bug, fixed here in passing).

get_input_groups() reads seed from the JSON case attrs (defaulting to 0) so existing JSON cases continue to work unchanged.

Backward compat

• seed is optional with default 0. Existing JSON cases that don't include {"name": "seed", ...} continue to work via attr_inputs.get("seed", 0).
• ModelNew / candidate implementations that don't pass seed will get seed=0 by default. They MUST also seed before any RNG-dependent construction (or extract weights from the seed-0 conv) to match ref.

Verification

Tested on A5 NPU (Ascend950PR_9589, container a5ops_dev3):

ref_preflight on op#6 ConvStandard1d: verdict=RUNNABLE
  (was FAILED with "Expected all tensors to be on the same device" before this fix)

Two consecutive `Model().forward(x, ..., seed=42)` calls — bit-exact identical output:
  6_ConvStandard1d: ✓ bit-exact
  7_ConvStandard2d: ✓ bit-exact
  9_ConvDepthwise2d: ✓ bit-exact
  10_ConvTranspose2d: ✓ bit-exact

(op#8 ConvStandard3d skipped in determinism test due to different x rank, but fix is identical pattern across all five files.)

Files changed

• benchmarks/NPUKernelBench/level3/6_ConvStandard1d.py
• benchmarks/NPUKernelBench/level3/7_ConvStandard2d.py
• benchmarks/NPUKernelBench/level3/8_ConvStandard3d.py
• benchmarks/NPUKernelBench/level3/9_ConvDepthwise2d.py
• benchmarks/NPUKernelBench/level3/10_ConvTranspose2d.py

🤖 Generated with Claude Code