MoE offload: decode overhead from padding masks (#83) and padding counted as expert 0 #89

Problem

After #83 (merged through #73), MoE offload decode throughput is about 5% lower. Padding ids are also counted as routing hits on expert 0.

Evidence

Qwen3-30B-A3B-Instruct-2507, moe_offload_generate.py --gpu-budget-gb 20 --kv-cache --max-new-tokens 128, 1× A100 80GB + host CPU, torch 2.13.0+cu130, transformers 5.15.0. Interleaved runs, 3 each:

• −5.7% mean, −3.0% median. Generated text is identical.
• A one-off adaptive run (--adapt-every 48) shows the same gap: 11.07 → 10.21.
• Likely cause: OffloadedLayer.forward now runs two compares, an and, a clamp and a mul on every call (about 5 kernel launches × 48 layers per token).
• _counts_d and the tier counts use the clamped ids, so padded rows count as expert 0.

Scope

1. Skip the mask/clamp when padding cannot occur, or fold it into the existing routing copy.
2. Count routing with the unclamped ids masked by valid.
3. In opt_grouped_mm, start from torch.zeros instead of hidden*0, so an inf in hidden does not become NaN.
4. Make tests/ skip, instead of fail, model families that the installed transformers does not have (on 4.46.2: solar_open, qwen3_moe, qwen3_next; rope_parameters).

Done when

• Throughput is back within noise of 41cfb2d on the setup above, with identical text.
• Padding no longer changes routing counts (test).
• pytest opt_prime/tests on torch 2.5.1 + transformers 4.46.2 has no failures, only skips.

Owner

@gspark-etri