Fused fprop scale/bias iterator advances without the split-k factor and drifts from activation channels

Description

In the fused (scale/bias) fprop mainloop, the scale/bias vector iterator and the activation/filter iterators disagree on the per-R*S-cycle channel stride when split_k_slices > 1, so scale/bias drift to the wrong channel blocks.

• Activation iterator on gemm-k wrap (include/cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h ~line 216, _optimized analogue ~314): filter_c_ += Shape::kColumn * problem_size_.split_k_slices;
• Scale/bias iterator on the same wrap (include/cutlass/conv/threadblock/predicated_scale_bias_vector_access_iterator.h:247-262): calls add_tile_offset(TensorCoord(1, 0)), which advances by a single ThreadblockShape::kContiguous, with no split-k factor.

Both start consistently at channel threadblock_tile_idx.k() * kK (implicit_gemm_convolution_fusion.h ~296-337), so the divergence accumulates once the loop wraps past RS. With interleaved slice partitioning, each additional RS cycle applies bias/scale from one kK-block earlier than the channels the activations actually carry. This needs split_k_slices > 1 and more than one C-channel block per slice (ceil(ceil(C/S)/kK) >= 2) to appear, e.g. split_k_slices=4, C=256, ThreadblockShape::kK=32, RS>=2: from the second cycle on, bias[64+j32] is applied where the activations come from channels 192+. The result is silent wrong numerics, not an OOB.

The same pattern exists in the conv3d fusion headers (conv3d_fprop_activation_tile_access_iterator_analytic/_optimized vs the shared scale/bias iterator). Wgrad fusion is unaffected: it loads scale/bias once, with no per-tile advance.

The device layer explicitly supports split-k for fused fprop (workspace handling in include/cutlass/conv/device/implicit_gemm_convolution_fusion.h ~148-175), so the combination is reachable through the public API.

This finding is from static verification of the two advance paths; I have not built an end-to-end failing kernel, but the stride mismatch is exactly the missing split_k_slices factor and the start offsets agree.

Suggested fix

Multiply the scale/bias iterator's wrap stride by problem_size.split_k_slices in advance(), mirroring the activation iterators.

Start with predicated_scale_bias_vector_access_iterator.h and compare its advance path with the 2D activation iterator headers named in the issue; then inspect the fused fprop device path in include/cutlass/conv/device/implicit_gemm_convolution_fusion.h. Check the corresponding conv3d iterator pair as well. Done means the scale/bias and activation iterators use matching split-k strides and split-k fused fprop numerics no longer drift.