[CUDA] Fix qmm_naive K-tail dispatch for FP quantized kernels

[Lyxot]

Fixes: #3444

Root Cause

In qmm.cu, qmm_naive selected the HasKResidue specialization using:

bool has_k_residue = k % group_size != 0;

but the kernel tiles the K dimension using max(64, group_size).

For FP quantization modes:

• mxfp4 / mxfp8 use group_size = 32
• nvfp4 uses group_size = 16

So shapes like K=544 satisfy:

• K % group_size == 0
• K % 64 != 0

The old dispatch therefore selected the no-residue specialization even when the kernel still had a real K tail.

Fix

Change the residue check to match the kernel's actual K tiling

[Copilot]

Pull request overview

Fixes a CUDA quantized-matmul (qmm_naive) dispatch bug where the wrong K-tail specialization was chosen for FP-quantized modes by aligning the residue check with the kernel’s actual K tiling (max(64, group_size)), addressing issue #3444.

Changes:

• Compute tile_k = max(64, group_size) inside qmm_naive.
• Change the HasKResidue dispatch condition from k % group_size to k % tile_k.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[zcbenz]

Nice fix, thanks!