[glennneuber]

The record of #312 on the CUDA side, and the rule it leaves behind.

Task doc (tasks/upstream-sync-0.34.1.md, new section "The MLX pin move and the vision encoder"): TestVisionGoldenParity at both ends of 2b95b4a5 → 8a7ba949 on MLX-CUDA, two runs per cell, every repeat to the digit. 12b and 26b identical across the fold; 31b moved slightly. Two causes, one per backend:

• Metal-only, a fix: Fix fp quantized matmul corruption when the quantized dim is not a multiple of 32 ml-explore/mlx#3912 (fp_qmm_t read past K when K % 32 == 16). gemma4's vision mlp.down_proj has K = 4304 in 26b and 31b; 12b has no such shape. The CUDA kernel never had the defect: 26b reads 0.0469 at both ends, which the fixed Metal kernel (0.0508) now agrees with. No campaign on this host is affected.
• Shared Go code, small: the fold's global-scale f32(f32(m × 2688) / 2688) round trip, one f32 ulp off for 17 of 31b's 191 vision global scales and exact for both of 12b's. A two-line control on the current fold reproduces the old 31b line to the digit. Inside the test's bounds; whether to keep the raw multiplier on the wrapper-applied paths is the maintainer's call.

Plus the recipe for running x/mlxrunner tests against the CUDA payload outside the image (the include symlink MLX needs for CCCL), which is what made the old side a 90-second run instead of a rebuild.

Rule for the next fold: list the pin range's commits and check every kernel fix's selection condition against the served models' quantized dims. Enforced by the new vision-suite/quant_dims.py (reads safetensors headers from the store, lists each quantized weight's K, flags K % 32 != 0; 8 offline tests), and a learnings-log entry with the cost.

Against the store today:

### gemma4:31b-nvfp4: 602 quantized weights, 27 with K % 32 != 0, 0 undetermined
### gemma4:26b-nvfp4: 455 quantized weights, 27 with K % 32 != 0, 0 undetermined
### gemma4:12b-nvfp4: 332 quantized weights, 0 with K % 32 != 0, 0 undetermined

Discussion: #312 (my two comments there carry the measurement and the reconciliation against the Metal session's prediction table).

🤖 Generated with Claude Code