Add NVFP4 four-over-six converter

[wolfcomos]

What

Wires TorchAO's NVFP4 four-over-six training prototype into torchtitan via
the model-converter pattern, opt-in through converter config.

Depends on the TorchAO PR wolfcomos/ao#7
(torchao.prototype.moe_training.nvfp4_training.four_over_six, branch
nvfp4-four-over-six-rowwise). This PR's NVFP4FourOverSixLinear is None
(and the converter raises ImportError) on torchao builds without it.

Four-over-six is an adaptive NVFP4 block-scaling recipe: every FP4 block is
encoded twice — the standard map-to-6 scale and a 1.5x-expanded map-to-4 scale
— and the candidate with the lower MAE/MSE dequantization error is stored.
Forward GEMM operands are four-over-six (activations 1x16, optionally
row-scaled with one FP32 global scale per row; weights 16x16). Gradients use
standard NVFP4, or bf16 GEMMs in row-scaled mode (a row-scaled four-over-six
tensor has no columnwise form for the wgrad operand). No RHT and no stochastic
rounding: the recipe targets RL and post-training.

• NVFP4FourOverSixLinear: pure Linear leaf swap (like MXFP8Linear),
  forward through four_over_six_linear. TP is not wired up.
• NVFP4FourOverSixLinearConverter: config knobs fqns, err_mode
  (mae/mse), e4m3_scale_bound (256/448), row_scaled_activation.
• llama3_debugmodel_nvfp4_four_over_six recipe (fqns=["layers"]; the
  lm_head stays bf16 since the vocab dim is not divisible by 128).

Evidence (2x GB200, FSDP=2, llama3 debugmodel, 50 steps)

step	bf16	four-over-six	four-over-six row-scaled
1	8.21253	8.15202	8.06924
10	3.71664	3.67363	3.59410
30	2.86240	2.87019	2.86336
50	2.87953	2.88080	2.87267

The underlying torchao quantizer is verified bitwise (384/384 comparisons)
against the reference CUDA kernels for this recipe; see the TorchAO PR for the
full parity evidence.

🤖 Generated with Claude Code