Add routed-expert MXFP4 QAT for Hy3

[ISEEKYAN]

Summary

• expose Megatron Lite's standard MXFP4 QAT primitive through ImplConfig.qat
• keep the model integration declarative and apply QAT before optimizer construction
• map Hy3's split Transformer Engine expert parameters onto BF16 QAT masters
• support canonical checkpoint load, BF16 round trip, and routed-only MXFP4 HF export
• provide reproducible scheduler entry points with pinned source revisions and import-path assertions

Quantization scope

The policy matches K3: quantize every routed-path linear and no other component. The concrete structures differ:

• Hy3 fc1.weightK is the fused gate-plus-up projection for local expert K; fc2.weightK is its down projection.
• These correspond to K3 experts.K.gate_up.weight and experts.K.down.weight.
• K3's latent-MoE additionally quantizes shared routed-path routed_expert_down_proj and routed_expert_up_proj linears and has a routed-expert norm. Hy3 has no equivalent latent bottleneck or norm.
• Hy3's K is the local-expert index because Megatron Lite passes num_experts / ep_size to each Transformer Engine GroupedLinear.

Attention, dense MLP, shared experts, embedding, router/correction bias, and lm_head remain BF16. Hy3 also has no K3-style standalone attention residual-projection modules: its attention output projection is excluded under attn, while residual additions have no parameters.

The Hy3 declaration therefore matches exactly moe.experts.fc1/fc2 modules and parameters named weight followed only by digits. This covers every local expert tensor without matching similarly named non-weight state.

Validation

• real Tencent Hy3 routed-expert weights: 100,663,296 elements across 16 tensors; fake quantization and serialized dequantization are bitwise equal to ModelOpt 0.43
• expected full-model routed-expert name set derived from the Hy3 checkpoint map: 46,080 tensors
• runtime parameter contract: the QAT BF16 master set must equal the pre-QAT routed fc1/fc2.weightK set; the two-expert tiny model has exactly 4 masters
• canonical checkpoint load and BF16 export: 33 tensors, bitwise round trip
• routed-only MXFP4 export: 6 packed HF weights with scales; excluded weights remain BF16
• forward/backward: finite loss and gradients present on every routed-expert master
• CPU suite: 21 passed; 8 scheduler/CUDA-only tests skipped locally and covered by the non-skipped scheduler acceptance above
• Ruff check, changed-file Ruff format check, shell syntax, and whitespace checks: passed

The repository currently has no .pre-commit-config.yaml or pinned Ruff version. The equivalent Ruff check and changed-file format checks were run directly without reformatting unrelated pre-existing files.