Non-record: PR315 repro on 1xH100 PCIe, int6+zstd (val_bpb=1.8338)

[sjp611]

Summary

• Based on PR Record: 11L Partial RoPE + LN Scale + EMA + XSA4 (val_bpb: 1.1248) #315 recipe (XSA+EMA+PartialRoPE+LateQAT), adapted for 1xH100 PCIe
• Flash Attention 2 instead of FA3
• Training schedule adjusted for 1GPU (WARMDOWN_ITERS=400, MOMENTUM_WARMUP=200)
• Sliding-window val_bpb=1.8338, artifact 10.0MB (int6+zstd)

Notes

• 1GPU limits to ~492 steps (vs ~6200 on 8xH100)
• QAT had only ~40 steps of adaptation due to limited training budget

Test plan

• Training completes within 10min wallclock
• Artifact under 16MB (10.0MB)
• int6+zstd roundtrip eval passes