Support Nemotron-H / Nemotron-Cascade-2 (hybrid Mamba2 + Attention + MoE)

[michael-rabe]

Goal

Enable AutoRound to quantize Nemotron-H family models — in particular nvidia/Nemotron-Cascade-2-30B-A3B — and produce a checkpoint whose decoded output is coherent without requiring launcher-side workarounds.

Target user flow:

from auto_round import AutoRound
AutoRound("nvidia/Nemotron-Cascade-2-30B-A3B", bits=4, group_size=128, sym=True) \
    .quantize_and_save(output_dir=..., format="auto_round")

Scope

• Nemotron-H architecture (hybrid Mamba2 SSM + Attention + routed MoE + shared expert)
• auto_round export format, W4A16 weight-only
• CPU + CUDA/XPU calibration paths

Acceptance

• Bare AutoRound("nvidia/...") call produces a checkpoint that loads and generates coherent text in a downstream inference engine.
• No regressions for other model_type values (non-NH models must be unaffected).
• CPU test suite covers registration, post-load behaviour, and export-time precision controls.

Out of scope (follow-ups)

• Pure-Mamba architectures (no attention layers)
• Other Nemotron-H variants beyond the Cascade-2 checkpoint family