Support NemotronH hybrid Mamba+Attention+MoE training (e.g. Nemotron Nano-30B-A3B)

[HelloWorldLTY]

Motivation

slime currently always builds a GPTModel (pure transformer). Hybrid Mamba-Attention-MoE models such as NVIDIA NemotronH (e.g. Nemotron Nano-30B-A3B: Mamba2 + attention + MoE, pattern MEMEM*E...) therefore get param names that don't match a Megatron-core NemotronH checkpoint (input_layernorm instead of mixer.*), and the training forward crashes because mcore MambaModel.forward() does not accept the loss_mask kwarg that GPTModel does.

This PR adds NemotronH (hybrid mamba arch) training support.

Modifications

• slime/backends/megatron_utils/model_provider.py: when --hybrid-override-pattern is set, build megatron.core.models.mamba.MambaModel with mamba_stack_spec instead of GPTModel. Param names (decoder.layers.N.mixer.*, GroupedMLP experts.experts, final_norm) then match a Megatron-core NemotronH checkpoint. The critic role still swaps in LinearForLastLayer.
• slime/backends/megatron_utils/model.py: drop the loss_mask kwarg via a signature check before calling the model. Loss masking already happens in slime's own loss function, so behavior is unchanged for GPT-family models.
• scripts/models/nemotron-nano-30B-A3B.sh: Megatron MODEL_ARGS for Nemotron Nano-30B-A3B (52 layers, hybrid pattern MEMEM*E..., 128 experts top-k 6, mamba 64 heads / state 128 / 8 groups, squared-ReLU).

Usage notes

• Weight sync to the rollout engine should use --megatron-to-hf-mode bridge (megatron-bridge ships a NemotronH bridge; the raw converters don't cover nemotron_h). SGLang serves nemotron_h natively.
• When initializing from a Megatron SFT checkpoint in bridge mode, pass --finetune --no-load-optim --no-load-rng explicitly (optimizer state saved under a different TP layout cannot be resharded).
• The training environment needs mamba-ssm and causal-conv1d.

Validation

Single-node 8-GPU GRPO on Nemotron-3 Nano-30B-A3B (from an SFT checkpoint; colocate mode, actor TP2/PP1/EP8 + sequence parallel, 2× TP4 SGLang engines, CPU-offloaded Adam, selective moe moe_act recompute): two full iterations of rollout → reward → GRPO update → weight sync completed cleanly. grad_norm 0.34/0.14, train-vs-rollout logprob abs diff ≈ 0.017 (weight sync consistent with the training model).