[NPU] Adapt LoRA and expert handling to Megatron Core 0.18 and TENPU

[addsubmuldiv]

Summary

Adapt the NPU model configuration and LoRA integration to Megatron Core 0.18 with MegatronAdaptor and TransformerEngineNPU (TENPU), removing the dependency on MindSpeed-specific APIs.

Companion integration: modelscope/ms-swift#10087

Changes

• Remove MindSpeed argument parsing, cached defaults, and dynamic configuration augmentation from ModelConfig.
• Use TELinear for NPU LoRA adapters instead of selecting an nn.Linear fallback by MindSpeed version.
• For NPU TELayerNormColumnParallelLinear, obtain the linear output and post-normalization activation from one base-layer forward using return_layernorm_output. Preserve the bias contract and restore the original flag in finally, rather than reconstructing normalization through MindSpeed private methods.
• Correct column-parallel LoRA output sizing at the local-to-global dimension boundary and use _tp_group as the TP-group fallback.
• Identify NPU expert layers through MCore TEGroupedLinear and expert configuration, rather than a MindSpeed module name. Pass expert tensor/data-parallel groups when constructing grouped LoRA sharded checkpoint tensors.

Validation

Tested together with the companion Swift change on Ascend 910B3, Python 3.12, CANN 9.1.0, PyTorch 2.10.0, torch_npu 2.10.0.post2, and Transformers 5.12.1. Backend snapshots: MCore ba7b5ebce12a (0.18.0), MegatronAdaptor 2a164d4548e2, and TENPU 0a389a677586.

• Dense Qwen3-0.6B LoRA training and checkpoint loading with TP2/PP2.
• Four-layer Qwen3.5 GDN full/LoRA SFT with TP2/PP2/CP2 and packing.
• Four-layer Qwen3.5 MoE full/LoRA SFT with TP2/PP2/EP2. The full smoke used no_save_optim; it does not establish MoE optimizer-resume correctness.
• Dense HF export/reload: maximum logits difference 7.82e-5.
• Dense FSDP LoRA save/resume with merge_lora=false: matching logged loss across continuous training and two successive resume stages.

These are bounded 2–4-step integration checks, not long-run or full-size-model qualification. GPU/FP8 paths were not tested.

Scope and known limitation

USE_MCORE_GDN and LoRA merge/export behavior are unchanged. BF16 in-place merge followed by subtraction-based unmerge can change training weights through rounding; exact restoration/non-mutating export is a separate follow-up, not fixed here. Use merge_lora=false when validating uninterrupted-versus-resumed LoRA training equivalence.