Support local and unfused transformer paths and LoRA

[hazelduan]

Summary

• derive Transformer Engine selection from config.transformer_impl for decoder, MTP, and Gemma 4 layer specs
• replace MindSpeed TE flash core with MCore DotProductAttention for the explicit NPU unfused backend while retaining TE Linear/Norm modules
• map input and pre-MLP norm weights to the actual fused TE or separate local-spec module layout
• dispatch LoRA to MCore ColumnParallelLinear and RowParallelLinear, using local parallel adapter layers while retaining the TE implementation
• add regression tests for spec selection, norm mapping, and local LoRA dispatch

Root cause

Layer-spec construction hard-coded use_transformer_engine=True, so even transformer_impl=local produced TEDotProductAttention.

On NPU, the MindSpeed Transformer Engine spec also constructs TEDotProductAttention -> FlashAttention for the explicit unfused backend. Disabling the ms-swift flash flag alone was insufficient: the first step still invoked aclnnFlashAttentionScoreV4. The NPU unfused path now replaces only the core attention module with MCore DotProductAttention, preserving TE Linear/Norm behavior and leaving the GPU spec unchanged.

Once true local specs were selected, weight conversion still assumed TE-fused norm parameters (linear_qkv.layer_norm_weight and linear_fc1.layer_norm_weight) although local specs expose input_layernorm.weight and pre_mlp_layernorm.weight. LoRA dispatch likewise only recognized TE linear classes. These assumptions prevented model loading and adapter injection before training.

Validation

• pre-commit hooks and Python byte-compilation passed for all changed files
• A3 regression tests: 6 passed (spec selection, NPU unfused core selection, local/TE norm mappings, and local linear LoRA dispatch)
• A3 validation used SSH user dxq, MindSpeed, Megatron Core 0.16.0, BF16, Qwen3-0.6B, a fixed 500-row dataset, padding_free=false, and identical seed/configuration for local and flash runs
• actual local model structure used MCore DotProductAttention, separate local RMSNorms, and MCore parallel linears
• every successful run completed 100 optimizer steps, kept loss/grad metrics finite, and saved checkpoint-100

Path	Backend	Steps	Mean loss	Last-10 mean loss	Final cumulative s/it	End-to-end wall time
SFT + LoRA	local	100	1.31381	1.04523	0.10189	45 s
SFT + LoRA	flash/TE	100	1.30809	1.04592	0.07201	42 s
SFT + LoRA	unfused	100	1.30841	1.04521	0.10207	54 s
GKD + LoRA, independent frozen teacher	local	100	0.16752	0.15426	0.10990	45 s
GKD + LoRA, independent frozen teacher	flash/TE	100	0.16662	0.15193	0.08711	43 s
GKD + LoRA, independent frozen teacher	unfused	100	0.16602	0.15295	0.10008	45 s

All six valid runs completed 100 optimizer steps. SFT unfused/flash per-step loss correlation was 0.99992 with mean absolute delta 0.00546. GKD unfused/flash total-loss correlation was 0.99912 with mean absolute delta 0.00197; JSD-loss correlation was 0.99419. The GKD runs executed separate student and frozen-teacher forwards on every step and logged total loss, JSD loss, SFT loss, and grad norm.

The A3 image required test-only import shims for unused optional acl/FLA paths; these are not repository changes. A separate Qwen3-30B-A3B teacher attempt reached local MoE construction but exposed an existing unsupported MindSpeedGmmExperts grouped-weight loading layout. The controlled dense student/teacher comparison above keeps the test focused on the spec, weight, and LoRA changes in this PR.

Paired PR: modelscope/ms-swift#9864