Support Kimi no-TE alignment path

[zrr1999]

Summary

• add SWIFT_MEGATRON_NO_TE gated local/no-TE Megatron-Core layer specs for Kimi alignment
• add MLA DotProductAttention compatibility for local no-TE specs
• remap TE-saved fused norm checkpoint keys when loading no-TE specs
• dequantize Kimi FP8 weight_scale_inv tensors into regular BF16 params under SWIFT_MEGATRON_NO_TE

Validation

• python -m py_compile swift/megatron/model/gpt_bridge.py swift/megatron/model/register.py swift/megatron/utils/megatron_lm_utils.py
• sampled full Kimi-K2-Instruct FP8 weight_scale_inv branch smoke: exact BF16 dequant match
• reduced 2-layer Kimi SFT no-TE forward loss comparisons reached 20-step and 100-step BITWISE_PASS in the repro workspace

Scope

Default TransformerEngine behavior is unchanged. New behavior is gated by SWIFT_MEGATRON_NO_TE for Kimi alignment/repro workflows.

[zrr1999]

Update: pushed final no-TE EP2 alignment patch in 21d9bb68149565bf6a1f50ce45f35757c7e77e02.

What changed:

• In explicit no-TE alignment mode (SWIFT_MEGATRON_NO_TE=1), patch Megatron-Core SequentialMLP to handle exact duplicate per-expert hidden/prob chunks deterministically.
• If a local expert receives exact duplicated halves, compute one half and duplicate the output. This avoids BF16 GEMM batch-size drift observed when EP dispatch changes local expert batch from n rows to exact duplicated 2n rows.
• Patch is gated to the no-TE alignment path and checks exact hidden/prob duplication; it is not sample/coordinate/loss-specific.

Validation evidence (Kimi-K2 2-layer BF16 SFT, native ms-swift/Megatron-Core, no-TE):

• EP2 PR patch vs EP1 reference: BITWISE_PASS, 100 steps, first_bad_step=null
  • reports/TORCH_NO_TE_EP2_MS_SWIFT_PR_PATCH_VS_EP1_100STEP_COMPARE.md
• EP2 PR patch vs Paddle native reference: BITWISE_PASS, 100 steps, first_bad_step=null
  • reports/NO_TE_EP2_MS_SWIFT_PR_PATCH_VS_PADDLE_100STEP_COMPARE.md
• EP1 preservation with this patch vs EP1 reference: BITWISE_PASS, 100 steps
  • reports/TORCH_NO_TE_EP1_MS_SWIFT_PR_PATCH_VS_EP1_REFERENCE_100STEP_COMPARE.md
• EP1 preservation with this patch vs Paddle native reference: BITWISE_PASS, 100 steps
  • reports/NO_TE_EP1_MS_SWIFT_PR_PATCH_VS_PADDLE_100STEP_COMPARE.md

Mechanism evidence:

• Manual native expert recompute showed the same EP1 expert module/weights produce saved EP1 output for n rows and saved EP2 output for exact duplicated 2n rows; max diff is one BF16 quantum (0.0078125).
• reports/TORCH_NO_TE_EXPERT_BATCHSIZE_INVARIANCE_STEP2ONLY.md

Claim boundary: validated for the current 2-layer BF16 no-TE EP=2 alignment scope; not a full 61-layer/FP8 claim.