[addsubmuldiv]

Summary

Fix TP-sharded LoRA A construction for TENPU row-parallel linears. PEFT reads a local in_features from TENPU, so pass the reconstructed global width to the constructor and let TENPU shard it once. Retain the native RowParallelLinear branch, where PEFT reads the global input_size.

- row_input_size = self.in_features if native_row or is_torch_npu_available() else in_features
+ row_input_size = self.in_features if native_row else in_features

Add a TP=1/2 regression test for CUDA and NPU that compares forward output, input gradients, and LoRA A/B gradients against an unsharded matrix reference.

Evidence

• Before: With global input width 2048 and TP=2, TENPU LoRA A was built with weight shape [8, 512] for a local input width of 1024, failing with GEMM not possible.
• After: NPU TP=1 and TP=2 numerical tests pass with rtol=0, atol=0 for forward output and input/LoRA A/B gradients: 2 passed, 2 deselected (pytest tests/test_lora_row_parallel.py -k npu -q -s).
• Qwen3-0.6B TP=2 all-linear LoRA SFT completes 4 optimizer steps on a local 32-record dataset and saves Megatron and safetensors adapter checkpoints. Final loss: 2.34315753; gradient norm: 5.59919882. All 196 zero-initialized LoRA B tensors are nonzero after training.
• Runtime validation used 2 Ascend 910B3 NPUs, CANN 9.2.0, torch 2.10.0, torch_npu 2.10.0.post4, Megatron Core 0.18.3, and TENPU 2.17.1. Pre-commit checks pass.
• CUDA and older MindSpeed stacks were not executed. MoE, sequence parallelism, long training, and checkpoint resume are outside this validation.

[Copilot]

🟢 Approval recommended

No active review issues were identified for the TENPU row-parallel LoRA input-sizing fix.

No open issues 🎉

What changed in this PR

Fixes TENPU row-parallel LoRA A construction by passing the global input width, preventing it from being sharded twice while preserving native-row and CUDA sizing behavior.

Changes:

• Removes the NPU-specific input-size override and clarifies the constructor’s sizing expectations.
• Adds CUDA/NPU regression tests for TP=1 and TP=2, comparing outputs and input/LoRA gradients against an unsharded reference.

File	Description
tests/​test_lora_row_parallel.py	Adds sharded LoRA shape, forward-output, and gradient checks.
src/​mcore_bridge/​tuners/​lora.py	Corrects TENPU LoRA A input sizing while retaining native-row handling.

_{Review effort: Balanced}