feat: Add npu megatron support

[UsernameFull]

Summary

This PR adds support for Huawei Ascend NPU devices with Megatron-Core backend, enabling ROLL framework to run reinforcement learning training on NPU hardware.

Key Changes

1. Platform Detection Priority

File: roll/platforms/__init__.py

Changes: Reordered platform detection to check NPU before CUDA.

Reason: NPU devices were incorrectly falling back to CUDA platform. Prioritizing NPU detection ensures NpuPlatform is properly initialized when torch_npu is available.

2. Device-Agnostic Operations

File: roll/pipeline/base_worker.py

Changes:

• Replaced "cuda" with current_platform.device_type
• Replaced torch.cuda.memory_allocated() with current_platform.memory_allocated()

3. MindSpeed Integration

File: mcore_adapter/src/mcore_adapter/training_args.py

Changes: Added optional import of mindspeed.megatron_adaptor .

Reason: MindSpeed is Huawei's library providing NPU-specific Megatron optimizations. The adaptor patches Megatron-Core for NPU compatibility while maintaining GPU compatibility via try-except.

4. NPU Attention Mask Format

File: roll/distributed/strategy/megatron_strategy.py

Changes: Added NPU-specific attention mask transformation to 4D format.

Reason: NPU requires 4D attention masks [B, 1, S, S] instead of standard 2D [B, S] . This hardware-specific requirement ensures correct attention computation on NPU.

if hasattr(torch, "npu") and torch.npu.is_available() and attention_mask is 
not None:
    attention_mask = attention_mask.bool()
    attention_mask = attention_mask[:, None, None, :].expand(B, 1, S, S)

5. Optimizer Compatibility

File: roll/third_party/megatron/optimizer.py

Changes: Added support for no_weight_decay_cond , scale_lr_cond , lr_mult parameters.

6. Example Configurations

Files:

- examples/ascend_examples/qwen3_4B_dpo_megatron.yaml
- examples/ascend_examples/qwen3_8b_rlvr_deepspeed.yaml
- examples/ascend_examples/run_dpo_pipeline.sh

Reason: Provides ready-to-use NPU training examples demonstrating proper device mapping and strategy configuration for both DPO and RLVR pipelines.

Impact

Benefits:

• Enables megatron on Huawei Ascend NPU hardware
• Maintains full backward compatibility with GPU systems
• Follows existing platform abstraction patterns

Requirements

• Huawei Ascend NPU with torch_npu installed
• MindSpeed(v0.15.3) library for NPU Megatron support

[HuangJoJo]

plz attach qwen3-8B rlvr megatron training curves

[UsernameFull]

plz attach qwen3-8B rlvr megatron training curves