ascend-ai-coding/awesome-ascend-skills - Seite 6

mindspeed-mm-generative

Universal MindSpeed-MM generative model training guide for Huawei Ascend NPU. Covers all backend patterns (Megatron, Megatron+FSDP2, FSDP2-native, Accelerate+DeepSpeed), feature extraction, weight conversion, and training for ALL supported generative models.…

Quellsprache: Englisch

mindspeed-mm-pipeline

MindSpeed-MM skill router and model index for Huawei Ascend NPU. Use when the user is uncertain which MindSpeed-MM skill to use, needs to choose between understanding/generative/omni/audio model categories, or wants an overview of the full training pipeline.…

Quellsprache: Englisch

mindspeed-mm-vlm

Universal VLM (vision-language understanding model) training guide for Huawei Ascend NPU using MindSpeed-MM. Covers all three framework patterns (Megatron, FSDP2, Custom trainers), weight conversion, dataset preparation (MLLM JSON format), fine-tuning,…

Quellsprache: Englisch

mindspeed-mm-weight-prep

MindSpeed-MM weight conversion guide using mm-convert CLI tool. Covers HuggingFace to MindSpeed-MM format conversion, reverse conversion, and PP weight resplitting. Supports Qwen2VLConverter, Qwen2_5_VLConverter, InternVLConverter, WanConverter and more. Use…

Quellsprache: Englisch

torch-npu-comm-test

通过 PyTorch torch.distributed 接口测试昇腾 NPU 通信算子性能。支持指定任意 tensor shape、dtype，使用 torchrun 启动，贴近真实训练场景的通信算子测试与性能分析。Use for testing collective communication operators (AllReduce, AllGather, ReduceScatter, etc.) with specific tensor shapes via torch.distributed on…

Quellsprache: Chinesisch

verl-quickstart

Generates an executable, end-to-end VERL reinforcement learning quickstart runbook for Ascend/NPU (docker image, dataset preprocessing, model setup, main_ppo training, and examples/run_*.sh flow). Use when users mention verl, quickstart, PPO/GRPO, gsm8k,…

Quellsprache: Chinesisch

create-pr

Creates GitHub pull requests with properly formatted titles that pass the check-pr-title CI validation. Use when creating PRs, submitting changes for review, or when the user says /pr or asks to create a pull request.

Quellsprache: Englisch