← All skills
mindspeed-mm-vlm
82 / 100 · Low risk
Fine to try.
By ascend-ai-coding/awesome-ascend-skills (GitHub) · other · 184 stars · scanned at commit 75ec89c
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, inference, and evaluation. Supports Qwen2.5VL, Qwen2VL, Qwen3VL, InternVL2.5/3/3.5, GLM4.1V, GLM4.5V, DeepSeekVL2, DeepSeekOCR, Ming, and more. Use when training or fine-tuning any multimodal understanding model on Ascend NPU.
Works with any agent that supports SKILL.md skills.
What this skill can do
Runs commands on your computer
This is disclosed by the author in the skill's SKILL.md or README.
Found in 21 places.
Show 17 more
- skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 68
pip install -e . - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 74
source /usr/local/Ascend/cann/set_env.sh - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 81
python3 -c " - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 133
torchrun $DISTRIBUTED_ARGS pretrain_vlm.py \ - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 144
CUDA_DEVICE_MAX_CONNECTIONS=2 torchrun $DISTRIBUTED_ARGS pretrain_transformers.py \ - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 151
torchrun $DISTRIBUTED_ARGS examples/<model>/finetune_<model>.py <args> - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 177
http_proxy=http://127.0.0.1:58232 https_proxy=http://127.0.0.1:58232 \ - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 187
mm-convert Qwen2_5_VLConverter hf_to_mm \ - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 232
mkdir -p data/COCO2017/train2017 - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 244
cd /root/workspace/MindSpeed-MM # Must run from repo root - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 323
LOAD_PATH="ckpt/mm_path/Qwen2.5-VL-3B-Instruct" # Converted MM weights - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 354
torchrun $DISTRIBUTED_ARGS pretrain_vlm.py $GPT_ARGS $MM_ARGS $OUTPUT_ARGS - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 363
MASTER_ADDR=localhost - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 382
bash examples/qwen2.5vl/finetune_qwen2_5_vl_3b.sh - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 396
bash examples/qwen2.5vl/inference_qwen2_5_vl_7b.sh - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 405
bash examples/qwen2.5vl/evaluate_qwen2_5_vl_7b.sh - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 424
bash examples/qwen2vl/finetune_qwen2vl_72b_dpo.sh
and 1 more places
Downloads and installs other software
The author does not mention this.
Found in 21 places.
Show 17 more
- skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 50
pip install numpy pyyaml scipy attrs decorator psutil - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 59
pip install -r requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 60
pip install -e . - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 68
pip install -e . - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 165
| Qwen3VL | `pip install -r examples/qwen3vl/requirements.txt` | Requires latest transformers; may conflict with other models | - skills/training/mindspeed-mm/mindspeed-mm-vlm/SKILL.md, line 166
| DeepSeekOCR | `pip install -r examples/deepseekocr/requirements.txt` | Downgrades transformers to 4.46.3; **must use isolated environment** | - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/model-delta-cards.md, line 17
| Extra deps | `pip install -e git+...transformers.git@c0dbe09` (git-installed, not pip release) | - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/model-delta-cards.md, line 30
bash scripts/install.sh --megatron --msid 96bc0a3bf3 && pip install -r examples/qwen3vl/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/model-delta-cards.md, line 123
pip install deepseekvl2 # from upstream repo - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/model-delta-cards.md, line 156
pip install -r examples/deepseekocr/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/model-delta-cards.md, line 192
pip install -r MindSpeed-MM/examples/ming/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/per-model-deps.md, line 43
pip install -r examples/qwen3vl/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/per-model-deps.md, line 70
pip install -r examples/glm4v/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/per-model-deps.md, line 91
pip install -r examples/deepseekocr/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/per-model-deps.md, line 112
pip install -r examples/qwen3vl/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/per-model-deps.md, line 116
pip install -r examples/deepseekocr/requirements.txt - skills/training/mindspeed-mm/mindspeed-mm-vlm/references/per-model-deps.md, line 130
pip install -r examples/qwen3vl/requirements.txt
and 1 more places
Install
Paste this into your terminal. It downloads exactly the version we scanned, not whatever is newest.
if [ -e ~/.agents/skills/'mindspeed-mm-vlm' ] || [ -L ~/.agents/skills/'mindspeed-mm-vlm' ]; then echo 'A skill folder named mindspeed-mm-vlm already exists — not replacing it' >&2; false; else d=$(mktemp -d) && git clone --quiet --filter=blob:none --no-checkout 'https://github.com/ascend-ai-coding/awesome-ascend-skills' "$d" && git -C "$d" checkout --quiet '75ec89ce9d401be8ac95980e1c3f42f8d4315057' -- 'skills/training/mindspeed-mm/mindspeed-mm-vlm' && mkdir -p ~/.agents/skills && cp -R "$d"/'skills/training/mindspeed-mm/mindspeed-mm-vlm' ~/.agents/skills/'mindspeed-mm-vlm' && rm -rf "$d"; fi
Codex reads ~/.agents/skills directly. For Claude Code, Cursor, Copilot, Gemini CLI, Antigravity or OpenCode, also link it. For Claude Code: ln -s ~/.agents/skills/'mindspeed-mm-vlm' ~/.claude/skills/'mindspeed-mm-vlm' (other agents: ~/.cursor/skills, ~/.copilot/skills, ~/.gemini/skills, ~/.gemini/config/skills, ~/.config/opencode/skills). Or install through the finder skill, which does both.
Badge for your README
The badge for ascend-ai-coding/awesome-ascend-skills shows the score of its lowest-scoring skill.

[](https://zalavijaysinh.github.io/skilltrust/repos/ascend-ai-coding/awesome-ascend-skills/)
Score history
| Date | Version | Score |
|---|
| 2026-11-08 | 75ec89c | 100 |