Auto-MLE-skills

Prepare SFT/chat datasets for training: convert between alpaca / sharegpt / openai-messages formats, check data quality, and verify tokenization & packing correctness — the framework-neutral data layer that serves any downstream trainer. Catches the silent failures that don't error but quietly hurt the model: empty/duplicate/malformed samples, answers truncated at max_len, wrong label masking (training on the prompt), missing EOS (model won't stop), double-BOS, and packing without a block-diagonal mask (cross-document contamination). Use when the user mentions: dataset format conversion, alpaca, sharegpt, openai messages / jsonl, data quality, dedup, length distribution, truncation, chat template, tokenize, packing, padding, label mask, loss mask, attention mask, preparing data for SFT/fine-tuning, LLaMA-Factory / axolotl / trl data format. Triggers: "数据格式", "格式转换", "alpaca", "sharegpt", "messages", "数据质量", "去重", "截断", "chat template", "packing", "label mask", "tokenize", "数据清洗", "微调数据", "训练数据准备".

Probe the machine's ML hardware/software environment and tell the agent what it can and cannot run BEFORE attempting any ML setup. Produces a JSON manifest (env_report.json) plus a human-readable capability report with tiered ✅/⚠️/❌ verdicts and concrete install commands. Use at the START of any MLE automation, or whenever the user asks about: CUDA / driver / PyTorch version compatibility, why flash-attention / xformers / apex / bitsandbytes / deepspeed won't install or build, which torch wheel to pick, how to set up an isolated env (uv/conda/venv), package conflicts (transformers↔tokenizers, numpy 2.x), or how to split a model across GPUs (tensor/pipeline/data parallel, FSDP). Covers NVIDIA, Apple MPS, CPU, Google TPU, and Chinese accelerators (昇腾 Ascend, 寒武纪 Cambricon, 海光 Hygon DCU, 沐曦 MetaX). Triggers: "环境探测", "能跑什么", "装不上", "编译失败", "CUDA 版本", "torch 版本", "flash-attn", "环境清单", "环境隔离", "并行策略", "tensor parallel".

Estimate GPU memory (VRAM) BEFORE launching, so a run doesn't OOM hours in. Given a model (HF config.json / model dir, or manual params) and a scenario (inference vs training, batch size, sequence length, dtype, parallelism), it computes a per-GPU breakdown — weights + KV-cache + gradients + optimizer + activations + overhead — and returns a fits ✅ / tight ⚠️ / OOM ❌ verdict with concrete fallbacks (smaller batch, shorter seq, quantize, grad checkpointing, more GPUs, FSDP/ZeRO, offload). Correctly handles GQA, MoE (all experts in VRAM), hybrid attention (KV-cache only for full-attention layers), and reads real on-disk weight size for quantized models. Use when the user asks: will this model fit, how much VRAM does X need, why did it OOM, what batch/seq can I run, how many GPUs do I need, KV cache size, activation memory, optimizer memory. Triggers: "显存", "会不会 OOM", "放得下吗", "显存估算", "VRAM", "OOM", "batch size 能开多大", "需要几张卡", "KV cache", "fit in memory", "memory footprint".