Health Test Model Attempts Failures | Generated by AI
Here’s what happened across all three attempts:
Health Test Summary
Attempt 1: unsloth/Qwen3-8B (4-bit, 7.5GB)
- Status: KILLED — download too slow
- Issue: HuggingFace download at ~300KB/s, would take ~7 hours for 7.5GB
- hf-mirror.com slightly faster (~470KB/s) but still ~4.5 hours
Attempt 2: Qwen/Qwen3.5-9B (cached, 18GB)
- Status: FAILED — wrong model type
- Issue: Qwen3.5-9B is a vision-language model (Qwen3VLProcessor), not a text-only model. Unsloth loaded it but
tokenizer.vocab_sizedoesn’t exist on the VL processor.
Attempt 3: Qwen/Qwen3-4B-FP8 (cached, 4.9GB)
- Status: KILLED — hung during load
- Issue: FP8 quantized models aren’t compatible with Unsloth’s 4-bit BNB loading pipeline. Process used 31% CPU but only 309MB GPU — stuck in weight conversion.
What Actually Works (confirmed)
- All Python imports: ✅ torch 2.10+cu128, unsloth, datasets, trl, peft, frontmatter, tiktoken
- CUDA: ✅ RTX 4070 detected, 11.5GB available
- Dataset: ✅ 21,234 train + 200 eval loaded correctly
- Chat template: ✅ works
Root Problem
No compatible model is cached locally. The correct model (unsloth/Qwen3-4B-unsloth-bnb-4bit or unsloth/Qwen3-8B-unsloth-bnb-4bit) needs to be downloaded, but HF is slow right now.
Options
- Download overnight:
huggingface-cli download unsloth/Qwen3-4B-unsloth-bnb-4bit— ~4GB, ~3-4hrs at current speed - Use Qwen3-0.6B (tiny, fast download, good for pipeline validation)
- Use a model that’s already cached — but none of the cached ones (Qwen3.5-9B-VL, Qwen3-4B-FP8, Mistral-Small-3.2) are compatible with the Unsloth SFT pipeline as-is
- Wait for better network conditions
Want me to start the Qwen3-4B download in the background so it’s ready by tomorrow? Or try the 0.6B model for a quick pipeline validation first?