Training pipeline

What this page covers

This page maps the full training workflow across two repositories:

• model preparation (qwen35-toolkit),
• adapter training (qwen-qlora-train),
• export back into GGUF workflow (qwen35-toolkit).

When to use

• You need a high-level map before starting end-to-end training.
• You want to understand where adapter inference and merge fit.
• You need to align training outputs with export/publishing steps.

Input -> Output

Input	Output
Qwen3.5 source VLM checkpoint	LoRA adapter (training artifact)
LoRA adapter + base model	merged fp16/bf16 model
merged fp16/bf16 model	GGUF quant files for inference/distribution

Diagram

Steps

1. Prepare source model into a text-only training-ready artifact.
2. Train LoRA adapter and validate with qlora-infer.
3. Optionally merge adapter into standalone fp16.
4. Export merged model into GGUF and quantize.
5. Upload validated artifacts to Hub.

Merge decision point

Use adapter-only path (no merge) when:

• you only need to evaluate or iterate quickly (qlora-infer).

Use merge path when:

• you need a standalone model directory,
• you need GGUF export,
• you need publishable merged weights.

Cross-repo ownership

qwen35-toolkit:
  convert / strip / verify / upload / GGUF conversion + quantization

qwen-qlora-train:
  train / adapter inference / CPU merge

After qlora-merge, GGUF conversion and upload are handled by qwen35-toolkit.

Phase gates

Gate 1 — Prep gate:
  - Training input checkpoint is text-only and verified.

Gate 2 — Train gate:
  - Adapter artifact is produced.
  - Basic `qlora-infer` checks pass.

Gate 3 — Merge gate (optional):
  - Standalone merged fp16/bf16 directory is created.

Gate 4 — Export gate:
  - GGUF f16 exists.
  - Required quant outputs are generated.

Gate 5 — Publish gate:
  - Upload dry-run looks correct.
  - Final push/pull sync completes without unexpected changes.

Phase map

Phase	Primary tools	Result
Model prep	qwen35-convert, qwen35-strip, qwen35-verify	text-only checkpoint
Training	qlora-train, qlora-infer	adapter + validation
Merge/export	qlora-merge, convert_hf_to_gguf.py, llama-quantize	merged fp16 + GGUF quants

Related

• Quickstart
• Inference
• Post-merge workflow