YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Week 08 Compression 20 — Submission 01
AutoRound W3A16/G128 compression for the behavioral-concise Qwen3.5-4B checkpoint. The recipe uses symmetric weights, four-block joint optimization, Qwen3.5-correct normalization/bias tuning, verified non-HMMT RAC calibration, and separate tied-embedding packing.
The existing uploaded artifact satisfies both official evaluator accounting
and strict accounting including sidecars below 0.20 of the 8.0585 GiB baseline.
This audit covers evaluation of that artifact, not regeneration with compress.py.
The uploaded checkpoint's official fraction is 0.1629705856; including the
packed embedding and fragile attention sidecars it is approximately 0.19185.
The standalone decompressor restores both sidecars and explicitly decodes the
48 narrow W3 attention projections, avoiding GPTQModel's NaN-producing export
path for 32-output projections. All reconstructed tensors must be finite.
The repaired checkpoint passed a three-problem official-evaluator/vLLM smoke:
3/3 correct, 100% parse rate, no completions at the 4096-token limit.
This small smoke is a compatibility check, not an HMMT accuracy estimate.
Install with uv sync --python 3.12. No external mathcomp package is required.
Run CPU packing regressions with
uv run python -m unittest decompression.test_restoration.
uv run python compress.py --model_name Qwen/Qwen3.5-4B \
--checkpoint_path <behavioral-bf16-checkpoint> \
--calibration_path <verified-rac-traces.jsonl> \
--output_path <compressed-checkpoint>
uv run python decompress.py --model_name Qwen-3.5-4B \
--checkpoint_path <compressed-checkpoint> \
--output_path <decompressed-checkpoint>
- Downloads last month
- 36