25M2111-Week02-Track2-20-Submission01

Week 02 — Track 2 — 20% target

This is the complete evaluation and reproduction package for enrollment 25M2111. It contains a losslessly packed W3/W4/W8 base plus a physically charged rank-8 grouped-JSD residual.

Evaluate the uploaded checkpoint

Run these commands from this Submission01 directory. A Conda environment name is not assumed.

python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[eval]'

hf download safffrron/25M2111-Week02-Track2-20-Submission01 \
  --local-dir hf_checkpoint
sha256sum hf_checkpoint/week02_20_base.ptz

EAIMATH_RESTORE_DEVICE=cuda:0 python dequantize_to_bf16.py \
  Qwen/Qwen3.5-4B hf_checkpoint restored_hf

python inference/smoke.py restored_hf
./evaluation/run_public_eval.sh \
  restored_hf 0 restored_hf/keep_ids.json

The digest must match the table. The dequantizer restores the token-predicted BF16 base, safely merges the submitted adapter, and writes an ordinary Hugging Face checkpoint. The evaluation defaults to max_new_tokens=65536 and repetition penalty 1.20; its artifact evidence used 32,768 tokens.

Required conversion API

code.py provides:

• convert_from_hf_checkpoint(model_name, output_path, sparsity=None)
• convert_to_hf_checkpoint(model_name, checkpoint_path, output_path)

The optional sparsity value exists only for starter-code compatibility and is ignored: this submission does not use pruning. checkpoint_path may be the download directory or its week02_20_base.ptz file. Missing companion files are fetched automatically from the submission's Hugging Face repository.

Reproduce compression and recovery

The optional full rebuild uses the verified trace corpus and a GPU. First run training/reproduce_source.sh, then compression/reproduce_gptq.sh, then training/reproduce_adapter.sh. Point the three EAIMATH_* variables printed by those scripts at the resulting GPTQ source, trace corpus, and JSD adapter, then run:

python -m pip install -e '.[compress,eval]'
python convert_from_hf_checkpoint.py Qwen/Qwen3.5-4B reproduced_week02_20

The method uses group-128 calibrated GPTQ (3-bit MLP, 4-bit Gated DeltaNet, 8-bit full attention), 30,000 selected 8-bit tied vocabulary rows, deterministic token-string prediction for omitted input rows, and a rank-8 residual trained after quantization against grouped Jensen–Shannon teacher distributions. Machine-readable settings and matched results are in configs/ and results/.

Track 2 CUDA rationale

This is byte-for-byte the corresponding Track 1 representation. CUDA_ACCELERATION.md explains how its physical operators map to packed CUDA execution and clearly separates that realizability argument from the submitted BF16-compatible evaluation path.