variant-e/quantization_config.txt

=== VPTQ Quantization Configuration ===
Date: 2026-02-22 19:40:24 UTC

Target: ~2-bit quantization with Residual VQ
Model: speakleash/Bielik-11B-v2.3-Instruct (MistralForCausalLM)

Parameters:
  --model_name /workspace/models/bielik-11b-instruct
  --output_dir /workspace/variant-e/output/
  --vector_lens -1 8          # Skip embeddings (-1), quantize in vectors of 8
  --group_num 1               # Single group
  --num_centroids -1 65536    # 2^16 centroids = 16 bits per vector index
  --num_res_centroids -1 256  # 2^8 residual centroids
  --npercent 0                # No outlier channels
  --blocksize 128             # Block size for quantization
  --new_eval                  # Use new evaluation method
  --seq_len 4096              # Sequence length for calibration
  --kmeans_mode hessian       # Hessian-weighted k-means clustering
  --num_gpus 1                # Single GPU
  --enable_perm               # Channel permutation for better quant
  --enable_norm               # Channel normalization
  --save_model                # Save quantized model
  --save_packed_model          # Save packed model for inference
  --hessian_path /workspace/hessians/quip-format/hessians
  --kiter 100                 # K-means iterations
  --ktol 1e-5                 # K-means convergence tolerance

Effective bitwidth calculation:
  Primary: log2(65536) / 8 = 16/8 = 2.0 bits/weight
  Residual: log2(256) / 8 = 8/8 = 1.0 bits/weight overhead
  Total: ~2.0 bits/weight (residual adds codebook storage overhead, not per-weight)

Hessian source: Jakubrd4/bielik-quip-e8p12
  Format: QuIP# (flatH + mu + n + ct per layer)
  Calibration: CulturaX-PL, 512 samples x 4096 tokens
  Layers: 50 x 4 projections (qkv, o, up, down) = 200 files