llm-quant-artifacts

• What: fake-quantized LLM checkpoints (weights 4-bit, group 128) used in a quantization study, packed losslessly.
• Use: transfer between our own machines; restore with the companion tool, which decodes .lqz back to the original model.pt and checks every tensor's sha256 against manifest.json.
• Layout
  • <model>/g_<method>/model.lqz: one packed checkpoint per quantization method
  • <model>/g_qoq/smooth.pt: per-channel smoothing scales used by that method at runtime
  • <model>/bases/, <model>/pca/: per-layer key-side bases and weight-side PCA bases used at runtime
  • <model>/qoq/: per-layer key-side smoothing bases of a SmoothAttention-style baseline (scales tied over rotary pairs)
  • isca1106_kv_materials.json: sha256 of the key-side bases of the edge models and the KV-store settings they are used with
  • <model>/manifest.json: file list, sha256 of every file and tensor, restore paths
  • <model>/ppl/: tokenized WikiText-2 / C4 evaluation windows (int32 token ids) with their manifest.json
  • <model>/moments/: calibration second-moment statistics the key-side bases were fitted from
  • <model>/mask_locality/cache/: inputs and intermediate products of a pruning-mask locality study (channel permutations from calibration, per-channel keep-frequency profiles); sha256 of every file is recorded in the study's ARTIFACTS.json
• <model>/mask_locality/cache/bases/beta1104/, <model>/mask_locality/cache/graded/: key-side bases with a fixed exponent of 1/2 and the per-channel key read-error statistics fitted in them (metadata paths are relative; each statistics file stores the sha256 of its basis file)
• Format (.lqz): a 4-bit group of 128 values holds at most 16 distinct fp16 values, stored as a 16-entry fp16 table plus 128 4-bit indices; all other tensors are stored raw. Decoding is exact (bit-identical tensors).
• Not an inference-ready release: research artifacts only, no quality guarantee.

Models and licenses

Each folder carries the license of the model it is derived from.

Folder	Derived from	License
q3_1p7b/	Qwen/Qwen3-1.7B	Apache-2.0
q3_4b/	Qwen/Qwen3-4B	Apache-2.0
q3_8b/	Qwen/Qwen3-8B	Apache-2.0
l31i/	meta-llama/Llama-3.1-8B-Instruct	Llama 3.1 Community License (l31i/LICENSE, l31i/USE_POLICY.md)
l31/	meta-llama/Llama-3.1-8B	Llama 3.1 Community License (l31/LICENSE, l31/USE_POLICY.md)
l32i/	meta-llama/Llama-3.2-3B-Instruct	Llama 3.2 Community License (l32i/LICENSE.txt, l32i/USE_POLICY.md)
l2/	meta-llama/Llama-2-7b-hf	Llama 2 Community License (l2/LICENSE.txt, l2/USE_POLICY.md)

Built with Llama. Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved. Use of the l31i/, l31/, l32i/, and l2/ artifacts is subject to those licenses and to Meta's Acceptable Use Policies included alongside.