license: llama3
tags:
- arxiv:2607.07964
- kronq
- quantization
- hessian
- fisher
task_categories:
- text-generation
Meta-Llama-3-8B — KronQ H_G (output-side gradient covariance)
Paper: arXiv:2607.07964 · Code: GitHub
Pre-computed H_G for Meta-Llama-3-8B, the output-side curvature factor used by KronQ under the K-FAC factorization H ≈ H_X ⊗ H_G. H_G is the per-sublayer empirical-Fisher gradient covariance (E[g gᵀ] over the layer output), distinct from the standard input-side Hessian H_X (built online during calibration).
Publishing this lets you reproduce KronQ quantization without the offline Fisher precompute step.
Contents (32 layers × 7 sublayers, ~56 GB)
layer_<i>/self_attn_{q,k,v,o}_proj_G.pt
layer_<i>/mlp_{gate,up,down}_proj_G.pt
metadata.pt
Each *_G.pt is the out_features × out_features gradient covariance for that sublayer.
Usage
Point KronQ's --grad_dir at the downloaded folder — it skips precompute_gradients.py:
python main.py --model meta-llama/Meta-Llama-3-8B \
--w_bits 4 --w_groupsize -1 --w_clip --w_asym --a_bits 16 --act_order \
--bi_calibration --use_gptaq --incoh_rotate --incoh_kernel had --incoh_mode full \
--alpha 0.25 --grad_dir <downloaded_HG_dir>
This is raw (unrotated) H_G, the form used by the weight-only recipe (H_G cancels in the OBS update, so raw and rotated give identical weights for per-channel weight-only).
License
Derived from Meta-Llama-3-8B — subject to the Llama 3 Community License.