reference: project: "FlatQuant: Flatness Matters for LLM Quantization" role: "Primary implementation cited as reference [35] by LoopQ" repository: "https://github.com/ruikangliu/FlatQuant" commit: "9d88ffcb7d2c6bda59fb5c44dad36adc101aadb1" commit_date: "2025-11-25T20:18:56+08:00" license: MIT inspected_on: "2026-09-06" files: - path: flatquant/trans_utils.py sha256: bbd12d6effa9b46d21875abd85ec8cfb91075bb0ed6151fa663c60411715bb93 relevance: "SVD/direct single and Kronecker-decomposed transform parameterizations" - path: flatquant/flat_utils.py sha256: 892bae4c29cd91e27e07034b75c2ed69a731eded35e833c91fb204a9ff9a0284 relevance: "Kronecker matmul and diagonal-scale folding into normalization" - path: flatquant/flat_linear.py sha256: e5977a76d076ad6fc0b2e737bdcc9e4f0fe62d1682b3df15613c0d948a8c904b relevance: "Activation-transform and inverse-transpose weight-folding contract" - path: flatquant/train_utils.py sha256: 52bd6a24a403a15be8a5d53413485ff49892796b642b4d46463140d5e5cb0104 relevance: "AdamW parameter groups, cosine schedule, and layer-wise reconstruction calibration" - path: flatquant/model_tools/llama_utils.py sha256: 4d5770f9f45bfc14b6d22060db8870989dfbd341b3f41ee90daab49c55665765 relevance: "QKV/O/Up-Gate/Down module placement and reparameterization" verified_contract: kronecker_matrix: "P = kron(P_left, P_right)" activation_side: "X_transformed = X @ P" weight_side: "W_folded = W @ P^{-T}" invariant: "(X @ P) @ (W @ P^{-T})^T = X @ W^T" diagonal_order: "When enabled, activation diagonal scaling precedes Kronecker transformation." module_groups: - "Attention input transform consumed by Q, K, and V projections" - "Attention output transform consumed by O projection" - "MLP input transform consumed by Gate and Up projections" - "MLP activation-output transform consumed by Down projection" not_vendored: >- No FlatQuant source code is copied into LoopQ. This manifest pins the primary reference; the local minimal module independently expresses only the folding invariant required by LoopQ LQ3.