PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Commits on main
fix: match layer names ending with a digit in ignore_layers (#2435)
5aef837 Alexandre Kohler committed 1 month ago
fix: forward pipeline call kwargs through diffusion entry (#2448)
63895eb Wang, Chang committed 1 month ago
fix: keep make_q3_quants sums in sync with accepted levels (#2439)
7059a8b Raashish Aggarwal committed 1 month ago
fix: keep 0 in the range of GGUF asymmetric K-quant groups (#2438)
5a113d1 Raashish Aggarwal committed 1 month ago
fix: give attn_v of LLM_TYPE_70B models Q5_K like llama.cpp (#2443)
ae21ef9 Raashish Aggarwal committed 1 month ago
chore: enable ruff default rules and fix violations (#2434)
c915fdc Sun, Xuehao committed 1 month ago
fix: apply debug layer limits to nested configs (#2431)
6afaecd Shubham Padkonde committed 1 month ago
docs: explain Windows Triton compilation prerequisites (#2432)
8cef3d9 Shubham Padkonde committed 1 month ago
fix: return the repaired string from escape_invalid_json_backslashes (#2419)
ba12c95 Shubham Padkonde committed 1 month ago
fix: honor AR_WORK_SPACE for the AudioCaps dataset cache (#2420)
1ace02c Shubham Padkonde committed 1 month ago
NeUQI grid search for optimized RTN (--enable_neuqi: joint asym scale/zp search, two-stage sym search, frozen-init anchor) (#2341)
44b48d6 avtc committed 1 month ago
fix(xpu): make ARK SDPA graph-capture safe for torch.xpu.XPUGraph (#2416)
a4f01f8 Luo Yu committed 1 month ago
bug fix: prune static kv/attention granularity keys when dtype not set (#2415)
a20f529 Yi Liu committed 1 month ago
enhance attention module with original implementation retrieval and context restoration (#2408)
5eceba6 Xin He committed 1 month ago
bug fix: layer_config mismatch due to transformers rename; fix model free stream mode detection (#2411)
24d1437 Xin He committed 1 month ago
Fix CWE-22 (path traversal) in checkpoint loading issues (#2389)
fe97352 Liang Lv committed 1 month ago
fix: avoid excessive memory during calibration dataset casting (#2405)
22d6e28 Wang, Chang committed 1 month ago
Revert "refactor: decouple datatype quantization from algorithms (#2406)
86e97ad Heng Guo committed 1 month ago