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: preserve diffusion mixed dtypes during loading (#2232)
7585f13 Wang, Chang committed 2 months ago
fix: remove stale quantizer access in dual-transformer diffusion (#2236)
68f8aa3 Wang, Chang committed 2 months ago
Update troubleshooting tips for quantization issues in README_CN.md (#2244)
7eb007a Sun, Xuehao committed 2 months ago
increase precision of loss logging in quantization tuning for debugging (#2229)
385a48c Xin He committed 2 months ago
refactor uts, remove duplicate uts and build common uts for all devices to increase coverage (#2157)
6410bc2 Heng Guo committed 2 months ago
fix: AutoScheme parallel scoring OOMs (MoE memory, VLM gate) and disk-stream tuning crash (#2220)
7da9bea avtc committed 2 months ago
Fix FP8 quantization config for native fp8 export (#2228)
017adc8 Zhenzhong Xu committed 2 months ago
add enable_torch_compile mark to SGLang and vLLM throughput tests (#2226)
4b7cc7e Xin He committed 2 months ago
add awq torch_compile fix, disable_opt_rtn by default for awq (#2208)
611d440 Weiwei committed 2 months ago
disable timeout in pytest until found a more stable method (#2222)
96ce448 chen, suyue committed 2 months ago
Preserve diffusion component dtypes during quantization (#2205)
d6ead47 Wang, Chang committed 2 months ago
refactor: optimize expert selection and output handling in moe forward (#2209)
b9f3d00 Xin He committed 3 months ago
add workaround for llmc issue and add unit tests for model_free_utils helpers (#2207)
a2c27b0 Xin He committed 3 months ago
fix: layer_config for lm_head takes effect in model-free mode (#2182)
1d05046 Copilot committed 3 months ago
feat: add support for FP8 model loading and ragged-tile dequantization fallback (#2158)
40635d0 Xin He committed 3 months ago
Increase memory limits for the build process in the CI pipeline (#2198)
8ae2b2e chen, suyue committed 3 months ago
[Experimental] Add NVFP4 E5M3 support and related components for model-free quantization (#2123)
0c0511b Xin He committed 3 months ago
adjust lrx2 logic to fix its issue in mixed bits scenarios and add more woq support in llmc (#2164)
7e35bd6 Wenhua Cheng committed 3 months ago