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: update get_predefined_ignore_layers to include shared_expert_gat… (#2397)
6db9435 Xin He committed 1 month ago
fix: update tensor summary creation to directly use tensors_to_restor… (#2401)
bff3dea Xin He committed 1 month ago
ark: add BF16/FP16 Sparge variants and reduce compilation RAM (#2145)
402b0ed Yi Liu committed 1 month ago
Honour the padding mask in SignRoundV2's plain loss (#2395)
0883217 Shubham Padkonde committed 2 months ago
Move the zero point to CPU for asymmetric layers too (#2394)
89c3dfe Shubham Padkonde committed 2 months ago
Apply autocast only when mixed precision is enabled (#2393)
3f6c819 Shubham Padkonde committed 2 months ago
Cast float32 weights to the export dtype when saving (#2385)
0aa5e35 Shubham Padkonde committed 2 months ago
fix model_free saving with single shard and MXFP fake format evaluation (#2391)
f4d2c16 Xin He committed 2 months ago
fix: update import to include get_model_path for model retrieval (#2388)
c29847e Xin He committed 2 months ago
fix:reduce memory usage when loading native fused MoE checkpoints (#2382)
a5f6ddd Heng Guo committed 2 months ago
fix: restore FP32 tensors saved as FP16/BF16 during tensor copying (#2375)
a6c612f Xin He committed 2 months ago
fix: share NVFP4 global scales for Wan QKV projections (#2379)
36079c4 Wang, Chang committed 2 months ago
[ARK] Improve XPU W4A16 WOQ decode with dense S4 DPAS path (#2331)
eaf57a2 Zhenzhong Xu committed 2 months ago
bug fix: fix NVFP4 evaluation with input_ global_scale handling and testing (#2374)
27daf7f Xin He committed 2 months ago
fix: recognize Modular Diffusers pipelines as diffusion models (#2120)
0d76712 Nikita Davidchuk committed 2 months ago
refactor: decouple datatype quantization from algorithms (#2360)
2e82448 Heng Guo committed 2 months ago
fix: save the config of diffusers models on export (#2119)
1983b65 Nikita Davidchuk committed 2 months ago
feat(ark): fused Hadamard + MXFP4 activation quantization kernel
ecc5b00 jijiaz committed 2 months ago
perf: avoid unused SVD reconstruction and select the CUDA driver in SVDQuant (#2356)
903ba0c Wang, Chang committed 2 months ago
unify target_bits/options into bits/schemes for AutoScheme (#2333)
2b5a3c5 Heng Guo committed 2 months ago
Refactor test timeouts and update dependencies for performance (#2361)
1965bdf Sun, Xuehao committed 2 months ago