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
Record AutoRound hyper-parameters in llm-compressor quantization_config (#2528)
4ae46af Cross2pro committed 16 days ago
support AWQ sample control and optimize AWQ+ARV2 for INT WOQ (#2484)
98ad5bb Weiwei committed 20 days ago
[Bug fix] Release offload temp-file mmap to fix cleanup issue (#2487)
e1ab33b Liang Lv committed 20 days ago
[Bug fix]: test/test_mlx/test_mlx_format.py failed on Mac OS (#2554)
9ffe65f Xin He committed 20 days ago
Fix native AutoRound INT4 packing with FP8 layer overrides (#2543)
3e18644 Wang, Chang committed 21 days ago
Prefetch next batch to CUDA for SignRound v1/v2 under low_gpu_mem_usage (#2549)
3f86dea Wenhua Cheng committed 21 days ago
Update stale issue and PR handling to 60 days of inactivity (#2553)
e4353de chen, suyue committed 21 days ago
feat: Add NVFP4_E5M3_GS scheme with global scale support (#2531)
91a9bbf Xin He committed 22 days ago
refactor: Simplify vllm evaluation by setting multiprocessing method directly (#2534)
e2e3f17 Xin He committed 22 days ago
Skip zero-point round-trip with compressed-tensors 0.20.0 (#2544)
76a114e Wang, Chang committed 22 days ago
fix: update scale handling in QuantizedKVParameterCache to use amax of all batches (#2533)
512872d Xin He committed 23 days ago
refactor: extract extensible shared smoothing engine (#2452)
9cc08ca Wang, Chang committed 24 days ago
feat: export SVDQuant NVFP4 checkpoints for vLLM-Omni (#2445)
7c0aed0 Wang, Chang committed 24 days ago
feat: enhance FP32 tensor handling in mixed-precision checkpoints and… (#2519)
1d623e2 Xin He committed 24 days ago
feat: W4A8 ARK XPU MoE kernel (int4 weight / int8 compute) with prefill + decode (#2143)
80b5f87 Copilot committed 27 days ago
enhance: add source config preservation in save_model and related tests (#2522)
710dcae Xin He committed 27 days ago
feat: multi-GPU quantization improvements -- batched device-local sea… (#2357)
a7656a2 avtc committed 28 days ago
enhance model_free functionality with strict parameter validation (#2520)
95eccea Xin He committed 28 days ago