Nov 19 Today 01:50 Card: cite the paper, arXiv:2611.23388 Nov 18 Yesterday 15:37 QDA-Qwen3.6-35B-A3B-Q4-GGUF recipe: back to the original (the published file holds the MTP head since its 11-18 upload) 15:24 Recipe headers: the measured build time, or that it was not measured, in place of up to tens of GPU hours 15:00 QDA-Qwen3.6-35B-A3B-Q4-GGUF recipe: drop the MTP-head tensor rule that mtp: false leaves without a tensor 14:32 QDA-Qwen3.6-35B-A3B-Q4-GGUF recipe: mtp: false, as the published file (no MTP head) 09:35 GGUF recipes: install line with the gguf extra; SHA256SUMS updated 01:14 Recipes: Qwen3.6-35B-A3B Q4 GGUF keeps the MTP head; drop the unused top-level seed, rename contamination to contamination_index 01:07 Recipes: Qwen3.8-27B Q4/Q6 GGUF drop the unused top-level seed, rename contamination to contamination_index; Q6 adds its KD traces tokenizer fingerprint Nov 17 2 days ago 21:23 W4A4 recipes: remove the top-level seed, rename contamination to contamination_index; folder READMEs with the calibration data and evaluation setup 21:06 Recipes: remove the unused top-level seed and rename the calibration contamination index key to contamination_index 10:18 README: add the five NVFP4 W4A4 recipe folders 10:18 Add the five NVFP4 W4A4 recipes 00:30 Recipe headers: state the GPUs a recipe was built on and that reproducing it may take up to tens of GPU hours; no wall-clock times Nov 16 3 days ago 06:20 Add QDA-Qwen3.8-27B-Q4-GGUF recipe and inputs Nov 15 4 days ago 22:41 README: QDA 0.2.0 version line 22:40 QDA-DeepSeek-R1-Distill-Llama-8B-W3A16: recipe and inputs of the new checkpoint 22:16 Add QDA-Qwen3.8-27B-W4A16-dsh recipe and calibration 19:54 README: add the QDA-Qwen3.8-27B-Q6-GGUF and QDA-Qwen3.6-35B-A3B-Q4-GGUF folders 19:42 Recipe and inputs for qda-lab/QDA-Qwen3.6-35B-A3B-Q4-GGUF 19:42 Recipe and inputs for qda-lab/QDA-Qwen3.8-27B-Q6-GGUF Nov 13 6 days ago 06:30 Recipe headers and README: give the output directory (--set out=...) and run from inside the folder; recipe bodies and data unchanged Nov 12 7 days ago 19:34 QDA-Qwen3.6-35B-A3B-W4A16: add repack_sglang.py (torch + safetensors) that makes the -SGLang repository from this checkpoint; output checked equal tensor for tensor 18:52 QDA-Qwen3.6-35B-A3B-W3A16: part -i uses sdpa attention so it fits 80 GB cards (same maths; not byte-identical to the original build) 16:00 35B-W4A16 recipe header and README: describe the 80 GB-card change without naming the original hardware 07:33 Add the four recipes of QDA-DeepSeek-R1-Distill-Llama-8B-W4A8 00:48 QDA-Qwen3.8-27B-W3A16: credit ISTA-DASLab for the 3-bit start (Apache-2.0 licence + README with our changes), derive_start.py rebuilds it from ISTA's release Nov 11 8 days ago 23:09 Add recipes and inputs: Qwen3.6-35B-A3B W4A16 and W3A16, Qwen3.5-9B and Qwen3.8-27B W3A16 (with their 3-bit starts) 22:39 QDA-Qwen3.6-35B-A3B-W3A16: add the 3-bit GSQ start this checkpoint was distilled from 22:39 QDA-Qwen3.8-27B-W3A16: add the 3-bit GSQ start this checkpoint was distilled from 22:39 QDA-Qwen3.5-9B-W3A16: add the 3-bit GSQ start this checkpoint was distilled from 22:31 Add each checkpoint's recipe to its folder; add Qwen3.6-35B-A3B-W4A8 22:20 Gemma-4-31B: add the contamination index the recipe reads (source paths reduced to file names) 20:42 QDA recipe inputs: calibration and KD files of the released checkpoints 20:42 initial commit