8 Pull requests merged by 4 people

• [Callbacks] Remove MagnitudePruningModifier.leave_enabled

  #1198 merged Mar 7, 2025

• [Docs] Add info on when to use which PTQ/Sparsification

  #1157 merged Mar 6, 2025

• [Training] Unifying Preprocess + Postprocessing logic for Train/Oneshot

  #1212 merged Mar 6, 2025

• [BugFix] Fix logging disabling bug and add tests

  #1218 merged Mar 5, 2025

• [Training] Datasets - update Module

  #1209 merged Mar 5, 2025

• [Cosmetic] Rename data_args to dataset_args

  #1206 merged Mar 5, 2025

• Remove MonkeyPatch for GPUs

  #1227 merged Mar 5, 2025

• [Training] Decouple Argument parser

  #1207 merged Mar 3, 2025

3 Pull requests opened by 2 people

• fixing reproducibility of lmeval tests

  #1220 opened Mar 4, 2025

• [Quantization] Channel-wise Output Activation Quantization for Attention QKV Modules + KV-cache channel quantization

  #1233 opened Mar 7, 2025

• wandb/tensorboard loggers set default init to False

  #1235 opened Mar 7, 2025

6 Issues closed by 3 people

• Bug of quantizing part of the Qwen model

  #1230 closed Mar 6, 2025

• Why quantized model of 640MB took almost 3GB of VRAM?

  #1229 closed Mar 6, 2025

• How to choose sparsification rates for different layers?

  #1231 closed Mar 6, 2025

• Does llmcompressor support hybrid sparsity?

  #1037 closed Mar 5, 2025

• run llmcompressor to transfer deepseek r1 bf16 to w8a8 occur error

  #1221 closed Mar 4, 2025

• [Clarification] Using FP8 Dynamic quantization saves space on disk, but not when loaded?

  #1210 closed Mar 3, 2025

10 Issues opened by 5 people

• Fix docker container to build llm-compressor, not sparseml

  #1236 opened Mar 7, 2025

• Cuda out of memory while loading the quantized model

  #1234 opened Mar 7, 2025

• [Bug] Sporadic test failure on sparsification test

  #1232 opened Mar 6, 2025

• Quantization Memory Requirements

  #1228 opened Mar 5, 2025

• Fail when invalid arguments are provided in the recipe for a modifier

  #1226 opened Mar 5, 2025

• Fail when faulty recipes are provided during `oneshot`

  #1225 opened Mar 5, 2025

• Expand tests for `ModuleSparsificationInfo`

  #1224 opened Mar 5, 2025

• Update to use `loguru`

  #1223 opened Mar 5, 2025

• Update observers to make `MSE` the default

  #1222 opened Mar 5, 2025

• Incomplete warning message during `oneshot` post process - `WARNING - Optimized model not saved. To save, please provide`

  #1219 opened Mar 3, 2025

12 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• does it support asymmetric ?

  #1190 commented on Mar 5, 2025 • 0 new comments

• W4A8 model larger than W4A16

  #1215 commented on Mar 5, 2025 • 0 new comments

• Significant Inference Performance Degradation After W8A8 Quantization on CommandR-35B Model

  #1196 commented on Mar 5, 2025 • 0 new comments

• Lazy Loading of Weights for Large Model Quantization

  #1216 commented on Mar 6, 2025 • 0 new comments

• [Question] Has anyone successfully quantinize Deepseek-V3 to int4-w4a16?

  #1203 commented on Mar 7, 2025 • 0 new comments

• Cannot load quantized Multimodal_audio model using whisper.load_model("quantized-model-path)

  #1204 commented on Mar 7, 2025 • 0 new comments

• Perplexity (ppl) Calculation of Local Sparse Model: NaN issue

  #853 commented on Mar 7, 2025 • 0 new comments

• OOM during save_pretrained of compressed model

  #1183 commented on Mar 9, 2025 • 0 new comments

• [Callbacks] Remove double initialization, replace with updating the state directly

  #1169 commented on Mar 6, 2025 • 0 new comments

• Bdellabe/awq modifier v3

  #1177 commented on Mar 5, 2025 • 0 new comments

• [Callbacks] Remove `on_update`

  #1199 commented on Mar 7, 2025 • 0 new comments

• [Train] Training Pipeline

  #1214 commented on Mar 7, 2025 • 0 new comments