onecomp 1.4.2__tar.gz → 1.4.3__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (336) hide show
  1. {onecomp-1.4.2/onecomp.egg-info → onecomp-1.4.3}/PKG-INFO +21 -7
  2. {onecomp-1.4.2 → onecomp-1.4.3}/README.md +20 -6
  3. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/__version__.py +1 -1
  4. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/qep/_quantize_with_qep_arch.py +7 -1
  5. {onecomp-1.4.2 → onecomp-1.4.3/onecomp.egg-info}/PKG-INFO +21 -7
  6. {onecomp-1.4.2 → onecomp-1.4.3}/LICENSE +0 -0
  7. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/jointq_benchmark/2bit_g32/quant_benchmark.py +0 -0
  8. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/jointq_benchmark/actorder_mse/quant_benchmark.py +0 -0
  9. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/jointq_benchmark/calibration_stability/quant_benchmark.py +0 -0
  10. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/jointq_benchmark/main/quant_benchmark.py +0 -0
  11. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/jointq_benchmark/unregularization/quant_benchmark.py +0 -0
  12. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama2-7b-gptq/quant_benchmark.py +0 -0
  13. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama2-7b-gptq/quant_benchmark_2bit_g32.py +0 -0
  14. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama3-8b-gptq/quant_benchmark.py +0 -0
  15. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama3-8b-gptq/quant_benchmark_2bit_g32.py +0 -0
  16. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama3-8b-lpcd-gptq/quant_benchmark.py +0 -0
  17. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama3-8b-qep-gptq/quant_benchmark.py +0 -0
  18. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/llama3-8b-various/quant_benchmark.py +0 -0
  19. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/qwen3-14b-gptq/quant_benchmark.py +0 -0
  20. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/qwen3-14b-gptq/quant_benchmark_2bit_g32.py +0 -0
  21. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/qwen3-8b-gptq/quant_benchmark.py +0 -0
  22. {onecomp-1.4.2 → onecomp-1.4.3}/benchmark/qwen3-8b-gptq/quant_benchmark_2bit_g32.py +0 -0
  23. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/__init__.py +0 -0
  24. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/api/__init__.py +0 -0
  25. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/api/jobs.py +0 -0
  26. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/constants.py +0 -0
  27. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/core/__init__.py +0 -0
  28. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/core/config.py +0 -0
  29. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/core/database.py +0 -0
  30. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/main.py +0 -0
  31. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/models/__init__.py +0 -0
  32. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/models/job.py +0 -0
  33. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/schemas/__init__.py +0 -0
  34. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/schemas/job.py +0 -0
  35. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/services/__init__.py +0 -0
  36. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/services/huggingface.py +0 -0
  37. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/services/inference.py +0 -0
  38. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/services/job_store.py +0 -0
  39. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/worker/__init__.py +0 -0
  40. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/worker/celery_app.py +0 -0
  41. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/app/worker/tasks.py +0 -0
  42. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/cpu_patch.py +0 -0
  43. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/start_backend.py +0 -0
  44. {onecomp-1.4.2 → onecomp-1.4.3}/dashboard/backend/start_worker.py +0 -0
  45. {onecomp-1.4.2 → onecomp-1.4.3}/envs/openvino/example_export_openvino.py +0 -0
  46. {onecomp-1.4.2 → onecomp-1.4.3}/envs/openvino/openvino_export_utils.py +0 -0
  47. {onecomp-1.4.2 → onecomp-1.4.3}/envs/vllm/v0_24_0_rocm/src/onecomp_vllm_v0_24_0_rocm/__init__.py +0 -0
  48. {onecomp-1.4.2 → onecomp-1.4.3}/envs/vllm/v0_24_0_rocm/src/onecomp_vllm_v0_24_0_rocm/patch.py +0 -0
  49. {onecomp-1.4.2 → onecomp-1.4.3}/example/cpu_inference/example_gptq_gguf_cpu.py +0 -0
  50. {onecomp-1.4.2 → onecomp-1.4.3}/example/cpu_inference/example_mixed_gptq_gguf_cpu.py +0 -0
  51. {onecomp-1.4.2 → onecomp-1.4.3}/example/cpu_inference/example_serve_cpu.py +0 -0
  52. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_auto_run.py +0 -0
  53. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_autobit.py +0 -0
  54. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_custom_calibration.py +0 -0
  55. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_gptq.py +0 -0
  56. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_jointq.py +0 -0
  57. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_lpcd_gptq.py +0 -0
  58. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_mdbf.py +0 -0
  59. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_qep_gptq.py +0 -0
  60. {onecomp-1.4.2 → onecomp-1.4.3}/example/example_save_load.py +0 -0
  61. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_blockwise_global_ptq.py +0 -0
  62. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_blockwise_global_ptq_staged.py +0 -0
  63. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_blockwise_ptq.py +0 -0
  64. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_global_ptq.py +0 -0
  65. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_global_ptq_dbf.py +0 -0
  66. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_global_ptq_distributed.py +0 -0
  67. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_global_ptq_mdbf.py +0 -0
  68. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_lora_gptq_vllm_inference.py +0 -0
  69. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_lora_sft.py +0 -0
  70. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_lora_sft_knowledge.py +0 -0
  71. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_lora_sft_knowledge_jointq.py +0 -0
  72. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_reload_post_process_resave.py +0 -0
  73. {onecomp-1.4.2 → onecomp-1.4.3}/example/post_process/example_router_fine_tuning.py +0 -0
  74. {onecomp-1.4.2 → onecomp-1.4.3}/example/pre_process/example_llama_preprocess_rtn.py +0 -0
  75. {onecomp-1.4.2 → onecomp-1.4.3}/example/pre_process/example_preprocess_save_load.py +0 -0
  76. {onecomp-1.4.2 → onecomp-1.4.3}/example/vllm_inference/example_autobit_vllm_inference.py +0 -0
  77. {onecomp-1.4.2 → onecomp-1.4.3}/example/vllm_inference/example_dbf_vllm_inference.py +0 -0
  78. {onecomp-1.4.2 → onecomp-1.4.3}/example/vllm_inference/example_gptq_vllm_gptoss_inference.py +0 -0
  79. {onecomp-1.4.2 → onecomp-1.4.3}/example/vllm_inference/example_gptq_vllm_inference.py +0 -0
  80. {onecomp-1.4.2 → onecomp-1.4.3}/example/vllm_inference/example_gptq_vllm_qwen36_inference.py +0 -0
  81. {onecomp-1.4.2 → onecomp-1.4.3}/example/vllm_inference/example_jointq_vllm_inference.py +0 -0
  82. {onecomp-1.4.2 → onecomp-1.4.3}/llamacpp_plugins/__init__.py +0 -0
  83. {onecomp-1.4.2 → onecomp-1.4.3}/llamacpp_plugins/gptq/__init__.py +0 -0
  84. {onecomp-1.4.2 → onecomp-1.4.3}/llamacpp_plugins/gptq/constants.py +0 -0
  85. {onecomp-1.4.2 → onecomp-1.4.3}/llamacpp_plugins/gptq/llamacpp_plugin.py +0 -0
  86. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/autobit/validate_autobit.py +0 -0
  87. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/autobit_qep/validate_autobit.py +0 -0
  88. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/gptq/validate_gptq.py +0 -0
  89. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/gptq/validate_load.py +0 -0
  90. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/gptq/validate_vllm.py +0 -0
  91. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/jointq/validate_jointq.py +0 -0
  92. {onecomp-1.4.2 → onecomp-1.4.3}/model_validation/qep_gptq/validate_gptq.py +0 -0
  93. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/__init__.py +0 -0
  94. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/__main__.py +0 -0
  95. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/analyzer/__init__.py +0 -0
  96. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/analyzer/cumulative_error.py +0 -0
  97. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/analyzer/output_squared_error.py +0 -0
  98. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/analyzer/quantization_error.py +0 -0
  99. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/analyzer/weight_outlier.py +0 -0
  100. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/__init__.py +0 -0
  101. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/_cache.py +0 -0
  102. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/c4.py +0 -0
  103. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/calibration_config.py +0 -0
  104. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/calibration_data_loader.py +0 -0
  105. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/chunking.py +0 -0
  106. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/custom.py +0 -0
  107. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/calibration/wikitext.py +0 -0
  108. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cli.py +0 -0
  109. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/__init__.py +0 -0
  110. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/cli.py +0 -0
  111. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/eval/__init__.py +0 -0
  112. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/eval/benchmark.py +0 -0
  113. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/eval/inspect_gguf.py +0 -0
  114. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/eval/parity.py +0 -0
  115. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/eval/perplexity.py +0 -0
  116. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/__init__.py +0 -0
  117. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/auto.py +0 -0
  118. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/blocks.py +0 -0
  119. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/checkpoint.py +0 -0
  120. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/dequantize.py +0 -0
  121. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/direct.py +0 -0
  122. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/fallback.py +0 -0
  123. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/rotation.py +0 -0
  124. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/skeleton.py +0 -0
  125. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/transforms/__init__.py +0 -0
  126. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/export/transforms/qwen35.py +0 -0
  127. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/inference.py +0 -0
  128. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/llama_tooling.py +0 -0
  129. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/cpu/serve.py +0 -0
  130. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/__init__.py +0 -0
  131. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/__main__.py +0 -0
  132. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/conf/__init__.py +0 -0
  133. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/conf/eval_config.yaml +0 -0
  134. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/__init__.py +0 -0
  135. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/base.py +0 -0
  136. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/__init__.py +0 -0
  137. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/adapter.py +0 -0
  138. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/data.py +0 -0
  139. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/gen_answer.py +0 -0
  140. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/judge.py +0 -0
  141. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/radar_chart.py +0 -0
  142. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/run.py +0 -0
  143. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/mt_bench/show_result.py +0 -0
  144. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/throughput/__init__.py +0 -0
  145. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/throughput/adapter.py +0 -0
  146. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/throughput/bench.py +0 -0
  147. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/evals/throughput/run.py +0 -0
  148. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/orchestrator/__init__.py +0 -0
  149. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/orchestrator/aggregator.py +0 -0
  150. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/orchestrator/runner.py +0 -0
  151. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/orchestrator/server.py +0 -0
  152. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/orchestrator/subprocess_runner.py +0 -0
  153. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/run_evaluate.py +0 -0
  154. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/schema.py +0 -0
  155. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/utils/__init__.py +0 -0
  156. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/utils/model_utils.py +0 -0
  157. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/utils/ports.py +0 -0
  158. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/utils/resources.py +0 -0
  159. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/eval/utils/secrets.py +0 -0
  160. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/export/__init__.py +0 -0
  161. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/export/gguf_export.py +0 -0
  162. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/export/gguf_reader.py +0 -0
  163. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/export/gguf_writer.py +0 -0
  164. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/export/hub.py +0 -0
  165. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/export/model_card.py +0 -0
  166. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/log.py +0 -0
  167. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/__init__.py +0 -0
  168. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/_gradient_solver.py +0 -0
  169. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/_lpcd_config.py +0 -0
  170. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/_lpcd_runner.py +0 -0
  171. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/_metric.py +0 -0
  172. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/_refiner.py +0 -0
  173. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/arch/_llama.py +0 -0
  174. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/arch/_llama_cf.py +0 -0
  175. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/lpcd/arch/_qwen3.py +0 -0
  176. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/model_config.py +0 -0
  177. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/__init__.py +0 -0
  178. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_base.py +0 -0
  179. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/__init__.py +0 -0
  180. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/dbf_block_optimizer.py +0 -0
  181. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/dbf_cbq_optimizer.py +0 -0
  182. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/generic_block_optimizer.py +0 -0
  183. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/gptq_block_optimizer.py +0 -0
  184. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/gptq_cbq_optimizer.py +0 -0
  185. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/helpers.py +0 -0
  186. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/onebit_block_optimizer.py +0 -0
  187. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_blockwise/onebit_cbq_optimizer.py +0 -0
  188. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/__init__.py +0 -0
  189. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/core.py +0 -0
  190. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/dbf_adapter.py +0 -0
  191. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/gptq_adapter.py +0 -0
  192. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/helpers.py +0 -0
  193. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/losses.py +0 -0
  194. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/mdbf_adapter.py +0 -0
  195. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_global_ptq/trainer.py +0 -0
  196. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/_runtime.py +0 -0
  197. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/blockwise_ptq.py +0 -0
  198. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/global_ptq.py +0 -0
  199. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/global_ptq_distributed.py +0 -0
  200. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/post_process_lora_sft.py +0 -0
  201. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/post_process/router_fine_tuning.py +0 -0
  202. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/__init__.py +0 -0
  203. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/hadamard_utils.py +0 -0
  204. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/modeling_llama.py +0 -0
  205. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/modeling_qwen3.py +0 -0
  206. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/optimizer.py +0 -0
  207. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/prepare_rotated_model.py +0 -0
  208. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/preprocess_args.py +0 -0
  209. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/quant_models.py +0 -0
  210. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/rotation_utils.py +0 -0
  211. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/pre_process/train_rotation.py +0 -0
  212. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/qep/__init__.py +0 -0
  213. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/qep/_qep_config.py +0 -0
  214. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/qep/_quantize_with_qep.py +0 -0
  215. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantized_model_loader.py +0 -0
  216. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/__init__.py +0 -0
  217. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/_quantizer.py +0 -0
  218. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/arb/__init__.py +0 -0
  219. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/arb/_arb.py +0 -0
  220. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/arb/arb_impl.py +0 -0
  221. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/__init__.py +0 -0
  222. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/_autobit.py +0 -0
  223. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/activation_stats.py +0 -0
  224. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/dbf_fallback.py +0 -0
  225. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/ilp.py +0 -0
  226. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/manual.py +0 -0
  227. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/autobit/visualize.py +0 -0
  228. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/cq/__init__.py +0 -0
  229. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/cq/_cq.py +0 -0
  230. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/cq/cq_impl.py +0 -0
  231. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/__init__.py +0 -0
  232. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/_dbf.py +0 -0
  233. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/admm_extended.py +0 -0
  234. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/balance.py +0 -0
  235. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/config.py +0 -0
  236. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/dbf_impl.py +0 -0
  237. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/dbf_layer.py +0 -0
  238. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/dbf_original.py +0 -0
  239. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/fine_tune.py +0 -0
  240. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/dbf/middle.py +0 -0
  241. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/gemlite.py +0 -0
  242. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/gptq/__init__.py +0 -0
  243. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/gptq/_gptq.py +0 -0
  244. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/gptq/config.py +0 -0
  245. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/gptq/gptq_layer.py +0 -0
  246. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/__init__.py +0 -0
  247. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/_jointq.py +0 -0
  248. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/__init__.py +0 -0
  249. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/__version__.py +0 -0
  250. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/clip.py +0 -0
  251. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/error_propagation/__init__.py +0 -0
  252. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/error_propagation/local_search_advanced.py +0 -0
  253. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/error_propagation/quantize_advanced.py +0 -0
  254. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/error_propagation/quantizer_advanced.py +0 -0
  255. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/group_swap.py +0 -0
  256. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/local_search.py +0 -0
  257. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/quantize.py +0 -0
  258. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/quantize_multi_gpu.py +0 -0
  259. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/quantizer.py +0 -0
  260. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/jointq/core/solution.py +0 -0
  261. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/__init__.py +0 -0
  262. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/_mdbf.py +0 -0
  263. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/admm.py +0 -0
  264. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/config.py +0 -0
  265. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/gradient_refine.py +0 -0
  266. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/initialize.py +0 -0
  267. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/mdbf_impl.py +0 -0
  268. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/mdbf_layer.py +0 -0
  269. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/mdbf/utils.py +0 -0
  270. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/onebit/__init__.py +0 -0
  271. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/onebit/_onebit.py +0 -0
  272. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/onebit/onebit_impl.py +0 -0
  273. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/onebit/onebit_layer.py +0 -0
  274. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/qbb/__init__.py +0 -0
  275. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/qbb/_qbb.py +0 -0
  276. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/qbb/qbb_impl.py +0 -0
  277. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/__init__.py +0 -0
  278. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/_quip.py +0 -0
  279. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/quant_quip.py +0 -0
  280. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/quip_impl.py +0 -0
  281. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/utils.py +0 -0
  282. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/utils_had.py +0 -0
  283. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/quip/vector_balance.py +0 -0
  284. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/rtn/__init__.py +0 -0
  285. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/rtn/_rtn.py +0 -0
  286. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/rtn/quantizer.py +0 -0
  287. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/quantizer/rtn/rtn_impl.py +0 -0
  288. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/rotated_model_config.py +0 -0
  289. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/runner.py +0 -0
  290. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/runner_methods/__init__.py +0 -0
  291. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/runner_methods/chunked_quantization.py +0 -0
  292. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/runner_methods/jointq_error_propagation.py +0 -0
  293. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/__init__.py +0 -0
  294. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/accuracy.py +0 -0
  295. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/activation_capture.py +0 -0
  296. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/activation_check.py +0 -0
  297. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/blockwise.py +0 -0
  298. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/device.py +0 -0
  299. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/dtype.py +0 -0
  300. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/lora.py +0 -0
  301. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/model_inputs.py +0 -0
  302. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/mxfp4_compat.py +0 -0
  303. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/perplexity.py +0 -0
  304. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/quant_config.py +0 -0
  305. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/quantization_progress.py +0 -0
  306. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/unfuse_moe.py +0 -0
  307. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp/utils/vram_estimator.py +0 -0
  308. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp.egg-info/SOURCES.txt +0 -0
  309. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp.egg-info/dependency_links.txt +0 -0
  310. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp.egg-info/entry_points.txt +0 -0
  311. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp.egg-info/requires.txt +0 -0
  312. {onecomp-1.4.2 → onecomp-1.4.3}/onecomp.egg-info/top_level.txt +0 -0
  313. {onecomp-1.4.2 → onecomp-1.4.3}/pyproject.toml +0 -0
  314. {onecomp-1.4.2 → onecomp-1.4.3}/scripts/check_copyright_header.py +0 -0
  315. {onecomp-1.4.2 → onecomp-1.4.3}/scripts/check_no_japanese.py +0 -0
  316. {onecomp-1.4.2 → onecomp-1.4.3}/scripts/check_readme_examples.py +0 -0
  317. {onecomp-1.4.2 → onecomp-1.4.3}/scripts/prepare_calibration_cache.py +0 -0
  318. {onecomp-1.4.2 → onecomp-1.4.3}/setup.cfg +0 -0
  319. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/__init__.py +0 -0
  320. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/dbf/__init__.py +0 -0
  321. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/dbf/modules/__init__.py +0 -0
  322. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/dbf/modules/gemlite_linear.py +0 -0
  323. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/dbf/modules/naive.py +0 -0
  324. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/dbf/vllm_plugin.py +0 -0
  325. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/gptq/__init__.py +0 -0
  326. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/gptq/constants.py +0 -0
  327. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/gptq/gptoss_wna16_moe.py +0 -0
  328. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/gptq/vllm_plugin.py +0 -0
  329. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/patches/__init__.py +0 -0
  330. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/patches/_paths.py +0 -0
  331. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/patches/apply_all.py +0 -0
  332. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/patches/gpt_oss_gptq_moe.py +0 -0
  333. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/patches/gpt_oss_wna16_bias.py +0 -0
  334. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/utils/__init__.py +0 -0
  335. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/utils/module.py +0 -0
  336. {onecomp-1.4.2 → onecomp-1.4.3}/vllm_plugins/utils/rotation.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: onecomp
3
- Version: 1.4.2
3
+ Version: 1.4.3
4
4
  Summary: Python package for LLM compression
5
5
  Author: Keiji Kimura
6
6
  License: MIT License
@@ -151,8 +151,8 @@ Full documentation is available at **[https://FujitsuResearch.github.io/OneCompr
151
151
  - **Layer-Projected Coordinate Descent (LPCD)**: A unified Post Training Quantization (PTQ) framework that extends layer-wise quantization to arbitrary submodules by optimising relaxed objectives and projecting the solutions with layer-wise quantizers. See [Ichikawa et al., 2025](https://arxiv.org/abs/2512.01546) for details.
152
152
  - **vLLM Plugin Integration**: Serve OneComp-quantized models with [vLLM](https://docs.vllm.ai/) via built-in plugins for DBF and Mixed-GPTQ quantization methods. Pair with [Open WebUI](https://github.com/open-webui/open-webui) for a ChatGPT-like chat experience on your local machine.
153
153
  - **AutoBit**: Mixed-precision quantization with ILP-based bitwidth assignment. Automatically estimates the target bitwidth from available VRAM and assigns per-layer bitwidths to minimize quantization error under the memory budget.
154
- - **JointQ**: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128).
155
- - **MDBF (Multi-Envelope Double Binary Factorization)**: A binary factorization quantizer that approximates each weight matrix as a sum of multi-path sign matrices with multi-scale FP16 envelopes, generalizing DBF and LittleBit for aggressive low-bit (sub-1-bit) compression. Supports ADMM/gradient refinement, activation-aware initialization, and a GemLite-accelerated 1-bit inference path. See the [MDBF guide](https://FujitsuResearch.github.io/OneCompression/algorithms/mdbf/) for details.
154
+ - **JointQ**: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128). See [Kimura & Ichikawa, 2026](https://arxiv.org/abs/2611.20384) for details.
155
+ - **MDBF (Multi-Envelope Double Binary Factorization)**: A binary factorization quantizer that approximates each weight matrix as a sum of multi-path sign matrices with multi-scale FP16 envelopes, generalizing DBF and LittleBit for aggressive low-bit (sub-1-bit) compression. Supports ADMM/gradient refinement, activation-aware initialization, and a GemLite-accelerated 1-bit inference path. See [Ichikawa et al., 2025](https://arxiv.org/abs/2512.24545) for details and the [MDBF guide](https://FujitsuResearch.github.io/OneCompression/algorithms/mdbf/) for usage.
156
156
  - **Block-wise PTQ**: Post-quantization block-wise distillation that minimises intermediate-representation MSE against an FP16 teacher model at Transformer-block granularity. Includes Phase 1 (greedy per-block optimisation) and Phase 2 CBQ (cross-block sliding-window optimisation). Supports GPTQ, DBF, and OneBit quantizers.
157
157
  - **Global PTQ**: Globally optimise GPTQ scales/zeros, DBF scaling factors, and MDBF amplitude parameters through KL distillation from a full-precision teacher. Supports optional MDBF binary-factor training with STE, single-device and distributed execution with DeepSpeed, and configurable teacher device placement. See the [Global PTQ guide](docs/user-guide/post-process.md#global-ptq-parameter-optimisation).
158
158
  - **MoE Router Fine-Tuning**: Adapt routing decisions after expert quantization by training only router parameters with next-token prediction loss, while experts and other weights remain frozen. Results use the normal quantized-model save/load workflow. See the [router fine-tuning guide](docs/user-guide/post-process.md#router-fine-tuning-for-quantized-moe-models).
@@ -492,7 +492,7 @@ terms of the installed versions as well.
492
492
 
493
493
  ## Citation
494
494
 
495
- OneComp technical report:
495
+ [OneComp technical report](https://arxiv.org/abs/2603.28845):
496
496
 
497
497
  ```
498
498
  @misc{ichikawa2026onecomponelinerevolutiongenerative,
@@ -506,7 +506,7 @@ OneComp technical report:
506
506
  }
507
507
  ```
508
508
 
509
- QEP (Quantization Error Propagation):
509
+ [QEP (Quantization Error Propagation)](https://openreview.net/forum?id=a3l3K9khbL):
510
510
 
511
511
  ```
512
512
  @inproceedings{
@@ -519,7 +519,7 @@ url={https://openreview.net/forum?id=a3l3K9khbL}
519
519
  }
520
520
  ```
521
521
 
522
- LPCD (Layer-Projected Coordinate Descent):
522
+ [LPCD (Layer-Projected Coordinate Descent)](https://arxiv.org/abs/2512.01546):
523
523
 
524
524
  ```
525
525
  @article{ichikawa2025lpcd,
@@ -531,7 +531,7 @@ url={https://arxiv.org/abs/2512.01546}
531
531
  }
532
532
  ```
533
533
 
534
- MDBF (Multi-Envelope Double Binary Factorization):
534
+ [MDBF (Multi-Envelope Double Binary Factorization)](https://arxiv.org/abs/2512.24545):
535
535
 
536
536
  ```
537
537
  @misc{ichikawa2025bitsmultienvelopedoublebinary,
@@ -544,3 +544,17 @@ MDBF (Multi-Envelope Double Binary Factorization):
544
544
  url={https://arxiv.org/abs/2512.24545},
545
545
  }
546
546
  ```
547
+
548
+ [JointQ (Joint Scale-and-Integer Quantization)](https://arxiv.org/abs/2611.20384):
549
+
550
+ ```bibtex
551
+ @misc{kimura2026jointqjointscaleandintegerquantization,
552
+ title={JointQ: Joint Scale-and-Integer Quantization for Large Language Models},
553
+ author={Keiji Kimura and Yuma Ichikawa},
554
+ year={2026},
555
+ eprint={2611.20384},
556
+ archivePrefix={arXiv},
557
+ primaryClass={cs.LG},
558
+ url={https://arxiv.org/abs/2611.20384},
559
+ }
560
+ ```
@@ -37,8 +37,8 @@ Full documentation is available at **[https://FujitsuResearch.github.io/OneCompr
37
37
  - **Layer-Projected Coordinate Descent (LPCD)**: A unified Post Training Quantization (PTQ) framework that extends layer-wise quantization to arbitrary submodules by optimising relaxed objectives and projecting the solutions with layer-wise quantizers. See [Ichikawa et al., 2025](https://arxiv.org/abs/2512.01546) for details.
38
38
  - **vLLM Plugin Integration**: Serve OneComp-quantized models with [vLLM](https://docs.vllm.ai/) via built-in plugins for DBF and Mixed-GPTQ quantization methods. Pair with [Open WebUI](https://github.com/open-webui/open-webui) for a ChatGPT-like chat experience on your local machine.
39
39
  - **AutoBit**: Mixed-precision quantization with ILP-based bitwidth assignment. Automatically estimates the target bitwidth from available VRAM and assigns per-layer bitwidths to minimize quantization error under the memory budget.
40
- - **JointQ**: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128).
41
- - **MDBF (Multi-Envelope Double Binary Factorization)**: A binary factorization quantizer that approximates each weight matrix as a sum of multi-path sign matrices with multi-scale FP16 envelopes, generalizing DBF and LittleBit for aggressive low-bit (sub-1-bit) compression. Supports ADMM/gradient refinement, activation-aware initialization, and a GemLite-accelerated 1-bit inference path. See the [MDBF guide](https://FujitsuResearch.github.io/OneCompression/algorithms/mdbf/) for details.
40
+ - **JointQ**: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128). See [Kimura & Ichikawa, 2026](https://arxiv.org/abs/2611.20384) for details.
41
+ - **MDBF (Multi-Envelope Double Binary Factorization)**: A binary factorization quantizer that approximates each weight matrix as a sum of multi-path sign matrices with multi-scale FP16 envelopes, generalizing DBF and LittleBit for aggressive low-bit (sub-1-bit) compression. Supports ADMM/gradient refinement, activation-aware initialization, and a GemLite-accelerated 1-bit inference path. See [Ichikawa et al., 2025](https://arxiv.org/abs/2512.24545) for details and the [MDBF guide](https://FujitsuResearch.github.io/OneCompression/algorithms/mdbf/) for usage.
42
42
  - **Block-wise PTQ**: Post-quantization block-wise distillation that minimises intermediate-representation MSE against an FP16 teacher model at Transformer-block granularity. Includes Phase 1 (greedy per-block optimisation) and Phase 2 CBQ (cross-block sliding-window optimisation). Supports GPTQ, DBF, and OneBit quantizers.
43
43
  - **Global PTQ**: Globally optimise GPTQ scales/zeros, DBF scaling factors, and MDBF amplitude parameters through KL distillation from a full-precision teacher. Supports optional MDBF binary-factor training with STE, single-device and distributed execution with DeepSpeed, and configurable teacher device placement. See the [Global PTQ guide](docs/user-guide/post-process.md#global-ptq-parameter-optimisation).
44
44
  - **MoE Router Fine-Tuning**: Adapt routing decisions after expert quantization by training only router parameters with next-token prediction loss, while experts and other weights remain frozen. Results use the normal quantized-model save/load workflow. See the [router fine-tuning guide](docs/user-guide/post-process.md#router-fine-tuning-for-quantized-moe-models).
@@ -378,7 +378,7 @@ terms of the installed versions as well.
378
378
 
379
379
  ## Citation
380
380
 
381
- OneComp technical report:
381
+ [OneComp technical report](https://arxiv.org/abs/2603.28845):
382
382
 
383
383
  ```
384
384
  @misc{ichikawa2026onecomponelinerevolutiongenerative,
@@ -392,7 +392,7 @@ OneComp technical report:
392
392
  }
393
393
  ```
394
394
 
395
- QEP (Quantization Error Propagation):
395
+ [QEP (Quantization Error Propagation)](https://openreview.net/forum?id=a3l3K9khbL):
396
396
 
397
397
  ```
398
398
  @inproceedings{
@@ -405,7 +405,7 @@ url={https://openreview.net/forum?id=a3l3K9khbL}
405
405
  }
406
406
  ```
407
407
 
408
- LPCD (Layer-Projected Coordinate Descent):
408
+ [LPCD (Layer-Projected Coordinate Descent)](https://arxiv.org/abs/2512.01546):
409
409
 
410
410
  ```
411
411
  @article{ichikawa2025lpcd,
@@ -417,7 +417,7 @@ url={https://arxiv.org/abs/2512.01546}
417
417
  }
418
418
  ```
419
419
 
420
- MDBF (Multi-Envelope Double Binary Factorization):
420
+ [MDBF (Multi-Envelope Double Binary Factorization)](https://arxiv.org/abs/2512.24545):
421
421
 
422
422
  ```
423
423
  @misc{ichikawa2025bitsmultienvelopedoublebinary,
@@ -430,3 +430,17 @@ MDBF (Multi-Envelope Double Binary Factorization):
430
430
  url={https://arxiv.org/abs/2512.24545},
431
431
  }
432
432
  ```
433
+
434
+ [JointQ (Joint Scale-and-Integer Quantization)](https://arxiv.org/abs/2611.20384):
435
+
436
+ ```bibtex
437
+ @misc{kimura2026jointqjointscaleandintegerquantization,
438
+ title={JointQ: Joint Scale-and-Integer Quantization for Large Language Models},
439
+ author={Keiji Kimura and Yuma Ichikawa},
440
+ year={2026},
441
+ eprint={2611.20384},
442
+ archivePrefix={arXiv},
443
+ primaryClass={cs.LG},
444
+ url={https://arxiv.org/abs/2611.20384},
445
+ }
446
+ ```
@@ -6,4 +6,4 @@ Author: Keiji Kimura
6
6
 
7
7
  """
8
8
 
9
- __version__ = "1.4.2"
9
+ __version__ = "1.4.3"
@@ -444,6 +444,9 @@ def run_quantize_with_qep_arch(
444
444
  expert_modules_q.extend(targets)
445
445
  else:
446
446
  regular_pairs.append((group_q, group_f))
447
+ # make_grouped_module records a Linear on every call, and the expert loop below pops each
448
+ # Hessian, so a Linear listed twice would raise KeyError.
449
+ expert_modules_q = list(dict.fromkeys(expert_modules_q))
447
450
 
448
451
  # make_grouped_module only sees modules whose forward hook fired for
449
452
  # its single calibration sample, so unrouted experts are missing
@@ -559,7 +562,9 @@ def run_quantize_with_qep_arch(
559
562
  )
560
563
  for module_q in expert_modules_q:
561
564
  name = quantizer.module_to_name[module_q]
562
- entry = expert_hessians[module_q]
565
+ # Pop instead of indexing, and del below: otherwise the dict and the loop
566
+ # variables keep this block's expert Hessians on the device into the next block.
567
+ entry = expert_hessians.pop(module_q)
563
568
  if entry is None:
564
569
  fallback_quantizer = _resolve_gptq_for_rtn_fallback(quantizer, module_q)
565
570
  if fallback_quantizer is not None:
@@ -597,6 +602,7 @@ def run_quantize_with_qep_arch(
597
602
  delta_hatX=None,
598
603
  nsamples=nsamples_expert if quantizer.flag_nsamples else None,
599
604
  )
605
+ del H, entry
600
606
  try:
601
607
  dtype = module_q.weight.data.dtype
602
608
  module_q.weight.data = (
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: onecomp
3
- Version: 1.4.2
3
+ Version: 1.4.3
4
4
  Summary: Python package for LLM compression
5
5
  Author: Keiji Kimura
6
6
  License: MIT License
@@ -151,8 +151,8 @@ Full documentation is available at **[https://FujitsuResearch.github.io/OneCompr
151
151
  - **Layer-Projected Coordinate Descent (LPCD)**: A unified Post Training Quantization (PTQ) framework that extends layer-wise quantization to arbitrary submodules by optimising relaxed objectives and projecting the solutions with layer-wise quantizers. See [Ichikawa et al., 2025](https://arxiv.org/abs/2512.01546) for details.
152
152
  - **vLLM Plugin Integration**: Serve OneComp-quantized models with [vLLM](https://docs.vllm.ai/) via built-in plugins for DBF and Mixed-GPTQ quantization methods. Pair with [Open WebUI](https://github.com/open-webui/open-webui) for a ChatGPT-like chat experience on your local machine.
153
153
  - **AutoBit**: Mixed-precision quantization with ILP-based bitwidth assignment. Automatically estimates the target bitwidth from available VRAM and assigns per-layer bitwidths to minimize quantization error under the memory budget.
154
- - **JointQ**: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128).
155
- - **MDBF (Multi-Envelope Double Binary Factorization)**: A binary factorization quantizer that approximates each weight matrix as a sum of multi-path sign matrices with multi-scale FP16 envelopes, generalizing DBF and LittleBit for aggressive low-bit (sub-1-bit) compression. Supports ADMM/gradient refinement, activation-aware initialization, and a GemLite-accelerated 1-bit inference path. See the [MDBF guide](https://FujitsuResearch.github.io/OneCompression/algorithms/mdbf/) for details.
154
+ - **JointQ**: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128). See [Kimura & Ichikawa, 2026](https://arxiv.org/abs/2611.20384) for details.
155
+ - **MDBF (Multi-Envelope Double Binary Factorization)**: A binary factorization quantizer that approximates each weight matrix as a sum of multi-path sign matrices with multi-scale FP16 envelopes, generalizing DBF and LittleBit for aggressive low-bit (sub-1-bit) compression. Supports ADMM/gradient refinement, activation-aware initialization, and a GemLite-accelerated 1-bit inference path. See [Ichikawa et al., 2025](https://arxiv.org/abs/2512.24545) for details and the [MDBF guide](https://FujitsuResearch.github.io/OneCompression/algorithms/mdbf/) for usage.
156
156
  - **Block-wise PTQ**: Post-quantization block-wise distillation that minimises intermediate-representation MSE against an FP16 teacher model at Transformer-block granularity. Includes Phase 1 (greedy per-block optimisation) and Phase 2 CBQ (cross-block sliding-window optimisation). Supports GPTQ, DBF, and OneBit quantizers.
157
157
  - **Global PTQ**: Globally optimise GPTQ scales/zeros, DBF scaling factors, and MDBF amplitude parameters through KL distillation from a full-precision teacher. Supports optional MDBF binary-factor training with STE, single-device and distributed execution with DeepSpeed, and configurable teacher device placement. See the [Global PTQ guide](docs/user-guide/post-process.md#global-ptq-parameter-optimisation).
158
158
  - **MoE Router Fine-Tuning**: Adapt routing decisions after expert quantization by training only router parameters with next-token prediction loss, while experts and other weights remain frozen. Results use the normal quantized-model save/load workflow. See the [router fine-tuning guide](docs/user-guide/post-process.md#router-fine-tuning-for-quantized-moe-models).
@@ -492,7 +492,7 @@ terms of the installed versions as well.
492
492
 
493
493
  ## Citation
494
494
 
495
- OneComp technical report:
495
+ [OneComp technical report](https://arxiv.org/abs/2603.28845):
496
496
 
497
497
  ```
498
498
  @misc{ichikawa2026onecomponelinerevolutiongenerative,
@@ -506,7 +506,7 @@ OneComp technical report:
506
506
  }
507
507
  ```
508
508
 
509
- QEP (Quantization Error Propagation):
509
+ [QEP (Quantization Error Propagation)](https://openreview.net/forum?id=a3l3K9khbL):
510
510
 
511
511
  ```
512
512
  @inproceedings{
@@ -519,7 +519,7 @@ url={https://openreview.net/forum?id=a3l3K9khbL}
519
519
  }
520
520
  ```
521
521
 
522
- LPCD (Layer-Projected Coordinate Descent):
522
+ [LPCD (Layer-Projected Coordinate Descent)](https://arxiv.org/abs/2512.01546):
523
523
 
524
524
  ```
525
525
  @article{ichikawa2025lpcd,
@@ -531,7 +531,7 @@ url={https://arxiv.org/abs/2512.01546}
531
531
  }
532
532
  ```
533
533
 
534
- MDBF (Multi-Envelope Double Binary Factorization):
534
+ [MDBF (Multi-Envelope Double Binary Factorization)](https://arxiv.org/abs/2512.24545):
535
535
 
536
536
  ```
537
537
  @misc{ichikawa2025bitsmultienvelopedoublebinary,
@@ -544,3 +544,17 @@ MDBF (Multi-Envelope Double Binary Factorization):
544
544
  url={https://arxiv.org/abs/2512.24545},
545
545
  }
546
546
  ```
547
+
548
+ [JointQ (Joint Scale-and-Integer Quantization)](https://arxiv.org/abs/2611.20384):
549
+
550
+ ```bibtex
551
+ @misc{kimura2026jointqjointscaleandintegerquantization,
552
+ title={JointQ: Joint Scale-and-Integer Quantization for Large Language Models},
553
+ author={Keiji Kimura and Yuma Ichikawa},
554
+ year={2026},
555
+ eprint={2611.20384},
556
+ archivePrefix={arXiv},
557
+ primaryClass={cs.LG},
558
+ url={https://arxiv.org/abs/2611.20384},
559
+ }
560
+ ```
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes