onecomp 1.3.4__tar.gz → 1.4.1__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.
- {onecomp-1.3.4/onecomp.egg-info → onecomp-1.4.1}/PKG-INFO +16 -6
- {onecomp-1.3.4 → onecomp-1.4.1}/README.md +15 -5
- onecomp-1.4.1/envs/openvino/example_export_openvino.py +66 -0
- onecomp-1.4.1/envs/openvino/openvino_export_utils.py +115 -0
- onecomp-1.4.1/example/post_process/example_global_ptq_mdbf.py +48 -0
- onecomp-1.4.1/example/post_process/example_router_fine_tuning.py +76 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/llamacpp_plugins/gptq/llamacpp_plugin.py +42 -6
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/__version__.py +1 -1
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cli.py +3 -1
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/__init__.py +1 -1
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/__init__.py +3 -3
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/auto.py +28 -11
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/checkpoint.py +8 -3
- onecomp-1.4.1/onecomp/cpu/export/dequantize.py +429 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/direct.py +32 -6
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/fallback.py +8 -6
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/skeleton.py +3 -2
- onecomp-1.4.1/onecomp/cpu/export/transforms/__init__.py +0 -0
- onecomp-1.4.1/onecomp/cpu/export/transforms/qwen35.py +258 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/llama_tooling.py +48 -5
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/export/hub.py +1 -1
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/export/model_card.py +38 -4
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/__init__.py +2 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/core.py +89 -16
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/dbf_adapter.py +30 -12
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/gptq_adapter.py +23 -12
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/helpers.py +23 -1
- onecomp-1.4.1/onecomp/post_process/_global_ptq/mdbf_adapter.py +236 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/trainer.py +36 -12
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/global_ptq.py +56 -1
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/global_ptq_distributed.py +116 -41
- onecomp-1.4.1/onecomp/post_process/router_fine_tuning.py +398 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/mdbf_layer.py +20 -11
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/runner.py +18 -99
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/quantization_progress.py +7 -8
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/unfuse_moe.py +91 -18
- {onecomp-1.3.4 → onecomp-1.4.1/onecomp.egg-info}/PKG-INFO +16 -6
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp.egg-info/SOURCES.txt +9 -1
- onecomp-1.4.1/scripts/check_readme_examples.py +68 -0
- onecomp-1.3.4/onecomp/cpu/export/dequantize.py +0 -166
- onecomp-1.3.4/onecomp/runner_methods/multi_gpu_quantization.py +0 -406
- {onecomp-1.3.4 → onecomp-1.4.1}/LICENSE +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/llama3-8b-gptq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/llama3-8b-jointq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/llama3-8b-lpcd-gptq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/llama3-8b-qep-gptq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/llama3-8b-various/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/qwen3-14b-gptq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/qwen3-14b-jointq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/qwen3-8b-gptq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/benchmark/qwen3-8b-jointq/quant_benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/api/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/api/jobs.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/constants.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/core/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/core/config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/core/database.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/main.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/models/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/models/job.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/schemas/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/schemas/job.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/services/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/services/huggingface.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/services/inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/services/job_store.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/worker/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/worker/celery_app.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/app/worker/tasks.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/cpu_patch.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/start_backend.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/dashboard/backend/start_worker.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/envs/vllm/v0_24_0_rocm/src/onecomp_vllm_v0_24_0_rocm/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/envs/vllm/v0_24_0_rocm/src/onecomp_vllm_v0_24_0_rocm/patch.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/cpu_inference/example_gptq_gguf_cpu.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/cpu_inference/example_mixed_gptq_gguf_cpu.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/cpu_inference/example_serve_cpu.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_auto_run.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_autobit.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_custom_calibration.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_gptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_jointq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_lpcd_gptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_mdbf.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_qep_gptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/example_save_load.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_blockwise_global_ptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_blockwise_global_ptq_staged.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_blockwise_ptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_global_ptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_global_ptq_dbf.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_global_ptq_distributed.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_lora_gptq_vllm_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_lora_sft.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_lora_sft_knowledge.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_lora_sft_knowledge_jointq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/post_process/example_reload_post_process_resave.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/pre_process/example_llama_preprocess_rtn.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/pre_process/example_preprocess_save_load.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/vllm_inference/example_autobit_vllm_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/vllm_inference/example_dbf_vllm_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/vllm_inference/example_gptq_vllm_gptoss_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/vllm_inference/example_gptq_vllm_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/vllm_inference/example_gptq_vllm_qwen36_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/example/vllm_inference/example_jointq_vllm_inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/llamacpp_plugins/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/llamacpp_plugins/gptq/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/llamacpp_plugins/gptq/constants.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/autobit/validate_autobit.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/autobit_qep/validate_autobit.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/gptq/validate_gptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/gptq/validate_load.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/gptq/validate_vllm.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/jointq/validate_jointq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/model_validation/qep_gptq/validate_gptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/__main__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/analyzer/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/analyzer/cumulative_error.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/analyzer/quantization_error.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/analyzer/weight_outlier.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/_cache.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/c4.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/calibration_config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/calibration_data_loader.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/chunking.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/custom.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/calibration/wikitext.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/cli.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/eval/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/eval/benchmark.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/eval/inspect_gguf.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/eval/parity.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/eval/perplexity.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/blocks.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/export/rotation.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/inference.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/cpu/serve.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/__main__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/conf/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/conf/eval_config.yaml +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/base.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/adapter.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/data.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/gen_answer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/judge.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/radar_chart.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/run.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/mt_bench/show_result.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/throughput/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/throughput/adapter.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/throughput/bench.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/evals/throughput/run.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/orchestrator/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/orchestrator/aggregator.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/orchestrator/runner.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/orchestrator/server.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/orchestrator/subprocess_runner.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/run_evaluate.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/schema.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/utils/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/utils/model_utils.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/utils/ports.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/utils/resources.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/eval/utils/secrets.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/export/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/export/gguf_export.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/export/gguf_reader.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/export/gguf_writer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/log.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/_gradient_solver.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/_lpcd_config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/_lpcd_runner.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/_metric.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/_refiner.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/arch/_llama.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/arch/_llama_cf.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/lpcd/arch/_qwen3.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/model_config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_base.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/dbf_block_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/dbf_cbq_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/generic_block_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/gptq_block_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/gptq_cbq_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/helpers.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/onebit_block_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_blockwise/onebit_cbq_optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_global_ptq/losses.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/_runtime.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/blockwise_ptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/post_process/post_process_lora_sft.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/hadamard_utils.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/modeling_llama.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/modeling_qwen3.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/optimizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/prepare_rotated_model.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/preprocess_args.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/quant_models.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/rotation_utils.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/pre_process/train_rotation.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/qep/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/qep/_qep_config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/qep/_quantize_with_qep.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/qep/_quantize_with_qep_arch.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantized_model_loader.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/_quantizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/arb/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/arb/_arb.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/arb/arb_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/_autobit.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/activation_stats.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/dbf_fallback.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/ilp.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/manual.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/autobit/visualize.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/cq/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/cq/_cq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/cq/cq_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/_dbf.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/admm_extended.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/balance.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/dbf_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/dbf_layer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/dbf_original.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/fine_tune.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/dbf/middle.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/gemlite.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/gptq/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/gptq/_gptq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/gptq/config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/gptq/gptq_layer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/_jointq.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/__version__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/clip.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/error_propagation/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/error_propagation/local_search_advanced.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/error_propagation/quantize_advanced.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/error_propagation/quantizer_advanced.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/local_search.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/quantize.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/quantize_multi_gpu.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/quantizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/jointq/core/solution.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/_mdbf.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/admm.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/gradient_refine.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/initialize.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/mdbf_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/mdbf/utils.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/onebit/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/onebit/_onebit.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/onebit/onebit_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/onebit/onebit_layer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/qbb/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/qbb/_qbb.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/qbb/qbb_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/_quip.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/quant_quip.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/quip_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/utils.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/utils_had.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/quip/vector_balance.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/rtn/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/rtn/_rtn.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/rtn/quantizer.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/quantizer/rtn/rtn_impl.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/rotated_model_config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/runner_methods/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/runner_methods/chunked_quantization.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/runner_methods/jointq_error_propagation.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/accuracy.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/activation_capture.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/activation_check.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/blockwise.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/device.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/dtype.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/lora.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/model_inputs.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/mxfp4_compat.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/perplexity.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/quant_config.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp/utils/vram_estimator.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp.egg-info/dependency_links.txt +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp.egg-info/entry_points.txt +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp.egg-info/requires.txt +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/onecomp.egg-info/top_level.txt +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/pyproject.toml +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/scripts/check_copyright_header.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/scripts/check_no_japanese.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/scripts/prepare_calibration_cache.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/setup.cfg +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/dbf/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/dbf/modules/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/dbf/modules/gemlite_linear.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/dbf/modules/naive.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/dbf/vllm_plugin.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/gptq/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/gptq/constants.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/gptq/gptoss_wna16_moe.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/gptq/vllm_plugin.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/patches/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/patches/_paths.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/patches/apply_all.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/patches/gpt_oss_gptq_moe.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/patches/gpt_oss_wna16_bias.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/utils/__init__.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/utils/module.py +0 -0
- {onecomp-1.3.4 → onecomp-1.4.1}/vllm_plugins/utils/rotation.py +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: onecomp
|
|
3
|
-
Version: 1.
|
|
3
|
+
Version: 1.4.1
|
|
4
4
|
Summary: Python package for LLM compression
|
|
5
5
|
Author: Keiji Kimura
|
|
6
6
|
License: MIT License
|
|
@@ -154,10 +154,13 @@ Full documentation is available at **[https://FujitsuResearch.github.io/OneCompr
|
|
|
154
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
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.
|
|
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
|
+
- **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
|
+
- **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).
|
|
157
159
|
- **LoRA SFT Post-Process**: Fine-tune quantized models with LoRA adapters for accuracy recovery or domain-specific knowledge injection. Supports SFT loss, teacher distillation, and intermediate block alignment.
|
|
158
160
|
- **Rotation Preprocessing**: SpinQuant/OstQuant-based rotation preprocessing that reduces quantization error by learning optimal rotation matrices before quantization. Rotation/scaling matrices are absorbed into model weights, with online Hadamard hooks automatically registered at load time. Supports Llama and Qwen3 architectures.
|
|
159
161
|
- **Web Dashboard (HPC)**: A browser-based dashboard for launching quantization jobs, deploying models, and validating chat-based inference in HPC environments. See [dashboard/README.md](dashboard/README.md) for details.
|
|
160
162
|
- **GGUF Export & Hugging Face Hub Integration**: Convert models to the GGUF v3 format (F16) for llama.cpp/Ollama with a dependency-free built-in writer, generate model cards with quantization recipes and evaluation results, and push save directories to the Hugging Face Hub. Supports Llama (SentencePiece or Llama-3-style BPE) and Qwen2 (BPE) architectures, including multi-EOS stop-token mapping. See the [GGUF Export guide](https://FujitsuResearch.github.io/OneCompression/user-guide/gguf-export/).
|
|
163
|
+
- **OpenVINO Export**: Export supported OneComp GPTQ 4-bit text-generation checkpoints to OpenVINO IR using an isolated, dependency-locked OpenVINO 2026.3.1 environment and conversion example. Checkpoint metadata is normalized in a temporary copy without modifying the source. See the [OpenVINO export guide](envs/openvino/README.md).
|
|
161
164
|
- (TBD)
|
|
162
165
|
|
|
163
166
|
## 🤖 Supported Models
|
|
@@ -168,10 +171,11 @@ Other Hugging Face-compatible models may work but are currently untested.
|
|
|
168
171
|
| # | Architecture | Verified Models | Status |
|
|
169
172
|
|---|-------------|-----------------|--------|
|
|
170
173
|
| 1 | Llama | TinyLlama, Llama-2, Llama-3 | ✅ Verified |
|
|
171
|
-
| 2 | Qwen3 | Qwen3-0.6B ~ 32B | ✅ Verified |
|
|
172
|
-
| 3 | Gemma | Gemma 2, Gemma 3, Gemma 4 | ✅ Verified |
|
|
174
|
+
| 2 | Qwen3 | Qwen3-0.6B ~ 32B, Qwen3-30B-A3B | ✅ Verified |
|
|
175
|
+
| 3 | Gemma | Gemma 2, Gemma 3, Gemma 4, Gemma 4-26B-A4B | ✅ Verified |
|
|
173
176
|
| 4 | [GPT-OSS](docs/user-guide/gptoss.md) | openai/gpt-oss-20b, gpt-oss-120b | ✅ Verified |
|
|
174
177
|
| 5 | Qwen3.6 | Qwen3.6-27B, Qwen3.6-35B-A3B | ✅ Verified |
|
|
178
|
+
| 6 | Qwen3.8 | Qwen3.8-27B | ✅ Verified |
|
|
175
179
|
|
|
176
180
|
|
|
177
181
|
> **Note:** Support for additional architectures is planned. Contributions and test reports are welcome.
|
|
@@ -240,7 +244,7 @@ Then install OneComp from PyPI (see step 2 below). GPTQ quantization and Hugging
|
|
|
240
244
|
> - GPTQ (`run_gptq`): Hessian and weights are moved to CPU for the full column-wise loop (including inverse-Hessian Cholesky). If that loop stayed on MPS, `quantize()` would call `maxq.item()` once per column; each call triggers **per-column host sync** (wait for pending MPS ops, then read one scalar—not a full Hessian/weight copy every column)—often several times slower than CPU on Apple Silicon (e.g. ~4× in internal benchmarks with PyTorch 2.12). Keeping GPTQ on CPU avoids that overhead. With `mse=True`, `find_params` also calls `quantize()` in a grid loop and benefits from the same CPU placement.
|
|
241
245
|
> - QEP weight correction (`adjust_weight`, when QEP correction runs—typically `qep=True` with error propagation enabled): Per-layer work stays on MPS (e.g. `weight @ delta_hatX`, diagonal damping). Only the Cholesky solve uses CPU via `_safe_cholesky_and_solve` (one solve per layer, not per column); moving all of QEP to CPU does not materially improve speed. The subsequent GPTQ step still uses the CPU path above.
|
|
242
246
|
>
|
|
243
|
-
> DBF-based AutoBit fallback
|
|
247
|
+
> DBF-based AutoBit fallback is not supported on MPS.
|
|
244
248
|
|
|
245
249
|
#### 2. Install `onecomp`
|
|
246
250
|
|
|
@@ -250,7 +254,7 @@ Once PyTorch is installed, you can install `onecomp`:
|
|
|
250
254
|
pip install onecomp
|
|
251
255
|
```
|
|
252
256
|
|
|
253
|
-
To enable multi-GPU training
|
|
257
|
+
To enable multi-GPU training for Global PTQ (DeepSpeed), install with the `distributed` extra:
|
|
254
258
|
|
|
255
259
|
```bash
|
|
256
260
|
pip install "onecomp[distributed]"
|
|
@@ -296,7 +300,7 @@ See the **MPS device placement (GPTQ vs QEP)** note under [macOS (MPS)](#macos-m
|
|
|
296
300
|
|
|
297
301
|
Adding `--extra dev` installs development tools (black, pre-commit, pytest, pylint).
|
|
298
302
|
Adding `--extra visualize` installs matplotlib for visualization features.
|
|
299
|
-
Adding `--extra distributed` installs DeepSpeed for multi-GPU training.
|
|
303
|
+
Adding `--extra distributed` installs DeepSpeed for Global PTQ multi-GPU training.
|
|
300
304
|
Adding `--extra hydra` installs `hydra-core` for the example scripts and `model_validation/` runners that use Hydra-based configuration.
|
|
301
305
|
|
|
302
306
|
To use vLLM for serving quantized models on Linux, add `--extra vllm` together with `--extra cu130`:
|
|
@@ -409,6 +413,7 @@ See [`notebook/README.md`](./notebook/README.md) for local setup, or the
|
|
|
409
413
|
| | [example_qep_gptq.py](./example/example_qep_gptq.py) | GPTQ + QEP (error propagation) |
|
|
410
414
|
| | [example_lpcd_gptq.py](./example/example_lpcd_gptq.py) | GPTQ + QEP + LPCD quantization |
|
|
411
415
|
| | [example_jointq.py](./example/example_jointq.py) | JointQ quantization |
|
|
416
|
+
| | [example_mdbf.py](./example/example_mdbf.py) | MDBF quantization |
|
|
412
417
|
| | [example_autobit.py](./example/example_autobit.py) | AutoBit mixed-precision quantization |
|
|
413
418
|
| | [example_auto_run.py](./example/example_auto_run.py) | AutoBit with automatic VRAM estimation |
|
|
414
419
|
| Calibration | [example_custom_calibration.py](./example/example_custom_calibration.py) | Custom calibration dataset with CalibrationConfig |
|
|
@@ -421,7 +426,9 @@ See [`notebook/README.md`](./notebook/README.md) for local setup, or the
|
|
|
421
426
|
| | [example_blockwise_global_ptq_staged.py](./example/post_process/example_blockwise_global_ptq_staged.py) | Staged BlockWisePTQ → GlobalPTQ across save/load boundaries with accumulated post-process metadata |
|
|
422
427
|
| | [example_global_ptq.py](./example/post_process/example_global_ptq.py) | Global PTQ with packed buffers by default and HF-compatible safetensors output |
|
|
423
428
|
| | [example_global_ptq_dbf.py](./example/post_process/example_global_ptq_dbf.py) | Global PTQ with the DBF backend and HF-compatible safetensors output |
|
|
429
|
+
| | [example_global_ptq_mdbf.py](./example/post_process/example_global_ptq_mdbf.py) | Global PTQ with the MDBF backend and HF-compatible safetensors output |
|
|
424
430
|
| | [example_global_ptq_distributed.py](./example/post_process/example_global_ptq_distributed.py) | Multi-GPU Global PTQ with DeepSpeed / torchrun and safetensors output |
|
|
431
|
+
| | [example_router_fine_tuning.py](./example/post_process/example_router_fine_tuning.py) | Router-only next-token fine-tuning for a quantized MoE model |
|
|
425
432
|
| | [example_lora_sft.py](./example/post_process/example_lora_sft.py) | LoRA SFT post-quantization fine-tuning |
|
|
426
433
|
| | [example_lora_sft_knowledge.py](./example/post_process/example_lora_sft_knowledge.py) | LoRA SFT knowledge injection |
|
|
427
434
|
| | [example_lora_sft_knowledge_jointq.py](./example/post_process/example_lora_sft_knowledge_jointq.py) | LoRA SFT knowledge injection on a JointQ-quantized model |
|
|
@@ -432,6 +439,9 @@ See [`notebook/README.md`](./notebook/README.md) for local setup, or the
|
|
|
432
439
|
| | [example_jointq_vllm_inference.py](./example/vllm_inference/example_jointq_vllm_inference.py) | JointQ quantization and vLLM inference |
|
|
433
440
|
| | [example_autobit_vllm_inference.py](./example/vllm_inference/example_autobit_vllm_inference.py) | AutoBit quantization and vLLM inference |
|
|
434
441
|
| | [example_dbf_vllm_inference.py](./example/vllm_inference/example_dbf_vllm_inference.py) | DBF quantization and vLLM inference |
|
|
442
|
+
| CPU Inference | [example_gptq_gguf_cpu.py](./example/cpu_inference/example_gptq_gguf_cpu.py) | Export GPTQ quantized models to GGUF and run CPU inference |
|
|
443
|
+
| | [example_mixed_gptq_gguf_cpu.py](./example/cpu_inference/example_mixed_gptq_gguf_cpu.py) | Export Mixed-GPTQ quantized models to GGUF and run CPU inference |
|
|
444
|
+
| | [example_serve_cpu.py](./example/cpu_inference/example_serve_cpu.py) | Serve a GGUF model with llama.cpp on CPU |
|
|
435
445
|
|
|
436
446
|
## 🔌 vLLM Inference
|
|
437
447
|
|
|
@@ -40,10 +40,13 @@ Full documentation is available at **[https://FujitsuResearch.github.io/OneCompr
|
|
|
40
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
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.
|
|
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
|
+
- **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
|
+
- **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).
|
|
43
45
|
- **LoRA SFT Post-Process**: Fine-tune quantized models with LoRA adapters for accuracy recovery or domain-specific knowledge injection. Supports SFT loss, teacher distillation, and intermediate block alignment.
|
|
44
46
|
- **Rotation Preprocessing**: SpinQuant/OstQuant-based rotation preprocessing that reduces quantization error by learning optimal rotation matrices before quantization. Rotation/scaling matrices are absorbed into model weights, with online Hadamard hooks automatically registered at load time. Supports Llama and Qwen3 architectures.
|
|
45
47
|
- **Web Dashboard (HPC)**: A browser-based dashboard for launching quantization jobs, deploying models, and validating chat-based inference in HPC environments. See [dashboard/README.md](dashboard/README.md) for details.
|
|
46
48
|
- **GGUF Export & Hugging Face Hub Integration**: Convert models to the GGUF v3 format (F16) for llama.cpp/Ollama with a dependency-free built-in writer, generate model cards with quantization recipes and evaluation results, and push save directories to the Hugging Face Hub. Supports Llama (SentencePiece or Llama-3-style BPE) and Qwen2 (BPE) architectures, including multi-EOS stop-token mapping. See the [GGUF Export guide](https://FujitsuResearch.github.io/OneCompression/user-guide/gguf-export/).
|
|
49
|
+
- **OpenVINO Export**: Export supported OneComp GPTQ 4-bit text-generation checkpoints to OpenVINO IR using an isolated, dependency-locked OpenVINO 2026.3.1 environment and conversion example. Checkpoint metadata is normalized in a temporary copy without modifying the source. See the [OpenVINO export guide](envs/openvino/README.md).
|
|
47
50
|
- (TBD)
|
|
48
51
|
|
|
49
52
|
## 🤖 Supported Models
|
|
@@ -54,10 +57,11 @@ Other Hugging Face-compatible models may work but are currently untested.
|
|
|
54
57
|
| # | Architecture | Verified Models | Status |
|
|
55
58
|
|---|-------------|-----------------|--------|
|
|
56
59
|
| 1 | Llama | TinyLlama, Llama-2, Llama-3 | ✅ Verified |
|
|
57
|
-
| 2 | Qwen3 | Qwen3-0.6B ~ 32B | ✅ Verified |
|
|
58
|
-
| 3 | Gemma | Gemma 2, Gemma 3, Gemma 4 | ✅ Verified |
|
|
60
|
+
| 2 | Qwen3 | Qwen3-0.6B ~ 32B, Qwen3-30B-A3B | ✅ Verified |
|
|
61
|
+
| 3 | Gemma | Gemma 2, Gemma 3, Gemma 4, Gemma 4-26B-A4B | ✅ Verified |
|
|
59
62
|
| 4 | [GPT-OSS](docs/user-guide/gptoss.md) | openai/gpt-oss-20b, gpt-oss-120b | ✅ Verified |
|
|
60
63
|
| 5 | Qwen3.6 | Qwen3.6-27B, Qwen3.6-35B-A3B | ✅ Verified |
|
|
64
|
+
| 6 | Qwen3.8 | Qwen3.8-27B | ✅ Verified |
|
|
61
65
|
|
|
62
66
|
|
|
63
67
|
> **Note:** Support for additional architectures is planned. Contributions and test reports are welcome.
|
|
@@ -126,7 +130,7 @@ Then install OneComp from PyPI (see step 2 below). GPTQ quantization and Hugging
|
|
|
126
130
|
> - GPTQ (`run_gptq`): Hessian and weights are moved to CPU for the full column-wise loop (including inverse-Hessian Cholesky). If that loop stayed on MPS, `quantize()` would call `maxq.item()` once per column; each call triggers **per-column host sync** (wait for pending MPS ops, then read one scalar—not a full Hessian/weight copy every column)—often several times slower than CPU on Apple Silicon (e.g. ~4× in internal benchmarks with PyTorch 2.12). Keeping GPTQ on CPU avoids that overhead. With `mse=True`, `find_params` also calls `quantize()` in a grid loop and benefits from the same CPU placement.
|
|
127
131
|
> - QEP weight correction (`adjust_weight`, when QEP correction runs—typically `qep=True` with error propagation enabled): Per-layer work stays on MPS (e.g. `weight @ delta_hatX`, diagonal damping). Only the Cholesky solve uses CPU via `_safe_cholesky_and_solve` (one solve per layer, not per column); moving all of QEP to CPU does not materially improve speed. The subsequent GPTQ step still uses the CPU path above.
|
|
128
132
|
>
|
|
129
|
-
> DBF-based AutoBit fallback
|
|
133
|
+
> DBF-based AutoBit fallback is not supported on MPS.
|
|
130
134
|
|
|
131
135
|
#### 2. Install `onecomp`
|
|
132
136
|
|
|
@@ -136,7 +140,7 @@ Once PyTorch is installed, you can install `onecomp`:
|
|
|
136
140
|
pip install onecomp
|
|
137
141
|
```
|
|
138
142
|
|
|
139
|
-
To enable multi-GPU training
|
|
143
|
+
To enable multi-GPU training for Global PTQ (DeepSpeed), install with the `distributed` extra:
|
|
140
144
|
|
|
141
145
|
```bash
|
|
142
146
|
pip install "onecomp[distributed]"
|
|
@@ -182,7 +186,7 @@ See the **MPS device placement (GPTQ vs QEP)** note under [macOS (MPS)](#macos-m
|
|
|
182
186
|
|
|
183
187
|
Adding `--extra dev` installs development tools (black, pre-commit, pytest, pylint).
|
|
184
188
|
Adding `--extra visualize` installs matplotlib for visualization features.
|
|
185
|
-
Adding `--extra distributed` installs DeepSpeed for multi-GPU training.
|
|
189
|
+
Adding `--extra distributed` installs DeepSpeed for Global PTQ multi-GPU training.
|
|
186
190
|
Adding `--extra hydra` installs `hydra-core` for the example scripts and `model_validation/` runners that use Hydra-based configuration.
|
|
187
191
|
|
|
188
192
|
To use vLLM for serving quantized models on Linux, add `--extra vllm` together with `--extra cu130`:
|
|
@@ -295,6 +299,7 @@ See [`notebook/README.md`](./notebook/README.md) for local setup, or the
|
|
|
295
299
|
| | [example_qep_gptq.py](./example/example_qep_gptq.py) | GPTQ + QEP (error propagation) |
|
|
296
300
|
| | [example_lpcd_gptq.py](./example/example_lpcd_gptq.py) | GPTQ + QEP + LPCD quantization |
|
|
297
301
|
| | [example_jointq.py](./example/example_jointq.py) | JointQ quantization |
|
|
302
|
+
| | [example_mdbf.py](./example/example_mdbf.py) | MDBF quantization |
|
|
298
303
|
| | [example_autobit.py](./example/example_autobit.py) | AutoBit mixed-precision quantization |
|
|
299
304
|
| | [example_auto_run.py](./example/example_auto_run.py) | AutoBit with automatic VRAM estimation |
|
|
300
305
|
| Calibration | [example_custom_calibration.py](./example/example_custom_calibration.py) | Custom calibration dataset with CalibrationConfig |
|
|
@@ -307,7 +312,9 @@ See [`notebook/README.md`](./notebook/README.md) for local setup, or the
|
|
|
307
312
|
| | [example_blockwise_global_ptq_staged.py](./example/post_process/example_blockwise_global_ptq_staged.py) | Staged BlockWisePTQ → GlobalPTQ across save/load boundaries with accumulated post-process metadata |
|
|
308
313
|
| | [example_global_ptq.py](./example/post_process/example_global_ptq.py) | Global PTQ with packed buffers by default and HF-compatible safetensors output |
|
|
309
314
|
| | [example_global_ptq_dbf.py](./example/post_process/example_global_ptq_dbf.py) | Global PTQ with the DBF backend and HF-compatible safetensors output |
|
|
315
|
+
| | [example_global_ptq_mdbf.py](./example/post_process/example_global_ptq_mdbf.py) | Global PTQ with the MDBF backend and HF-compatible safetensors output |
|
|
310
316
|
| | [example_global_ptq_distributed.py](./example/post_process/example_global_ptq_distributed.py) | Multi-GPU Global PTQ with DeepSpeed / torchrun and safetensors output |
|
|
317
|
+
| | [example_router_fine_tuning.py](./example/post_process/example_router_fine_tuning.py) | Router-only next-token fine-tuning for a quantized MoE model |
|
|
311
318
|
| | [example_lora_sft.py](./example/post_process/example_lora_sft.py) | LoRA SFT post-quantization fine-tuning |
|
|
312
319
|
| | [example_lora_sft_knowledge.py](./example/post_process/example_lora_sft_knowledge.py) | LoRA SFT knowledge injection |
|
|
313
320
|
| | [example_lora_sft_knowledge_jointq.py](./example/post_process/example_lora_sft_knowledge_jointq.py) | LoRA SFT knowledge injection on a JointQ-quantized model |
|
|
@@ -318,6 +325,9 @@ See [`notebook/README.md`](./notebook/README.md) for local setup, or the
|
|
|
318
325
|
| | [example_jointq_vllm_inference.py](./example/vllm_inference/example_jointq_vllm_inference.py) | JointQ quantization and vLLM inference |
|
|
319
326
|
| | [example_autobit_vllm_inference.py](./example/vllm_inference/example_autobit_vllm_inference.py) | AutoBit quantization and vLLM inference |
|
|
320
327
|
| | [example_dbf_vllm_inference.py](./example/vllm_inference/example_dbf_vllm_inference.py) | DBF quantization and vLLM inference |
|
|
328
|
+
| CPU Inference | [example_gptq_gguf_cpu.py](./example/cpu_inference/example_gptq_gguf_cpu.py) | Export GPTQ quantized models to GGUF and run CPU inference |
|
|
329
|
+
| | [example_mixed_gptq_gguf_cpu.py](./example/cpu_inference/example_mixed_gptq_gguf_cpu.py) | Export Mixed-GPTQ quantized models to GGUF and run CPU inference |
|
|
330
|
+
| | [example_serve_cpu.py](./example/cpu_inference/example_serve_cpu.py) | Serve a GGUF model with llama.cpp on CPU |
|
|
321
331
|
|
|
322
332
|
## 🔌 vLLM Inference
|
|
323
333
|
|
|
@@ -0,0 +1,66 @@
|
|
|
1
|
+
"""Export a local OneComp GPTQ 4-bit checkpoint to OpenVINO IR.
|
|
2
|
+
|
|
3
|
+
Copyright 2025-2026 Fujitsu Ltd.
|
|
4
|
+
"""
|
|
5
|
+
|
|
6
|
+
from __future__ import annotations
|
|
7
|
+
|
|
8
|
+
from pathlib import Path
|
|
9
|
+
|
|
10
|
+
from openvino_export_utils import (
|
|
11
|
+
modules_shape,
|
|
12
|
+
read_checkpoint_quantization,
|
|
13
|
+
save_openvino_tokenizer,
|
|
14
|
+
temporary_model_path_for_openvino_export,
|
|
15
|
+
validate_ir_files,
|
|
16
|
+
)
|
|
17
|
+
|
|
18
|
+
# Replace this placeholder with the path to your local OneComp GPTQ 4-bit model.
|
|
19
|
+
MODEL_PATH = "CHANGE_TO_ONECOMP_GPTQ_MODEL_PATH"
|
|
20
|
+
|
|
21
|
+
# Directory that receives the OpenVINO IR and tokenizer files.
|
|
22
|
+
OUT_DIR = Path("ov_gptq_int4_model_from_onecomp")
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+
def main() -> None:
|
|
26
|
+
from optimum.intel.openvino import OVModelForCausalLM
|
|
27
|
+
from transformers import AutoTokenizer
|
|
28
|
+
|
|
29
|
+
model_path = Path(MODEL_PATH)
|
|
30
|
+
quantization = read_checkpoint_quantization(model_path)
|
|
31
|
+
print(
|
|
32
|
+
"[INFO] checkpoint modules_in_block_to_quantize shape: "
|
|
33
|
+
f"{modules_shape(quantization.get('modules_in_block_to_quantize'))}"
|
|
34
|
+
)
|
|
35
|
+
|
|
36
|
+
OUT_DIR.mkdir(parents=True, exist_ok=True)
|
|
37
|
+
with temporary_model_path_for_openvino_export(model_path) as prepared_model_path:
|
|
38
|
+
# For an already-GPTQ-quantized model, do not pass OVWeightQuantizationConfig
|
|
39
|
+
# here. Keep the existing GPTQ 4-bit weights when exporting to OpenVINO IR.
|
|
40
|
+
model = OVModelForCausalLM.from_pretrained(
|
|
41
|
+
prepared_model_path,
|
|
42
|
+
export=True,
|
|
43
|
+
compile=False,
|
|
44
|
+
local_files_only=True,
|
|
45
|
+
trust_remote_code=True,
|
|
46
|
+
load_in_8bit=False,
|
|
47
|
+
)
|
|
48
|
+
model.save_pretrained(OUT_DIR)
|
|
49
|
+
|
|
50
|
+
# Save the Hugging Face tokenizer, plus the OpenVINO tokenizer and
|
|
51
|
+
# detokenizer required by OpenVINO GenAI.
|
|
52
|
+
tokenizer = AutoTokenizer.from_pretrained(
|
|
53
|
+
prepared_model_path,
|
|
54
|
+
local_files_only=True,
|
|
55
|
+
trust_remote_code=True,
|
|
56
|
+
)
|
|
57
|
+
tokenizer.save_pretrained(OUT_DIR)
|
|
58
|
+
save_openvino_tokenizer(tokenizer, OUT_DIR)
|
|
59
|
+
|
|
60
|
+
parsed_models = validate_ir_files(OUT_DIR)
|
|
61
|
+
print(f"[INFO] Parsed OpenVINO IR files: {', '.join(parsed_models)}")
|
|
62
|
+
print(f"[INFO] Export completed: {OUT_DIR.resolve()}")
|
|
63
|
+
|
|
64
|
+
|
|
65
|
+
if __name__ == "__main__":
|
|
66
|
+
main()
|
|
@@ -0,0 +1,115 @@
|
|
|
1
|
+
"""Shared helpers for exporting OneComp GPTQ checkpoints to OpenVINO IR.
|
|
2
|
+
|
|
3
|
+
Copyright 2025-2026 Fujitsu Ltd.
|
|
4
|
+
"""
|
|
5
|
+
|
|
6
|
+
from __future__ import annotations
|
|
7
|
+
|
|
8
|
+
import json
|
|
9
|
+
import shutil
|
|
10
|
+
import tempfile
|
|
11
|
+
from contextlib import contextmanager
|
|
12
|
+
from pathlib import Path
|
|
13
|
+
from typing import Any, Iterator
|
|
14
|
+
|
|
15
|
+
|
|
16
|
+
def read_checkpoint_quantization(model_path: Path) -> dict[str, Any]:
|
|
17
|
+
"""Return the checkpoint quantization metadata without modifying it."""
|
|
18
|
+
config_path = model_path / "config.json"
|
|
19
|
+
if not config_path.is_file():
|
|
20
|
+
raise FileNotFoundError(f"Checkpoint config was not found: {config_path}")
|
|
21
|
+
|
|
22
|
+
with config_path.open(encoding="utf-8") as config_file:
|
|
23
|
+
config = json.load(config_file)
|
|
24
|
+
|
|
25
|
+
quantization = config.get("quantization_config")
|
|
26
|
+
if not isinstance(quantization, dict):
|
|
27
|
+
raise ValueError(f"quantization_config is missing from {config_path}")
|
|
28
|
+
return quantization
|
|
29
|
+
|
|
30
|
+
|
|
31
|
+
def modules_shape(modules: Any) -> str:
|
|
32
|
+
"""Describe modules_in_block_to_quantize as flat, nested, empty, or invalid."""
|
|
33
|
+
if not isinstance(modules, list):
|
|
34
|
+
return "invalid"
|
|
35
|
+
if not modules:
|
|
36
|
+
return "empty"
|
|
37
|
+
if all(isinstance(name, str) for name in modules):
|
|
38
|
+
return "flat"
|
|
39
|
+
if all(
|
|
40
|
+
isinstance(group, list) and group and all(isinstance(name, str) for name in group)
|
|
41
|
+
for group in modules
|
|
42
|
+
):
|
|
43
|
+
return "nested"
|
|
44
|
+
return "invalid"
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
@contextmanager
|
|
48
|
+
def temporary_model_path_for_openvino_export(model_path: Path) -> Iterator[Path]:
|
|
49
|
+
"""Yield a model path with flat GPTQ module metadata normalized.
|
|
50
|
+
|
|
51
|
+
Some OneComp checkpoints store modules_in_block_to_quantize as List[str],
|
|
52
|
+
while current Transformers and Optimum expect List[List[str]]. For that
|
|
53
|
+
shape, a temporary checkpoint copy is patched and removed after use.
|
|
54
|
+
The source checkpoint is never changed.
|
|
55
|
+
"""
|
|
56
|
+
source = model_path.expanduser().resolve()
|
|
57
|
+
if not source.is_dir():
|
|
58
|
+
raise NotADirectoryError(f"Checkpoint directory was not found: {source}")
|
|
59
|
+
|
|
60
|
+
quantization = read_checkpoint_quantization(source)
|
|
61
|
+
modules_key = "modules_in_block_to_quantize"
|
|
62
|
+
if modules_key not in quantization:
|
|
63
|
+
raise ValueError(f"{modules_key} is missing from {source / 'config.json'}")
|
|
64
|
+
|
|
65
|
+
modules = quantization[modules_key]
|
|
66
|
+
shape = modules_shape(modules)
|
|
67
|
+
if shape == "invalid":
|
|
68
|
+
raise ValueError("modules_in_block_to_quantize has an unsupported shape")
|
|
69
|
+
if shape != "flat":
|
|
70
|
+
yield source
|
|
71
|
+
return
|
|
72
|
+
|
|
73
|
+
with tempfile.TemporaryDirectory(prefix="onecomp_ov_export_compat_") as temp_root:
|
|
74
|
+
temporary_model = Path(temp_root) / source.name
|
|
75
|
+
shutil.copytree(source, temporary_model, symlinks=True)
|
|
76
|
+
|
|
77
|
+
temporary_config_path = temporary_model / "config.json"
|
|
78
|
+
with temporary_config_path.open(encoding="utf-8") as config_file:
|
|
79
|
+
temporary_config = json.load(config_file)
|
|
80
|
+
temporary_config["quantization_config"]["modules_in_block_to_quantize"] = [modules]
|
|
81
|
+
with temporary_config_path.open("w", encoding="utf-8") as config_file:
|
|
82
|
+
json.dump(temporary_config, config_file, indent=2, ensure_ascii=False)
|
|
83
|
+
config_file.write("\n")
|
|
84
|
+
|
|
85
|
+
print(
|
|
86
|
+
"[INFO] Using a temporary checkpoint copy with "
|
|
87
|
+
"modules_in_block_to_quantize normalized from flat to nested."
|
|
88
|
+
)
|
|
89
|
+
print(f"[INFO] Temporary checkpoint: {temporary_model}")
|
|
90
|
+
yield temporary_model
|
|
91
|
+
|
|
92
|
+
|
|
93
|
+
def save_openvino_tokenizer(tokenizer: Any, output_dir: Path) -> None:
|
|
94
|
+
"""Save OpenVINO tokenizer and detokenizer models for OpenVINO GenAI."""
|
|
95
|
+
import openvino as ov
|
|
96
|
+
from openvino_tokenizers import convert_tokenizer
|
|
97
|
+
|
|
98
|
+
ov_tokenizer, ov_detokenizer = convert_tokenizer(tokenizer, with_detokenizer=True)
|
|
99
|
+
ov.save_model(ov_tokenizer, output_dir / "openvino_tokenizer.xml")
|
|
100
|
+
ov.save_model(ov_detokenizer, output_dir / "openvino_detokenizer.xml")
|
|
101
|
+
|
|
102
|
+
|
|
103
|
+
def validate_ir_files(output_dir: Path) -> list[str]:
|
|
104
|
+
"""Parse every exported non-tokenizer IR with OpenVINO Core."""
|
|
105
|
+
import openvino as ov
|
|
106
|
+
|
|
107
|
+
excluded = {"openvino_tokenizer.xml", "openvino_detokenizer.xml"}
|
|
108
|
+
model_files = sorted(path for path in output_dir.glob("*.xml") if path.name not in excluded)
|
|
109
|
+
if not model_files:
|
|
110
|
+
raise FileNotFoundError(f"No model IR XML was generated in {output_dir}")
|
|
111
|
+
|
|
112
|
+
core = ov.Core()
|
|
113
|
+
for model_file in model_files:
|
|
114
|
+
core.read_model(model_file)
|
|
115
|
+
return [path.name for path in model_files]
|
|
@@ -0,0 +1,48 @@
|
|
|
1
|
+
"""Example: MDBF quantization followed by Global PTQ.
|
|
2
|
+
|
|
3
|
+
Usage:
|
|
4
|
+
python example/post_process/example_global_ptq_mdbf.py
|
|
5
|
+
|
|
6
|
+
Copyright 2025-2026 Fujitsu Ltd.
|
|
7
|
+
"""
|
|
8
|
+
|
|
9
|
+
import torch
|
|
10
|
+
|
|
11
|
+
from onecomp import MDBF, CalibrationConfig, GlobalPTQ, ModelConfig, Runner, setup_logger
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
def main():
|
|
15
|
+
setup_logger()
|
|
16
|
+
model_id = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
|
|
17
|
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
|
18
|
+
model_config = ModelConfig(model_id=model_id, device=device)
|
|
19
|
+
|
|
20
|
+
global_ptq = GlobalPTQ(
|
|
21
|
+
epochs=3,
|
|
22
|
+
dbf_lr=5e-4,
|
|
23
|
+
optimize_binary=True,
|
|
24
|
+
mdbf_ste_k=2.0,
|
|
25
|
+
calibration_config=CalibrationConfig(num_calibration_samples=32, max_length=512),
|
|
26
|
+
)
|
|
27
|
+
runner = Runner(
|
|
28
|
+
model_config=model_config,
|
|
29
|
+
quantizer=MDBF(target_bits=1.0),
|
|
30
|
+
calibration_config=CalibrationConfig(max_length=512, num_calibration_samples=128),
|
|
31
|
+
post_processes=[global_ptq],
|
|
32
|
+
)
|
|
33
|
+
runner.run()
|
|
34
|
+
|
|
35
|
+
original_ppl, _, quantized_ppl = runner.calculate_perplexity(
|
|
36
|
+
original_model=True,
|
|
37
|
+
quantized_model=True,
|
|
38
|
+
)
|
|
39
|
+
print(f"\nOriginal PPL: {original_ppl:.4f}")
|
|
40
|
+
print(f"MDBF + Global PTQ PPL: {quantized_ppl:.4f}")
|
|
41
|
+
|
|
42
|
+
save_dir = "./tinyllama-mdbf-globalptq"
|
|
43
|
+
runner.save_quantized_model(save_dir)
|
|
44
|
+
print(f"\nModel saved to {save_dir}")
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
if __name__ == "__main__":
|
|
48
|
+
main()
|
|
@@ -0,0 +1,76 @@
|
|
|
1
|
+
"""Example: MoE quantization + router-only fine-tuning.
|
|
2
|
+
|
|
3
|
+
End-to-end demonstration of the RouterFineTuning post-process workflow:
|
|
4
|
+
1. Quantize a mixture-of-experts model with GPTQ
|
|
5
|
+
2. Fine-tune only its routers with next-token prediction loss
|
|
6
|
+
3. Evaluate PPL (original vs quantized + router fine-tuning)
|
|
7
|
+
4. Save the fine-tuned model to HF-compatible safetensors
|
|
8
|
+
|
|
9
|
+
Copyright 2025-2026 Fujitsu Ltd.
|
|
10
|
+
|
|
11
|
+
Usage:
|
|
12
|
+
python example/post_process/example_router_fine_tuning.py
|
|
13
|
+
"""
|
|
14
|
+
|
|
15
|
+
from onecomp import (
|
|
16
|
+
GPTQ,
|
|
17
|
+
CalibrationConfig,
|
|
18
|
+
ModelConfig,
|
|
19
|
+
RouterFineTuning,
|
|
20
|
+
Runner,
|
|
21
|
+
setup_logger,
|
|
22
|
+
)
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+
def main():
|
|
26
|
+
setup_logger()
|
|
27
|
+
|
|
28
|
+
save_dir = "./gpt-oss-20b-mixed_gptq_router_ft"
|
|
29
|
+
|
|
30
|
+
model_config = ModelConfig(
|
|
31
|
+
model_id="openai/gpt-oss-20b",
|
|
32
|
+
)
|
|
33
|
+
quantizer = GPTQ(wbits=4, groupsize=128)
|
|
34
|
+
print(
|
|
35
|
+
f"Quantizer: {type(quantizer).__name__} (wbits={quantizer.wbits}, groupsize={quantizer.groupsize})"
|
|
36
|
+
)
|
|
37
|
+
|
|
38
|
+
router_fine_tuning = RouterFineTuning(
|
|
39
|
+
dataset_name="Salesforce/wikitext",
|
|
40
|
+
dataset_config_name="wikitext-2-raw-v1",
|
|
41
|
+
train_split="train",
|
|
42
|
+
text_column="text",
|
|
43
|
+
max_train_samples=512,
|
|
44
|
+
max_length=512,
|
|
45
|
+
epochs=1,
|
|
46
|
+
batch_size=1,
|
|
47
|
+
gradient_accumulation_steps=8,
|
|
48
|
+
lr=1e-5,
|
|
49
|
+
logging_steps=10,
|
|
50
|
+
)
|
|
51
|
+
|
|
52
|
+
runner = Runner(
|
|
53
|
+
model_config=model_config,
|
|
54
|
+
quantizer=quantizer,
|
|
55
|
+
calibration_config=CalibrationConfig(
|
|
56
|
+
max_length=512,
|
|
57
|
+
num_calibration_samples=128,
|
|
58
|
+
),
|
|
59
|
+
moe_quant_experts=True,
|
|
60
|
+
post_processes=[router_fine_tuning],
|
|
61
|
+
)
|
|
62
|
+
runner.run()
|
|
63
|
+
|
|
64
|
+
original_ppl, _, fine_tuned_ppl = runner.calculate_perplexity(
|
|
65
|
+
original_model=True,
|
|
66
|
+
quantized_model=True,
|
|
67
|
+
)
|
|
68
|
+
print(f"\nOriginal MoE PPL: {original_ppl:.4f}")
|
|
69
|
+
print(f"Quantized MoE + router FT PPL: {fine_tuned_ppl:.4f}")
|
|
70
|
+
|
|
71
|
+
runner.save_quantized_model(save_dir)
|
|
72
|
+
print(f"\nModel saved (safetensors) to {save_dir}")
|
|
73
|
+
|
|
74
|
+
|
|
75
|
+
if __name__ == "__main__":
|
|
76
|
+
main()
|
|
@@ -51,6 +51,12 @@ from onecomp.cpu.export.skeleton import (
|
|
|
51
51
|
skeleton_logical_shapes,
|
|
52
52
|
stitch_gguf,
|
|
53
53
|
)
|
|
54
|
+
from onecomp.cpu.export.transforms.qwen35 import (
|
|
55
|
+
Qwen35LinearAttentionLayout,
|
|
56
|
+
is_qwen35_model,
|
|
57
|
+
load_qwen35_layout,
|
|
58
|
+
transform_qwen35_gptq,
|
|
59
|
+
)
|
|
54
60
|
from onecomp.cpu.llama_tooling import run_llama_quantize_per_tensor
|
|
55
61
|
|
|
56
62
|
logger = getLogger(__name__)
|
|
@@ -99,7 +105,9 @@ def _direct_pack_feasible(layer: GPTQLayer) -> bool:
|
|
|
99
105
|
return False
|
|
100
106
|
|
|
101
107
|
|
|
102
|
-
def _route_layer(
|
|
108
|
+
def _route_layer(
|
|
109
|
+
layer: GPTQLayer, model_layout: Optional[Qwen35LinearAttentionLayout] = None
|
|
110
|
+
) -> Tuple[str, str, str]:
|
|
103
111
|
"""Decide (route, ggml_type_name, reason) for one unpacked GPTQ layer."""
|
|
104
112
|
route, gtype = select_gguf_route(layer.wbits, layer.sym)
|
|
105
113
|
|
|
@@ -118,10 +126,26 @@ def _route_layer(layer: GPTQLayer) -> Tuple[str, str, str]:
|
|
|
118
126
|
kq = _ACTORDER_KQUANT.get(layer.wbits, "Q4_K")
|
|
119
127
|
return ROUTE_KQUANT, kq, "direct pack infeasible (zero-point/shape): K-quant fallback"
|
|
120
128
|
|
|
129
|
+
if route == ROUTE_DIRECT and model_layout is not None:
|
|
130
|
+
try:
|
|
131
|
+
transform_qwen35_gptq(
|
|
132
|
+
layer_name=layer.name,
|
|
133
|
+
q_int=layer.q_int,
|
|
134
|
+
scales=layer.scales,
|
|
135
|
+
zeros=layer.zeros,
|
|
136
|
+
groupsize=layer.groupsize,
|
|
137
|
+
layout=model_layout,
|
|
138
|
+
)
|
|
139
|
+
except UnsupportedGPTQLayout as exc:
|
|
140
|
+
kq = _ACTORDER_KQUANT.get(layer.wbits, "Q4_K")
|
|
141
|
+
return ROUTE_KQUANT, kq, str(exc)
|
|
142
|
+
|
|
121
143
|
return route, gtype.name, ""
|
|
122
144
|
|
|
123
145
|
|
|
124
|
-
def plan_mixed_export(
|
|
146
|
+
def plan_mixed_export(
|
|
147
|
+
quantized_dir: str, model_layout: Optional[Qwen35LinearAttentionLayout] = None
|
|
148
|
+
) -> List[ModulePlan]:
|
|
125
149
|
"""Return the per-module GGUF routing plan for a (mixed_)gptq checkpoint.
|
|
126
150
|
|
|
127
151
|
Does not build a skeleton or pack anything; useful for previewing how a
|
|
@@ -129,7 +153,7 @@ def plan_mixed_export(quantized_dir: str) -> List[ModulePlan]:
|
|
|
129
153
|
"""
|
|
130
154
|
plans: List[ModulePlan] = []
|
|
131
155
|
for layer in iter_gptq_layers(quantized_dir):
|
|
132
|
-
route, gtype_name, reason = _route_layer(layer)
|
|
156
|
+
route, gtype_name, reason = _route_layer(layer, model_layout=model_layout)
|
|
133
157
|
plans.append(
|
|
134
158
|
ModulePlan(
|
|
135
159
|
name=layer.name,
|
|
@@ -178,8 +202,10 @@ def export_mixed_gptq_gguf(
|
|
|
178
202
|
Returns:
|
|
179
203
|
Summary dict with the routing breakdown and replacement counts.
|
|
180
204
|
"""
|
|
181
|
-
|
|
205
|
+
model_layout = load_qwen35_layout(quantized_dir)
|
|
206
|
+
plans = plan_mixed_export(quantized_dir, model_layout=model_layout)
|
|
182
207
|
has_kquant = any(p.route == ROUTE_KQUANT for p in plans)
|
|
208
|
+
model_layout = load_qwen35_layout(quantized_dir)
|
|
183
209
|
|
|
184
210
|
owns_workdir = work_dir is None
|
|
185
211
|
work_dir = work_dir or tempfile.mkdtemp(prefix="onecomp_mixed_gguf_")
|
|
@@ -189,7 +215,8 @@ def export_mixed_gptq_gguf(
|
|
|
189
215
|
# *dequantized* values (not the original FP weights) so llama-quantize
|
|
190
216
|
# re-quantizes the GPTQ-optimized weights. Force the dequantize path.
|
|
191
217
|
skeleton_src = None if has_kquant else original_model
|
|
192
|
-
|
|
218
|
+
no_mtp = is_qwen35_model(quantized_dir)
|
|
219
|
+
skeleton = build_skeleton_gguf(quantized_dir, skeleton_src, work_dir, no_mtp=no_mtp)
|
|
193
220
|
|
|
194
221
|
import gguf
|
|
195
222
|
|
|
@@ -209,7 +236,7 @@ def export_mixed_gptq_gguf(
|
|
|
209
236
|
if gguf_name is None:
|
|
210
237
|
unmapped[layer.name] = "no GGUF tensor-name mapping"
|
|
211
238
|
continue
|
|
212
|
-
route, gtype_name, _reason = _route_layer(layer)
|
|
239
|
+
route, gtype_name, _reason = _route_layer(layer, model_layout=model_layout)
|
|
213
240
|
|
|
214
241
|
if route == ROUTE_DIRECT:
|
|
215
242
|
q_int, scales, zeros = layer.q_int, layer.scales, layer.zeros
|
|
@@ -217,6 +244,15 @@ def export_mixed_gptq_gguf(
|
|
|
217
244
|
if n_perm is not None:
|
|
218
245
|
q_int, scales, zeros = permute_gptq_rows(q_int, scales, zeros, n_perm)
|
|
219
246
|
try:
|
|
247
|
+
if model_layout is not None:
|
|
248
|
+
q_int, scales, zeros = transform_qwen35_gptq(
|
|
249
|
+
layer_name=layer.name,
|
|
250
|
+
q_int=q_int,
|
|
251
|
+
scales=scales,
|
|
252
|
+
zeros=zeros,
|
|
253
|
+
groupsize=layer.groupsize,
|
|
254
|
+
layout=model_layout,
|
|
255
|
+
)
|
|
220
256
|
packed, qtype = pack_gptq_linear(
|
|
221
257
|
q_int=q_int.cpu().numpy(),
|
|
222
258
|
scales=scales.cpu().numpy(),
|
|
@@ -144,5 +144,7 @@ def main():
|
|
|
144
144
|
"wbits": args.wbits if args.wbits is not None else "auto",
|
|
145
145
|
"groupsize": args.groupsize,
|
|
146
146
|
}
|
|
147
|
-
card = generate_model_card(
|
|
147
|
+
card = generate_model_card(
|
|
148
|
+
args.model_id, recipe=recipe, results={}, repo_id=args.push_to_hub
|
|
149
|
+
)
|
|
148
150
|
push_to_hub(save_dir, args.push_to_hub, model_card=card)
|
|
@@ -11,7 +11,7 @@ Public API:
|
|
|
11
11
|
export_to_gguf -- single entry: routes any supported checkpoint to GGUF
|
|
12
12
|
convert_gptq_to_gguf -- direct, lossless GPTQ -> GGUF (preserves QEP codes)
|
|
13
13
|
export_via_dequantize -- fallback: dequantize -> convert -> llama-quantize
|
|
14
|
-
dequantize_to_hf -- reconstruct a dense
|
|
14
|
+
dequantize_to_hf -- reconstruct a dense HF model (GPTQ/DBF/MDBF/rotated)
|
|
15
15
|
LlamaCppModel -- CPU text generation on a GGUF model
|
|
16
16
|
inspect_gguf -- per-tensor quant types / size / effective bit-width
|
|
17
17
|
perplexity -- CPU perplexity of a GGUF model on text
|
|
@@ -1,12 +1,12 @@
|
|
|
1
|
-
"""GGUF export for OneComp
|
|
1
|
+
"""GGUF export for OneComp quantized checkpoints.
|
|
2
2
|
|
|
3
3
|
Modules:
|
|
4
4
|
blocks -- lossless packing of GPTQ codes into GGUF legacy blocks
|
|
5
5
|
checkpoint -- read an OneComp GPTQ checkpoint into ``GPTQLayer`` objects
|
|
6
|
-
dequantize -- reconstruct a dense
|
|
6
|
+
dequantize -- reconstruct a dense HF model from GPTQ/DBF/MDBF weights
|
|
7
7
|
skeleton -- build a metadata/tokenizer skeleton GGUF and stitch tensors
|
|
8
8
|
direct -- direct, lossless GPTQ -> GGUF export (preferred)
|
|
9
|
-
fallback -- dequantize -> llama-quantize export (re-quantizes
|
|
9
|
+
fallback -- dequantize -> llama-quantize export (re-quantizes)
|
|
10
10
|
rotation -- fold a rotated model's online down_proj Hadamard into weights
|
|
11
11
|
auto -- single entry point that routes by quant_method / rotation
|
|
12
12
|
|