├── .gitignore ├── INSTALL.md ├── LICENSE ├── README.md ├── assets ├── aimet_config.json ├── aimet_per_channel_config.json ├── sm8550_htp_basic_config.json ├── sm8550_htp_ext_config.json ├── sm8650_htp_basic_config.json └── sm8650_htp_ext_config.json ├── capp ├── Makefile ├── README.md ├── api │ └── libllmod.h ├── make │ ├── Android.mk │ └── Application.mk ├── scripts │ ├── export_bin.py │ └── tokenizer.py ├── src │ ├── buffer.cpp │ ├── buffer.h │ ├── context.cpp │ ├── context.h │ ├── errors.cpp │ ├── errors.h │ ├── libllmod.cpp │ ├── llm.cpp │ ├── llm.h │ ├── logging.cpp │ ├── logging.h │ ├── qnn_context.cpp │ ├── qnn_context.h │ ├── utils.cpp │ └── utils.h └── test │ └── simple_app.cpp ├── device ├── README.md ├── calibrate.py ├── convert_sim.py ├── debug.py ├── export.py ├── harness_aimet_ctx.py └── utils.py ├── eval ├── README.md ├── collect_mmlu_result.py ├── harness_eval.py └── simple_eval.py ├── experiments ├── smoothquant │ ├── smoothquant_modified_baseline_gemma.sh │ ├── smoothquant_modified_baseline_llama.sh │ ├── smoothquant_modified_baseline_stablelm.sh │ ├── smoothquant_original_baseline_gemma.sh │ ├── smoothquant_original_baseline_llama.sh │ └── smoothquant_original_baseline_stablelm.sh ├── w4a8 │ ├── baselines │ │ ├── lrl_gemma.sh │ │ ├── lrl_llama.sh │ │ ├── lrl_stablelm.sh │ │ ├── omniquant_modified_baseline_gemma.sh │ │ ├── omniquant_modified_baseline_llama.sh │ │ ├── omniquant_modified_baseline_stablelm.sh │ │ ├── omniquant_original_baseline_gemma.sh │ │ ├── omniquant_original_baseline_llama.sh │ │ └── omniquant_original_baseline_stablelm.sh │ └── main │ │ ├── e2e_gemma-s1024-ep60-sym.sh │ │ ├── e2e_gemma-s1024-ep60.sh │ │ ├── e2e_llama-s1024-ep60-sym.sh │ │ ├── e2e_llama-s1024-ep60.sh │ │ ├── e2e_stablelm-s1024-ep60-sym.sh │ │ └── e2e_stablelm-s1024-ep60.sh └── w8a8 │ ├── baselines │ ├── lrl_gemma.sh │ ├── lrl_llama.sh │ ├── lrl_stablelm.sh │ ├── omniquant_modified_baseline_gemma.sh │ ├── omniquant_modified_baseline_llama.sh │ ├── omniquant_modified_baseline_stablelm.sh │ ├── omniquant_original_baseline_gemma.sh │ ├── omniquant_original_baseline_llama.sh │ └── omniquant_original_baseline_stablelm.sh │ └── main │ ├── e2e_gemma-s1024-ep60.sh │ ├── e2e_llama-s1024-ep60.sh │ └── e2e_stablelm-s1024-ep60.sh ├── mobilellm ├── data │ └── datautils.py ├── model │ ├── hf_config.py │ ├── hf_model.py │ ├── ops.py │ ├── sim_layers.py │ └── sim_model.py ├── quantization │ ├── algorithm.py │ └── qmodule.py └── utils │ ├── bench.py │ ├── io.py │ ├── optim.py │ └── parallel_utils.py ├── ptq ├── generate_act_range.py ├── generate_act_scale_shift.py ├── generate_qcfg.py ├── mobilequant.py └── smoothquant.py ├── requirements.txt ├── scripts └── convert_ckpt.py └── setup.py /.gitignore: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/.gitignore -------------------------------------------------------------------------------- /INSTALL.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/INSTALL.md -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/LICENSE -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/README.md -------------------------------------------------------------------------------- /assets/aimet_config.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/assets/aimet_config.json -------------------------------------------------------------------------------- /assets/aimet_per_channel_config.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/assets/aimet_per_channel_config.json -------------------------------------------------------------------------------- /assets/sm8550_htp_basic_config.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/assets/sm8550_htp_basic_config.json -------------------------------------------------------------------------------- /assets/sm8550_htp_ext_config.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/assets/sm8550_htp_ext_config.json -------------------------------------------------------------------------------- /assets/sm8650_htp_basic_config.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/assets/sm8650_htp_basic_config.json -------------------------------------------------------------------------------- /assets/sm8650_htp_ext_config.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/assets/sm8650_htp_ext_config.json -------------------------------------------------------------------------------- /capp/Makefile: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/Makefile -------------------------------------------------------------------------------- /capp/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/README.md -------------------------------------------------------------------------------- /capp/api/libllmod.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/api/libllmod.h -------------------------------------------------------------------------------- /capp/make/Android.mk: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/make/Android.mk -------------------------------------------------------------------------------- /capp/make/Application.mk: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/make/Application.mk -------------------------------------------------------------------------------- /capp/scripts/export_bin.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/scripts/export_bin.py -------------------------------------------------------------------------------- /capp/scripts/tokenizer.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/scripts/tokenizer.py -------------------------------------------------------------------------------- /capp/src/buffer.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/buffer.cpp -------------------------------------------------------------------------------- /capp/src/buffer.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/buffer.h -------------------------------------------------------------------------------- /capp/src/context.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/context.cpp -------------------------------------------------------------------------------- /capp/src/context.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/context.h -------------------------------------------------------------------------------- /capp/src/errors.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/errors.cpp -------------------------------------------------------------------------------- /capp/src/errors.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/errors.h -------------------------------------------------------------------------------- /capp/src/libllmod.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/libllmod.cpp -------------------------------------------------------------------------------- /capp/src/llm.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/llm.cpp -------------------------------------------------------------------------------- /capp/src/llm.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/llm.h -------------------------------------------------------------------------------- /capp/src/logging.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/logging.cpp -------------------------------------------------------------------------------- /capp/src/logging.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/logging.h -------------------------------------------------------------------------------- /capp/src/qnn_context.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/qnn_context.cpp -------------------------------------------------------------------------------- /capp/src/qnn_context.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/qnn_context.h -------------------------------------------------------------------------------- /capp/src/utils.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/utils.cpp -------------------------------------------------------------------------------- /capp/src/utils.h: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/src/utils.h -------------------------------------------------------------------------------- /capp/test/simple_app.cpp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/capp/test/simple_app.cpp -------------------------------------------------------------------------------- /device/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/README.md -------------------------------------------------------------------------------- /device/calibrate.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/calibrate.py -------------------------------------------------------------------------------- /device/convert_sim.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/convert_sim.py -------------------------------------------------------------------------------- /device/debug.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/debug.py -------------------------------------------------------------------------------- /device/export.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/export.py -------------------------------------------------------------------------------- /device/harness_aimet_ctx.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/harness_aimet_ctx.py -------------------------------------------------------------------------------- /device/utils.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/device/utils.py -------------------------------------------------------------------------------- /eval/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/eval/README.md -------------------------------------------------------------------------------- /eval/collect_mmlu_result.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/eval/collect_mmlu_result.py -------------------------------------------------------------------------------- /eval/harness_eval.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/eval/harness_eval.py -------------------------------------------------------------------------------- /eval/simple_eval.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/eval/simple_eval.py -------------------------------------------------------------------------------- /experiments/smoothquant/smoothquant_modified_baseline_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/smoothquant/smoothquant_modified_baseline_gemma.sh -------------------------------------------------------------------------------- /experiments/smoothquant/smoothquant_modified_baseline_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/smoothquant/smoothquant_modified_baseline_llama.sh -------------------------------------------------------------------------------- /experiments/smoothquant/smoothquant_modified_baseline_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/smoothquant/smoothquant_modified_baseline_stablelm.sh -------------------------------------------------------------------------------- /experiments/smoothquant/smoothquant_original_baseline_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/smoothquant/smoothquant_original_baseline_gemma.sh -------------------------------------------------------------------------------- /experiments/smoothquant/smoothquant_original_baseline_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/smoothquant/smoothquant_original_baseline_llama.sh -------------------------------------------------------------------------------- /experiments/smoothquant/smoothquant_original_baseline_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/smoothquant/smoothquant_original_baseline_stablelm.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/lrl_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/lrl_gemma.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/lrl_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/lrl_llama.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/lrl_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/lrl_stablelm.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/omniquant_modified_baseline_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/omniquant_modified_baseline_gemma.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/omniquant_modified_baseline_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/omniquant_modified_baseline_llama.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/omniquant_modified_baseline_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/omniquant_modified_baseline_stablelm.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/omniquant_original_baseline_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/omniquant_original_baseline_gemma.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/omniquant_original_baseline_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/omniquant_original_baseline_llama.sh -------------------------------------------------------------------------------- /experiments/w4a8/baselines/omniquant_original_baseline_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/baselines/omniquant_original_baseline_stablelm.sh -------------------------------------------------------------------------------- /experiments/w4a8/main/e2e_gemma-s1024-ep60-sym.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/main/e2e_gemma-s1024-ep60-sym.sh -------------------------------------------------------------------------------- /experiments/w4a8/main/e2e_gemma-s1024-ep60.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/main/e2e_gemma-s1024-ep60.sh -------------------------------------------------------------------------------- /experiments/w4a8/main/e2e_llama-s1024-ep60-sym.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/main/e2e_llama-s1024-ep60-sym.sh -------------------------------------------------------------------------------- /experiments/w4a8/main/e2e_llama-s1024-ep60.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/main/e2e_llama-s1024-ep60.sh -------------------------------------------------------------------------------- /experiments/w4a8/main/e2e_stablelm-s1024-ep60-sym.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/main/e2e_stablelm-s1024-ep60-sym.sh -------------------------------------------------------------------------------- /experiments/w4a8/main/e2e_stablelm-s1024-ep60.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w4a8/main/e2e_stablelm-s1024-ep60.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/lrl_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/lrl_gemma.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/lrl_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/lrl_llama.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/lrl_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/lrl_stablelm.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/omniquant_modified_baseline_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/omniquant_modified_baseline_gemma.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/omniquant_modified_baseline_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/omniquant_modified_baseline_llama.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/omniquant_modified_baseline_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/omniquant_modified_baseline_stablelm.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/omniquant_original_baseline_gemma.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/omniquant_original_baseline_gemma.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/omniquant_original_baseline_llama.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/omniquant_original_baseline_llama.sh -------------------------------------------------------------------------------- /experiments/w8a8/baselines/omniquant_original_baseline_stablelm.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/baselines/omniquant_original_baseline_stablelm.sh -------------------------------------------------------------------------------- /experiments/w8a8/main/e2e_gemma-s1024-ep60.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/main/e2e_gemma-s1024-ep60.sh -------------------------------------------------------------------------------- /experiments/w8a8/main/e2e_llama-s1024-ep60.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/main/e2e_llama-s1024-ep60.sh -------------------------------------------------------------------------------- /experiments/w8a8/main/e2e_stablelm-s1024-ep60.sh: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/experiments/w8a8/main/e2e_stablelm-s1024-ep60.sh -------------------------------------------------------------------------------- /mobilellm/data/datautils.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/data/datautils.py -------------------------------------------------------------------------------- /mobilellm/model/hf_config.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/model/hf_config.py -------------------------------------------------------------------------------- /mobilellm/model/hf_model.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/model/hf_model.py -------------------------------------------------------------------------------- /mobilellm/model/ops.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/model/ops.py -------------------------------------------------------------------------------- /mobilellm/model/sim_layers.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/model/sim_layers.py -------------------------------------------------------------------------------- /mobilellm/model/sim_model.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/model/sim_model.py -------------------------------------------------------------------------------- /mobilellm/quantization/algorithm.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/quantization/algorithm.py -------------------------------------------------------------------------------- /mobilellm/quantization/qmodule.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/quantization/qmodule.py -------------------------------------------------------------------------------- /mobilellm/utils/bench.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/utils/bench.py -------------------------------------------------------------------------------- /mobilellm/utils/io.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/utils/io.py -------------------------------------------------------------------------------- /mobilellm/utils/optim.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/utils/optim.py -------------------------------------------------------------------------------- /mobilellm/utils/parallel_utils.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/mobilellm/utils/parallel_utils.py -------------------------------------------------------------------------------- /ptq/generate_act_range.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/ptq/generate_act_range.py -------------------------------------------------------------------------------- /ptq/generate_act_scale_shift.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/ptq/generate_act_scale_shift.py -------------------------------------------------------------------------------- /ptq/generate_qcfg.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/ptq/generate_qcfg.py -------------------------------------------------------------------------------- /ptq/mobilequant.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/ptq/mobilequant.py -------------------------------------------------------------------------------- /ptq/smoothquant.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/ptq/smoothquant.py -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/requirements.txt -------------------------------------------------------------------------------- /scripts/convert_ckpt.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/scripts/convert_ckpt.py -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/saic-fi/MobileQuant/HEAD/setup.py --------------------------------------------------------------------------------