ms-swift 4.4.1__tar.gz → 4.5.0.dev0__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.
- {ms_swift-4.4.1/ms_swift.egg-info → ms_swift-4.5.0.dev0}/PKG-INFO +6 -5
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/README.md +3 -2
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0/ms_swift.egg-info}/PKG-INFO +6 -5
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/ms_swift.egg-info/SOURCES.txt +2 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/ms_swift.egg-info/requires.txt +2 -2
- ms_swift-4.5.0.dev0/requirements/megatron.txt +3 -0
- ms_swift-4.5.0.dev0/swift/agent_template/kimi_k3.py +151 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/mapping.py +2 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/rlhf_args.py +97 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/packing.py +2 -2
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/protocol.py +1 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/transformers_engine.py +11 -2
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/vllm_engine.py +19 -11
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/__init__.py +2 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/megatron_args.py +63 -10
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/init.py +3 -5
- ms_swift-4.5.0.dev0/swift/megatron/model/utils.py +293 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/train/sft.py +4 -8
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/base.py +2 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/gkd_trainer.py +17 -30
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/grpo_trainer.py +17 -19
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/rollout_mixin.py +38 -4
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/trainer.py +14 -6
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/utils.py +1 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/convert_utils.py +2 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/megatron_lm_utils.py +16 -4
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/utils.py +13 -2
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/constant.py +3 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/model_arch.py +7 -4
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/model_meta.py +6 -3
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/deepseek.py +47 -24
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/moonshot.py +29 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/qwen.py +106 -53
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/__init__.py +2 -2
- ms_swift-4.5.0.dev0/swift/model/npu_patch/mindspeed.py +335 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/model.py +29 -53
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/vllm_ascend.py +40 -4
- ms_swift-4.5.0.dev0/swift/model/npu_patch/vllm_ascend_lora.py +198 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/vllm_ascend_moe.py +128 -62
- ms_swift-4.5.0.dev0/swift/model/npu_patcher.py +8 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/infer/rollout.py +1 -4
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/sampling/vanilla_sampler.py +15 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/train/rlhf.py +8 -2
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/train/tuner.py +7 -3
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/utils.py +9 -7
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rl_core/advantage.py +133 -2
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/args_mixin.py +23 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/grpo_trainer.py +88 -14
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/rollout_mixin.py +70 -31
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/utils.py +17 -12
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/sequence_parallel.py +65 -7
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/base.py +5 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/constant.py +2 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/deepseek.py +140 -14
- ms_swift-4.5.0.dev0/swift/template/templates/moonshot.py +439 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/qwen.py +11 -14
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/vision_utils.py +1 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/torch_utils.py +1 -1
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/version.py +2 -2
- ms_swift-4.4.1/requirements/megatron.txt +0 -3
- ms_swift-4.4.1/swift/megatron/model/utils.py +0 -87
- ms_swift-4.4.1/swift/model/chunk_gated_delta_rule.py +0 -360
- ms_swift-4.4.1/swift/model/npu_patch/mindspeed.py +0 -52
- ms_swift-4.4.1/swift/model/npu_patcher.py +0 -8
- ms_swift-4.4.1/swift/template/templates/moonshot.py +0 -198
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/LICENSE +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/MANIFEST.in +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/ms_swift.egg-info/dependency_links.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/ms_swift.egg-info/entry_points.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/ms_swift.egg-info/not-zip-safe +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/ms_swift.egg-info/top_level.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/docs.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/eval.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/framework.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/npu.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/ray.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/swanlab.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/requirements/tests.txt +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/setup.cfg +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/setup.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/deepseek_v3_1.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/deepseek_v4.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/extra.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/gemma4.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/glm4.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/hermes.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/hy_v3.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/kimi_k25.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/llama.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/minicpm5.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/minimax_m2.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/minimax_m3.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/mistral.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/qwen.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/qwen3_coder.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/react.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/seed_oss.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/toolbench.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/agent_template/youtu.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/app_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/base_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/data_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/generation_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/model_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/quant_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/base_args/template_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/deploy_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/eval_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/export_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/infer_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/merge_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/pretrain_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/sampling_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/sft_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/tuner_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/arguments/webui_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/activation_cpu_offload.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/adalora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/deepspeed_elastic.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/early_stop.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/lisa.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/callbacks/perf_log.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/_megatron/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/_megatron/export.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/_megatron/main.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/_megatron/pt.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/_megatron/rlhf.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/_megatron/sft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/app.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/deploy.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/eval.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/export.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/infer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/main.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/merge_lora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/pt.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/rlhf.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/rollout.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/sample.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/sft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/cli/web_ui.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/fsdp2.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/zero0.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/zero1.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/zero2.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/zero2_offload.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/zero3.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/config/zero3_offload.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataloader/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataloader/dispatcher.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataloader/shard.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/data/dataset_info.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/dataset/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/dataset/llm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/dataset/mllm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/dataset_meta.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/dataset_syntax.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/indexed_dataset.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/loader.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/media.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/preprocessor/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/preprocessor/core.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/preprocessor/extra.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/register.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/dataset/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/hub/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/hub/constant.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/hub/hub.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/grpo_vllm_engine.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/infer_client.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/infer_engine.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/lmdeploy_engine.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/patch.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/sglang_engine.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/infer_engine/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss/causal_lm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss/embedding.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss/reranker.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/agent.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/config/agentflan.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/config/alpha_umi.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/config/hermes.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/config/ignore_empty_think.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/config/qwen.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/config/react.json +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/other.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/loss_scale/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/export_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/megatron_base_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/pretrain_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/rlhf_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/arguments/sft_args.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/default_flow.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/print.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/swanlab.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/tensorboard.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/callbacks/wandb.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/convert.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/model/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/export/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/export/export.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/train/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/train/pretrain.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/pipelines/train/rlhf.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/batch_sampler.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/dpo_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/embedding_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/gkd_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/kto_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/reranker_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/reward_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/rlhf_mixin.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/trainers/vocab_parallel_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/parallel_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/patcher.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/megatron/utils/router_replay_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/acc.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/embedding.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/nlg.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/reranker.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/metrics/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/baai.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/baichuan.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/baidu.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/bert.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/codefuse.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/gemma.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/glm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/internlm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/llama.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/llava.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/llm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/mamba.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/microsoft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/minicpm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/minimax.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/mistral.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/mllm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/mplug.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/openbuddy.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/seed.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/skywork.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/stepfun.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/telechat.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/tencent.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/valley.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/models/yi.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/env.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/fsdp.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/megatron_checkpoint.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/npu_patch/vllm_ascend_memory.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/patcher.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/register.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/model/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/galore/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/galore/adafactor.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/galore/adamw.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/galore/adamw8bit.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/galore/galore_projector.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/galore/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/lorap.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/multimodal.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/muon.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/optimizers/muonclip.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/app/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/app/app.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/app/build_ui.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/app/locale.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/eval/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/eval/eval.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/eval/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/export/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/export/cached_dataset.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/export/export.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/export/merge_lora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/export/ollama.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/export/quant.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/infer/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/infer/deploy.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/infer/infer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/infer/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/sampling/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/sampling/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/sampling/distill_sampler.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/sampling/sampling.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/sampling/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/train/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/train/kto.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/train/pretrain.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/pipelines/train/sft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/base_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/checkpoint_engine/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/checkpoint_engine/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/checkpoint_engine/hccl.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/checkpoint_engine/manager.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/checkpoint_engine/mixin.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/checkpoint_engine/nccl.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/driver_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/gkd_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/grpo_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/loss/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/loss/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/loss/gkd.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/loss/grpo.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/megatron_worker.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/model_worker.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/pipeline.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/resource_pool.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/rollout/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/rollout/adapter.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/rollout/ray_vllm_engine.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/rollout/replica.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/rollout/vllm_server.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/rollout/weight_transfer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray/megatron/worker_group.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray_utils/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray_utils/arguments.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray_utils/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ray_utils/resource_manager.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rewards/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rewards/orm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rewards/prm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rewards/rm_plugin.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rl_core/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rl_core/data.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rl_core/grpo_algorithm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rl_core/resample.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/arguments.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/base_rollout_mixin.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/cpo_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/dpo_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/gkd_helpers.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/gkd_loss.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/gkd_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/kto_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/orpo_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/ppo_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/reward_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/rlhf_mixin.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rlhf_trainers/vllm_client.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rollout/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rollout/agent_loop.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rollout/gym_env.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rollout/multi_turn.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/rollout/openenv_wrapper.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/ring_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/ulysses.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/zigzag_ring_attn.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/sequence_parallel/zigzag_ring_attn_npu.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/grounding.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/register.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/template_inputs.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/template_meta.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/baai.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/baidu.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/bert.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/dots.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/gemma.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/glm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/idefics3.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/internlm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/internvl.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/kwai.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/llama.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/llava.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/llm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/megrez.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/microsoft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/midashenglm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/minicpm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/minimax.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/minimind.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/mistral.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/molmo.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/mplug.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/openbuddy.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/pixtral.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/seed.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/stepfun.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/tencent.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/valley.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/templates/yi.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/template/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/arguments.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/embedding_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/mixin.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/patcher.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/reranker_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/seq2seq_trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/trainer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/trainer_factory.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/trainers/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuner_plugin/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuner_plugin/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuner_plugin/dummy.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuner_plugin/ia3.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuner_plugin/lora_llm.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuner_plugin/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/adapter.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/llamapro.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/longlora/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/longlora/llama.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/longlora/longlora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/lora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/lora_layers.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/mapping.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/neftune.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/part.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/peft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/prompt.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/reft.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/restuning.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/restuning_components.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/scetuning/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/scetuning/scetuning.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/scetuning/scetuning_components.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/side.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/tuners/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/app.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/base.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_eval/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_eval/eval.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_eval/llm_eval.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_eval/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_eval/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_export/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_export/export.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_export/llm_export.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_export/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_export/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/advanced.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/dataset.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/external_rollout.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/external_runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/grpo_advanced.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/hyper.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/llm_grpo.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/lora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/optimizer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/quantization.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/report_to.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/reward.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/rollout.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/save.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/target.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_grpo/tuner.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_infer/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_infer/generate.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_infer/llm_infer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_infer/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_infer/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/advanced.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/dataset.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/hyper.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/llm_rlhf.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/lora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/optimizer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/quantization.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/report_to.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/rlhf.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/save.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/target.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_rlhf/tuner.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_sample/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_sample/llm_sample.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_sample/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_sample/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_sample/sample.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/advanced.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/dataset.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/hyper.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/llm_train.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/lora.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/model.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/optimizer.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/quantization.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/report_to.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/runtime.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/save.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/self_cog.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/target.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/task.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/tuner.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/ui/llm_train/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/__init__.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/constants.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/env.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/hf_config.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/hub_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/import_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/io_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/logger.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/np_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/processor_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/shutdown_manager.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/tb_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/transformers_utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/swift/utils/utils.py +0 -0
- {ms_swift-4.4.1 → ms_swift-4.5.0.dev0}/tests/test_utils.py +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.1
|
|
2
2
|
Name: ms_swift
|
|
3
|
-
Version: 4.
|
|
3
|
+
Version: 4.5.0.dev0
|
|
4
4
|
Summary: Swift: Scalable lightWeight Infrastructure for Fine-Tuning
|
|
5
5
|
Home-page: https://github.com/modelscope/ms-swift
|
|
6
6
|
Author: DAMO ModelScope teams
|
|
@@ -59,8 +59,8 @@ Requires-Dist: trl<1.0,>=0.15
|
|
|
59
59
|
Requires-Dist: uvicorn
|
|
60
60
|
Requires-Dist: zstandard
|
|
61
61
|
Provides-Extra: megatron
|
|
62
|
-
Requires-Dist: mcore-bridge>=1.
|
|
63
|
-
Requires-Dist: megatron-core>=0.
|
|
62
|
+
Requires-Dist: mcore-bridge>=1.5.0; extra == "megatron"
|
|
63
|
+
Requires-Dist: megatron-core>=0.16; extra == "megatron"
|
|
64
64
|
Requires-Dist: peft>=0.15; extra == "megatron"
|
|
65
65
|
Provides-Extra: eval
|
|
66
66
|
Requires-Dist: evalscope>=1.0.0; extra == "eval"
|
|
@@ -169,7 +169,7 @@ You can contact us and communicate with us by adding our group:
|
|
|
169
169
|
|
|
170
170
|
|
|
171
171
|
## 📝 Introduction
|
|
172
|
-
🍲 **ms-swift** is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 400+ multimodal large models. Large models include: Qwen3, Qwen3.5, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.
|
|
172
|
+
🍲 **ms-swift** is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 400+ multimodal large models. Large models include: Qwen3, Qwen3.5, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Kimi-K3, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.
|
|
173
173
|
|
|
174
174
|
🍔 In addition, ms-swift integrates the latest training technologies, including Megatron parallelism techniques such as TP, PP, CP, EP to accelerate training, as well as numerous GRPO algorithm family reinforcement learning algorithms including: GRPO, DAPO, GSPO, SAPO, CISPO, RLOO, Reinforce++, etc. to enhance model intelligence. ms-swift supports a wide range of training tasks, including preference learning algorithms such as DPO, KTO, RM, CPO, SimPO, ORPO, as well as Embedding, Reranker, and sequence classification tasks. ms-swift provides full-pipeline support for large model training, including acceleration for inference, evaluation, and deployment modules using vLLM, SGLang, and LMDeploy, as well as model quantization using GPTQ, AWQ, BNB, and FP8 technologies.
|
|
175
175
|
|
|
@@ -195,6 +195,7 @@ You can contact us and communicate with us by adding our group:
|
|
|
195
195
|
|
|
196
196
|
|
|
197
197
|
## 🎉 News
|
|
198
|
+
- 🎁 2026.07.22: Support inference for the moonshotai multimodal model [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3), including the XTML chat template, thinking channel (`reasoning_effort`) and tool calling (`--agent_template kimi_k3`).
|
|
198
199
|
- 🎁 2026.06.10: Megatron-Ray now supports GRPO and GKD training. See [docs](./docs/source_en/Instruction/Ray.md) and [examples](examples/ray).
|
|
199
200
|
- 🎁 2026.03.03: **ms-swift v4.0** major version is officially released. For release notes, please refer to [here](https://github.com/modelscope/ms-swift/releases/tag/v4.0.0). You can provide your suggestions to us in [this issue](https://github.com/modelscope/ms-swift/issues/7250). Thank you for your support.
|
|
200
201
|
- 🎁 2025.11.14: Megatron GRPO is now available! Check out the [docs](./docs/source_en/Megatron-SWIFT/GRPO.md) and [examples](examples/megatron/grpo).
|
|
@@ -438,7 +439,7 @@ Supported Training Methods:
|
|
|
438
439
|
| [PPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/ppo) | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
|
|
439
440
|
| [DPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/dpo) | ✅ | ✅ | ✅ | ✅ | ✅ | [✅](https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/dpo) |
|
|
440
441
|
| [KTO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/kto.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | [✅](https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/kto.sh) |
|
|
441
|
-
| [Reward Model](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/rm
|
|
442
|
+
| [Reward Model](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/rm) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
442
443
|
| [CPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/cpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
443
444
|
| [SimPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/simpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
444
445
|
| [ORPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/orpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
@@ -51,7 +51,7 @@ You can contact us and communicate with us by adding our group:
|
|
|
51
51
|
|
|
52
52
|
|
|
53
53
|
## 📝 Introduction
|
|
54
|
-
🍲 **ms-swift** is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 400+ multimodal large models. Large models include: Qwen3, Qwen3.5, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.
|
|
54
|
+
🍲 **ms-swift** is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 400+ multimodal large models. Large models include: Qwen3, Qwen3.5, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Kimi-K3, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.
|
|
55
55
|
|
|
56
56
|
🍔 In addition, ms-swift integrates the latest training technologies, including Megatron parallelism techniques such as TP, PP, CP, EP to accelerate training, as well as numerous GRPO algorithm family reinforcement learning algorithms including: GRPO, DAPO, GSPO, SAPO, CISPO, RLOO, Reinforce++, etc. to enhance model intelligence. ms-swift supports a wide range of training tasks, including preference learning algorithms such as DPO, KTO, RM, CPO, SimPO, ORPO, as well as Embedding, Reranker, and sequence classification tasks. ms-swift provides full-pipeline support for large model training, including acceleration for inference, evaluation, and deployment modules using vLLM, SGLang, and LMDeploy, as well as model quantization using GPTQ, AWQ, BNB, and FP8 technologies.
|
|
57
57
|
|
|
@@ -77,6 +77,7 @@ You can contact us and communicate with us by adding our group:
|
|
|
77
77
|
|
|
78
78
|
|
|
79
79
|
## 🎉 News
|
|
80
|
+
- 🎁 2026.07.22: Support inference for the moonshotai multimodal model [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3), including the XTML chat template, thinking channel (`reasoning_effort`) and tool calling (`--agent_template kimi_k3`).
|
|
80
81
|
- 🎁 2026.06.10: Megatron-Ray now supports GRPO and GKD training. See [docs](./docs/source_en/Instruction/Ray.md) and [examples](examples/ray).
|
|
81
82
|
- 🎁 2026.03.03: **ms-swift v4.0** major version is officially released. For release notes, please refer to [here](https://github.com/modelscope/ms-swift/releases/tag/v4.0.0). You can provide your suggestions to us in [this issue](https://github.com/modelscope/ms-swift/issues/7250). Thank you for your support.
|
|
82
83
|
- 🎁 2025.11.14: Megatron GRPO is now available! Check out the [docs](./docs/source_en/Megatron-SWIFT/GRPO.md) and [examples](examples/megatron/grpo).
|
|
@@ -320,7 +321,7 @@ Supported Training Methods:
|
|
|
320
321
|
| [PPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/ppo) | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
|
|
321
322
|
| [DPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/dpo) | ✅ | ✅ | ✅ | ✅ | ✅ | [✅](https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/dpo) |
|
|
322
323
|
| [KTO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/kto.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | [✅](https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/kto.sh) |
|
|
323
|
-
| [Reward Model](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/rm
|
|
324
|
+
| [Reward Model](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/rm) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
324
325
|
| [CPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/cpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
325
326
|
| [SimPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/simpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
326
327
|
| [ORPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/orpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.1
|
|
2
2
|
Name: ms_swift
|
|
3
|
-
Version: 4.
|
|
3
|
+
Version: 4.5.0.dev0
|
|
4
4
|
Summary: Swift: Scalable lightWeight Infrastructure for Fine-Tuning
|
|
5
5
|
Home-page: https://github.com/modelscope/ms-swift
|
|
6
6
|
Author: DAMO ModelScope teams
|
|
@@ -59,8 +59,8 @@ Requires-Dist: trl<1.0,>=0.15
|
|
|
59
59
|
Requires-Dist: uvicorn
|
|
60
60
|
Requires-Dist: zstandard
|
|
61
61
|
Provides-Extra: megatron
|
|
62
|
-
Requires-Dist: mcore-bridge>=1.
|
|
63
|
-
Requires-Dist: megatron-core>=0.
|
|
62
|
+
Requires-Dist: mcore-bridge>=1.5.0; extra == "megatron"
|
|
63
|
+
Requires-Dist: megatron-core>=0.16; extra == "megatron"
|
|
64
64
|
Requires-Dist: peft>=0.15; extra == "megatron"
|
|
65
65
|
Provides-Extra: eval
|
|
66
66
|
Requires-Dist: evalscope>=1.0.0; extra == "eval"
|
|
@@ -169,7 +169,7 @@ You can contact us and communicate with us by adding our group:
|
|
|
169
169
|
|
|
170
170
|
|
|
171
171
|
## 📝 Introduction
|
|
172
|
-
🍲 **ms-swift** is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 400+ multimodal large models. Large models include: Qwen3, Qwen3.5, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.
|
|
172
|
+
🍲 **ms-swift** is a large model and multimodal large model fine-tuning and deployment framework provided by the ModelScope community. It now supports training (pre-training, fine-tuning, human alignment), inference, evaluation, quantization, and deployment for 600+ text-only large models and 400+ multimodal large models. Large models include: Qwen3, Qwen3.5, InternLM3, GLM4.5, Mistral, DeepSeek-R1, Llama4, etc. Multimodal large models include: Qwen3-VL, Qwen3-Omni, Kimi-K3, Llava, InternVL3.5, MiniCPM-V-4, Ovis2.5, GLM4.5-V, DeepSeek-VL2, etc.
|
|
173
173
|
|
|
174
174
|
🍔 In addition, ms-swift integrates the latest training technologies, including Megatron parallelism techniques such as TP, PP, CP, EP to accelerate training, as well as numerous GRPO algorithm family reinforcement learning algorithms including: GRPO, DAPO, GSPO, SAPO, CISPO, RLOO, Reinforce++, etc. to enhance model intelligence. ms-swift supports a wide range of training tasks, including preference learning algorithms such as DPO, KTO, RM, CPO, SimPO, ORPO, as well as Embedding, Reranker, and sequence classification tasks. ms-swift provides full-pipeline support for large model training, including acceleration for inference, evaluation, and deployment modules using vLLM, SGLang, and LMDeploy, as well as model quantization using GPTQ, AWQ, BNB, and FP8 technologies.
|
|
175
175
|
|
|
@@ -195,6 +195,7 @@ You can contact us and communicate with us by adding our group:
|
|
|
195
195
|
|
|
196
196
|
|
|
197
197
|
## 🎉 News
|
|
198
|
+
- 🎁 2026.07.22: Support inference for the moonshotai multimodal model [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3), including the XTML chat template, thinking channel (`reasoning_effort`) and tool calling (`--agent_template kimi_k3`).
|
|
198
199
|
- 🎁 2026.06.10: Megatron-Ray now supports GRPO and GKD training. See [docs](./docs/source_en/Instruction/Ray.md) and [examples](examples/ray).
|
|
199
200
|
- 🎁 2026.03.03: **ms-swift v4.0** major version is officially released. For release notes, please refer to [here](https://github.com/modelscope/ms-swift/releases/tag/v4.0.0). You can provide your suggestions to us in [this issue](https://github.com/modelscope/ms-swift/issues/7250). Thank you for your support.
|
|
200
201
|
- 🎁 2025.11.14: Megatron GRPO is now available! Check out the [docs](./docs/source_en/Megatron-SWIFT/GRPO.md) and [examples](examples/megatron/grpo).
|
|
@@ -438,7 +439,7 @@ Supported Training Methods:
|
|
|
438
439
|
| [PPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/ppo) | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
|
|
439
440
|
| [DPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/dpo) | ✅ | ✅ | ✅ | ✅ | ✅ | [✅](https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/dpo) |
|
|
440
441
|
| [KTO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/kto.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | [✅](https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/kto.sh) |
|
|
441
|
-
| [Reward Model](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/rm
|
|
442
|
+
| [Reward Model](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/rm) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
442
443
|
| [CPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/cpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
443
444
|
| [SimPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/simpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
444
445
|
| [ORPO](https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/orpo.sh) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
|
@@ -30,6 +30,7 @@ swift/agent_template/glm4.py
|
|
|
30
30
|
swift/agent_template/hermes.py
|
|
31
31
|
swift/agent_template/hy_v3.py
|
|
32
32
|
swift/agent_template/kimi_k25.py
|
|
33
|
+
swift/agent_template/kimi_k3.py
|
|
33
34
|
swift/agent_template/llama.py
|
|
34
35
|
swift/agent_template/mapping.py
|
|
35
36
|
swift/agent_template/minicpm5.py
|
|
@@ -211,7 +212,6 @@ swift/metrics/nlg.py
|
|
|
211
212
|
swift/metrics/reranker.py
|
|
212
213
|
swift/metrics/utils.py
|
|
213
214
|
swift/model/__init__.py
|
|
214
|
-
swift/model/chunk_gated_delta_rule.py
|
|
215
215
|
swift/model/constant.py
|
|
216
216
|
swift/model/model_arch.py
|
|
217
217
|
swift/model/model_meta.py
|
|
@@ -257,6 +257,7 @@ swift/model/npu_patch/mindspeed.py
|
|
|
257
257
|
swift/model/npu_patch/model.py
|
|
258
258
|
swift/model/npu_patch/utils.py
|
|
259
259
|
swift/model/npu_patch/vllm_ascend.py
|
|
260
|
+
swift/model/npu_patch/vllm_ascend_lora.py
|
|
260
261
|
swift/model/npu_patch/vllm_ascend_memory.py
|
|
261
262
|
swift/model/npu_patch/vllm_ascend_moe.py
|
|
262
263
|
swift/optimizers/__init__.py
|
|
@@ -0,0 +1,151 @@
|
|
|
1
|
+
# Copyright (c) ModelScope Contributors. All rights reserved.
|
|
2
|
+
import json
|
|
3
|
+
import re
|
|
4
|
+
from typing import List, Optional, Tuple, Union
|
|
5
|
+
|
|
6
|
+
from swift.infer_engine import Function
|
|
7
|
+
from swift.template import Prompt
|
|
8
|
+
from .base import BaseAgentTemplate
|
|
9
|
+
|
|
10
|
+
|
|
11
|
+
class KimiK3AgentTemplate(BaseAgentTemplate):
|
|
12
|
+
"""Agent template for Kimi K3 models (XTML format).
|
|
13
|
+
|
|
14
|
+
Aligned with `encoding_k3.py` in the model repo:
|
|
15
|
+
- Tools are declared in a separate system message with `type="tool-declare"`,
|
|
16
|
+
carrying the deep-sorted JSONSchema in a compact ```json block.
|
|
17
|
+
- Tool calls follow the closed response channel:
|
|
18
|
+
<|close|>response<|sep|><|open|>tools<|sep|>
|
|
19
|
+
<|open|>call tool="{name}" index="{i}"<|sep|>
|
|
20
|
+
<|open|>argument key="{key}" type="{type}"<|sep|>{value}<|close|>argument<|sep|>
|
|
21
|
+
<|close|>call<|sep|><|close|>tools<|sep|>
|
|
22
|
+
- Tool response:
|
|
23
|
+
<|open|>message role="tool" tool="{name}" index="{i}"<|sep|>{content}<|close|>message<|sep|><|end_of_msg|>
|
|
24
|
+
"""
|
|
25
|
+
|
|
26
|
+
@staticmethod
|
|
27
|
+
def _deep_sort(obj):
|
|
28
|
+
if isinstance(obj, dict):
|
|
29
|
+
return {k: KimiK3AgentTemplate._deep_sort(v) for k, v in sorted(obj.items())}
|
|
30
|
+
if isinstance(obj, list):
|
|
31
|
+
return [KimiK3AgentTemplate._deep_sort(item) for item in obj]
|
|
32
|
+
return obj
|
|
33
|
+
|
|
34
|
+
@staticmethod
|
|
35
|
+
def _escape_attr(value) -> str:
|
|
36
|
+
return str(value).replace('&', '&').replace('"', '"')
|
|
37
|
+
|
|
38
|
+
@staticmethod
|
|
39
|
+
def _unescape_attr(value: str) -> str:
|
|
40
|
+
return value.replace('"', '"').replace('&', '&')
|
|
41
|
+
|
|
42
|
+
@staticmethod
|
|
43
|
+
def _xtml_type(value) -> str:
|
|
44
|
+
if isinstance(value, bool):
|
|
45
|
+
return 'boolean'
|
|
46
|
+
if value is None:
|
|
47
|
+
return 'null'
|
|
48
|
+
if isinstance(value, (int, float)):
|
|
49
|
+
return 'number'
|
|
50
|
+
if isinstance(value, str):
|
|
51
|
+
return 'string'
|
|
52
|
+
if isinstance(value, dict):
|
|
53
|
+
return 'object'
|
|
54
|
+
return 'array'
|
|
55
|
+
|
|
56
|
+
@staticmethod
|
|
57
|
+
def _xtml_value(value) -> str:
|
|
58
|
+
if isinstance(value, str):
|
|
59
|
+
return value
|
|
60
|
+
return json.dumps(value, ensure_ascii=False)
|
|
61
|
+
|
|
62
|
+
def get_toolcall(self, response: str) -> List[Function]:
|
|
63
|
+
call_pattern = r'<\|open\|>call tool="(.*?)"(?: index="\d+")?<\|sep\|>(.*?)<\|close\|>call<\|sep\|>'
|
|
64
|
+
arg_pattern = r'<\|open\|>argument key="(.*?)" type="(.*?)"<\|sep\|>(.*?)<\|close\|>argument<\|sep\|>'
|
|
65
|
+
json_pattern = r'<\|open\|>json type="object"<\|sep\|>(.*?)<\|close\|>json<\|sep\|>'
|
|
66
|
+
functions = []
|
|
67
|
+
for name, body in re.findall(call_pattern, response, re.DOTALL):
|
|
68
|
+
name = self._unescape_attr(name)
|
|
69
|
+
json_match = re.search(json_pattern, body, re.DOTALL)
|
|
70
|
+
if json_match:
|
|
71
|
+
arguments = self._parse_json(json_match.group(1))
|
|
72
|
+
if arguments is None:
|
|
73
|
+
arguments = json_match.group(1)
|
|
74
|
+
else:
|
|
75
|
+
arguments = {}
|
|
76
|
+
for key, arg_type, value in re.findall(arg_pattern, body, re.DOTALL):
|
|
77
|
+
key = self._unescape_attr(key)
|
|
78
|
+
if arg_type == 'string':
|
|
79
|
+
arguments[key] = value
|
|
80
|
+
elif arg_type == 'null':
|
|
81
|
+
arguments[key] = None
|
|
82
|
+
else:
|
|
83
|
+
parsed = self._parse_json(value)
|
|
84
|
+
arguments[key] = value if parsed is None else parsed
|
|
85
|
+
functions.append(Function(name=name, arguments=arguments))
|
|
86
|
+
if len(functions) == 0:
|
|
87
|
+
# compat react_en
|
|
88
|
+
return super().get_toolcall(response)
|
|
89
|
+
return functions
|
|
90
|
+
|
|
91
|
+
def _format_tools(self, tools: List[Union[str, dict]], system: Optional[str] = None, user_message=None) -> str:
|
|
92
|
+
tools = self._deep_sort([self.wrap_tool(tool) for tool in tools])
|
|
93
|
+
tools_json = json.dumps(tools, ensure_ascii=False, separators=(',', ':'))
|
|
94
|
+
tool_content = ('# Tools\nHere are the available tools, described in JSONSchema.\n\n'
|
|
95
|
+
f'```json\n{tools_json}\n```')
|
|
96
|
+
# Hijack the `<|open|>message {{SYSTEM}}<|close|>message<|sep|><|end_of_msg|>` system
|
|
97
|
+
# prefix so the tool-declare message is rendered before the plain system message.
|
|
98
|
+
res = f'role="system" type="tool-declare"<|sep|>{tool_content}'
|
|
99
|
+
if system:
|
|
100
|
+
res += f'<|close|>message<|sep|><|end_of_msg|><|open|>message role="system"<|sep|>{system}'
|
|
101
|
+
return res
|
|
102
|
+
|
|
103
|
+
def _format_tool_calls(self, tool_call_messages) -> str:
|
|
104
|
+
calls = []
|
|
105
|
+
for index, message in enumerate(tool_call_messages, start=1):
|
|
106
|
+
tool_call = self._parse_tool_call(message['content'])
|
|
107
|
+
name, arguments = tool_call['name'], tool_call['arguments']
|
|
108
|
+
call = [f'<|open|>call tool="{self._escape_attr(name)}" index="{index}"<|sep|>']
|
|
109
|
+
if isinstance(arguments, dict):
|
|
110
|
+
for key, value in arguments.items():
|
|
111
|
+
call.append(f'<|open|>argument key="{self._escape_attr(key)}" '
|
|
112
|
+
f'type="{self._xtml_type(value)}"<|sep|>'
|
|
113
|
+
f'{self._xtml_value(value)}<|close|>argument<|sep|>')
|
|
114
|
+
else:
|
|
115
|
+
# non-object arguments fall back to a raw json block
|
|
116
|
+
json_block = arguments if isinstance(arguments, str) else json.dumps(arguments, ensure_ascii=False)
|
|
117
|
+
call.append(f'<|open|>json type="object"<|sep|>{json_block}<|close|>json<|sep|>')
|
|
118
|
+
call.append('<|close|>call<|sep|>')
|
|
119
|
+
calls.append(''.join(call))
|
|
120
|
+
return f'<|open|>tools<|sep|>{"".join(calls)}<|close|>tools<|sep|>'
|
|
121
|
+
|
|
122
|
+
def _add_tool_call_prefix(self, tool_content: str, pre_message=None) -> str:
|
|
123
|
+
# Without a preceding assistant text message (whose converted content already ends
|
|
124
|
+
# with `<|close|>response<|sep|>`), the empty response channel must be closed here.
|
|
125
|
+
if pre_message is None or pre_message['role'] != 'assistant':
|
|
126
|
+
tool_content = '<|close|>response<|sep|>' + tool_content
|
|
127
|
+
return tool_content
|
|
128
|
+
|
|
129
|
+
def _format_tool_responses(
|
|
130
|
+
self,
|
|
131
|
+
assistant_content,
|
|
132
|
+
tool_messages,
|
|
133
|
+
) -> Tuple[str, 'Prompt']:
|
|
134
|
+
contents = assistant_content if isinstance(assistant_content, list) else [assistant_content]
|
|
135
|
+
joined = ''.join(c for c in contents if isinstance(c, str))
|
|
136
|
+
with_action = self.keyword.action in joined and self.keyword.action_input in joined
|
|
137
|
+
if with_action:
|
|
138
|
+
return super()._format_tool_responses(assistant_content, tool_messages)
|
|
139
|
+
# `tool`/`index` attributes are resolved from the preceding assistant tool calls,
|
|
140
|
+
# in call order (aligned with the `tool_index` enumeration in encoding_k3.py).
|
|
141
|
+
names = re.findall(r'<\|open\|>call tool="(.*?)" index="\d+"<\|sep\|>', joined)
|
|
142
|
+
res = ['<|close|>message<|sep|><|end_of_msg|>']
|
|
143
|
+
for index, tool_message in enumerate(tool_messages, start=1):
|
|
144
|
+
if index <= len(names):
|
|
145
|
+
name = self._unescape_attr(names[index - 1])
|
|
146
|
+
else:
|
|
147
|
+
name = tool_message.get('tool') or tool_message.get('name') or ''
|
|
148
|
+
res.append(f'<|open|>message role="tool" tool="{self._escape_attr(name)}" index="{index}"<|sep|>'
|
|
149
|
+
f'{tool_message["content"]}<|close|>message<|sep|><|end_of_msg|>')
|
|
150
|
+
res.append('<|open|>message role="assistant"<|sep|>')
|
|
151
|
+
return assistant_content, res
|
|
@@ -7,6 +7,7 @@ from .glm4 import (ChatGLM4AgentTemplate, GLM4_5AgentTemplate, GLM4_7AgentTempla
|
|
|
7
7
|
GLM5_1AgentTemplate)
|
|
8
8
|
from .hermes import HermesAgentTemplate, HunyuanHermesAgentTemplate
|
|
9
9
|
from .hy_v3 import HyV3AgentTemplate, HyV3PreviewAgentTemplate
|
|
10
|
+
from .kimi_k3 import KimiK3AgentTemplate
|
|
10
11
|
from .kimi_k25 import KimiK25AgentTemplate
|
|
11
12
|
from .llama import Llama3AgentTemplate, Llama4AgentTemplate
|
|
12
13
|
from .minicpm5 import MiniCPM5AgentTemplate
|
|
@@ -57,5 +58,6 @@ agent_template_map = {
|
|
|
57
58
|
'mistral': MistralAgentTemplate,
|
|
58
59
|
'youtu': YoutuAgentTemplate,
|
|
59
60
|
'kimi_k25': KimiK25AgentTemplate,
|
|
61
|
+
'kimi_k3': KimiK3AgentTemplate,
|
|
60
62
|
'minicpm5': MiniCPM5AgentTemplate,
|
|
61
63
|
}
|
|
@@ -568,6 +568,103 @@ class RLHFArguments(TeacherModelArguments, GRPOArguments, PPOArguments, RewardMo
|
|
|
568
568
|
raise NotImplementedError('Currently, async_generate is not supported with multi-turn functionality.')
|
|
569
569
|
|
|
570
570
|
self._check_opd_rl()
|
|
571
|
+
self._check_rlsd()
|
|
572
|
+
self._check_sdar()
|
|
573
|
+
|
|
574
|
+
def _check_rlsd(self):
|
|
575
|
+
"""Validate RLSD (Self-Distilled RLVR) advantage reweighting parameters.
|
|
576
|
+
|
|
577
|
+
RLSD reweights the per-sequence GRPO advantage per token using the teacher-vs-student
|
|
578
|
+
logprob gap. It relies on an OPSD teacher forward driven by a per-sample ``teacher_prompt``
|
|
579
|
+
column (privileged context, e.g. question + reference solution / GT answer). It also needs
|
|
580
|
+
an environment reward (``reward_funcs``) because the reweight direction uses ``sign(A)``.
|
|
581
|
+
|
|
582
|
+
Two teacher modes are supported (mirroring the reference's ``freeze_teacher_model`` flag):
|
|
583
|
+
|
|
584
|
+
* **Dynamic self-distillation** (no ``--teacher_model``, default): teacher = current policy
|
|
585
|
+
evaluated on the privileged prompt (paper's same-model, two-contexts Δ_t; equivalent to
|
|
586
|
+
the reference's ``freeze_teacher_model=false``).
|
|
587
|
+
* **Frozen local teacher** (``--teacher_model <ckpt>``): a separate model scores the
|
|
588
|
+
privileged prompt under ``no_grad`` and is never in the optimizer, so it stays frozen
|
|
589
|
+
throughout training. Point ``--teacher_model`` at the initial policy checkpoint to
|
|
590
|
+
reproduce the reference's ``freeze_teacher_model=true`` (frozen θ₀ anchor).
|
|
591
|
+
"""
|
|
592
|
+
if self.advantage_reweight != 'rlsd':
|
|
593
|
+
return
|
|
594
|
+
if not (0.0 <= self.rlsd_lambda <= 1.0):
|
|
595
|
+
raise ValueError(f'rlsd_lambda must be in [0, 1], got {self.rlsd_lambda}.')
|
|
596
|
+
if self.rlsd_reweight_clip_range < 0:
|
|
597
|
+
raise ValueError(f'rlsd_reweight_clip_range (eps_w) must be >= 0, got {self.rlsd_reweight_clip_range}.')
|
|
598
|
+
if not self.reward_funcs:
|
|
599
|
+
raise ValueError('advantage_reweight=rlsd requires reward_funcs: the reweight direction uses '
|
|
600
|
+
'sign(A) of the GRPO advantage, which needs an environment reward.')
|
|
601
|
+
if self.use_liger_kernel:
|
|
602
|
+
raise ValueError('advantage_reweight=rlsd is not compatible with use_liger_kernel '
|
|
603
|
+
'(the fused loss path bypasses the per-token RLSD reweight).')
|
|
604
|
+
if self.loss_type in ['real', 'fipo']:
|
|
605
|
+
raise ValueError(f'advantage_reweight=rlsd does not support loss_type={self.loss_type!r} '
|
|
606
|
+
'(it reduces the advantage to a per-sequence scalar).')
|
|
607
|
+
if self.off_policy_sequence_mask_delta is not None:
|
|
608
|
+
raise ValueError('advantage_reweight=rlsd does not support off_policy_sequence_mask_delta.')
|
|
609
|
+
# Local frozen teacher (--teacher_model) is allowed: it drives the OPSD teacher forward on
|
|
610
|
+
# the privileged prompt via the same code path as OPD-RL, and use_rlsd bypasses OPD-RL's
|
|
611
|
+
# additive term (grpo_trainer.py:266) so there is no double-count. The API-teacher path
|
|
612
|
+
# (--teacher_model_server) scores tokens over a remote inference call whose tokenizer
|
|
613
|
+
# alignment with the student is not guaranteed, so keep it gated for RLSD's per-token reweight.
|
|
614
|
+
if self.teacher_model_server is not None:
|
|
615
|
+
raise ValueError('advantage_reweight=rlsd does not support --teacher_model_server (remote API '
|
|
616
|
+
'teacher). Use --teacher_model <local ckpt> for a frozen local teacher, or omit '
|
|
617
|
+
'both to use dynamic self-distillation (teacher = current policy).')
|
|
618
|
+
if self.beta not in (0, 0.0):
|
|
619
|
+
logger.warning(f'advantage_reweight=rlsd: the reference RLSD config disables KL (beta=0), '
|
|
620
|
+
f'but beta={self.beta}. Consider setting --beta 0.')
|
|
621
|
+
if self.teacher_model is not None:
|
|
622
|
+
logger.info(f'advantage_reweight=rlsd: frozen-teacher mode — --teacher_model={self.teacher_model!r} scores '
|
|
623
|
+
f'the `teacher_prompt` column under no_grad and stays frozen throughout training (matches the '
|
|
624
|
+
f'RLSD reference freeze_teacher_model=true when this checkpoint is the initial policy).')
|
|
625
|
+
else:
|
|
626
|
+
logger.info('advantage_reweight=rlsd: dynamic self-distillation mode — expecting a `teacher_prompt` '
|
|
627
|
+
'column in the dataset; teacher = current policy on the privileged prompt.')
|
|
628
|
+
|
|
629
|
+
def _check_sdar(self):
|
|
630
|
+
"""Validate SDAR (Self-Distilled Agentic RL) confidence-gated distillation parameters.
|
|
631
|
+
|
|
632
|
+
SDAR adds a confidence-gated teacher distillation auxiliary loss to the GRPO policy loss:
|
|
633
|
+
``loss = policy_loss + sdar_loss_coef * token-mean(sigmoid(gate_beta*Δ) * Δ)`` with
|
|
634
|
+
``Δ = logP_T - logP_S``. It reuses the OPSD teacher forward driven by a per-sample
|
|
635
|
+
``teacher_prompt`` column (privileged / skill-augmented context). Unlike RLSD it does NOT
|
|
636
|
+
modify the advantage, and the reference keeps a small KL (so KL is not forced off).
|
|
637
|
+
|
|
638
|
+
Two teacher modes are supported (mirroring the reference's self-distillation setup):
|
|
639
|
+
|
|
640
|
+
* **Dynamic self-distillation** (no ``--teacher_model``, default): teacher = current policy
|
|
641
|
+
on the privileged ``teacher_prompt`` (the reference ALFWorld/Search/WebShop setting).
|
|
642
|
+
* **Frozen local teacher** (``--teacher_model <ckpt>``): a separate model scores the
|
|
643
|
+
privileged prompt under ``no_grad`` and stays frozen throughout training.
|
|
644
|
+
"""
|
|
645
|
+
if self.sdar_loss_coef <= 0:
|
|
646
|
+
return
|
|
647
|
+
if self.sdar_gate_beta <= 0:
|
|
648
|
+
raise ValueError(f'sdar_gate_beta must be > 0, got {self.sdar_gate_beta}.')
|
|
649
|
+
if self.use_liger_kernel:
|
|
650
|
+
raise ValueError('sdar_loss_coef>0 is not compatible with use_liger_kernel (the fused Liger '
|
|
651
|
+
'loss path bypasses the per-token loss where the SDAR term is added).')
|
|
652
|
+
if self.advantage_reweight == 'rlsd':
|
|
653
|
+
raise ValueError('sdar_loss_coef>0 cannot be combined with advantage_reweight=rlsd: SDAR adds a '
|
|
654
|
+
'confidence-gated distillation loss while RLSD reweights the advantage — pick one '
|
|
655
|
+
'teacher-distillation mechanism.')
|
|
656
|
+
if self.teacher_model_server is not None:
|
|
657
|
+
raise ValueError('sdar_loss_coef>0 does not support --teacher_model_server (remote API teacher). '
|
|
658
|
+
'Use --teacher_model <local ckpt> for a frozen local teacher, or omit both to use '
|
|
659
|
+
'dynamic self-distillation (teacher = current policy on the `teacher_prompt`).')
|
|
660
|
+
if self.teacher_model is not None:
|
|
661
|
+
logger.info(f'SDAR (sdar_loss_coef={self.sdar_loss_coef}): frozen-teacher mode — '
|
|
662
|
+
f'--teacher_model={self.teacher_model!r} scores the `teacher_prompt` column under no_grad '
|
|
663
|
+
f'and stays frozen throughout training.')
|
|
664
|
+
else:
|
|
665
|
+
logger.info(f'SDAR (sdar_loss_coef={self.sdar_loss_coef}): dynamic self-distillation mode — expecting '
|
|
666
|
+
'a `teacher_prompt` column in the dataset; teacher = current policy on the privileged '
|
|
667
|
+
'prompt. Without teacher logps the SDAR loss is skipped (behaves as plain GRPO).')
|
|
571
668
|
|
|
572
669
|
def _check_opd_rl(self):
|
|
573
670
|
"""Fail-fast OPD-RL (teacher distillation on GRPO) parameter compatibility.
|
|
@@ -194,8 +194,8 @@ class IterablePackingDataset(IterableDataset):
|
|
|
194
194
|
i, data = self._out_queue.get()
|
|
195
195
|
if not data:
|
|
196
196
|
continue
|
|
197
|
-
res[i] = (data,
|
|
198
|
-
res = [
|
|
197
|
+
res[i] = data if isinstance(data, list) else [data]
|
|
198
|
+
res = [(item, len(item['input_ids'])) for group in res if group for item in group]
|
|
199
199
|
last_res += res
|
|
200
200
|
return last_res
|
|
201
201
|
|
|
@@ -191,6 +191,7 @@ class RequestConfig:
|
|
|
191
191
|
temperature: Optional[float] = None
|
|
192
192
|
top_k: Optional[int] = None
|
|
193
193
|
top_p: Optional[float] = None
|
|
194
|
+
min_p: Optional[float] = None
|
|
194
195
|
repetition_penalty: Optional[float] = None
|
|
195
196
|
num_beams: int = 1
|
|
196
197
|
stop: Optional[List[str]] = field(default_factory=list)
|
|
@@ -55,6 +55,7 @@ class TransformersEngine(InferEngine):
|
|
|
55
55
|
*,
|
|
56
56
|
template: Optional[Template] = None,
|
|
57
57
|
adapters: Optional[List[str]] = None,
|
|
58
|
+
adapter_names: Optional[List[str]] = None,
|
|
58
59
|
max_batch_size: int = 1, # 0/1: no limit
|
|
59
60
|
reranker_use_activation: bool = True,
|
|
60
61
|
# model kwargs
|
|
@@ -104,8 +105,16 @@ class TransformersEngine(InferEngine):
|
|
|
104
105
|
if template is None:
|
|
105
106
|
raise ValueError('`template` is required when `model` is a nn.Module')
|
|
106
107
|
super().__init__(template)
|
|
107
|
-
|
|
108
|
-
|
|
108
|
+
if isinstance(adapter_names, str):
|
|
109
|
+
adapter_names = [adapter_names]
|
|
110
|
+
if adapter_names and len(adapter_names) != len(self.adapters):
|
|
111
|
+
raise ValueError(f'The length of adapter_names ({len(adapter_names)}) must match the length of '
|
|
112
|
+
f'adapters ({len(self.adapters)})')
|
|
113
|
+
for i, adapter in enumerate(self.adapters):
|
|
114
|
+
adapter_name = None if adapter_names is None else adapter_names[i]
|
|
115
|
+
self._add_adapter(
|
|
116
|
+
safe_snapshot_download(adapter, use_hf=self.use_hf, hub_token=self.hub_token),
|
|
117
|
+
adapter_name=adapter_name)
|
|
109
118
|
self.engine = self.model # dummy
|
|
110
119
|
self.generation_config = getattr(self.model, 'generation_config', None)
|
|
111
120
|
self._queue = Queue()
|
|
@@ -284,18 +284,28 @@ class VllmEngine(InferEngine):
|
|
|
284
284
|
def _prepare_engine_kwargs(self, max_model_len, engine_kwargs) -> None:
|
|
285
285
|
if engine_kwargs is None:
|
|
286
286
|
engine_kwargs = {}
|
|
287
|
-
|
|
288
|
-
|
|
289
|
-
|
|
290
|
-
|
|
291
|
-
|
|
292
|
-
|
|
287
|
+
encode_task_mapping = {
|
|
288
|
+
'embedding': 'embed',
|
|
289
|
+
'seq_cls': 'classify',
|
|
290
|
+
'reranker': 'score',
|
|
291
|
+
'generative_reranker': 'score',
|
|
292
|
+
}
|
|
293
|
+
vllm_task = encode_task_mapping.get(self.task_type)
|
|
293
294
|
disable_log_stats = engine_kwargs.pop('disable_log_stats', True)
|
|
294
295
|
if self.use_async_engine:
|
|
295
296
|
engine_cls = AsyncEngineArgs
|
|
296
297
|
else:
|
|
297
298
|
engine_cls = EngineArgs
|
|
298
299
|
parameters = inspect.signature(engine_cls).parameters
|
|
300
|
+
if vllm_task:
|
|
301
|
+
if 'runner' in parameters:
|
|
302
|
+
engine_kwargs['runner'] = 'pooling'
|
|
303
|
+
elif 'task' in parameters:
|
|
304
|
+
engine_kwargs['task'] = vllm_task
|
|
305
|
+
else:
|
|
306
|
+
raise ValueError(
|
|
307
|
+
f'task_type={self.task_type} requires a vLLM version that supports `runner` or `task`. '
|
|
308
|
+
'Please upgrade vLLM.')
|
|
299
309
|
if self.use_async_engine and 'disable_log_requests' in parameters:
|
|
300
310
|
engine_kwargs['disable_log_requests'] = True
|
|
301
311
|
if 'enable_lora' in parameters and self.enable_lora:
|
|
@@ -321,10 +331,8 @@ class VllmEngine(InferEngine):
|
|
|
321
331
|
engine_kwargs[key] = value
|
|
322
332
|
else:
|
|
323
333
|
logger.warning(f'The current version of vLLM does not support `{key}`. Ignored.')
|
|
324
|
-
|
|
325
|
-
|
|
326
|
-
if val is not None:
|
|
327
|
-
engine_kwargs[key] = val
|
|
334
|
+
if self.seed is not None:
|
|
335
|
+
engine_kwargs['seed'] = self.seed
|
|
328
336
|
|
|
329
337
|
model_info = self.model_info
|
|
330
338
|
arch_mapping = {'deepseek_vl2': ['DeepseekVLV2ForCausalLM'], 'chatglm4v': ['GLM4VForCausalLM']}
|
|
@@ -507,7 +515,7 @@ class VllmEngine(InferEngine):
|
|
|
507
515
|
|
|
508
516
|
def _prepare_generation_config(self, request_config: RequestConfig) -> SamplingParams:
|
|
509
517
|
kwargs = {'max_tokens': request_config.max_tokens}
|
|
510
|
-
for key in ['temperature', 'top_k', 'top_p', 'repetition_penalty']:
|
|
518
|
+
for key in ['temperature', 'top_k', 'top_p', 'min_p', 'repetition_penalty']:
|
|
511
519
|
new_value = getattr(request_config, key)
|
|
512
520
|
if new_value is None:
|
|
513
521
|
kwargs[key] = getattr(self.generation_config, key)
|
|
@@ -4,6 +4,8 @@ try:
|
|
|
4
4
|
from transformers.utils import is_torch_npu_available
|
|
5
5
|
|
|
6
6
|
if is_torch_npu_available():
|
|
7
|
+
from swift.model.npu_patch.mindspeed import prepare_mindspeed_gdn_import
|
|
8
|
+
prepare_mindspeed_gdn_import()
|
|
7
9
|
# Enable Megatron on Ascend NPU
|
|
8
10
|
import mindspeed.megatron_adaptor # F401
|
|
9
11
|
from .init import init_megatron_env
|