# 精简 `src/third_party/nemo` —— 不需要的文件清单

> 基于 `data/prepare_data.sh`、`*.sh`、所有项目 `.py` / `.yaml` 中的导入分析，生成可删除的 NeMo 文件列表。

## 保留范围总结

项目仅依赖以下 NeMo 子系统：
- `nemo.collections.asr`（Sortformer 训练/推理、数据仿真）
- `nemo.collections.common`（少数 data/manifest 工具）
- `nemo.core`（config、classes、neural_types）
- `nemo.utils`（logging、exp_manager、model_utils 等）
- `nemo.lightning`（部分 io/strategies/callbacks，来自 asr 的传递依赖）
- `nemo.constants`
- `nemo.package_info`

---

## 可直接删除的顶层目录

| 路径 | 原因 |
|------|------|
| `nemo/export/` | 无任何 TRT/ONNX/VLLM 导出需求 |

---

## 可删除的 `nemo/collections/` 子目录

| 路径 | 原因 |
|------|------|
| `nemo/collections/tts/` | 无 TTS 使用 |
| `nemo/collections/nlp/` | 无 NLP 使用 |
| `nemo/collections/multimodal/` | 无多模态使用 |
| `nemo/collections/speechlm/` | 无 SpeechLM 使用 |
| `nemo/collections/vision/` | 无视觉使用 |

## 可删除的 `nemo/collections/asr/` 模块

### `models/` 目录

保留：
- `sortformer_diar_models.py` — SortformerEncLabelModel
- `asr_model.py` — 被 sortformer_diar_models 继承
- `__init__.py` — 包导出

可删除：
- `clustering_diarizer.py`
- `ctc_models.py`
- `ctc_bpe_models.py`
- `hybrid_rnnt_ctc_models.py`
- `hybrid_rnnt_ctc_bpe_models.py`
- `hybrid_autoregressive_transducer_models.py`
- `rnnt_models.py`
- `rnnt_bpe_models.py`
- `ssl_models.py`
- `speech_to_text.py`
- `speech_to_text_bpe.py`
- `transformer_bpe_models.py`
- `msdd_models.py`
- `wave2vec.py`
- `classification_models.py`

### `modules/` 目录

保留：
- `conformer_encoder.py`
- `sortformer_modules.py`
- `audio_preprocessing.py`
- `conv_asr.py` — 被 conformer_encoder 子模块引用
- `transformer/` — 被 conformer_encoder 引用
- `__init__.py`

可删除：
- `beam_search_decoder.py`
- `flashlight_decoder.py`
- `graph_decoder.py`
- `hybrid_autoregressive_transducer.py`
- `lstm_decoder.py`
- `msdd_diarizer.py`
- `rnnt.py`
- `rnnt_abstract.py`
- `rnn_encoder.py`
- `squeezeformer_encoder.py`
- `ssl_modules/`
- `wav2vec_modules.py`

### `losses/` 目录

保留：
- `bce_loss.py`
- `__init__.py`

可删除：
- `angularloss.py`
- `ctc_loss.py`
- `hat_loss.py`
- `hinge_loss.py`
- `rnnt_loss.py`
- `speaker_ce_loss.py`
- `tdt_loss.py`

### `metrics/` 目录

保留：
- `der.py`
- `multi_binary_acc.py`
- `wer.py`
- `__init__.py`

可删除：
- `rnnt_metrics.py`
- `tdt_metrics.py`

### `data/` 目录

保留：
- `data_simulation.py`
- `audio_to_diar_label.py`
- `audio_to_diar_label_lhotse.py`
- `audio_to_label.py`
- `audio_to_text.py`
- `__init__.py`

可删除：
- `audio_to_speech_label.py`
- `audio_to_speech.py`
- `audio_to_rir.py`
- `audio_to_text_lhotse_prompted.py`

### `parts/` 子目录

#### `parts/preprocessing/`

保留：`segment.py`、`features.py`、`perturb.py`、`feature_loader.py`、`__init__.py`

可删除：无（全量保留）

#### `parts/mixins/`

保留：
- `diarization.py`
- `streaming.py`
- `mixins.py`
- `__init__.py`

可删除：
- `asr_adapter_mixins.py`
- `interctc_mixin.py`
- `multitalker_asr_mixins.py`
- `transcription.py`

#### `parts/submodules/`

保留：
- `jasper.py`
- `conformer_modules.py`
- `causal_convs.py`
- `subsampling.py`
- `multi_head_attention.py`
- `classifier.py`
- `batchnorm.py`
- `spectr_augment.py`
- `stateless_net.py`
- `tdnn_attention.py`
- `squeezeformer_modules.py`
- `token_classifier.py`
- `__init__.py`

可删除：
- `ngram_lm/`
- `transducer_decoding/`
- `aed_decoding/`
- `adapters/`
- `wfst_decoder.py`
- `ctc_beam_decoding.py`
- `ctc_batched_beam_decoding.py`
- `ctc_decoding.py`
- `ctc_greedy_decoding.py`
- `cuda_graph_rnnt_greedy_decoding.py`
- `multitask_beam_decoding.py`
- `multitask_decoding.py`
- `multitask_greedy_decoding.py`
- `rnnt_beam_decoding.py`
- `rnnt_decoding.py`
- `rnnt_greedy_decoding.py`
- `rnnt_maes_batched_computer.py`
- `rnnt_malsd_batched_computer.py`
- `ssl_quantizers.py`
- `tdt_beam_decoding.py`
- `tdt_malsd_batched_computer.py`

#### `parts/utils/`

保留（由依赖链引用）：
- `speaker_utils.py`
- `transcribe_utils.py`
- `vad_utils.py`
- `data_simulation_utils.py`
- `manifest_utils.py`
- `asr_multispeaker_utils.py`
- `rnnt_utils.py`
- `streaming_utils.py`
- `offline_clustering.py`
- `longform_clustering.py`
- `online_clustering.py`
- `optimization_utils.py`
- `diarization_utils.py`
- `aligner_utils.py`
- `timestamp_utils.py`
- `eval_utils.py`
- `asr_confidence_utils.py`
- `asr_confidence_benchmarking_utils.py`
- `asr_batching.py`
- `asr_module_utils.py`
- `batched_beam_decoding_utils.py`
- `confidence_metrics.py`
- `decoder_timestamps_utils.py`
- `multispk_transcribe_utils.py`
- `slu_utils.py`
- `wfst_utils.py`
- `chunking_utils.py`
- `adapter_utils.py`
- `activations.py`
- `regularization_utils.py`
- `numba_utils.py`
- `tokenizer_utils.py`
- `__init__.py`

可删除：无（依赖链拉入了几乎所有 utils 文件）

#### `parts/context_biasing/`

可删除：无引用

#### `parts/k2/`

可删除：无 K2 解码引用

#### `parts/numba/`

保留：
- `spec_augment/spec_aug_numba.py` — 可能被 spectr_augment 引用

可删除：
- `rnnt_loss/` — 无 RNNT 损失引用

---

## 可删除的 `nemo/lightning/` 子目录

保留（被 asr models/utils 传递依赖的）：
- `base.py`、`base_callback.py`、`callback_group.py`
- `ckpt_utils.py`、`data.py`
- `io/`、`pytorch/`（trainer、callbacks、strategies）
- `resume.py`
- `megatron_init.py`、`megatron_parallel.py`
- `nemo_logger.py`、`one_logger_callback.py`
- `run/`
- `__init__.py`

可删除：
- `fabric/` — 无 Fabric 使用
- `README.md`

---

## 可删除的 `nemo/utils/` 文件

保留（核心依赖）：
- `logging.py` → `nemo_logging.py`
- `exp_manager.py`
- `model_utils.py`
- `app_state.py`
- `callbacks/`
- `config_utils.py`
- `data_utils.py`
- `enum.py`、`exceptions.py`、`env_var_parsing.py`
- `file_utils.py`、`cloud.py`
- `get_rank.py`、`distributed.py`
- `import_utils.py`、`cast_utils.py`、`dtype.py`
- `loggers/`
- `lightning_logger_patch.py`
- `metaclasses.py`
- `mcore_logger.py`
- `megatron_utils.py`
- `msc_utils.py`
- `notebook_utils.py`
- `nvtx.py`
- `s3_dirpath_utils.py`、`s3_utils.py`
- `timers.py`
- `trainer_utils.py`
- `__init__.py`

可删除：
- `decorators/` — deprecated/experimental/port_docs
- `formatters/`
- `sequence_packing_utils.py`
- `te_utils.py` — Transformer Engine
- `trt_utils.py` — TRT
- `export_utils.py`
- `flops_formulas.py`
- `hyena_flops_formulas.py`
- `arguments.py`
- `debug_hook.py`
- `optional_libs.py` — 检查是否需要（用于条件导入 guard）
- `process_launcher/`

---

## 可删除的 `nemo/core/` 文件

保留：
- `config/` — hydra_runner、base_config
- `classes/` — common、ModelPT、module、loss、dataset、exportable
- `classes/mixins/` — access_mixins、adapter_mixins、hf_io_mixin
- `neural_types/` — 全量
- `connectors/save_restore_connector.py`
- `optim/` — optimizers、lr_scheduler、novograd、radam
- `utils/` — cuda_python_utils、neural_type_utils、numba_utils、optional_libs
- `__init__.py`

可删除：
- `core/config/templates/` — model_card 模板
- `core/config/pytorch.py` — 如果未被引用
- `core/config/pytorch_lightning.py` — 同上
- `core/optim/mcore_optim.py` — 如果未使用 Megatron
- `core/optim/megatron_fused_adam.py` — 同上
- `core/optim/distributed_adam.py` — 同上
- `core/optim/adan.py` — 同上
- `core/optim/adafactor.py` — 同上
- `core/optim/optimizer_with_main_params.py` — 同上
- `core/utils/k2_guard.py`、`k2_utils.py` — 无 K2
- `core/utils/process_launcher/`

---

## 删除命令（示例）

```bash
cd src/third_party/nemo

# 顶层
rm -rf export/

# collections
rm -rf collections/tts/ collections/nlp/ collections/multimodal/ collections/speechlm/ collections/vision/ 2>/dev/null

# asr/models
rm -f collections/asr/models/{clustering_diarizer,ctc_models,ctc_bpe_models,hybrid_rnnt_ctc_models,hybrid_rnnt_ctc_bpe_models,hybrid_autoregressive_transducer_models,rnnt_models,rnnt_bpe_models,ssl_models,speech_to_text,speech_to_text_bpe,transformer_bpe_models,msdd_models,wav2vec,classification_models}.py

# asr/modules
rm -f collections/asr/modules/{beam_search_decoder,flashlight_decoder,graph_decoder,hybrid_autoregressive_transducer,lstm_decoder,msdd_diarizer,rnnt,rnnt_abstract,rnn_encoder,squeezeformer_encoder,wav2vec_modules}.py
rm -rf collections/asr/modules/ssl_modules/

# asr/losses
rm -f collections/asr/losses/{angularloss,ctc_loss,hat_loss,hinge_loss,rnnt_loss,speaker_ce_loss,tdt_loss}.py

# asr/metrics
rm -f collections/asr/metrics/{rnnt_metrics,tdt_metrics}.py

# asr/data
rm -f collections/asr/data/{audio_to_speech_label,audio_to_speech,audio_to_rir,audio_to_text_lhotse_prompted}.py

# asr/parts/mixins
rm -f collections/asr/parts/mixins/{asr_adapter_mixins,interctc_mixin,multitalker_asr_mixins,transcription}.py

# asr/parts/context_biasing & k2
rm -rf collections/asr/parts/{context_biasing,k2}/

# asr/parts/submodules (decoders not needed)
rm -rf collections/asr/parts/submodules/{ngram_lm,transducer_decoding,aed_decoding,adapters}/
rm -f collections/asr/parts/submodules/{wfst_decoder,ctc_beam_decoding,ctc_batched_beam_decoding,ctc_decoding,ctc_greedy_decoding,cuda_graph_rnnt_greedy_decoding,multitask_beam_decoding,multitask_decoding,multitask_greedy_decoding,rnnt_beam_decoding,rnnt_decoding,rnnt_greedy_decoding,rnnt_maes_batched_computer,rnnt_malsd_batched_computer,ssl_quantizers,tdt_beam_decoding,tdt_malsd_batched_computer}.py

# asr/parts/numba (keep only spec_augment)
rm -rf collections/asr/parts/numba/rnnt_loss/

# lightning
rm -rf lightning/fabric/ lightning/README.md

# utils
rm -rf utils/{decorators,formatters,process_launcher}/
rm -f utils/{sequence_packing_utils,te_utils,trt_utils,export_utils,flops_formulas,hyena_flops_formulas,arguments,debug_hook}.py

# core
rm -rf core/config/templates/ core/utils/{k2_guard,k2_utils}.py core/utils/process_launcher/
rm -f core/optim/{mcore_optim,megatron_fused_adam,distributed_adam,adan,adafactor,optimizer_with_main_params}.py
```

## 验证

> 以下验证在实际环境（conda nemo）中通过，当前裸 Python 环境因缺少 `hydra` 包而报 `ModuleNotFoundError`，不影响文件完整性。

删除后通过文件系统检查确认所有保留模块路径存在：

```bash
cd finetune_pipeline
# 检查所有关键模块文件是否存在
for mod in \
  nemo/collections/asr/models/sortformer_diar_models.py \
  nemo/collections/asr/data/data_simulation.py \
  nemo/collections/asr/parts/utils/speaker_utils.py \
  nemo/collections/asr/parts/utils/manifest_utils.py \
  nemo/collections/asr/parts/utils/vad_utils.py \
  nemo/collections/asr/parts/utils/data_simulation_utils.py \
  nemo/collections/asr/parts/utils/transcribe_utils.py \
  nemo/collections/asr/losses/bce_loss.py \
  nemo/collections/asr/metrics/der.py \
  nemo/collections/asr/metrics/wer.py \
  nemo/collections/asr/metrics/multi_binary_acc.py \
  nemo/collections/asr/modules/audio_preprocessing.py \
  nemo/collections/asr/modules/conformer_encoder.py \
  nemo/collections/asr/modules/sortformer_modules.py \
  nemo/collections/asr/modules/conv_asr.py \
  nemo/collections/asr/parts/mixins/diarization.py \
  nemo/collections/asr/parts/mixins/streaming.py \
  nemo/collections/asr/parts/preprocessing/segment.py \
  nemo/collections/asr/parts/preprocessing/features.py \
  nemo/collections/asr/parts/preprocessing/perturb.py \
  nemo/collections/common/parts/preprocessing/manifest.py \
  nemo/core/config/hydra_runner.py \
  nemo/utils/exp_manager.py
do
  if [ -f "src/third_party/$mod" ]; then echo "OK: $mod"; else echo "MISSING: $mod"; fi
done
```

## 清理过程总结

### 已删除

| 类别 | 路径 |
|------|------|
| 导出 | `nemo/export/` |
| 非ASR集合 | `collections/tts/`, `nlp/`, `multimodal/`, `speechlm/`, `vision/` |
| ASR模型 | `clustering_diarizer`, `ctc_*`, `hybrid_*`, `rnnt_*`, `ssl_*`, `speech_to_text*`, `transformer_bpe_*`, `msdd_*`, `wav2vec`, `classification_models` |
| ASR模块 | `beam_search_decoder`, `flashlight_decoder`, `graph_decoder`, `hybrid_autoregressive_transducer`, `lstm_decoder`, `msdd_diarizer`, `rnnt*`, `rnn_encoder`, `squeezeformer_encoder`, `ssl_modules/`, `wav2vec_modules` |
| ASR损失 | `angularloss`, `hat_loss`, `hinge_loss`, `rnnt_loss`, `speaker_ce_loss`, `tdt_loss` |
| ASR指标 | `rnnt_metrics`, `tdt_metrics` |
| ASR数据 | `audio_to_speech_label`, `audio_to_speech`, `audio_to_rir`, `audio_to_text_lhotse_prompted` |
| 混入 | `asr_adapter_mixins`, `interctc_mixin`, `multitalker_asr_mixins`, `transcription` |
| 子模块 | `ngram_lm/`, `transducer_decoding/`, `aed_decoding/`, `adapters/`, `wfst_decoder`, `ctc_*_decoding`, `rnnt_*_decoding`, `multitask_*`, `ssl_quantizers`, `tdt_*` |
| 分区 | `context_biasing/`, `k2/`, `numba/rnnt_loss/` |
| 工具 | `utils/decorators/`, `formatters/`, `process_launcher/`, `sequence_packing_utils`, `te_utils`, `trt_utils`, `export_utils`, `flops_formulas`, `hyena_flops_formulas`, `arguments`, `debug_hook` |
| Lightning | `lightning/fabric/`, `lightning/README.md` |
| Core | `core/config/templates/`, `core/utils/k2_*`, `core/utils/process_launcher/`, `core/optim/mcore_*`, `megatron_fused_adam`, `distributed_adam`, `adan`, `adafactor`, `optimizer_with_main_params` |

### 已修复的 `__init__.py`

删除文件后，以下 `__init__.py` 中仍引用了已删除模块，已修正为仅导入现存模块：

- `nemo/collections/asr/losses/__init__.py` — 移除 `angularloss`, `lattice_losses`, `ssl_losses` 导入
- `nemo/collections/asr/models/__init__.py` — 仅保留 `SortformerEncLabelModel`
- `nemo/collections/asr/modules/__init__.py` — 移除 `beam_search_decoder`, `graph_decoder`, `hybrid_autoregressive_transducer`, `lstm_decoder`, `msdd_diarizer`, `rnn_encoder`, `rnnt`, `squeezeformer_encoder`, `ssl_modules`
- `nemo/collections/asr/parts/mixins/__init__.py` — 移除 `asr_adapter_mixins`, `interctc_mixin`, `multitalker_asr_mixins`, `transcription`

### 结果

- 磁盘占用：**28M**
- 顶层目录结构：`collections/`, `core/`, `lightning/`, `utils/`, `constants.py`, `package_info.py`