Tyl3rDrden commited on
Commit
578b6a8
·
verified ·
1 Parent(s): 8f2b310

Upload folder using huggingface_hub

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +15 -0
  2. Dockerfile +25 -0
  3. NOTICES.txt +2 -0
  4. Parakeet-RNNT-XXL-1.1b_merged_universal_spe8.5k_1.0.nemo +3 -0
  5. README.md +60 -0
  6. checkRemoteDir.ipynb +0 -0
  7. docker-examples/Dockerfile.addpackages +16 -0
  8. docker-examples/Dockerfile.custompytorch +49 -0
  9. envCheck3.0.txt +251 -0
  10. environment.yml +281 -0
  11. lazy.py +829 -0
  12. lazyShar.py +356 -0
  13. license.txt +7 -0
  14. makeyaml.ipynb +0 -0
  15. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/cmd-args.log +1 -0
  16. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/lightning_logs.txt +0 -0
  17. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/nemo_error_log.txt +45 -0
  18. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/nemo_log_globalrank-0_localrank-0.txt +0 -0
  19. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/cmd-args.log +1 -0
  20. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/lightning_logs.txt +0 -0
  21. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/nemo_error_log.txt +45 -0
  22. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/nemo_log_globalrank-0_localrank-0.txt +0 -0
  23. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/cmd-args.log +1 -0
  24. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/lightning_logs.txt +0 -0
  25. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/nemo_error_log.txt +45 -0
  26. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/nemo_log_globalrank-0_localrank-0.txt +0 -0
  27. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/cmd-args.log +1 -0
  28. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/lightning_logs.txt +0 -0
  29. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/nemo_error_log.txt +45 -0
  30. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/nemo_log_globalrank-0_localrank-0.txt +0 -0
  31. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/cmd-args.log +1 -0
  32. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/lightning_logs.txt +0 -0
  33. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/nemo_error_log.txt +45 -0
  34. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/nemo_log_globalrank-0_localrank-0.txt +0 -0
  35. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/cmd-args.log +1 -0
  36. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/events.out.tfevents.1775305282.d12a7902a35c.540.0 +3 -0
  37. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/hparams.yaml +0 -0
  38. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/lightning_logs.txt +19 -0
  39. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/nemo_error_log.txt +48 -0
  40. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/nemo_log_globalrank-0_localrank-0.txt +0 -0
  41. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/cmd-args.log +1 -0
  42. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/events.out.tfevents.1775305542.d12a7902a35c.1101.0 +3 -0
  43. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/hparams.yaml +0 -0
  44. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/lightning_logs.txt +19 -0
  45. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/nemo_error_log.txt +49 -0
  46. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/nemo_log_globalrank-0_localrank-0.txt +0 -0
  47. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/cmd-args.log +1 -0
  48. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/lightning_logs.txt +0 -0
  49. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/nemo_error_log.txt +4 -0
  50. nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/nemo_log_globalrank-0_localrank-0.txt +106 -0
.gitattributes CHANGED
@@ -58,3 +58,18 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ Parakeet-RNNT-XXL-1.1b_merged_universal_spe8.5k_1.0.nemo filter=lfs diff=lfs merge=lfs -text
62
+ nemo_experiments/wandb/run-20260404_122123-2026-04-04_12-20-06/run-2026-04-04_12-20-06.wandb filter=lfs diff=lfs merge=lfs -text
63
+ nemo_experiments/wandb/run-20260404_122542-2026-04-04_12-24-26/run-2026-04-04_12-24-26.wandb filter=lfs diff=lfs merge=lfs -text
64
+ nemo_experiments/wandb/run-20260404_123043-2026-04-04_12-29-27/run-2026-04-04_12-29-27.wandb filter=lfs diff=lfs merge=lfs -text
65
+ nemo_experiments/wandb/run-20260404_124317-2026-04-04_12-42-01/run-2026-04-04_12-42-01.wandb filter=lfs diff=lfs merge=lfs -text
66
+ nemo_experiments/wandb/run-20260404_125341-2026-04-04_12-49-00/run-2026-04-04_12-49-00.wandb filter=lfs diff=lfs merge=lfs -text
67
+ nemo_experiments/wandb/run-20260417_122220-ay1fljdh/run-ay1fljdh.wandb filter=lfs diff=lfs merge=lfs -text
68
+ nemo_experiments/wandb/run-20260417_122553-qpeskb0g/run-qpeskb0g.wandb filter=lfs diff=lfs merge=lfs -text
69
+ nemo_experiments/wandb/run-20260417_122923-x5r87ql5/run-x5r87ql5.wandb filter=lfs diff=lfs merge=lfs -text
70
+ nemo_experiments/wandb/run-20260417_123233-2bafvzqk/run-2bafvzqk.wandb filter=lfs diff=lfs merge=lfs -text
71
+ nemo_experiments/wandb/run-20260417_123742-6chcp0l8/run-6chcp0l8.wandb filter=lfs diff=lfs merge=lfs -text
72
+ nemo_experiments/wandb/run-20260417_124319-7dpbglx5/run-7dpbglx5.wandb filter=lfs diff=lfs merge=lfs -text
73
+ nemo_experiments/wandb/run-20260417_125022-dqr4y0zs/run-dqr4y0zs.wandb filter=lfs diff=lfs merge=lfs -text
74
+ nemo_experiments/wandb/run-20260417_125445-uap4pyah/run-uap4pyah.wandb filter=lfs diff=lfs merge=lfs -text
75
+ tutorials/_static/img/thumbnails/cropped/Tutorials_Card_Template.psd filter=lfs diff=lfs merge=lfs -text
Dockerfile ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Use the specific NVIDIA NeMo release
2
+ FROM nvcr.io/nvidia/nemo:26.02
3
+
4
+ # 1. Set the working directory
5
+ WORKDIR /workspace
6
+ COPY . .
7
+ # We point uv specifically to the NeMo virtual environment
8
+ RUN uv pip install --upgrade --python /opt/venv/bin/python huggingface_hub fsspec && \
9
+ sed -i 's/require_version_core(deps\[pkg\])/pass/g' /opt/venv/lib/python3.12/site-packages/transformers/dependency_versions_check.py
10
+ # This overwrites the broken file in the pre-installed library
11
+ COPY rnnt_models.py /opt/NeMo/nemo/collections/asr/models/rnnt_models.py
12
+ RUN LHOTSE_DIR=$(python -c 'import lhotse, os; print(os.path.dirname(lhotse.__file__))') && \
13
+ cp serialization.py $LHOTSE_DIR/serialization.py && \
14
+ cp lazy.py $LHOTSE_DIR/lazy.py && \
15
+ cp lazyShar.py $LHOTSE_DIR/shar/readers/lazy.py
16
+ # 3. Handle WANDB (Passed at runtime for security)
17
+ # We leave the variable name here so the script can find it
18
+ ENV WANDB_API_KEY="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"
19
+
20
+ # 4. Copy your local training scripts/configs
21
+
22
+ # 5. Setup execution
23
+ RUN chmod +x /workspace/runTrain.sh
24
+
25
+ CMD ["/bin/bash", "-c", "sleep infinity"]
NOTICES.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ This distribution includes open source which is archived at the following URL: https://opensource.nvidia.com/oss/teams/nvidia/nemo/26.02.01:linux-amd64/index.html
2
+ For further inquiries or assistance, contact us at oss-requests@nvidia.com
Parakeet-RNNT-XXL-1.1b_merged_universal_spe8.5k_1.0.nemo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:52332e96ef68ff8cfefd1d8d7b8c5d7b5333faa3cfac87ed4cc7b5ec3d5821c0
3
+ size 4011233560
README.md ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # NeMo Framework Docker Container
2
+
3
+ ## Repository Overview
4
+
5
+ - Megatron-Bridge (/opt/Megatron-Bridge/)
6
+ - Megatron-LM (/opt/Megatron-Bridge/3rdparty/Megatron-LM/)
7
+ - Evaluator (/opt/Evaluator/)
8
+ - Export-Deploy (/opt/Export-Deploy/)
9
+ - NeMo (/opt/NeMo/)
10
+ - Run (/opt/Run/)
11
+
12
+ ---
13
+
14
+ ## Installed Packged
15
+
16
+ Pip installed packages:
17
+
18
+ - DeepEP
19
+ - vLLM
20
+ - TRT-LLM
21
+
22
+ Execute `pip list` to see full list of installed packages via pip.
23
+
24
+ Uv virtualenv installed packages (/opt/venv/)
25
+
26
+ - TransformerEngine
27
+ - nvidia-resiliency
28
+ - TensorRT-Model-Optimizer
29
+
30
+ Execute `uv pip list` to see full list of installed packages. Packages installed in /opt/venv take precedence over pip installed packages.
31
+
32
+ See /opt/NeMo-FW/pyproject.toml for additonal uv configurations.
33
+
34
+ ---
35
+
36
+ ## Development
37
+
38
+ ### Mounting and syncing local repository into the container
39
+
40
+ Local working directories can be mounted via docker run:
41
+
42
+ ```bash
43
+ docker run -v <local-folder-path>:<container-folder-path> <container-image>
44
+ ```
45
+
46
+ #### 1 Overwrite existing directory
47
+
48
+ 1. Define the `<container-folder-path>` to correspond to the path defined in Repository Overview. (ie. /my-path/Megatron-Bridge/:/opt/Megatron-Bridge/)
49
+
50
+ 2. `cd /opt/NeMo-FW` and run `uv sync --no-cache-dir --all-groups --inexact`
51
+
52
+ 3. Local development directory is synced to run in the container
53
+
54
+ #### 2 Mount directory to a different path
55
+
56
+ 1. Modify `/opt/NeMo-FW/pyproject.toml` sections `tool.uv.sources` and `tool.uv; override-dependencies` to reflect new path. (ie. "megatron-bridge" = { path = `<container-folder-path>`, editable = true })
57
+
58
+ 2. `cd /opt/NeMo-FW` and run `uv sync --no-cache-dir --all-groups --inexact`
59
+
60
+ 3. Local development directory is synced to run in the container
checkRemoteDir.ipynb ADDED
The diff for this file is too large to render. See raw diff
 
docker-examples/Dockerfile.addpackages ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #
2
+ # This example Dockerfile illustrates a method to install
3
+ # additional packages on top of NVIDIA's PyTorch container image.
4
+ #
5
+ # To use this Dockerfile, use the `docker build` command.
6
+ # See https://docs.docker.com/engine/reference/builder/
7
+ # for more information.
8
+ #
9
+ FROM nvcr.io/nvidia/pytorch:25.11-py3
10
+
11
+ # Install my-extra-package-1 and my-extra-package-2
12
+ RUN apt-get update && apt-get install -y --no-install-recommends \
13
+ my-extra-package-1 \
14
+ my-extra-package-2 \
15
+ && \
16
+ rm -rf /var/lib/apt/lists/
docker-examples/Dockerfile.custompytorch ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #
2
+ # This example Dockerfile illustrates a method to apply
3
+ # patches to the source code in NVIDIA's PyTorch
4
+ # container image and to rebuild PyTorch. The RUN command
5
+ # included below will rebuild PyTorch in the same way as
6
+ # it was built in the original image.
7
+ #
8
+ # By applying customizations through a Dockerfile and
9
+ # `docker build` in this manner rather than modifying the
10
+ # container interactively, it will be straightforward to
11
+ # apply the same changes to later versions of the PyTorch
12
+ # container image.
13
+ #
14
+ # https://docs.docker.com/engine/reference/builder/
15
+ #
16
+ FROM nvcr.io/nvidia/pytorch:25.11-py3
17
+
18
+ # Bring in changes from outside container to /tmp
19
+ # (assumes my-pytorch-modifications.patch is in same directory as Dockerfile)
20
+ COPY my-pytorch-modifications.patch /tmp
21
+
22
+ # Change working directory to PyTorch source path
23
+ WORKDIR /opt/pytorch
24
+
25
+ # Apply modifications
26
+ RUN patch -p1 < /tmp/my-pytorch-modifications.patch
27
+
28
+ # Rebuild PyTorch
29
+ RUN cd pytorch && \
30
+ USE_CUPTI_SO=1 \
31
+ USE_KINETO=1 \
32
+ CMAKE_PREFIX_PATH="/usr/local" \
33
+ NCCL_ROOT="/usr" \
34
+ USE_SYSTEM_NCCL=1 \
35
+ USE_UCC=1 \
36
+ USE_SYSTEM_UCC=1 \
37
+ UCC_HOME="/opt/hpcx/ucc" \
38
+ # UCC_DIR is for PyTorch to find ucc-config.cmake
39
+ UCC_DIR="/opt/hpcx/ucc/lib/cmake/ucc" \
40
+ UCX_HOME="/opt/hpcx/ucx" \
41
+ UCX_DIR="/opt/hpcx/ucx/lib/cmake/ucx" \
42
+ CFLAGS='-fno-gnu-unique' \
43
+ DEFAULT_INTEL_MKL_DIR="/usr/local" \
44
+ INTEL_MKL_DIR="/usr/local" \
45
+ python setup.py install \
46
+ && python setup.py clean
47
+
48
+ # Reset default working directory
49
+ WORKDIR /workspace
envCheck3.0.txt ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ absl-py==2.4.0
2
+ accelerate==1.13.0
3
+ aiobotocore==3.3.0
4
+ aiohappyeyeballs==2.6.1
5
+ aiohttp==3.13.3
6
+ aioitertools==0.13.0
7
+ aiosignal==1.4.0
8
+ aistore==1.23.0
9
+ alembic==1.18.4
10
+ annotated-doc==0.0.4
11
+ annotated-types==0.7.0
12
+ antlr4-python3-runtime==4.9.3
13
+ anyio==4.13.0
14
+ asttokens @ file:///home/conda/feedstock_root/build_artifacts/asttokens_1763409923949/work
15
+ async-timeout==5.0.1
16
+ attrs==26.1.0
17
+ audioread==3.1.0
18
+ backports-datetime-fromisoformat==2.0.3
19
+ botocore==1.42.70
20
+ braceexpand==0.1.7
21
+ certifi==2026.2.25
22
+ cffi==2.0.0
23
+ charset-normalizer==3.4.6
24
+ click==8.3.1
25
+ cloudpickle==3.1.2
26
+ colorama==0.4.6
27
+ colorlog==6.10.1
28
+ comm @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_comm_1753453984/work
29
+ contourpy==1.3.2
30
+ cuda-bindings==13.2.0
31
+ cuda-pathfinder==1.5.0
32
+ cuda-toolkit==13.0.2
33
+ cycler==0.12.1
34
+ cytoolz==1.1.0
35
+ datasets==4.8.2
36
+ debugpy @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_debugpy_1769744974/work
37
+ decorator @ file:///home/conda/feedstock_root/build_artifacts/decorator_1740384970518/work
38
+ dill==0.4.1
39
+ diskcache==5.6.3
40
+ editdistance==0.8.1
41
+ einops==0.8.2
42
+ exceptiongroup @ file:///home/conda/feedstock_root/build_artifacts/exceptiongroup_1763918002538/work
43
+ executing @ file:///home/conda/feedstock_root/build_artifacts/executing_1756729339227/work
44
+ fiddle==0.3.0
45
+ filelock==3.25.2
46
+ fonttools==4.62.1
47
+ frozenlist==1.8.0
48
+ fsspec==2026.3.0
49
+ gitdb==4.0.12
50
+ GitPython==3.1.46
51
+ graphviz==0.21
52
+ greenlet==3.3.2
53
+ grpcio==1.80.0
54
+ grpcio-tools==1.80.0
55
+ h11==0.16.0
56
+ hf-xet==1.4.2
57
+ httpcore==1.0.9
58
+ httpx==0.28.1
59
+ huggingface==0.0.1
60
+ huggingface_hub==1.8.0
61
+ humanize==4.15.0
62
+ hydra-core==1.3.2
63
+ idna==3.11
64
+ importlib_metadata==9.0.0
65
+ indic_numtowords==1.1.0
66
+ inflect==7.5.0
67
+ intervaltree==3.2.1
68
+ ipykernel @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_ipykernel_1770566133/work
69
+ ipython @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_ipython_1748711175/work
70
+ jedi @ file:///home/conda/feedstock_root/build_artifacts/jedi_1733300866624/work
71
+ Jinja2==3.1.6
72
+ jiwer==3.1.0
73
+ jmespath==1.1.0
74
+ joblib==1.5.3
75
+ jupyter_client @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_jupyter_client_1767954655/work
76
+ jupyter_core @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_jupyter_core_1760643864/work
77
+ kaldi-python-io==1.2.2
78
+ kaldialign==0.9.1
79
+ kiwisolver==1.5.0
80
+ lazy-loader==0.5
81
+ lhotse==1.32.2
82
+ libcst==1.8.6
83
+ librosa==0.11.0
84
+ lightning==2.4.0
85
+ lightning-utilities==0.15.3
86
+ lilcom==1.8.2
87
+ llvmlite==0.46.0
88
+ lxml==6.0.2
89
+ Mako==1.3.10
90
+ Markdown==3.10.2
91
+ markdown-it-py==4.0.0
92
+ MarkupSafe==3.0.3
93
+ marshmallow==4.2.2
94
+ matplotlib==3.10.8
95
+ matplotlib-inline @ file:///home/conda/feedstock_root/build_artifacts/matplotlib-inline_1761214490209/work
96
+ mdurl==0.1.2
97
+ mediapy==1.1.6
98
+ ml_dtypes==0.5.4
99
+ more-itertools==10.8.0
100
+ mpmath==1.3.0
101
+ msgpack==1.1.2
102
+ msgspec==0.20.0
103
+ multidict==6.7.1
104
+ multiprocess==0.70.19
105
+ nemo-toolkit==2.7.0
106
+ nest_asyncio @ file:///home/conda/feedstock_root/build_artifacts/nest-asyncio_1733325553580/work
107
+ networkx==3.4.2
108
+ ninja==1.13.0
109
+ numba==0.64.0
110
+ numba-cuda==0.15.1
111
+ numexpr==2.13.1
112
+ numpy==2.2.6
113
+ nv-one-logger-core==2.3.1
114
+ nv-one-logger-pytorch-lightning-integration==2.3.1
115
+ nv-one-logger-training-telemetry==2.3.1
116
+ nvdlfw_inspect==0.2.2
117
+ nvidia-cublas==13.1.0.3
118
+ nvidia-cublas-cu12==12.1.3.1
119
+ nvidia-cuda-cupti==13.0.85
120
+ nvidia-cuda-cupti-cu12==12.1.105
121
+ nvidia-cuda-nvrtc==13.0.88
122
+ nvidia-cuda-nvrtc-cu12==12.1.105
123
+ nvidia-cuda-runtime==13.0.96
124
+ nvidia-cuda-runtime-cu12==12.1.105
125
+ nvidia-cudnn-cu12==9.1.0.70
126
+ nvidia-cudnn-cu13==9.19.0.56
127
+ nvidia-cufft==12.0.0.61
128
+ nvidia-cufft-cu12==11.0.2.54
129
+ nvidia-cufile==1.15.1.6
130
+ nvidia-cufile-cu12==1.13.1.3
131
+ nvidia-curand==10.4.0.35
132
+ nvidia-curand-cu12==10.3.2.106
133
+ nvidia-cusolver==12.0.4.66
134
+ nvidia-cusolver-cu12==11.4.5.107
135
+ nvidia-cusparse==12.6.3.3
136
+ nvidia-cusparse-cu12==12.1.0.106
137
+ nvidia-cusparselt-cu12==0.7.1
138
+ nvidia-cusparselt-cu13==0.8.0
139
+ nvidia-nccl-cu12==2.21.5
140
+ nvidia-nccl-cu13==2.28.9
141
+ nvidia-nvjitlink==13.0.88
142
+ nvidia-nvjitlink-cu12==12.8.93
143
+ nvidia-nvshmem-cu12==3.3.20
144
+ nvidia-nvshmem-cu13==3.4.5
145
+ nvidia-nvtx==13.0.85
146
+ nvidia-nvtx-cu12==12.1.105
147
+ nvidia-riva-client==2.25.1
148
+ omegaconf==2.3.0
149
+ onnx==1.20.1
150
+ onnx-ir==0.2.0
151
+ onnxscript==0.6.2
152
+ optuna==4.8.0
153
+ overrides==7.7.0
154
+ packaging==26.0
155
+ pandas==2.3.3
156
+ parso @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_parso_1770676243/work
157
+ pathlib_abc==0.5.2
158
+ peft==0.18.1
159
+ pexpect @ file:///home/conda/feedstock_root/build_artifacts/pexpect_1733301927746/work
160
+ pickleshare @ file:///home/conda/feedstock_root/build_artifacts/pickleshare_1733327343728/work
161
+ pillow==12.1.1
162
+ platformdirs @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_platformdirs_1773199142/work
163
+ pooch==1.9.0
164
+ portalocker==3.2.0
165
+ prompt_toolkit @ file:///home/conda/feedstock_root/build_artifacts/prompt-toolkit_1756321756983/work
166
+ propcache==0.4.1
167
+ protobuf==6.33.6
168
+ psutil @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_psutil_1769678154/work
169
+ ptyprocess @ file:///home/conda/feedstock_root/build_artifacts/ptyprocess_1733302279685/work/dist/ptyprocess-0.7.0-py2.py3-none-any.whl#sha256=92c32ff62b5fd8cf325bec5ab90d7be3d2a8ca8c8a3813ff487a8d2002630d1f
170
+ pure_eval @ file:///home/conda/feedstock_root/build_artifacts/pure_eval_1733569405015/work
171
+ pyannote-core==6.0.1
172
+ pyannote-database==6.1.1
173
+ pyannote-metrics==4.0.0
174
+ pyarrow==23.0.1
175
+ pycparser==3.0
176
+ pydantic==2.12.5
177
+ pydantic_core==2.41.5
178
+ pydub==0.25.1
179
+ Pygments @ file:///home/conda/feedstock_root/build_artifacts/pygments_1750615794071/work
180
+ pyloudnorm==0.2.0
181
+ pyparsing==3.3.2
182
+ python-dateutil @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_python-dateutil_1751104122/work
183
+ pytorch-lightning==2.6.1
184
+ pytz==2026.1.post1
185
+ PyYAML==6.0.3
186
+ pyzmq @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_pyzmq_1771716952/work
187
+ RapidFuzz==3.14.3
188
+ regex==2026.3.32
189
+ requests==2.32.5
190
+ resampy==0.4.3
191
+ rich==14.3.3
192
+ ruamel.yaml==0.19.1
193
+ s3fs==2026.3.0
194
+ sacrebleu==2.6.0
195
+ sacremoses==0.1.1
196
+ safetensors==0.7.0
197
+ scikit-learn==1.7.2
198
+ scipy==1.15.3
199
+ sentencepiece==0.2.1
200
+ sentry-sdk==2.55.0
201
+ shellingham==1.5.4
202
+ six @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_six_1753199211/work
203
+ smart_open==7.5.1
204
+ smmap==5.0.3
205
+ sortedcontainers==2.4.0
206
+ soundfile==0.13.1
207
+ sox==1.5.0
208
+ soxr==1.0.0
209
+ SQLAlchemy==2.0.48
210
+ stack_data @ file:///home/conda/feedstock_root/build_artifacts/stack_data_1733569443808/work
211
+ StrEnum==0.4.15
212
+ sympy==1.13.1
213
+ tabulate==0.10.0
214
+ tenacity==9.1.4
215
+ tensorboard==2.20.0
216
+ tensorboard-data-server==0.7.2
217
+ text-unidecode==1.3
218
+ threadpoolctl==3.6.0
219
+ tokenizers==0.22.2
220
+ toml==0.10.2
221
+ tomli==2.4.0
222
+ toolz==1.1.0
223
+ torch==2.5.1+cu121
224
+ torchaudio==2.5.1+cu121
225
+ TorchCodec==0.1.1+cu121
226
+ torchmetrics==1.9.0
227
+ tornado @ file:///home/conda/feedstock_root/build_artifacts/tornado_1765836373149/work
228
+ tqdm==4.67.3
229
+ traitlets @ file:///home/conda/feedstock_root/build_artifacts/traitlets_1733367359838/work
230
+ transformer_engine==2.13.0
231
+ transformer_engine_cu12==2.13.0
232
+ transformers==5.4.0
233
+ triton==3.1.0
234
+ typeguard==4.5.1
235
+ typer==0.24.1
236
+ typing-inspection==0.4.2
237
+ typing_extensions @ file:///home/conda/feedstock_root/build_artifacts/bld/rattler-build_typing_extensions_1756220668/work
238
+ tzdata==2025.3
239
+ universal_pathlib==0.3.10
240
+ urllib3==2.6.3
241
+ wandb==0.25.1
242
+ wcwidth @ file:///home/conda/feedstock_root/build_artifacts/wcwidth_1770634525285/work
243
+ webdataset==1.0.2
244
+ websockets==15.0.1
245
+ Werkzeug==3.1.6
246
+ wget==3.2
247
+ whisper_normalizer==0.1.12
248
+ wrapt==2.1.2
249
+ xxhash==3.6.0
250
+ yarl==1.23.0
251
+ zipp==3.23.0
environment.yml ADDED
@@ -0,0 +1,281 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: canaryFineTune
2
+ channels:
3
+ - defaults
4
+ - conda-forge
5
+ dependencies:
6
+ - _libgcc_mutex=0.1=main
7
+ - _openmp_mutex=5.1=1_gnu
8
+ - asttokens=3.0.1=pyhd8ed1ab_0
9
+ - bzip2=1.0.8=h5eee18b_6
10
+ - ca-certificates=2026.2.25=hbd8a1cb_0
11
+ - comm=0.2.3=pyhe01879c_0
12
+ - cudatoolkit=11.8.0=h6a678d5_0
13
+ - debugpy=1.8.20=py310h25320af_0
14
+ - decorator=5.2.1=pyhd8ed1ab_0
15
+ - exceptiongroup=1.3.1=pyhd8ed1ab_0
16
+ - executing=2.2.1=pyhd8ed1ab_0
17
+ - ipykernel=7.2.0=pyha191276_1
18
+ - jedi=0.19.2=pyhd8ed1ab_1
19
+ - jupyter_client=8.8.0=pyhcf101f3_0
20
+ - jupyter_core=5.9.1=pyhc90fa1f_0
21
+ - keyutils=1.6.3=hb9d3cd8_0
22
+ - krb5=1.22.2=ha1258a1_0
23
+ - ld_impl_linux-64=2.44=h9e0c5a2_3
24
+ - libedit=3.1.20250104=pl5321h7949ede_0
25
+ - libexpat=2.7.4=h7354ed3_0
26
+ - libffi=3.4.4=h6a678d5_1
27
+ - libgcc=15.2.0=h69a1729_7
28
+ - libgcc-ng=15.2.0=h166f726_7
29
+ - libgomp=15.2.0=h4751f2c_7
30
+ - libnsl=2.0.0=h5eee18b_0
31
+ - libsodium=1.0.21=h280c20c_3
32
+ - libstdcxx=15.2.0=h39759b7_7
33
+ - libstdcxx-ng=15.2.0=hc03a8fd_7
34
+ - libuuid=1.41.5=h5eee18b_0
35
+ - libxcb=1.17.0=h9b100fa_0
36
+ - libzlib=1.3.1=hb25bd0a_0
37
+ - matplotlib-inline=0.2.1=pyhd8ed1ab_0
38
+ - ncurses=6.5=h7934f7d_0
39
+ - nest-asyncio=1.6.0=pyhd8ed1ab_1
40
+ - openssl=3.6.1=h35e630c_1
41
+ - packaging=26.0=pyhcf101f3_0
42
+ - parso=0.8.6=pyhcf101f3_0
43
+ - pexpect=4.9.0=pyhd8ed1ab_1
44
+ - pickleshare=0.7.5=pyhd8ed1ab_1004
45
+ - pip=26.0.1=pyhc872135_0
46
+ - platformdirs=4.9.4=pyhcf101f3_0
47
+ - prompt-toolkit=3.0.52=pyha770c72_0
48
+ - psutil=7.2.2=py310h139afa4_0
49
+ - pthread-stubs=0.3=h0ce48e5_1
50
+ - ptyprocess=0.7.0=pyhd8ed1ab_1
51
+ - pure_eval=0.2.3=pyhd8ed1ab_1
52
+ - pygments=2.19.2=pyhd8ed1ab_0
53
+ - python=3.10.20=h741d88c_0
54
+ - python-dateutil=2.9.0.post0=pyhe01879c_2
55
+ - python_abi=3.10=2_cp310
56
+ - pyzmq=27.1.0=py310hc4bea81_2
57
+ - readline=8.3=hc2a1206_0
58
+ - six=1.17.0=pyhe01879c_1
59
+ - sqlite=3.51.2=h3e8d24a_0
60
+ - stack_data=0.6.3=pyhd8ed1ab_1
61
+ - tk=8.6.15=h54e0aa7_0
62
+ - tornado=6.5.4=py310ha78b2d2_0
63
+ - traitlets=5.14.3=pyhd8ed1ab_1
64
+ - typing_extensions=4.15.0=pyhcf101f3_0
65
+ - wcwidth=0.6.0=pyhd8ed1ab_0
66
+ - wheel=0.46.3=py310h06a4308_0
67
+ - xorg-libx11=1.8.12=h9b100fa_1
68
+ - xorg-libxau=1.0.12=h9b100fa_0
69
+ - xorg-libxdmcp=1.1.5=h9b100fa_0
70
+ - xorg-xorgproto=2024.1=h5eee18b_1
71
+ - xz=5.8.2=h448239c_0
72
+ - zeromq=4.3.5=h41580af_10
73
+ - zlib=1.3.1=hb25bd0a_0
74
+ - pip:
75
+ - absl-py==2.4.0
76
+ - accelerate==1.13.0
77
+ - aiohappyeyeballs==2.6.1
78
+ - aiohttp==3.13.3
79
+ - aiosignal==1.4.0
80
+ - aistore==1.23.0
81
+ - alembic==1.18.4
82
+ - annotated-doc==0.0.4
83
+ - annotated-types==0.7.0
84
+ - antlr4-python3-runtime==4.9.3
85
+ - anyio==4.13.0
86
+ - async-timeout==5.0.1
87
+ - attrs==26.1.0
88
+ - audioread==3.1.0
89
+ - backports-datetime-fromisoformat==2.0.3
90
+ - braceexpand==0.1.7
91
+ - certifi==2026.2.25
92
+ - cffi==2.0.0
93
+ - charset-normalizer==3.4.6
94
+ - click==8.3.1
95
+ - cloudpickle==3.1.2
96
+ - colorama==0.4.6
97
+ - colorlog==6.10.1
98
+ - contourpy==1.3.2
99
+ - cuda-bindings==13.2.0
100
+ - cuda-pathfinder==1.5.0
101
+ - cuda-toolkit==13.0.2
102
+ - cycler==0.12.1
103
+ - cytoolz==1.1.0
104
+ - datasets==4.8.2
105
+ - dill==0.4.1
106
+ - diskcache==5.6.3
107
+ - editdistance==0.8.1
108
+ - einops==0.8.2
109
+ - fiddle==0.3.0
110
+ - filelock==3.25.2
111
+ - fonttools==4.62.1
112
+ - frozenlist==1.8.0
113
+ - fsspec==2026.3.0
114
+ - gitdb==4.0.12
115
+ - gitpython==3.1.46
116
+ - graphviz==0.21
117
+ - greenlet==3.3.2
118
+ - grpcio==1.78.0
119
+ - h11==0.16.0
120
+ - hf-xet==1.4.2
121
+ - httpcore==1.0.9
122
+ - httpx==0.28.1
123
+ - huggingface==0.0.1
124
+ - huggingface-hub==1.8.0
125
+ - humanize==4.15.0
126
+ - hydra-core==1.3.2
127
+ - idna==3.11
128
+ - indic-numtowords==1.1.0
129
+ - inflect==7.5.0
130
+ - intervaltree==3.2.1
131
+ - ipython==8.38.0
132
+ - jinja2==3.1.6
133
+ - jiwer==3.1.0
134
+ - joblib==1.5.3
135
+ - kaldi-python-io==1.2.2
136
+ - kaldialign==0.9.1
137
+ - kiwisolver==1.5.0
138
+ - lazy-loader==0.5
139
+ - lhotse==1.32.2
140
+ - libcst==1.8.6
141
+ - librosa==0.11.0
142
+ - lightning==2.4.0
143
+ - lightning-utilities==0.15.3
144
+ - lilcom==1.8.2
145
+ - llvmlite==0.46.0
146
+ - lxml==6.0.2
147
+ - mako==1.3.10
148
+ - markdown==3.10.2
149
+ - markdown-it-py==4.0.0
150
+ - markupsafe==3.0.3
151
+ - marshmallow==4.2.2
152
+ - matplotlib==3.10.8
153
+ - mdurl==0.1.2
154
+ - mediapy==1.1.6
155
+ - ml-dtypes==0.5.4
156
+ - more-itertools==10.8.0
157
+ - mpmath==1.3.0
158
+ - msgpack==1.1.2
159
+ - msgspec==0.20.0
160
+ - multidict==6.7.1
161
+ - multiprocess==0.70.19
162
+ - nemo-toolkit==2.7.0
163
+ - networkx==3.4.2
164
+ - numba==0.64.0
165
+ - numba-cuda==0.15.1
166
+ - numexpr==2.13.1
167
+ - numpy==2.2.6
168
+ - nv-one-logger-core==2.3.1
169
+ - nv-one-logger-pytorch-lightning-integration==2.3.1
170
+ - nv-one-logger-training-telemetry==2.3.1
171
+ - nvidia-cublas==13.1.0.3
172
+ - nvidia-cublas-cu12==12.1.3.1
173
+ - nvidia-cuda-cupti==13.0.85
174
+ - nvidia-cuda-cupti-cu12==12.1.105
175
+ - nvidia-cuda-nvrtc==13.0.88
176
+ - nvidia-cuda-nvrtc-cu12==12.1.105
177
+ - nvidia-cuda-runtime==13.0.96
178
+ - nvidia-cuda-runtime-cu12==12.1.105
179
+ - nvidia-cudnn-cu12==9.1.0.70
180
+ - nvidia-cudnn-cu13==9.19.0.56
181
+ - nvidia-cufft==12.0.0.61
182
+ - nvidia-cufft-cu12==11.0.2.54
183
+ - nvidia-cufile==1.15.1.6
184
+ - nvidia-cufile-cu12==1.13.1.3
185
+ - nvidia-curand==10.4.0.35
186
+ - nvidia-curand-cu12==10.3.2.106
187
+ - nvidia-cusolver==12.0.4.66
188
+ - nvidia-cusolver-cu12==11.4.5.107
189
+ - nvidia-cusparse==12.6.3.3
190
+ - nvidia-cusparse-cu12==12.1.0.106
191
+ - nvidia-cusparselt-cu12==0.7.1
192
+ - nvidia-cusparselt-cu13==0.8.0
193
+ - nvidia-nccl-cu12==2.21.5
194
+ - nvidia-nccl-cu13==2.28.9
195
+ - nvidia-nvjitlink==13.0.88
196
+ - nvidia-nvjitlink-cu12==12.8.93
197
+ - nvidia-nvshmem-cu12==3.3.20
198
+ - nvidia-nvshmem-cu13==3.4.5
199
+ - nvidia-nvtx==13.0.85
200
+ - nvidia-nvtx-cu12==12.1.105
201
+ - omegaconf==2.3.0
202
+ - onnx==1.20.1
203
+ - optuna==4.8.0
204
+ - overrides==7.7.0
205
+ - pandas==2.3.3
206
+ - pathlib-abc==0.5.2
207
+ - peft==0.18.1
208
+ - pillow==12.1.1
209
+ - pooch==1.9.0
210
+ - portalocker==3.2.0
211
+ - propcache==0.4.1
212
+ - protobuf==5.29.6
213
+ - pyannote-core==6.0.1
214
+ - pyannote-database==6.1.1
215
+ - pyannote-metrics==4.0.0
216
+ - pyarrow==23.0.1
217
+ - pycparser==3.0
218
+ - pydantic==2.12.5
219
+ - pydantic-core==2.41.5
220
+ - pydub==0.25.1
221
+ - pyloudnorm==0.2.0
222
+ - pyparsing==3.3.2
223
+ - pytorch-lightning==2.6.1
224
+ - pytz==2026.1.post1
225
+ - pyyaml==6.0.3
226
+ - rapidfuzz==3.14.3
227
+ - regex==2026.3.32
228
+ - requests==2.32.5
229
+ - resampy==0.4.3
230
+ - rich==14.3.3
231
+ - ruamel-yaml==0.19.1
232
+ - sacrebleu==2.6.0
233
+ - sacremoses==0.1.1
234
+ - safetensors==0.7.0
235
+ - scikit-learn==1.7.2
236
+ - scipy==1.15.3
237
+ - sentencepiece==0.2.1
238
+ - sentry-sdk==2.55.0
239
+ - setuptools==81.0.0
240
+ - shellingham==1.5.4
241
+ - smart-open==7.5.1
242
+ - smmap==5.0.3
243
+ - sortedcontainers==2.4.0
244
+ - soundfile==0.13.1
245
+ - sox==1.5.0
246
+ - soxr==1.0.0
247
+ - sqlalchemy==2.0.48
248
+ - strenum==0.4.15
249
+ - sympy==1.13.1
250
+ - tabulate==0.10.0
251
+ - tenacity==9.1.4
252
+ - tensorboard==2.20.0
253
+ - tensorboard-data-server==0.7.2
254
+ - text-unidecode==1.3
255
+ - threadpoolctl==3.6.0
256
+ - tokenizers==0.22.2
257
+ - toml==0.10.2
258
+ - tomli==2.4.0
259
+ - toolz==1.1.0
260
+ - torch==2.5.1+cu121
261
+ - torchaudio==2.5.1+cu121
262
+ - torchcodec==0.1.1+cu121
263
+ - torchmetrics==1.9.0
264
+ - tqdm==4.67.3
265
+ - transformers==5.4.0
266
+ - triton==3.1.0
267
+ - typeguard==4.5.1
268
+ - typer==0.24.1
269
+ - typing-inspection==0.4.2
270
+ - tzdata==2025.3
271
+ - universal-pathlib==0.3.10
272
+ - urllib3==2.6.3
273
+ - wandb==0.25.1
274
+ - webdataset==1.0.2
275
+ - werkzeug==3.1.6
276
+ - wget==3.2
277
+ - whisper-normalizer==0.1.12
278
+ - wrapt==2.1.2
279
+ - xxhash==3.6.0
280
+ - yarl==1.23.0
281
+ prefix: /home/bsoft/miniconda3/envs/canaryFineTune
lazy.py ADDED
@@ -0,0 +1,829 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import random
3
+ import types
4
+ import warnings
5
+ from contextlib import contextmanager
6
+ from functools import partial
7
+ from typing import Any, Callable, Iterable, List, Literal, Optional, TypeVar, Union
8
+
9
+ from lhotse.serialization import (
10
+ LazyMixin,
11
+ decode_json_line,
12
+ deserialize_item,
13
+ open_best,
14
+ )
15
+ from lhotse.utils import Pathlike, fastcopy, is_module_available, streaming_shuffle
16
+
17
+ T = TypeVar("T")
18
+
19
+
20
+ class AlgorithmMixin(LazyMixin, Iterable):
21
+ """
22
+ Helper base class with methods that are supposed to work identically
23
+ on Lhotse manifest classes such as CutSet, RecordingSet, etc.
24
+ """
25
+
26
+ def filter(self, predicate: Callable[[T], bool]):
27
+ """
28
+ Return a new manifest containing only the items that satisfy ``predicate``.
29
+ If the manifest is lazy, the filtering will also be applied lazily.
30
+
31
+ :param predicate: a function that takes a cut as an argument and returns bool.
32
+ :return: a filtered manifest.
33
+ """
34
+ cls = type(self)
35
+
36
+ if self.is_lazy:
37
+ return cls(LazyFilter(self, predicate=predicate))
38
+
39
+ return cls.from_items(cut for cut in self if predicate(cut))
40
+
41
+ def map(self, transform_fn: Callable[[T], T]):
42
+ """
43
+ Apply `transform_fn` to each item in this manifest and return a new manifest.
44
+ If the manifest is opened lazy, the transform is also applied lazily.
45
+
46
+ :param transform_fn: A callable (function) that accepts a single item instance
47
+ and returns a new (or the same) instance of the same type.
48
+ E.g. with CutSet, callable accepts ``Cut`` and returns also ``Cut``.
49
+ :return: a new ``CutSet`` with transformed cuts.
50
+ """
51
+ cls = type(self)
52
+ ans = cls(LazyMapper(self.data, fn=transform_fn))
53
+ if self.is_lazy:
54
+ return ans
55
+ return ans.to_eager()
56
+
57
+ @classmethod
58
+ def mux(
59
+ cls,
60
+ *manifests,
61
+ stop_early: bool = False,
62
+ weights: Optional[List[Union[int, float]]] = None,
63
+ seed: Union[int, Literal["trng", "randomized"]] = 0,
64
+ ):
65
+ """
66
+ Merges multiple manifest iterables into a new iterable by lazily multiplexing them during iteration time.
67
+ If one of the iterables is exhausted before the others, we will keep iterating until all iterables
68
+ are exhausted. This behavior can be changed with ``stop_early`` parameter.
69
+
70
+ :param manifests: iterables to be multiplexed.
71
+ They can be either lazy or eager, but the resulting manifest will always be lazy.
72
+ :param stop_early: should we stop the iteration as soon as we exhaust one of the manifests.
73
+ :param weights: an optional weight for each iterable, affects the probability of it being sampled.
74
+ The weights are uniform by default.
75
+ If lengths are known, it makes sense to pass them here for uniform distribution of
76
+ items in the expectation.
77
+ :param seed: the random seed, ensures deterministic order across multiple iterations.
78
+ """
79
+ return cls(
80
+ LazyIteratorMultiplexer(
81
+ *manifests, stop_early=stop_early, weights=weights, seed=seed
82
+ )
83
+ )
84
+
85
+ @classmethod
86
+ def infinite_mux(
87
+ cls,
88
+ *manifests,
89
+ weights: Optional[List[Union[int, float]]] = None,
90
+ seed: Union[int, Literal["trng", "randomized"]] = 0,
91
+ max_open_streams: Optional[int] = None,
92
+ ):
93
+ """
94
+ Merges multiple manifest iterables into a new iterable by lazily multiplexing them during iteration time.
95
+ Unlike ``mux()``, this method allows to limit the number of max open sub-iterators at any given time.
96
+
97
+ To enable this, it performs 2-stage sampling.
98
+ First, it samples with replacement the set of iterators ``I`` to construct a subset ``I_sub``
99
+ of size ``max_open_streams``.
100
+ Then, for each iteration step, it samples an iterator ``i`` from ``I_sub``,
101
+ fetches the next item from it, and yields it.
102
+ Once ``i`` becomes exhausted, it is replaced with a new iterator ``j`` sampled from ``I_sub``.
103
+
104
+ .. caution:: Do not use this method with inputs that are infinitely iterable as they will
105
+ silently break the multiplexing property by only using a subset of the input iterables.
106
+
107
+ .. caution:: This method is not recommended for multiplexing for a small amount of iterations,
108
+ as it may be much less accurate than ``mux()`` depending on the number of open streams,
109
+ iterable sizes, and the random seed.
110
+
111
+ :param manifests: iterables to be multiplexed.
112
+ They can be either lazy or eager, but the resulting manifest will always be lazy.
113
+ :param weights: an optional weight for each iterable, affects the probability of it being sampled.
114
+ The weights are uniform by default.
115
+ If lengths are known, it makes sense to pass them here for uniform distribution of
116
+ items in the expectation.
117
+ :param seed: the random seed, ensures deterministic order across multiple iterations.
118
+ :param max_open_streams: the number of iterables that can be open simultaneously at any given time.
119
+ """
120
+ return cls(
121
+ LazyInfiniteApproximateMultiplexer(
122
+ *manifests,
123
+ weights=weights,
124
+ seed=seed,
125
+ max_open_streams=max_open_streams,
126
+ )
127
+ )
128
+
129
+ def shuffle(
130
+ self,
131
+ rng: Optional[random.Random] = None,
132
+ buffer_size: int = 10000,
133
+ ):
134
+ """
135
+ Shuffles the elements and returns a shuffled variant of self.
136
+ If the manifest is opened lazily, performs shuffling on-the-fly with a fixed buffer size.
137
+
138
+ :param rng: an optional instance of ``random.Random`` for precise control of randomness.
139
+ :return: a shuffled copy of self, or a manifest that is shuffled lazily.
140
+ """
141
+ cls = type(self)
142
+
143
+ if rng is None:
144
+ rng = random
145
+
146
+ if self.is_lazy:
147
+ return cls(LazyShuffler(self.data, buffer_size=buffer_size, rng=rng))
148
+ else:
149
+ new: List = self.data.copy()
150
+ rng.shuffle(new)
151
+ return cls(new)
152
+
153
+ def repeat(self, times: Optional[int] = None, preserve_id: bool = False):
154
+ """
155
+ Return a new, lazily evaluated manifest that iterates over the original elements ``times``
156
+ number of times.
157
+
158
+ :param times: how many times to repeat (infinite by default).
159
+ :param preserve_id: when ``True``, we won't update the element ID with repeat number.
160
+ :return: a repeated manifest.
161
+ """
162
+ cls = type(self)
163
+ return cls(LazyRepeater(self, times=times, preserve_id=preserve_id))
164
+
165
+ def __add__(self, other):
166
+ cls = type(self)
167
+ return cls(LazyIteratorChain(self.data, other.data))
168
+
169
+
170
+ class Dillable:
171
+ """
172
+ Mix-in that will leverage ``dill`` instead of ``pickle``
173
+ when pickling an object.
174
+
175
+ It is useful when the user can't avoid ``pickle`` (e.g. in multiprocessing),
176
+ but needs to use unpicklable objects such as lambdas.
177
+
178
+ If ``dill`` is not installed, it defers to what ``pickle`` does by default.
179
+ """
180
+
181
+ _ENABLED_VALUES = {"1", "True", "true", "yes"}
182
+
183
+ def __getstate__(self):
184
+ if is_dill_enabled():
185
+ import dill
186
+
187
+ return dill.dumps(self.__dict__)
188
+ else:
189
+ return self.__dict__
190
+
191
+ def __setstate__(self, state):
192
+ if is_dill_enabled():
193
+ import dill
194
+
195
+ self.__dict__ = dill.loads(state)
196
+ else:
197
+ self.__dict__ = state
198
+
199
+
200
+ def is_dill_enabled(_ENABLED_VALUES=frozenset(("1", "True", "true", "yes"))) -> bool:
201
+ """Returns bool indicating if dill-based pickling in Lhotse is enabled or not."""
202
+ return (
203
+ is_module_available("dill")
204
+ and os.environ.get("LHOTSE_DILL_ENABLED", "0") in _ENABLED_VALUES
205
+ )
206
+
207
+
208
+ def set_dill_enabled(value: bool) -> None:
209
+ """Enable or disable dill-based pickling in Lhotse."""
210
+ assert is_module_available("dill"), (
211
+ "Cannot enable dill because dill is not installed. "
212
+ "Please run 'pip install dill' and try again."
213
+ )
214
+ # We use os.environ here so that sub-processes / forks will inherit this value
215
+ os.environ["LHOTSE_DILL_ENABLED"] = "1" if value else "0"
216
+
217
+
218
+ @contextmanager
219
+ def dill_enabled(value: bool):
220
+ """
221
+ Context manager that overrides the setting of Lhotse's dill-backed pickling
222
+ and restores the previous value after exit.
223
+
224
+ Example::
225
+
226
+ >>> import pickle
227
+ ... with dill_enabled(True):
228
+ ... pickle.dump(CutSet(...).filter(lambda c: c.duration < 5), open("cutset.pickle", "wb"))
229
+ """
230
+ previous = is_dill_enabled()
231
+ set_dill_enabled(value)
232
+ yield
233
+ set_dill_enabled(previous)
234
+
235
+
236
+ class LazyTxtIterator:
237
+ """
238
+ LazyTxtIterator is a thin wrapper over builtin ``open`` function to
239
+ iterate over lines in a (possibly compressed) text file.
240
+ It can also provide the number of lines via __len__ via fast newlines counting.
241
+ """
242
+
243
+ def __init__(self, path: Pathlike, as_text_example: bool = True) -> None:
244
+ self.path = path
245
+ self.as_text_example = as_text_example
246
+ self._len = None
247
+
248
+ def __iter__(self):
249
+ from lhotse.cut.text import TextExample
250
+
251
+ tot = 0
252
+ with open_best(self.path, "r") as f:
253
+ for line in f:
254
+ line = line.strip()
255
+ if self.as_text_example:
256
+ line = TextExample(line)
257
+ yield line
258
+ tot += 1
259
+ if self._len is None:
260
+ self._len = tot
261
+
262
+ def __len__(self) -> int:
263
+ if self._len is None:
264
+ self._len = count_newlines_fast(self.path)
265
+ return self._len
266
+
267
+
268
+ class LazyJsonlIterator:
269
+ """
270
+ LazyJsonlIterator provides the ability to read JSON lines as Python dicts.
271
+ It can also provide the number of lines via __len__ via fast newlines counting.
272
+ """
273
+
274
+ def __init__(self, path: Pathlike) -> None:
275
+ self.path = path
276
+ self._len = None
277
+
278
+ def __iter__(self):
279
+ tot = 0
280
+ try:
281
+ with open_best(self.path, "r") as f:
282
+ for line in f:
283
+ try:
284
+ data = decode_json_line(line)
285
+ yield data
286
+ tot += 1
287
+ except EOFError:
288
+ # Log it so you know which shard is the "landmine"
289
+ print(f"\n⚠️ Caught EOFError in {self.path} - stopping shard early.")
290
+ break
291
+ except Exception:
292
+ # Handle actual JSON corruption
293
+ continue
294
+ except Exception as e:
295
+ print(f"❌ Could not open {self.path}: {e}")
296
+
297
+ if self._len is None:
298
+ self._len = tot
299
+
300
+
301
+ class LazyManifestIterator(Dillable):
302
+ """
303
+ LazyManifestIterator provides the ability to read Lhotse objects from a
304
+ JSONL file on-the-fly, without reading its full contents into memory.
305
+
306
+ This class is designed to be a partial "drop-in" replacement for ordinary dicts
307
+ to support lazy loading of RecordingSet, SupervisionSet and CutSet.
308
+ Since it does not support random access reads, some methods of these classes
309
+ might not work properly.
310
+ """
311
+
312
+ def __init__(self, path: Pathlike) -> None:
313
+ self.source = LazyJsonlIterator(path)
314
+
315
+ @property
316
+ def path(self) -> Pathlike:
317
+ return self.source.path
318
+
319
+ def __iter__(self):
320
+ yield from map(deserialize_item, self.source)
321
+
322
+ def __len__(self) -> int:
323
+ return len(self.source)
324
+
325
+ def __add__(self, other) -> "LazyIteratorChain":
326
+ return LazyIteratorChain(self, other)
327
+
328
+
329
+ class LazyIteratorChain(Dillable):
330
+ """
331
+ A thin wrapper over multiple iterators that enables to combine lazy manifests
332
+ in Lhotse. It iterates all underlying iterables sequentially.
333
+
334
+ It also supports shuffling the sub-iterators when it's iterated over.
335
+ This can be used to implement sharding (where each iterator is a shard)
336
+ with randomized shard order. Every iteration of this object will increment
337
+ an internal counter so that the next time it's iterated, the order of shards
338
+ is again randomized.
339
+
340
+ .. note:: if any of the input iterables is a dict, we'll iterate only its values.
341
+ """
342
+
343
+ def __init__(
344
+ self,
345
+ *iterators: Iterable,
346
+ shuffle_iters: bool = False,
347
+ seed: Optional[Union[int, Literal["trng", "randomized"]]] = None,
348
+ ) -> None:
349
+ self.iterators = []
350
+ self.shuffle_iters = shuffle_iters
351
+ self.seed = seed
352
+ self.num_iters = 0
353
+ for it in iterators:
354
+ # Auto-flatten LazyIteratorChain instances if any are passed
355
+ if isinstance(it, LazyIteratorChain):
356
+ for sub_it in it.iterators:
357
+ self.iterators.append(sub_it)
358
+ else:
359
+ self.iterators.append(it)
360
+
361
+ def __iter__(self):
362
+ from lhotse.dataset.dataloading import resolve_seed
363
+
364
+ iterators = self.iterators
365
+ if self.shuffle_iters:
366
+ if self.seed is None:
367
+ rng = random # global Python RNG
368
+ else:
369
+ rng = random.Random(resolve_seed(self.seed) + self.num_iters)
370
+ rng.shuffle(iterators)
371
+ self.num_iters += 1
372
+ for it in iterators:
373
+ if isinstance(it, dict):
374
+ it = it.values()
375
+ yield from it
376
+
377
+ def __len__(self) -> int:
378
+ return sum(len(it) for it in self.iterators)
379
+
380
+ def __add__(self, other) -> "LazyIteratorChain":
381
+ return LazyIteratorChain(self, other)
382
+
383
+
384
+ class LazyIteratorMultiplexer(Dillable):
385
+ """
386
+ A wrapper over multiple iterators that enables to combine lazy manifests in Lhotse.
387
+ During iteration, unlike :class:`.LazyIteratorChain`, :class:`.LazyIteratorMultiplexer`
388
+ at each step randomly selects the iterable used to yield an item.
389
+
390
+ Since the iterables might be of different length, we provide a ``weights`` parameter
391
+ to let the user decide which iterables should be sampled more frequently than others.
392
+ When an iterable is exhausted, we will keep sampling from the other iterables, until
393
+ we exhaust them all, unless ``stop_early`` is set to ``True``.
394
+ """
395
+
396
+ def __init__(
397
+ self,
398
+ *iterators: Iterable,
399
+ stop_early: bool = False,
400
+ weights: Optional[List[Union[int, float]]] = None,
401
+ seed: Union[int, Literal["trng", "randomized"]] = 0,
402
+ ) -> None:
403
+ self.iterators = list(iterators)
404
+ self.stop_early = stop_early
405
+ self.seed = seed
406
+
407
+ assert (
408
+ len(self.iterators) > 1
409
+ ), "There have to be at least two iterables to multiplex."
410
+
411
+ if weights is None:
412
+ self.weights = [1] * len(self.iterators)
413
+ else:
414
+ self.weights = weights
415
+
416
+ assert len(self.iterators) == len(self.weights)
417
+
418
+ def __iter__(self):
419
+ from lhotse.dataset.dataloading import resolve_seed
420
+
421
+ rng = random.Random(resolve_seed(self.seed))
422
+ iters = [iter(it) for it in self.iterators]
423
+ exhausted = [False for _ in range(len(iters))]
424
+
425
+ def should_continue():
426
+ if self.stop_early:
427
+ return not any(exhausted)
428
+ else:
429
+ return not all(exhausted)
430
+
431
+ while should_continue():
432
+ active_indexes, active_weights = zip(
433
+ *[
434
+ (i, w)
435
+ for i, (is_exhausted, w) in enumerate(zip(exhausted, self.weights))
436
+ if not is_exhausted
437
+ ]
438
+ )
439
+ idx = rng.choices(active_indexes, weights=active_weights, k=1)[0]
440
+ selected = iters[idx]
441
+ try:
442
+ item = next(selected)
443
+ yield item
444
+ except StopIteration:
445
+ exhausted[idx] = True
446
+ continue
447
+
448
+ def __len__(self) -> int:
449
+ return sum(len(it) for it in self.iterators)
450
+
451
+ def __add__(self, other) -> "LazyIteratorChain":
452
+ return LazyIteratorChain(self, other)
453
+
454
+
455
+ class LazyInfiniteApproximateMultiplexer(Dillable):
456
+ """
457
+ A variant of :class:`.LazyIteratorMultiplexer` that allows to control the number of
458
+ iterables that are simultaneously open.
459
+
460
+ It is useful for large-scale data sets where opening multiple file handles in
461
+ many processes leads to exhaustion of the operating system resources.
462
+
463
+ If the data sets are sharded, it is recommended to pass each shard as a separate iterator
464
+ when creating objects of this class. It is OK to assign a dataset-level weight to each shard
465
+ (e.g., if a dataset has a weight of 0.5, assign weight 0.5 to each of its shards).
466
+
467
+ There are several differences between this class and :class:`.LazyIteratorMultiplexer`:
468
+ * Objects of this class are infinite iterators.
469
+ * We hold a list of ``max_open_streams`` open iterators at any given time.
470
+ This list is filled by sampling input iterators with replacement.
471
+
472
+ These differences are necessary to guarantee the weighted sampling property.
473
+ If we did not sample with replacement or make it infinite, we would simply
474
+ exhaust highly-weighted iterators towards the beginning of each "epoch"
475
+ and keep sampling only lowly-weighted iterators towards the end of each "epoch".
476
+ """
477
+
478
+ def __init__(
479
+ self,
480
+ *iterators: Iterable,
481
+ stop_early: bool = False,
482
+ weights: Optional[List[Union[int, float]]] = None,
483
+ seed: Union[int, Literal["trng", "randomized"]] = 0,
484
+ max_open_streams: Optional[int] = None,
485
+ ) -> None:
486
+ self.iterators = list(iterators)
487
+ self.stop_early = stop_early
488
+ self.seed = seed
489
+ self.max_open_streams = max_open_streams
490
+ if max_open_streams is None or max_open_streams > len(self.iterators):
491
+ self.max_open_streams = len(self.iterators)
492
+
493
+ assert len(self.iterators) > 0
494
+ self.weights = weights
495
+ if weights is None:
496
+ self.weights = [1] * len(self.iterators)
497
+ assert len(self.iterators) == len(self.weights)
498
+ assert (
499
+ self.max_open_streams is None or self.max_open_streams >= 1
500
+ ), f"{self.max_open_streams=}"
501
+
502
+ def __iter__(self):
503
+ """
504
+ Assumptions
505
+ - we have N streams but can only open M at the time (M < N)
506
+ - the streams are finite
507
+ - each stream needs to be "short" to ensure the mux property
508
+ - each stream may be interpreted as a shard belonging to some larger group of streams
509
+ (e.g. multiple shards of a given dataset).
510
+ """
511
+ from lhotse.dataset.dataloading import resolve_seed
512
+
513
+ rng = random.Random(resolve_seed(self.seed))
514
+
515
+ def shuffled_streams():
516
+ # Create an infinite iterable of our streams.
517
+ # Assume N is "small" enough that shuffling it will be quick
518
+ #
519
+ # we need to incorporate weights into shuffling here
520
+ # and sample iterators with replacement.
521
+ # consider it0=[shard00, shard01] with weight 0.95
522
+ # and it1=[shard10, shard11] with weight 0.05
523
+ # so we have 4 streams [shard{01}{01}]
524
+ # if we just shuffle randomly and sample without replacement
525
+ # per each "epoch" (epoch = 4 shards) then we would have
526
+ # ignored the weights because we'll just exhaust it0 shards
527
+ # towards the beginning of an "epoch" and then keep yielding
528
+ # from it1 shards until the epoch is finished and we can sample
529
+ # from it0 again...
530
+ indexes = list(range(len(self.iterators)))
531
+ while True:
532
+ selected = rng.choices(indexes, self.weights, k=1)[0]
533
+ yield self.iterators[selected], self.weights[selected]
534
+
535
+ # Initialize an infinite sequence of finite streams.
536
+ # It is sampled with weights and replacement from ``self.iterators``,
537
+ # which are of length N.
538
+ stream_source = shuffled_streams()
539
+
540
+ # Sample the first M active streams to be multiplexed.
541
+ # As streams get depleted, we will replace them with
542
+ # new streams sampled from the stream source.
543
+ active_streams = [None] * self.max_open_streams
544
+ active_weights = [None] * self.max_open_streams
545
+ stream_indexes = list(range(self.max_open_streams))
546
+
547
+ def sample_new_stream_at(pos: int) -> None:
548
+ sampled_stream, sampled_weight = next(stream_source)
549
+ active_streams[pos] = iter(sampled_stream)
550
+ active_weights[pos] = sampled_weight
551
+
552
+ for stream_pos in range(self.max_open_streams):
553
+ sample_new_stream_at(stream_pos)
554
+
555
+ # The actual multiplexing loop.
556
+ while True:
557
+ # Select a stream from the currently active streams.
558
+ # We actually sample an index so that we know which position
559
+ # to replace if a stream is exhausted.
560
+ stream_pos = rng.choices(
561
+ stream_indexes,
562
+ weights=active_weights if sum(active_weights) > 0 else None,
563
+ k=1,
564
+ )[0]
565
+ selected = active_streams[stream_pos]
566
+ try:
567
+ # Sample from the selected stream.
568
+ item = next(selected)
569
+ yield item
570
+ except StopIteration:
571
+ # The selected stream is exhausted. Replace it with another one,
572
+ # and return a sample from the newly opened stream.
573
+ sample_new_stream_at(stream_pos)
574
+ item = next(active_streams[stream_pos])
575
+ yield item
576
+
577
+
578
+ class LazyShuffler(Dillable):
579
+ """
580
+ A wrapper over an iterable that enables lazy shuffling.
581
+ The shuffling algorithm is reservoir-sampling based.
582
+ See :func:`lhotse.utils.streaming_shuffle` for details.
583
+ """
584
+
585
+ def __init__(
586
+ self,
587
+ iterator: Iterable,
588
+ buffer_size: int = 10000,
589
+ rng: Optional[random.Random] = None,
590
+ ) -> None:
591
+ self.iterator = iterator
592
+ self.buffer_size = buffer_size
593
+ self.rng = rng
594
+
595
+ def __iter__(self):
596
+ return iter(
597
+ streaming_shuffle(
598
+ iter(self.iterator),
599
+ bufsize=self.buffer_size,
600
+ rng=self.rng,
601
+ )
602
+ )
603
+
604
+ def __len__(self) -> int:
605
+ return len(self.iterator)
606
+
607
+ def __add__(self, other) -> "LazyIteratorChain":
608
+ return LazyIteratorChain(self, other)
609
+
610
+
611
+ class LazyFilter(Dillable):
612
+ """
613
+ A wrapper over an iterable that enables lazy filtering.
614
+ It works like Python's `filter` built-in by applying the filter predicate
615
+ to each element and yielding it further if predicate returned ``True``.
616
+ """
617
+
618
+ def __init__(self, iterator: Iterable, predicate: Callable[[Any], bool]) -> None:
619
+ self.iterator = iterator
620
+ self.predicate = predicate
621
+ assert callable(
622
+ self.predicate
623
+ ), f"LazyFilter: 'predicate' arg must be callable (got {predicate})."
624
+ if (
625
+ isinstance(self.predicate, types.LambdaType)
626
+ and self.predicate.__name__ == "<lambda>"
627
+ and not is_module_available("dill")
628
+ ):
629
+ warnings.warn(
630
+ "A lambda was passed to LazyFilter: it may prevent you from forking this process. "
631
+ "If you experience issues with num_workers > 0 in torch.utils.data.DataLoader, "
632
+ "try passing a regular function instead."
633
+ )
634
+
635
+ def __iter__(self):
636
+ return filter(self.predicate, self.iterator)
637
+
638
+ def __add__(self, other) -> "LazyIteratorChain":
639
+ return LazyIteratorChain(self, other)
640
+
641
+ def __len__(self) -> int:
642
+ raise TypeError(
643
+ "LazyFilter does not support __len__ because it would require "
644
+ "iterating over the whole iterator, which is not possible in a lazy fashion. "
645
+ "If you really need to know the length, convert to eager mode first using "
646
+ "`.to_eager()`. Note that this will require loading the whole iterator into memory."
647
+ )
648
+
649
+
650
+ class LazyMapper(Dillable):
651
+ """
652
+ A wrapper over an iterable that enables lazy function evaluation on each item.
653
+ It works like Python's `map` built-in by applying a callable ``fn``
654
+ to each element ``x`` and yielding the result of ``fn(x)`` further.
655
+
656
+ New in Lhotse v1.22.0: ``apply_fn`` can be provided to decide whether ``fn`` should be applied
657
+ to a given example or not (in which case it will return it as-is, i.e., it does not filter).
658
+ """
659
+
660
+ def __init__(
661
+ self,
662
+ iterator: Iterable,
663
+ fn: Callable[[Any], Any],
664
+ apply_fn: Optional[Callable[[Any], bool]] = None,
665
+ ) -> None:
666
+ self.iterator = iterator
667
+ self.fn = fn
668
+ self.apply_fn = apply_fn
669
+ assert callable(self.fn), f"LazyMapper: 'fn' arg must be callable (got {fn})."
670
+ if self.apply_fn is not None:
671
+ assert callable(
672
+ self.apply_fn
673
+ ), f"LazyMapper: 'apply_fn' arg must be callable (got {fn})."
674
+ if (
675
+ (isinstance(self.fn, types.LambdaType) and self.fn.__name__ == "<lambda>")
676
+ or (
677
+ isinstance(self.apply_fn, types.LambdaType)
678
+ and self.apply_fn.__name__ == "<lambda>"
679
+ )
680
+ and not is_dill_enabled()
681
+ ):
682
+ warnings.warn(
683
+ "A lambda was passed to LazyMapper: it may prevent you from forking this process. "
684
+ "If you experience issues with num_workers > 0 in torch.utils.data.DataLoader, "
685
+ "try passing a regular function instead."
686
+ )
687
+
688
+ def __iter__(self):
689
+ if self.apply_fn is None:
690
+ yield from map(self.fn, self.iterator)
691
+ else:
692
+ for item in self.iterator:
693
+ if self.apply_fn(item):
694
+ ans = self.fn(item)
695
+ else:
696
+ ans = item
697
+ yield ans
698
+
699
+ def __len__(self) -> int:
700
+ return len(self.iterator)
701
+
702
+ def __add__(self, other) -> "LazyIteratorChain":
703
+ return LazyIteratorChain(self, other)
704
+
705
+
706
+ class LazyFlattener(Dillable):
707
+ """
708
+ A wrapper over an iterable of collections that flattens it to an iterable of items.
709
+
710
+ Example::
711
+
712
+ >>> list_of_cut_sets: List[CutSet] = [CutSet(...), CutSet(...)]
713
+ >>> list_of_cuts: List[Cut] = list(LazyFlattener(list_of_cut_sets))
714
+ """
715
+
716
+ def __init__(self, iterator: Iterable) -> None:
717
+ self.iterator = iterator
718
+
719
+ def __iter__(self):
720
+ for cuts in self.iterator:
721
+ yield from cuts
722
+
723
+ def __add__(self, other) -> "LazyIteratorChain":
724
+ return LazyIteratorChain(self, other)
725
+
726
+ def __len__(self) -> int:
727
+ raise TypeError(
728
+ "LazyFlattener does not support __len__ because it would require "
729
+ "iterating over the whole iterator, which is not possible in a lazy fashion. "
730
+ "If you really need to know the length, convert to eager mode first using "
731
+ "`.to_eager()`. Note that this will require loading the whole iterator into memory."
732
+ )
733
+
734
+
735
+ class LazyRepeater(Dillable):
736
+ """
737
+ A wrapper over an iterable that enables to repeat it N times or infinitely (default).
738
+ """
739
+
740
+ def __init__(
741
+ self, iterator: Iterable, times: Optional[int] = None, preserve_id: bool = False
742
+ ) -> None:
743
+ self.iterator = iterator
744
+ self.times = times
745
+ self.preserve_id = preserve_id
746
+ assert self.times is None or self.times > 0
747
+
748
+ def __iter__(self):
749
+ epoch = 0
750
+ while self.times is None or epoch < self.times:
751
+ if self.preserve_id:
752
+ iterator = self.iterator
753
+ else:
754
+ iterator = LazyMapper(
755
+ self.iterator, partial(attach_repeat_idx_to_id, idx=epoch)
756
+ )
757
+ at_least_once = False
758
+ for item in iterator:
759
+ at_least_once = True
760
+ yield item
761
+ if not at_least_once:
762
+ return # Detect empty iterables to avoid hanging the program.
763
+ epoch += 1
764
+
765
+ def __len__(self) -> int:
766
+ if self.times is None:
767
+ raise TypeError(
768
+ f"object of type '{type(self).__name__}' is an infinite iterator"
769
+ )
770
+ return len(self.iterator) * self.times
771
+
772
+ def __add__(self, other) -> "LazyIteratorChain":
773
+ return LazyIteratorChain(self, other)
774
+
775
+
776
+ class LazySlicer(Dillable):
777
+ """
778
+ A wrapper over an iterable that enables selecting k-th element every n elements.
779
+ """
780
+
781
+ def __init__(self, iterator: Iterable, k: int, n: int) -> None:
782
+ self.iterator = iterator
783
+ assert (
784
+ k < n
785
+ ), f"When selecting k-th element every n elements, k must be less than n (got k={k} n={n})."
786
+ self.k = k
787
+ self.n = n
788
+
789
+ def __iter__(self):
790
+ for idx, item in enumerate(self.iterator):
791
+ if idx % self.n == self.k:
792
+ yield item
793
+
794
+ def __add__(self, other) -> "LazyIteratorChain":
795
+ return LazyIteratorChain(self, other)
796
+
797
+ def __len__(self) -> int:
798
+ raise TypeError(
799
+ "LazySlicer does not support __len__ because it would require "
800
+ "iterating over the whole iterator, which is not possible in a lazy fashion. "
801
+ "If you really need to know the length, convert to eager mode first using "
802
+ "`.to_eager()`. Note that this will require loading the whole iterator into memory."
803
+ )
804
+
805
+
806
+ def attach_repeat_idx_to_id(item: Any, idx: int) -> Any:
807
+ if not hasattr(item, "id"):
808
+ return item
809
+ return fastcopy(item, id=f"{item.id}_repeat{idx}")
810
+
811
+
812
+ def count_newlines_fast(path: Pathlike):
813
+ """
814
+ Counts newlines in a file using buffered chunk reads.
815
+ The fastest possible option in Python according to:
816
+ https://stackoverflow.com/a/68385697/5285891
817
+ (This is a slightly modified variant of that answer.)
818
+ """
819
+
820
+ def _make_gen(reader):
821
+ b = reader(2**16)
822
+ while b:
823
+ yield b
824
+ b = reader(2**16)
825
+
826
+ read_mode = "rb" if not str(path) == "-" else "r"
827
+ with open_best(path, read_mode) as f:
828
+ count = sum(buf.count(b"\n") for buf in _make_gen(f.read))
829
+ return count
lazyShar.py ADDED
@@ -0,0 +1,356 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import random
2
+ import fsspec
3
+ from pathlib import Path
4
+ from typing import (
5
+ Callable,
6
+ Dict,
7
+ Generator,
8
+ List,
9
+ Literal,
10
+ Optional,
11
+ Sequence,
12
+ Tuple,
13
+ Union,
14
+ )
15
+
16
+ from lhotse.cut import Cut
17
+ from lhotse.dataset.dataloading import resolve_seed
18
+ from lhotse.lazy import (
19
+ Dillable,
20
+ LazyIteratorChain,
21
+ LazyJsonlIterator,
22
+ LazyManifestIterator,
23
+ count_newlines_fast,
24
+ )
25
+ from lhotse.serialization import extension_contains
26
+ from lhotse.shar.readers.tar import TarIterator
27
+ from lhotse.utils import Pathlike, exactly_one_not_null, ifnone
28
+
29
+
30
+ class LazySharIterator(Dillable):
31
+ """
32
+ LazySharIterator reads cuts and their corresponding data from multiple shards,
33
+ also recognized as the Lhotse Shar format.
34
+ Each shard is numbered and represented as a collection of one text manifest and
35
+ one or more binary tarfiles.
36
+ Each tarfile contains a single type of data, e.g., recordings, features, or custom fields.
37
+
38
+ Given an example directory named ``some_dir``, its expected layout is
39
+ ``some_dir/cuts.000000.jsonl.gz``, ``some_dir/recording.000000.tar``,
40
+ ``some_dir/features.000000.tar``, and then the same names but numbered with ``000001``, etc.
41
+ There may also be other files if the cuts have custom data attached to them.
42
+
43
+ The main idea behind Lhotse Shar format is to optimize dataloading with sequential reads,
44
+ while keeping the data composition more flexible than e.g. WebDataset tar archives do.
45
+ To achieve this, Lhotse Shar keeps each data type in a separate archive, along a single
46
+ CutSet JSONL manifest.
47
+ This way, the metadata can be investigated without iterating through the binary data.
48
+ The format also allows iteration over a subset of fields, or extension of existing data
49
+ with new fields.
50
+
51
+ As you iterate over cuts from ``LazySharIterator``, it keeps a file handle open for the
52
+ JSONL manifest and all of the tar files that correspond to the current shard.
53
+ The tar files are read item by item together, and their binary data is attached to
54
+ the cuts.
55
+ It can be normally accessed using methods such as ``cut.load_audio()``.
56
+
57
+ We can simply load a directory created by :class:`~lhotse.shar.writers.shar.SharWriter`.
58
+ Example::
59
+
60
+ >>> cuts = LazySharIterator(in_dir="some_dir")
61
+ ... for cut in cuts:
62
+ ... print("Cut", cut.id, "has duration of", cut.duration)
63
+ ... audio = cut.load_audio()
64
+ ... fbank = cut.load_features()
65
+
66
+ :class:`.LazySharIterator` can also be initialized from a dict, where the keys
67
+ indicate fields to be read, and the values point to actual shard locations.
68
+ This is useful when only a subset of data is needed, or it is stored in different
69
+ directories. Example::
70
+
71
+ >>> cuts = LazySharIterator({
72
+ ... "cuts": ["some_dir/cuts.000000.jsonl.gz"],
73
+ ... "recording": ["another_dir/recording.000000.tar"],
74
+ ... "features": ["yet_another_dir/features.000000.tar"],
75
+ ... })
76
+ ... for cut in cuts:
77
+ ... print("Cut", cut.id, "has duration of", cut.duration)
78
+ ... audio = cut.load_audio()
79
+ ... fbank = cut.load_features()
80
+
81
+ We also support providing shell commands as shard sources, inspired by WebDataset.
82
+ Example::
83
+
84
+ >>> cuts = LazySharIterator({
85
+ ... "cuts": ["pipe:curl https://my.page/cuts.000000.jsonl.gz"],
86
+ ... "recording": ["pipe:curl https://my.page/recording.000000.tar"],
87
+ ... })
88
+ ... for cut in cuts:
89
+ ... print("Cut", cut.id, "has duration of", cut.duration)
90
+ ... audio = cut.load_audio()
91
+
92
+ Finally, we allow specifying URLs or cloud storage URIs for the shard sources.
93
+ We defer to ``smart_open`` library to handle those.
94
+ Example::
95
+
96
+ >>> cuts = LazySharIterator({
97
+ ... "cuts": ["s3://my-bucket/cuts.000000.jsonl.gz"],
98
+ ... "recording": ["s3://my-bucket/recording.000000.tar"],
99
+ ... })
100
+ ... for cut in cuts:
101
+ ... print("Cut", cut.id, "has duration of", cut.duration)
102
+ ... audio = cut.load_audio()
103
+
104
+ :param fields: a dict whose keys specify which fields to load,
105
+ and values are lists of shards (either paths or shell commands).
106
+ The field "cuts" pointing to CutSet shards always has to be present.
107
+ :param in_dir: path to a directory created with ``SharWriter`` with
108
+ all the shards in a single place. Can be used instead of ``fields``.
109
+ :param split_for_dataloading: bool, by default ``False`` which does nothing.
110
+ Setting it to ``True`` is intended for PyTorch training with multiple
111
+ dataloader workers and possibly multiple DDP nodes.
112
+ It results in each node+worker combination receiving a unique subset
113
+ of shards from which to read data to avoid data duplication.
114
+ This is mutually exclusive with ``seed='randomized'``.
115
+ :param shuffle_shards: bool, by default ``False``. When ``True``, the shards
116
+ are shuffled (in case of multi-node training, the shuffling is the same
117
+ on each node given the same seed).
118
+ :param seed: When ``shuffle_shards`` is ``True``, we use this number to
119
+ seed the RNG.
120
+ Seed can be set to ``'randomized'`` in which case we expect that the user provided
121
+ :func:`lhotse.dataset.dataloading.worker_init_fn` as DataLoader's ``worker_init_fn``
122
+ argument. It will cause the iterator to shuffle shards differently on each node
123
+ and dataloading worker in PyTorch training. This is mutually exclusive with
124
+ ``split_for_dataloading=True``.
125
+ Seed can be set to ``'trng'`` which, like ``'randomized'``, shuffles the shards
126
+ differently on each iteration, but is not possible to control (and is not reproducible).
127
+ ``trng`` mode is mostly useful when the user has limited control over the training loop
128
+ and may not be able to guarantee internal Shar epoch is being incremented, but needs
129
+ randomness on each iteration (e.g. useful with PyTorch Lightning).
130
+ :param stateful_shuffle: bool, by default ``True``. When ``True``, every
131
+ time this object is fully iterated, it increments an internal epoch counter
132
+ and triggers shard reshuffling with RNG seeded by ``seed`` + ``epoch``.
133
+ Doesn't have any effect when ``shuffle_shards`` is ``False``.
134
+ :param cut_map_fns: optional sequence of callables that accept cuts and return cuts.
135
+ It's expected to have the same length as the number of shards, so each function
136
+ corresponds to a specific shard.
137
+ It can be used to attach shard-specific custom attributes to cuts.
138
+ :param slice_length: optional int, when set enables random slicing of shards that
139
+ may improve sampling randomness for many-dataset-with-many-large-shards setups
140
+ at the cost of efficiency. In this mode, we randomly select K to skip first K examples
141
+ and read only ``slice_length`` examples from each shard, then move to the next one.
142
+
143
+ See also: :class:`~lhotse.shar.writers.shar.SharWriter`
144
+ """
145
+
146
+ def __init__(
147
+ self,
148
+ fields: Optional[Dict[str, Sequence[Pathlike]]] = None,
149
+ in_dir: Optional[Pathlike] = None,
150
+ split_for_dataloading: bool = False,
151
+ shuffle_shards: bool = False,
152
+ stateful_shuffle: bool = True,
153
+ seed: Union[int, Literal["randomized"], Literal["trng"]] = 42,
154
+ cut_map_fns: Optional[Sequence[Callable[[Cut], Cut]]] = None,
155
+ slice_length: Optional[int] = None,
156
+ ) -> None:
157
+ assert exactly_one_not_null(
158
+ fields, in_dir
159
+ ), "To read Lhotse Shar format, provide either 'in_dir' or 'fields' argument."
160
+ if split_for_dataloading:
161
+ assert seed not in ("randomized", "trng"), (
162
+ "Error: setting seed to 'randomized' or 'trng' and using split_for_dataloading=True "
163
+ "are mutually exclusive options as they would result in data loss."
164
+ )
165
+
166
+ self.split_for_dataloading = split_for_dataloading
167
+ self.shuffle_shards = shuffle_shards
168
+ self.stateful_shuffle = stateful_shuffle
169
+ self.seed = seed
170
+ self.slice_length = slice_length
171
+ self.epoch = 0
172
+
173
+ self._len = None
174
+ if in_dir is not None:
175
+ self._init_from_dir(in_dir)
176
+ else:
177
+ self._init_from_inputs(fields)
178
+
179
+ self.num_shards = len(self.streams["cuts"])
180
+ for field in self.fields:
181
+ assert (
182
+ len(self.streams[field]) == self.num_shards
183
+ ), f"Expected {self.num_shards} shards available for field '{field}' but found {len(self.streams[field])}: {self.streams[field]}"
184
+
185
+ self.shards = [
186
+ {field: self.streams[field][shard_idx] for field in self.streams}
187
+ for shard_idx in range(self.num_shards)
188
+ ]
189
+
190
+ self.cut_map_fns = ifnone(cut_map_fns, [None] * self.num_shards)
191
+
192
+ def _init_from_inputs(self, fields: Optional[Dict[str, Sequence[str]]] = None):
193
+ assert (
194
+ "cuts" in fields
195
+ ), "To initialize Shar reader, please provide the value for key 'cuts' in 'fields'."
196
+ self.fields = set(fields.keys())
197
+ self.fields.remove("cuts")
198
+ self.streams = fields
199
+
200
+ def _init_from_dir(self, in_dir: Pathlike):
201
+ """
202
+ Patched to support fsspec cloud URIs (hf://, s3://, gs://, etc.)
203
+ as well as local directories.
204
+ """
205
+ self.in_dir = str(in_dir)
206
+
207
+ # 1. Resolve filesystem using fsspec
208
+ fs, fs_path = fsspec.core.url_to_fs(self.in_dir)
209
+ protocol = self.in_dir.split("://")[0] + "://" if "://" in self.in_dir else ""
210
+
211
+ # 2. List all items in the directory/bucket
212
+ raw_ls = fs.ls(fs_path)
213
+
214
+ all_paths = []
215
+ for item in raw_ls:
216
+ # fsspec returns dicts for remote objects, strings for local
217
+ path_str = item['name'] if isinstance(item, dict) else item
218
+
219
+ # Reattach the protocol if fsspec stripped it
220
+ if protocol and not path_str.startswith(protocol):
221
+ full_path = protocol + path_str
222
+ else:
223
+ full_path = path_str
224
+ all_paths.append(full_path)
225
+
226
+ # 3. Helper to extract prefix (stem equivalent) without pathlib breaking on URIs
227
+ def get_prefix(p_str: str) -> str:
228
+ filename = p_str.split("/")[-1]
229
+ return filename.split(".")[0]
230
+
231
+ # 4. Discover fields and validate
232
+ self.fields = set(get_prefix(p) for p in all_paths)
233
+ print(f"Found the following fields in '{self.in_dir}': {self.fields}")
234
+ assert "cuts" in self.fields, f"Could not find 'cuts' in {self.in_dir}"
235
+ self.fields.remove("cuts")
236
+
237
+ # 5. Populate streams
238
+ self.streams = {
239
+ "cuts": sorted(
240
+ p for p in all_paths
241
+ if get_prefix(p) == "cuts" and extension_contains(".jsonl", p)
242
+ )
243
+ }
244
+ for field in self.fields:
245
+ self.streams[field] = sorted(
246
+ p for p in all_paths if get_prefix(p) == field
247
+ )
248
+
249
+ def _maybe_split_for_dataloading(self, shards: List) -> List:
250
+ from .utils import split_by_node, split_by_worker
251
+
252
+ if self.split_for_dataloading:
253
+ return split_by_worker(split_by_node(shards))
254
+ else:
255
+ return shards
256
+
257
+ def _get_rng(self) -> random.Random:
258
+ seed = resolve_seed(self.seed)
259
+ if self.stateful_shuffle:
260
+ seed += self.epoch
261
+ return random.Random(seed)
262
+
263
+ def _maybe_shuffle_shards(self, shards: List) -> List:
264
+ if self.shuffle_shards:
265
+ shards = shards.copy()
266
+ self._get_rng().shuffle(shards)
267
+ return shards
268
+
269
+ def __iter__(self):
270
+ shards, map_fns = self.shards, self.cut_map_fns
271
+ rng = self._get_rng()
272
+ shards = self._maybe_shuffle_shards(shards)
273
+ shards = self._maybe_split_for_dataloading(shards)
274
+ if map_fns is not None:
275
+ # The functions also need to be shuffled/split, if present.
276
+ map_fns = self._maybe_shuffle_shards(map_fns)
277
+ map_fns = self._maybe_split_for_dataloading(map_fns)
278
+
279
+ for shard, cut_map_fn in zip(shards, map_fns):
280
+ # Iterate over cuts for the current shard
281
+ cuts = LazyManifestIterator(shard["cuts"])
282
+ if self.slice_length is not None:
283
+ # Sampling a slicing offset requires to know the length
284
+ cuts = list(cuts)
285
+
286
+ # Iterate over tarfiles/jsonl containing data for specific fields of each cut
287
+ field_paths = {
288
+ field: path for field, path in shard.items() if field != "cuts"
289
+ }
290
+
291
+ # Open every tarfile/jsonl so it's ready for streaming
292
+ field_iters = {
293
+ field: TarIterator(path)
294
+ if extension_contains(".tar", path)
295
+ else _jsonl_tar_adaptor(LazyJsonlIterator(path), field=field)
296
+ for field, path in field_paths.items()
297
+ }
298
+
299
+ # *field_data contains all fields for a single cut (recording, features, array, etc.)
300
+ yielded_cntr = 0
301
+ slice_offset = (
302
+ rng.randint(0, len(cuts) - self.slice_length)
303
+ if self.slice_length is not None and self.slice_length < len(cuts)
304
+ else -1
305
+ )
306
+ for idx, (cut, *field_data) in enumerate(zip(cuts, *field_iters.values())):
307
+ if idx < slice_offset:
308
+ continue
309
+ elif yielded_cntr == self.slice_length:
310
+ break
311
+
312
+ # Filling shar placeholders with actual data from tar files etc.
313
+ for (field, (maybe_manifest, data_path)) in zip(
314
+ field_iters.keys(),
315
+ field_data,
316
+ ):
317
+ if maybe_manifest is None:
318
+ continue # No value available for the current field for this cut.
319
+ assert (
320
+ str(data_path.parent / data_path.stem) == cut.id
321
+ ), f"Mismatched IDs: cut ID is '{cut.id}' but found data with name '{data_path}' for field {field}"
322
+ setattr(cut, field, maybe_manifest)
323
+
324
+ cut.shard_origin = shard["cuts"]
325
+ cut.shar_epoch = self.epoch
326
+ if cut_map_fn is not None:
327
+ cut = cut_map_fn(cut)
328
+ yield cut
329
+ yielded_cntr += 1
330
+
331
+ self.epoch += 1
332
+
333
+ def __len__(self) -> int:
334
+ if self._len is None:
335
+ self._len = sum(count_newlines_fast(p) for p in self.streams["cuts"])
336
+ return self._len
337
+
338
+ def __add__(self, other) -> "LazyIteratorChain":
339
+ return LazyIteratorChain(self, other)
340
+
341
+
342
+ def _jsonl_tar_adaptor(
343
+ jsonl_iter: LazyJsonlIterator, field: str
344
+ ) -> Generator[Tuple[Optional[dict], Path], None, None]:
345
+ """
346
+ Used to adapt the iteration output of LazyJsonlIterator to mimic that of TarIterator.
347
+ """
348
+ for item in jsonl_iter:
349
+ # Add extension to make sure Path.stem works OK...
350
+ pseudo_path = Path(f"{item['cut_id']}.dummy")
351
+ if field not in item:
352
+ # We got a placeholder
353
+ item = None
354
+ else:
355
+ item = item[field]
356
+ yield item, pseudo_path
license.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+
2
+ Various files include modifications (c) NVIDIA CORPORATION & AFFILIATES. All rights reserved.
3
+
4
+ GOVERNING TERMS: The software and materials are governed by the NVIDIA Software License Agreement
5
+ (found at https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-license-agreement/)
6
+ and the Product-Specific Terms for NVIDIA AI Products
7
+ (found at https://www.nvidia.com/en-us/agreements/enterprise-software/product-specific-terms-for-ai-products/).
makeyaml.ipynb ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/lightning_logs.txt ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/nemo_error_log.txt ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 11:18:02 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 11:18:03 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 11:23:01 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 11:23:01 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 11:23:05 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 11:23:05 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 11:23:05 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 11:23:09 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 11:23:10 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: allow_missing_data,is_tarred
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-18-03/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/lightning_logs.txt ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/nemo_error_log.txt ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 11:42:26 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 11:42:27 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 11:46:53 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 11:46:53 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 11:46:57 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 11:46:57 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 11:46:57 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 11:47:01 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 11:47:02 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: allow_missing_data,is_tarred
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-42-27/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/lightning_logs.txt ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/nemo_error_log.txt ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 11:48:34 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 11:48:34 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 11:49:38 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 11:49:38 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 11:49:42 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 11:49:42 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 11:49:42 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 11:49:46 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 11:49:47 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: allow_missing_data,is_tarred
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-48-34/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/lightning_logs.txt ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/nemo_error_log.txt ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 11:54:04 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 11:54:05 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 11:55:09 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 11:55:09 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 11:55:12 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 11:55:12 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 11:55:12 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 11:55:16 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 11:55:17 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: is_tarred,allow_missing_data
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_11-54-05/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/lightning_logs.txt ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/nemo_error_log.txt ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 12:12:11 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 12:12:11 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 12:16:38 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 12:16:38 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 12:16:41 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 12:16:41 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 12:16:41 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 12:16:46 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 12:16:47 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: is_tarred,allow_missing_data
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-12-11/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/events.out.tfevents.1775305282.d12a7902a35c.540.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:274e9dfb7ac1204275503319612ed5da784fcf5001dd5ff78a1bbeaa47750342
3
+ size 626485
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/hparams.yaml ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/lightning_logs.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
2
+
3
+ | Name | Type | Params | Mode
4
+ --------------------------------------------------------------------------------
5
+ 0 | preprocessor | AudioToMelSpectrogramPreprocessor | 0 | train
6
+ 1 | encoder | ConformerEncoder | 608 M | train
7
+ 2 | spec_augmentation | SpectrogramAugmentation | 0 | train
8
+ 3 | wer | WER | 0 | train
9
+ 4 | joint | RNNTJoint | 22.1 M | train
10
+ 5 | decoder | RNNTDecoder | 27.5 M | train
11
+ 6 | loss | RNNTLoss | 0 | train
12
+ 7 | spec_augment | SpectrogramAugmentation | 0 | train
13
+ --------------------------------------------------------------------------------
14
+ 658 M Trainable params
15
+ 0 Non-trainable params
16
+ 658 M Total params
17
+ 2,633.960 Total estimated model params size (MB)
18
+ 708 Modules in train mode
19
+ 0 Modules in eval mode
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/nemo_error_log.txt ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 12:20:06 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 12:20:06 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 12:21:10 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 12:21:10 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 12:21:14 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 12:21:14 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 12:21:14 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 12:21:18 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 12:21:19 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: allow_missing_data,is_tarred
46
+ [NeMo W 2026-04-04 12:21:21 audio_to_text_dataset:833] Could not load dataset as `manifest_filepath` was None. Provided config : {'manifest_filepath': None, 'sample_rate': 16000, 'batch_size': 16, 'shuffle': False, 'use_start_end_token': False, 'num_workers': 8, 'pin_memory': True}
47
+ [NeMo W 2026-04-04 12:21:27 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/lightning/pytorch/loops/fit_loop.py:298: The number of training batches (1) is smaller than the logging interval Trainer(log_every_n_steps=10). Set a lower value for log_every_n_steps if you want to see logs for the training epoch.
48
+
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-20-06/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/events.out.tfevents.1775305542.d12a7902a35c.1101.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e78d7b83a7c4161e082ac67c8fc8d0d3081eaef37e5cd333c5017ed9926aac6
3
+ size 626485
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/hparams.yaml ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/lightning_logs.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
2
+
3
+ | Name | Type | Params | Mode
4
+ --------------------------------------------------------------------------------
5
+ 0 | preprocessor | AudioToMelSpectrogramPreprocessor | 0 | train
6
+ 1 | encoder | ConformerEncoder | 608 M | train
7
+ 2 | spec_augmentation | SpectrogramAugmentation | 0 | train
8
+ 3 | wer | WER | 0 | train
9
+ 4 | joint | RNNTJoint | 22.1 M | train
10
+ 5 | decoder | RNNTDecoder | 27.5 M | train
11
+ 6 | loss | RNNTLoss | 0 | train
12
+ 7 | spec_augment | SpectrogramAugmentation | 0 | train
13
+ --------------------------------------------------------------------------------
14
+ 658 M Trainable params
15
+ 0 Non-trainable params
16
+ 658 M Total params
17
+ 2,633.960 Total estimated model params size (MB)
18
+ 708 Modules in train mode
19
+ 0 Modules in eval mode
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/nemo_error_log.txt ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 12:24:26 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 12:24:26 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
5
+ [NeMo W 2026-04-04 12:25:30 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
6
+ Train config :
7
+ use_lhotse: true
8
+ skip_missing_manifest_entries: true
9
+ input_cfg: null
10
+ tarred_audio_filepaths: null
11
+ manifest_filepath: null
12
+ sample_rate: 16000
13
+ shuffle: true
14
+ num_workers: 2
15
+ pin_memory: true
16
+ max_duration: 10.0
17
+ min_duration: 1.0
18
+ text_field: answer
19
+ batch_duration: null
20
+ max_tps: null
21
+ use_bucketing: true
22
+ bucket_duration_bins: null
23
+ bucket_batch_size: null
24
+ num_buckets: 30
25
+ bucket_buffer_size: 20000
26
+ shuffle_buffer_size: 10000
27
+
28
+ [NeMo W 2026-04-04 12:25:30 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s).
29
+ Validation config :
30
+ use_lhotse: true
31
+ manifest_filepath: null
32
+ sample_rate: 16000
33
+ batch_size: 16
34
+ shuffle: false
35
+ max_duration: 40.0
36
+ min_duration: 0.1
37
+ num_workers: 2
38
+ pin_memory: true
39
+ text_field: answer
40
+
41
+ [NeMo W 2026-04-04 12:25:33 modelPT:300] You tried to register an artifact under config key=tokenizer.model_path but an artifact for it has already been registered.
42
+ [NeMo W 2026-04-04 12:25:33 modelPT:300] You tried to register an artifact under config key=tokenizer.vocab_path but an artifact for it has already been registered.
43
+ [NeMo W 2026-04-04 12:25:33 modelPT:300] You tried to register an artifact under config key=tokenizer.spe_tokenizer_vocab but an artifact for it has already been registered.
44
+ [NeMo W 2026-04-04 12:25:38 speech_to_text_finetune:167] The vocabulary size of the new tokenizer differs from that of the loaded model. As a result, finetuning will proceed with the new vocabulary, and the decoder will be reinitialized.
45
+ [NeMo W 2026-04-04 12:25:38 dataloader:826] The following configuration keys are ignored by Lhotse dataloader: is_tarred,allow_missing_data
46
+ [NeMo W 2026-04-04 12:25:40 audio_to_text_dataset:833] Could not load dataset as `manifest_filepath` was None. Provided config : {'manifest_filepath': None, 'sample_rate': 16000, 'batch_size': 16, 'shuffle': False, 'use_start_end_token': False, 'num_workers': 8, 'pin_memory': True}
47
+ [NeMo W 2026-04-04 12:25:46 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/lightning/pytorch/loops/fit_loop.py:298: The number of training batches (1) is smaller than the logging interval Trainer(log_every_n_steps=10). Set a lower value for log_every_n_steps if you want to see logs for the training epoch.
48
+
49
+ [NeMo W 2026-04-04 12:25:51 audio_preprocessing:85] AudioPreprocessor received an input signal of dtype torch.bfloat16, rather than torch.float32. In sweeps across multiple datasets, we have found that the preprocessor is not robust to low precision mathematics. As such, it runs in float32. Your input will be cast to float32, but this is not necessarily enough to recovery full accuracy. For example, simply casting input_signal from torch.float32 to torch.bfloat16, then back to torch.float32 before running AudioPreprocessor causes drops in absolute WER of up to 0.1%. torch.bfloat16 simply does not have enough mantissa bits to represent enough values in the range [-1.0,+1.0] correctly.
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-24-26/nemo_log_globalrank-0_localrank-0.txt ADDED
The diff for this file is too large to render. See raw diff
 
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/cmd-args.log ADDED
@@ -0,0 +1 @@
 
 
1
+ ./speech_to_text_finetune.py --config-path=. --config-name=speech_to_text_finetune trainer.accumulate_grad_batches=16 trainer.devices=1 trainer.accelerator=gpu +trainer.limit_train_batches=null exp_manager.create_wandb_logger=True exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/lightning_logs.txt ADDED
File without changes
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/nemo_error_log.txt ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ [NeMo W 2026-04-04 12:27:59 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo W 2026-04-04 12:28:00 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59/nemo_log_globalrank-0_localrank-0.txt ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [NeMo W 2026-04-04 12:27:59 nemo_logging:364] /opt/venv/lib/python3.12/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
2
+ warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
3
+
4
+ [NeMo I 2026-04-04 12:27:59 speech_to_text_finetune:198] Hydra config: name: Speech_To_Text_Finetuning
5
+ init_from_pretrained_model: nvidia/parakeet-tdt-0.6b-v3
6
+ model:
7
+ sample_rate: 16000
8
+ activation_checkpointing: true
9
+ train_ds:
10
+ shar_path: hf://buckets/Tyl3rDrden/IvritKnessetRecordingsData
11
+ use_lhotse: true
12
+ is_tarred: true
13
+ batch_size: 1
14
+ allow_missing_data: true
15
+ shuffle: false
16
+ max_duration: 30.1
17
+ num_workers: 8
18
+ pin_memory: true
19
+ validation_ds:
20
+ manifest_filepath: null
21
+ sample_rate: ${model.sample_rate}
22
+ batch_size: 16
23
+ shuffle: false
24
+ use_start_end_token: false
25
+ num_workers: 8
26
+ pin_memory: true
27
+ test_ds:
28
+ manifest_filepath: null
29
+ sample_rate: ${model.sample_rate}
30
+ batch_size: 16
31
+ shuffle: false
32
+ use_start_end_token: false
33
+ num_workers: 8
34
+ pin_memory: true
35
+ char_labels:
36
+ update_labels: false
37
+ labels: null
38
+ tokenizer:
39
+ update_tokenizer: true
40
+ dir: ./tokenizer_spe_bpe_v32768_pad_bos_eos
41
+ type: bpe
42
+ spec_augment:
43
+ _target_: nemo.collections.asr.modules.SpectrogramAugmentation
44
+ freq_masks: 2
45
+ time_masks: 10
46
+ freq_width: 27
47
+ time_width: 0.05
48
+ optim:
49
+ name: adamw
50
+ lr: 0.0001
51
+ betas:
52
+ - 0.9
53
+ - 0.98
54
+ weight_decay: 0.001
55
+ sched:
56
+ name: CosineAnnealing
57
+ warmup_steps: 5000
58
+ warmup_ratio: null
59
+ min_lr: 5.0e-06
60
+ trainer:
61
+ devices: 1
62
+ num_nodes: 1
63
+ max_epochs: -1
64
+ max_steps: 1000000
65
+ val_check_interval: 2000
66
+ accelerator: gpu
67
+ strategy:
68
+ _target_: lightning.pytorch.strategies.DDPStrategy
69
+ gradient_as_bucket_view: true
70
+ accumulate_grad_batches: 16
71
+ gradient_clip_val: 0.0
72
+ precision: bf16
73
+ log_every_n_steps: 10
74
+ enable_progress_bar: true
75
+ num_sanity_val_steps: 0
76
+ check_val_every_n_epoch: null
77
+ sync_batchnorm: true
78
+ enable_checkpointing: false
79
+ logger: false
80
+ benchmark: false
81
+ limit_train_batches: null
82
+ exp_manager:
83
+ exp_dir: null
84
+ name: ${name}
85
+ create_tensorboard_logger: true
86
+ create_checkpoint_callback: true
87
+ checkpoint_callback_params:
88
+ monitor: val_wer
89
+ mode: min
90
+ save_top_k: 5
91
+ always_save_nemo: true
92
+ resume_if_exists: false
93
+ resume_ignore_no_checkpoint: false
94
+ create_wandb_logger: true
95
+ wandb_logger_kwargs:
96
+ name: parakeet_v3_finetune_fixed
97
+ project: parakeet-hebrew-asr
98
+
99
+ [NeMo I 2026-04-04 12:27:59 exp_manager:594] ExpManager schema
100
+ [NeMo I 2026-04-04 12:27:59 exp_manager:595] {'explicit_log_dir': None, 'exp_dir': None, 'name': None, 'version': None, 'use_datetime_version': True, 'resume_if_exists': False, 'resume_past_end': False, 'resume_ignore_no_checkpoint': False, 'resume_from_checkpoint': None, 'create_tensorboard_logger': True, 'summary_writer_kwargs': None, 'create_wandb_logger': False, 'wandb_logger_kwargs': None, 'create_mlflow_logger': False, 'mlflow_logger_kwargs': {'experiment_name': None, 'run_name': None, 'tracking_uri': None, 'tags': None, 'save_dir': './mlruns', 'prefix': '', 'artifact_location': None, 'run_id': None, 'log_model': False}, 'create_dllogger_logger': False, 'dllogger_logger_kwargs': {'verbose': False, 'stdout': False, 'json_file': './dllogger.json'}, 'create_clearml_logger': False, 'clearml_logger_kwargs': {'project': None, 'task': None, 'connect_pytorch': False, 'model_name': None, 'tags': None, 'log_model': False, 'log_cfg': False, 'log_metrics': False}, 'create_neptune_logger': False, 'neptune_logger_kwargs': None, 'create_checkpoint_callback': True, 'checkpoint_callback_params': {'filepath': None, 'dirpath': None, 'filename': None, 'monitor': 'val_loss', 'verbose': True, 'save_last': True, 'save_top_k': 3, 'save_weights_only': False, 'mode': 'min', 'auto_insert_metric_name': True, 'every_n_epochs': 1, 'every_n_train_steps': None, 'train_time_interval': None, 'prefix': None, 'postfix': '.nemo', 'save_best_model': False, 'always_save_nemo': False, 'save_nemo_on_train_end': True, 'model_parallel_size': None, 'save_on_train_epoch_end': False, 'async_save': False, 'save_last_n_optim_states': -1}, 'create_early_stopping_callback': False, 'create_ipl_epoch_stopper_callback': False, 'early_stopping_callback_params': {'monitor': 'val_loss', 'mode': 'min', 'min_delta': 0.001, 'patience': 10, 'verbose': True, 'strict': True, 'check_finite': True, 'stopping_threshold': None, 'divergence_threshold': None, 'check_on_train_epoch_end': None, 'log_rank_zero_only': False}, 'ipl_epoch_stopper_callback_params': {'enable_stop': True, 'stop_every_n_epochs': 1}, 'create_preemption_callback': True, 'files_to_copy': None, 'log_step_timing': True, 'log_delta_step_timing': False, 'step_timing_kwargs': {'reduction': 'mean', 'sync_cuda': False, 'buffer_size': 1}, 'log_local_rank_0_only': False, 'log_global_rank_0_only': False, 'disable_validation_on_resume': True, 'ema': {'enable': False, 'decay': 0.999, 'cpu_offload': False, 'validate_original_weights': False, 'every_n_steps': 1}, 'max_time_per_run': None, 'seconds_to_sleep': 5.0, 'create_straggler_detection_callback': False, 'straggler_detection_params': {'report_time_interval': 300.0, 'calc_relative_gpu_perf': True, 'calc_individual_gpu_perf': True, 'num_gpu_perf_scores_to_log': 5, 'gpu_relative_perf_threshold': 0.7, 'gpu_individual_perf_threshold': 0.7, 'stop_if_detected': False}, 'create_fault_tolerance_callback': False, 'fault_tolerance': {'workload_check_interval': 5.0, 'initial_rank_heartbeat_timeout': 3600.0, 'rank_heartbeat_timeout': 2700.0, 'calculate_timeouts': True, 'safety_factor': 5.0, 'rank_termination_signal': <Signals.SIGKILL: 9>, 'log_level': 'INFO', 'max_rank_restarts': 0, 'max_subsequent_job_failures': 0, 'additional_ft_launcher_args': '', 'simulated_fault': None}, 'log_tflops_per_sec_per_gpu': True}
101
+ [NeMo I 2026-04-04 12:27:59 exp_manager:655] Experiments will be logged at /workspace/nemo_experiments/Speech_To_Text_Finetuning/2026-04-04_12-27-59
102
+ [NeMo I 2026-04-04 12:28:00 exp_manager:1262] TensorboardLogger has been set up
103
+ [NeMo I 2026-04-04 12:28:00 exp_manager:1277] WandBLogger has been set up
104
+ [NeMo W 2026-04-04 12:28:00 exp_manager:1413] The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to 1000000. Please ensure that max_steps will run for at least None epochs to ensure that checkpointing will not error out.
105
+ [NeMo I 2026-04-04 12:28:00 exp_manager:804] TFLOPs per sec per GPU will be calculated, conditioned on supported models. Defaults to -1 upon failure.
106
+ [NeMo I 2026-04-04 12:28:00 speech_to_text_finetune:106] Sleeping for at least 60 seconds to wait for model download to finish.