mosaicml / composer

Supercharge Your Model Training

http://docs.mosaicml.com

mosaicml/composer Issues

Make Composer compatible with Python 3.12.
Closed 21 days ago2
"ValueError: Reused local directory" How should I reuse local file with streaming?
Closed 21 days ago1
resuming BERT training takes forever...
Closed a month ago5
Logs from all ranks
Closed a month ago1
self.labels not set in eval_forward() for huggingfacemodel if self.use_logits=False
Updated 2 months ago1
Finetune Composer model with Streaming dataset example
Updated 2 months ago1
Documentation: Dead links to images stored in GCS buckets in "Analyzing Traces" Tutorial
Updated 2 months ago
Unable to script model
Updated 2 months ago5
Torch script export not working with hugging face model
Updated 2 months ago1
FSDP Wrapping Alters Optimizer's Parameter Tracking Behavior
Closed 3 months ago8
Eval Loss
Updated 3 months ago1
Multi-TPU Training Support for Composer
Closed 3 months ago4
How to use composer under python3.8
Closed 3 months ago1
Save interval not working
Closed 3 months ago9
ONNX export with `dynamic_axes` does not work when applying `BlurPool`
Updated 4 months ago1
Resuming Training from Load Path
Closed 4 months ago1
Epoch length incorrectly calculated when using DDP
Closed 4 months ago5
Training stops after first pass of Evaluation when pretraining MosaicBert
Closed 4 months ago8
EVAL_STANDALONE_END documentation missing
Updated 4 months ago1
Support DDP with rank-dependent dataloader lengths
Closed 5 months ago2
Computing train metrics at a given frequency
Updated 5 months ago1
CUDA OOM error not caught with auto microbatching
Closed 5 months ago3
Autoresume and duration mismatch on reload
Updated 5 months ago12
Optional `CheckpointSaver` instantiation inside the `Trainer`
Closed 5 months ago9
TypeError: Subscripted generics cannot be used with class and instance checks
Updated 5 months ago2
NUMA affinity control
Closed 6 months ago2
Error when using `Trainer.compile_config={}` in DDP mode
Updated 6 months ago4
TORCH_NCCL_ASYNC_ERROR_HANDLING vs NCCL_ASYNC_ERROR_HANDLING at launcher.py
Closed 6 months ago3
Enable control over MLFlowLogger run_name str to match a pre-existing tag run_name in MLFlow and resume model training
Updated 6 months ago1
Allow default credentials for accessing GCS
Closed 6 months ago4
Memory leak due to the copy of Metric objects in Composer's trainer
Closed 6 months ago2
Composer default lr scheduler create spurious warnings when max training duration is in epochs
Closed 7 months ago2
Remove logic that pops labels from batch in composer/models/huggingface.py
Closed 7 months ago
Allow smaller `t_max` in schedulers
Updated 7 months ago8
LookSAM Optimizer
Closed 7 months ago3
Streaming Dataloader with Multiple Workers causes out-of-memory error
Closed 7 months ago
Safer Checkpoint File Format
Updated 7 months ago1
Remote file name in `MemorySnapshot` not being formatted
Updated 7 months ago1
Set `torch.utils.data.DataLoader.batch_sampler` epoch if defined
Closed 7 months ago2
Augment training batches with "on-the-fly" features
Updated 7 months ago
Crash during import when signal handler cannot be set
Closed 7 months ago
override serialized fields when resuming from checkpoint
Closed 7 months ago1
Different device error when running evaluation_loop on "mps" device with ComposeClassifier
Closed 8 months ago
Resuming from a monolithic FSDP checkpoint times out on multinode
Closed 8 months ago9
More compression options using CLI tools
Closed 8 months ago
The `Profiler` should upload generated traces to the logger.
Closed 8 months ago3
Update packaging from <23 to <24
Closed 8 months ago2
Add support for Pytorch 2.2.0
Closed 9 months ago1
`0.18.1` not available in PyPI
Closed 9 months ago3
Saving checkpoints on network drive fails due to symlinks
Updated 9 months ago1