composer

composer copied to clipboard

# What does this PR do? Allows `load_fsdp_monolith_rank0_only` and `use_orig_params` to both be set to `True` - this was initially disabled due to a pytorch issue I think, but has...

milocress

Fix TRT-LLM Multigpu Compatibility

1

# [Wip] What does this PR do? We need to use Composer to run our evaluation framework on TRT-LLM models. Unfortunately, this breaks in the Multi-GPU case. These fixes allow...

nik-mosaic

use flash attn fuse cross entropy loss to reduce metric memory usage

2

This PR uses fused cross entropy loss from flash attention in the metric LanguageCrossEntropy (also LanguagePerplexity). The current torch.nn.CrossEntropyLoss call needs `6 * seq_len * vocab_size` GPU memory, and can...

cli99

# What does this PR do? We are migrating ICL tasks from composer to foundry and need to deprecate the existing composer implementations. The migration PR is here: https://github.com/mosaicml/llm-foundry/pull/936 #...

bmosaicml

Set `torch.utils.data.DataLoader.batch_sampler` epoch if defined

1

## 🚀 Feature Request Current implementation of composer trainer call the `DistributedSampler.set_epoch` method only on the `Dataloader.sampler` attribute but not on the `Dataloader.batch_sampler`, even if it is defined. One example...

Ghelfi

TODO # What does this PR do? # What issue(s) does this change relate to? # Before submitting - [ ] Have you read the [contributor guidelines](https://github.com/mosaicml/composer/blob/dev/CONTRIBUTING.md)? - [ ]...

ShashankMosaicML

Device mismatch during evaluation when training on mps

7

I ran into an issue trying to train flan-t5 on an M1 using torchmetrics. Training metrics worked fine, but I got the following stacktrace when calculating evaluation metrics: ```commandline ......

erosenthal-square

Fix error on HYBRID_SHARD misconfiguraiton

1

# What does this PR do? If you use `HYBRID_SHARD` but only specify a single element list for `device_mesh`, it gets to this ValueError, but then the message within the...

abhi-mosaic

Support multi-node logging with mlflow run as child experiment

1

shitaoli-db

Forward metric name to `update_metric`

7

## 🚀 Feature Request ## Motivation `get_metrics` returns a mapping between strings and `Metric`s. However the name isn't available in the `update_metric`. Checking for types might not be sufficient to...

antoinebrl