[huggingface_pytorch] Training - update for Transformers to 4.41.2 PyTorch 2.2

[JingyaHuang]

GitHub Issue #3870:

This PR updates Hugginface's PyTorch DLC for inference. Here are the corresponding updated dependencies versions:

• transformers: 4.41.2
• datasets: 2.19.0
• evaluate: 0.4.2
• accelerate: 0.31.0
• torch: 2.2.0
• diffusers: 0.28.2
• trl: 0.9.4
• peft: 0.11.1
• flash-attn:2.5.8

Note:

• If merging this PR should also close the associated Issue, please also add that Issue # to the Linked Issues section on the right.

• All PR's are checked weekly for staleness. This PR will be closed if not updated in 30 days.

Description

Tests run

NOTE: By default, docker builds are disabled. In order to build your container, please update dlc_developer_config.toml and specify the framework to build in "build_frameworks"

• I have run builds/tests on commit for my changes.

NOTE: If you are creating a PR for a new framework version, please ensure success of the standard, rc, and efa sagemaker remote tests by updating the dlc_developer_config.toml file:

Expand

• sagemaker_remote_tests = true
• sagemaker_efa_tests = true
• sagemaker_rc_tests = true

Additionally, please run the sagemaker local tests in at least one revision:

• sagemaker_local_tests = true

Formatting

• I have run black -l 100 on my code (formatting tool: https://black.readthedocs.io/en/stable/getting_started.html)

DLC image/dockerfile

Builds to Execute

Expand

Click the checkbox to enable a build to execute upon merge.

Note: By default, pipelines are set to "latest". Replace with major.minor framework version if you do not want "latest".

• build_pytorch_training_latest
• build_pytorch_inference_latest
• build_tensorflow_training_latest
• build_tensorflow_inference_latest

Additional context

PR Checklist

Expand

• I've prepended PR tag with frameworks/job this applies to : [mxnet, tensorflow, pytorch] | [ei/neuron/graviton] | [build] | [test] | [benchmark] | [ec2, ecs, eks, sagemaker]
• If the PR changes affects SM test, I've modified dlc_developer_config.toml in my PR branch by setting sagemaker_tests = true and efa_tests = true
• If this PR changes existing code, the change fully backward compatible with pre-existing code. (Non backward-compatible changes need special approval.)
• (If applicable) I've documented below the DLC image/dockerfile this relates to
• (If applicable) I've documented below the tests I've run on the DLC image
• (If applicable) I've reviewed the licenses of updated and new binaries and their dependencies to make sure all licenses are on the Apache Software Foundation Third Party License Policy Category A or Category B license list. See https://www.apache.org/legal/resolved.html.
• (If applicable) I've scanned the updated and new binaries to make sure they do not have vulnerabilities associated with them.

NEURON/GRAVITON Testing Checklist

• When creating a PR:

• I've modified dlc_developer_config.toml in my PR branch by setting neuron_mode = true or graviton_mode = true

Benchmark Testing Checklist

• When creating a PR:

• I've modified dlc_developer_config.toml in my PR branch by setting ec2_benchmark_tests = true or sagemaker_benchmark_tests = true

Pytest Marker Checklist

Expand

• (If applicable) I have added the marker @pytest.mark.model("<model-type>") to the new tests which I have added, to specify the Deep Learning model that is used in the test (use "N/A" if the test doesn't use a model)
• (If applicable) I have added the marker @pytest.mark.integration("<feature-being-tested>") to the new tests which I have added, to specify the feature that will be tested
• (If applicable) I have added the marker @pytest.mark.multinode(<integer-num-nodes>) to the new tests which I have added, to specify the number of nodes used on a multi-node test
• (If applicable) I have added the marker @pytest.mark.processor(<"cpu"/"gpu"/"eia"/"neuron">) to the new tests which I have added, if a test is specifically applicable to only one processor type

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license. I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[JingyaHuang]

I temporarily set fp16=False to pass the smdp gpu training test, since otherwise we will get the following error:

ValueError: FP16 Mixed precision training with AMP or APEX (`--fp16`) and FP16 half precision evaluation 
(`--fp16_full_eval`) can only be used on CUDA or MLU devices or NPU devices or certain XPU devices (with IPEX).

Which hints that the cuda environment is not correctly setup. The HF DLC is based on the PT training DLC in which the cuda shall be setup, I wonder if half precision is working in the base DLC? And maybe I need to do something in our dlc to activate the cuda env?

[github-actions]

This PR has been marked stale as a result of being open for 30 days without activity or updates. Please remove the stale label or comment in order to keep this open, otherwise the PR will be closed in 5 days.

[github-actions]

This PR has been marked stale as a result of being open for 30 days without activity or updates. Please remove the stale label or comment in order to keep this open, otherwise the PR will be closed in 5 days.

[github-actions]

This PR has had no activity or updates in the last 5 days since being marked stale. Closing this PR as a result.