End-to-End LLM Model Development with Torchtitan and Torchtune #341
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
SMHP: Remove 14k log lines from efa exporter LCC
Add conda and docker environment setups for 16.pytorch-capu-ddp test case.
Bump dcgm exporter version to correctly capture GPU utilization
NCCL 2.19.4 has performance regression.
Change nccl version to 2.20.3
Update 3.container-train.sbatch
This reverts commit da7a51d.
Typo in the name.
Rename 0.crate-conda-env.sh to 0.create-conda-env.sh
Updating CF template for HyperPod to support second private subnet
smp v2 llama2 training example using fp8
Update 1.conda-train.sbatch
Signed-off-by: Sean Smith <[email protected]>
Validate Json in preflight check
KeitaW
force-pushed
the
torchtitan-torchtune
branch
from
3ae455a to
64e0724
Compare
…distributed-training into torchtitan-torchtune
KeitaW
force-pushed
the
torchtitan-torchtune
branch
from
64e0724 to
00dfbf5
Compare
KeitaW
force-pushed
the
main
branch
2 times, most recently
from
44e448e to
1209815
Compare
…distributed-training into torchtitan-torchtune
KeitaW
force-pushed
the
torchtitan-torchtune
branch
from
436b58c to
952eba3
Compare
|
Basic functionalities have been implemented. Allow me to iterate on the other PRs... |
pbelevich
requested changes
Jun 11, 2024
pbelevich
reviewed
Jun 11, 2024
3.test_cases/torchtune/slurm/tutorials/e2e-llama3-70b-development/README.md
Outdated
Show resolved
Hide resolved
Co-authored-by: Pavel Belevich <[email protected]>
…ent/README.md Co-authored-by: Pavel Belevich <[email protected]>
Co-authored-by: Pavel Belevich <[email protected]>
pbelevich
requested changes
Jun 11, 2024
| * Evaluation | ||
| * Deployment | ||
|
|
||
| for details of each step, refer the [overview documentation](../../README.md). |
There was a problem hiding this comment.
Suggested change
| for details of each step, refer the [overview documentation](../../README.md). | |
| for details of each step, refer to the [overview documentation](../../README.md). |
| In this step, you will fine-tune the Llama3 model starting from the original checkpoint using the WikiText dataset. This process, known as Full-Parameter Finetuning, updates all the parameters in the original model. The configuration file used for this process is `./tutorials/e2e-llama3-70b-development/full_finetune_distributed.yaml`. | ||
|
|
||
| ### Memory Consumption Challenges | ||
| One of the primary challenges during such training is memory consumption. A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter plus activation memory (6 bytes for parameters in mixed precision training, 8 bytes for AdamW, and 4 bytes for other overheads). For more details on the anatomy, see the [Hugging Face blog post](https://huggingface.co/docs/transformers/model_memory_anatomy) blog post. This means that training a 70B parameter model would require more than 1.12 TB of accelerated memory, which far exceeds the 80 GB capacity of H100 accelerated memory. To address this issue, torchtune integrates PyTorch Fully Sharded Data Parallel (FSDP). |
There was a problem hiding this comment.
Suggested change
| One of the primary challenges during such training is memory consumption. A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter plus activation memory (6 bytes for parameters in mixed precision training, 8 bytes for AdamW, and 4 bytes for other overheads). For more details on the anatomy, see the [Hugging Face blog post](https://huggingface.co/docs/transformers/model_memory_anatomy) blog post. This means that training a 70B parameter model would require more than 1.12 TB of accelerated memory, which far exceeds the 80 GB capacity of H100 accelerated memory. To address this issue, torchtune integrates PyTorch Fully Sharded Data Parallel (FSDP). | |
| One of the primary challenges during such training is memory consumption. A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter(6 bytes for parameter in mixed precision training, 4 bytes for gradient and 8 bytes for AdamW optimizer states) plus activation memory. For more details on the anatomy, see the [Hugging Face blog post](https://huggingface.co/docs/transformers/model_memory_anatomy) blog post. This means that training a 70B parameter model would require more than 1.12 TiB of accelerator's memory, which far exceeds the 80 GB capacity of H100 memory. To address this issue, torchtune integrates PyTorch Fully Sharded Data Parallel (FSDP). |
| In this step, you will fine-tune the Llama3 model starting from the original checkpoint using the WikiText dataset. This process, known as Full-Parameter Finetuning, updates all the parameters in the original model. The configuration file used for this process is `./tutorials/e2e-llama3-70b-development/full_finetune_distributed.yaml`. | ||
|
|
||
| ### Memory Consumption Challenges | ||
| One of the primary challenges during such training is memory consumption. A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter plus activation memory (6 bytes for parameters in mixed precision training, 8 bytes for AdamW, and 4 bytes for other overheads). For more details on the anatomy, see the [Hugging Face blog post](https://huggingface.co/docs/transformers/model_memory_anatomy) blog post. This means that training a 70B parameter model would require more than 1.12 TB of accelerated memory, which far exceeds the 80 GB capacity of H100 accelerated memory. To address this issue, torchtune integrates PyTorch Fully Sharded Data Parallel (FSDP). |
There was a problem hiding this comment.
How was 1.12 TiB calculated?
70_000_000_000 * 18 = 1_260_000_000_000
1_260_000_000_000 / 1024 / 1024 / 1024 / 1024 = 1.15TiB
pbelevich
reviewed
Jun 11, 2024
| In this step, you will fine-tune the Llama3 model starting from the original checkpoint using the WikiText dataset. This process, known as Full-Parameter Finetuning, updates all the parameters in the original model. The configuration file used for this process is `./tutorials/e2e-llama3-70b-development/full_finetune_distributed.yaml`. | ||
|
|
||
| ### Memory Consumption Challenges | ||
| One of the primary challenges during such training is memory consumption. A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter plus activation memory (6 bytes for parameters in mixed precision training, 8 bytes for AdamW, and 4 bytes for other overheads). For more details on the anatomy, see the [Hugging Face blog post](https://huggingface.co/docs/transformers/model_memory_anatomy) blog post. This means that training a 70B parameter model would require more than 1.12 TB of accelerated memory, which far exceeds the 80 GB capacity of H100 accelerated memory. To address this issue, torchtune integrates PyTorch Fully Sharded Data Parallel (FSDP). |
There was a problem hiding this comment.
memory is not accelerated itself
pbelevich
reviewed
Jun 11, 2024
|
|
||
| ### Basic concepts and relevant configuration | ||
|
|
||
| **FSDP** is a distributed training feature designed to efficiently handle large model training by sharding model parameters, gradients, and optimizer states across multiple devices. This approach significantly reduces memory consumption and optimizes resource utilization, making it possible to train models that are too large to fit on a single GPU. In `torchtune` users can launch FSDP training job with command `tune run full_finetune_distributed`. |
There was a problem hiding this comment.
Suggested change
| **FSDP** is a distributed training feature designed to efficiently handle large model training by sharding model parameters, gradients, and optimizer states across multiple devices. This approach significantly reduces memory consumption and optimizes resource utilization, making it possible to train models that are too large to fit on a single GPU. In `torchtune` users can launch FSDP training job with command `tune run full_finetune_distributed`. | |
| **FSDP** is a distributed training technique designed to efficiently handle large model training by sharding model parameters, gradients, and optimizer states across multiple devices. This approach significantly reduces memory consumption and optimizes resource utilization, making it possible to train models that are too large to fit on a single GPU. In `torchtune` users can launch FSDP training job with command `tune run full_finetune_distributed`. |
pbelevich
reviewed
Jun 11, 2024
| --master_port $RANDOM | ||
| --nproc_per_node=8 | ||
| --nnodes $NNODES | ||
| --nnodes=$SLURM_JOB_NUM_NODES |
pbelevich
reviewed
Jun 11, 2024
| --master_port $RANDOM | ||
| --nproc_per_node=8 | ||
| --nnodes $NNODES | ||
| --nnodes=$SLURM_JOB_NUM_NODES |
pbelevich
reviewed
Jun 11, 2024
| sbatch tutorials/e2e-llama3-70b-development/full_finetune_distributed.sbatch | ||
| ``` | ||
|
|
||
| By default, this script launches the FSDP training job with two instances. Once the job has been scheduled, you will see the following outputs in the log file named `logs/full-finetuning*`: |
There was a problem hiding this comment.
I don't see where two instances are specified by default, I see only --nnodes 1 --nnodes=1 in sbatch files
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Issue #, if available:
Description of changes:
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.