Llama 3.1 8B #813
Conversation
|
MLCommons CLA bot: |
|
|
||
| <!-- # 6. Other | ||
|
|
||
| #### Run model conversion |
There was a problem hiding this comment.
Can we drop this section because there is no starting checkpoint for this benchmark. Submitters should not start from the HF checkpoint and instead they need to start from randomly initialized weights
| # This is the checkpoint that we want to start with. | ||
| # Each checkpoint should be a folder containing two sub-folders: context and weights. | ||
| # And we need to pass this folder's path (the folder containing context and weights) here. | ||
| export MODEL_CKPT="/data/llama3_8b/model/Llama-3.1-8B_nemo" |
There was a problem hiding this comment.
There is no model ckpt, this should not be set.
| export USE_CKPT=0 | ||
| # Model: Whether we are resuming from a NeMo-formatted HuggingFace checkpoint (weights only). | ||
| # If set to 1, then checkpoint resuming code will not try to load the optimizer states. | ||
| export FROM_HF=1 |
There was a problem hiding this comment.
Can we remove these flags since they are not relevant
| # This is the checkpoint that we want to start with. | ||
| # Each checkpoint should be a folder containing two sub-folders: context and weights. | ||
| # And we need to pass this folder's path (the folder containing context and weights) here. | ||
| export MODEL_CKPT="/data/llama3_8b/model/Llama-3.1-8B_nemo" |
| export USE_CKPT=0 | ||
| # Model: Whether we are resuming from a NeMo-formatted HuggingFace checkpoint (weights only). | ||
| # If set to 1, then checkpoint resuming code will not try to load the optimizer states. | ||
| export FROM_HF=1 |
| @@ -0,0 +1,10 @@ | |||
| if __name__ == "__main__": | |||
There was a problem hiding this comment.
This file is not necessary
| @@ -0,0 +1,23 @@ | |||
| #!/bin/bash | |||
There was a problem hiding this comment.
This file is not necessary
|
There is an issue with the CLA again - this needs to be fixed #813 (comment) |
2 participants
Continuing from the previous PR #799 with some cleanups.