bugfix: inv_freq buffer in Llama RotaryEmbedding shouldn't be persistent #21
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
* In a more recent version of Transformers, inv_freq buffer is no longer persistent huggingface/transformers@95f96b4 * This causes a crash when someone tries to load a (converted) Llama checkpoint in NeMo (ie.: CodeLlama 7b)
|
Hi, Does anyone have an update on this? Basically, loading any of the recent HF checkpoints that rely on rotary position embeddings (ie.: Mistral, Llama), will result in a crash: Because HF is not serializing |
|
… -Adarsh
From: Eduardo Vellasques ***@***.***>
Reply-To: aws-neuron/neuronx-nemo-megatron ***@***.***>
Date: Thursday, April 4, 2024 at 11:41 AM
To: aws-neuron/neuronx-nemo-megatron ***@***.***>
Cc: Subscribed ***@***.***>
Subject: Re: [aws-neuron/neuronx-nemo-megatron] bugfix: inv_freq buffer in Llama RotaryEmbedding shouldn't be persistent (PR #21)
Hi,
Does anyone have an update on this? Basically, loading any of the recent HF checkpoints that rely on rotary position embeddings (ie.: Mistral, Llama), will result in a crash:
Missing key(s) in state_dict: "model.language_model.encoder.layers.0.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.1.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.2.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.3.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.4.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.5.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.6.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.7.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.8.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.9.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.10.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.11.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.12.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.13.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.14.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.15.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.16.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.17.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.18.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.19.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.20.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.21.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.22.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.23.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.24.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.25.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.26.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.27.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.28.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.29.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.30.self_attention.core_attention.rotary_emb.inv_freq", "model.language_model.encoder.layers.31.self_attention.core_attention.rotary_emb.inv_freq".
Because HF is not serializing inv_freq.
—
Reply to this email directly, view it on GitHub<#21 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AWRRYCJKTPS6WHOQUA247GTY3WNERAVCNFSM6AAAAABFDUQZXCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAMZXHEZTKOBQGM>.
You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
|
|
Can you resolve the merge conflicts? Then we can merge it |
Quick question, solving the conflict will involve replacing nemo/nemo/collections/nlp/modules/common/megatron/llama_module.py with nemo/nemo/collections/nlp/modules/common/megatron/falcon_module.py. I opened an issue about a bug in the Llama conversion scripts (basically they should aggregate |
I noticed that in a more recent release you guys handle that during checkpoint loading. I created another PR for the issue with checkpoint conversion (#26) so I'll close this one. |
|
Main issue was solved by another PR. |
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: N/A
Description of changes: Switches off persistent flag when registering inv_freq buffer
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.