Makes Llama checkpoint convertion compatible with fused up/gate projection #26
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
This was referenced Apr 17, 2024
|
Thanks @evellasques for the PR. Going over it! |
HolySahar
reviewed
Apr 22, 2024
| @@ -125,8 +126,7 @@ def convert_checkpoint(config_file, | |||
| "self_attention.dense.weight": (1, "self_attn.o_proj.weight", 1, 0), | |||
| "post_attention_layernorm.weight": (0, "post_attention_layernorm.weight", None, 0), | |||
| "self_attention.core_attention.rotary_emb.inv_freq": (0, "self_attn.rotary_emb.inv_freq", None, 0), | |||
| "mlp.dense_h_to_4h.weight": (1, "mlp.gate_proj.weight", 0, 0), | |||
| "mlp.dense_h_to_4h_2.weight": (1, "mlp.up_proj.weight", 0, 0), | |||
| "mlp.dense_h_to_4h.weight": (1, "mlp.gate_proj_up_proj.weight", 0, 0), | |||
There was a problem hiding this comment.
why considering "gate" and "up" proj to be fused for HF checkpoint? Shouldn't you split them from nemo checkpoint instead and then save as separate "gate" and "up" params for HF?
| @@ -217,6 +217,13 @@ def convert_checkpoint(config_file, | |||
| hf_model[hf_key_q], hf_model[hf_key_k], hf_model[hf_key_v] = torch.split(hf_model[hf_key], size_per_seg, dim=0) | |||
| hf_model.pop(hf_key) | |||
|
|
|||
| if "dense_h_to_4h" in k: | |||
There was a problem hiding this comment.
I think this is not accurate. "gate" and "proj" fusion is per tp rank in the nemo checkpoint. So, you can't first concatenate all tps and then split to "gate" and "proj". Instead you should split them for each tp rank.
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Outdated suggestions cannot be applied.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 #, 24
Description of changes:
Recent merging of up/down projection in Llama requires the equivalent merging in the HF to NeMo conversion scripts (and subsequent splitting in the NeMo to HF script).
This change fixes that for the following converters:
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.