Fix the rotary embedding computation in LLaMA #1544
Merged
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
Also run the reverse embedding stage in compute_dtype instead of full-precision. This is how HF does it, so helps get the numerics closer
mattdangerw
reviewed
Apr 2, 2024
| # If `cache_update_index` is a tensor, RotaryEmbedding expects it | ||
| # to have dtype `self.compute_dtype`. | ||
| start_index = ops.cast( | ||
| start_index, self.rotary_embedding_layer.compute_dtype |
There was a problem hiding this comment.
Won't this run into the same problems as #5 in https://unsloth.ai/blog/gemma-bugs? float16 or bfloat16 are both bad for an incrementing integer.
mattdangerw
reviewed
Apr 2, 2024
|
|
||
| # [batch_shape, seq_len, num_key_value_heads, head_dim] | ||
| # -> [batch_shape, seq_len, num_heads, head_dim] | ||
| key = ops.repeat(key, repeats=self.num_key_value_groups, axis=2) | ||
| value = ops.repeat(value, repeats=self.num_key_value_groups, axis=2) | ||
| key = ops.repeat(key, repeats=self._num_key_value_groups, axis=2) |
There was a problem hiding this comment.
should not be underscore I think
mattdangerw
approved these changes
Apr 2, 2024
SamanehSaadat
pushed a commit
to SamanehSaadat/keras-nlp
that referenced
this pull request
Apr 10, 2024
* Fix rotary embedding computation in LLaMA Also run the reverse embedding stage in compute_dtype instead of full-precision. This is how HF does it, so helps get the numerics closer * Don't cast start_index; save rope keys * Remove underscore from num_key_value_heads
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Outdated suggestions cannot be applied.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.
LLaMA backbone ignored the start_index parameter when computing the rotary embeddings which lead to numerical issues during generation. This PR fixes it along with the reverse embedding layer in both Mistral and LLaMA: run the reverse embedding stage in
compute_dtypeinstead of full-precision. This is how HF does it, so helps get the numerics closer.