Support Qwen3 and Gemma3 #81
Open
+60
−15
Conversation
This file contains hidden or 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
9ab6c0f to
e9d2ad9
Compare
Signed-off-by: alessiodevoto <devoto.alessio@gmail.com>
Signed-off-by: alessiodevoto <devoto.alessio@gmail.com>
e9d2ad9 to
ec53981
Compare
SimJeg
reviewed
Jun 12, 2025
kvpress/presses/base_press.py
Outdated
| @@ -127,13 +129,23 @@ def __call__(self, model: PreTrainedModel) -> Generator: | |||
| model : PreTrainedModel | |||
| Model to apply the compression method to | |||
| """ | |||
|
|
|||
| if not isinstance(model, (LlamaForCausalLM, MistralForCausalLM, Phi3ForCausalLM, Qwen2ForCausalLM)): | |||
| supported_models = ( | |||
There was a problem hiding this comment.
Maybe move this variable at the beginning of the file as a global variable with name SUPPORTED_MODELS ?
SimJeg
reviewed
Jun 12, 2025
| logger.warning(f"Model {type(model)} not tested") | ||
|
|
||
| hooks = [] | ||
| try: | ||
| for layer in model.model.layers: | ||
| if isinstance(model, Gemma3ForCausalLM) and layer.is_sliding: | ||
| # Skip layers with sliding window attention, only for Gemma3 |
There was a problem hiding this comment.
Should we add
logger.warning(f"Compression in Gemma3 is only applied to layer without sliding window attention")?
SimJeg
reviewed
Jun 12, 2025
| @@ -161,12 +163,16 @@ def duo_attention_on_the_fly(model, num_samples=50, q_len=500): | |||
| # Mean query | |||
| q = module.self_attn.q_proj(h) | |||
| q = q.view(1, q.shape[1], -1, d) | |||
| if isinstance(module, (Gemma3Attention, Qwen3Attention)): | |||
There was a problem hiding this comment.
much better to use modules than name of layers, thanks!
SimJeg
reviewed
Jun 12, 2025
| @@ -12,6 +12,10 @@ | |||
|
|
|||
| from kvpress.presses.scorer_press import ScorerPress | |||
There was a problem hiding this comment.
please put kvpress import after transformers import (good practice)
SimJeg
reviewed
Jun 12, 2025
kvpress/presses/snapkv_press.py
Outdated
| @@ -12,6 +12,10 @@ | |||
|
|
|||
There was a problem hiding this comment.
same here, kvpress imports should come after
|
Also @alessiodevoto please check if issue #80 involves other changes in this PR |
Signed-off-by: alessiodevoto <devoto.alessio@gmail.com>
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.
This addresses #76 to support the QK normalization used in Gemma3 and Qwen3 (+ updates library version).