Fix GPTQ for inputs with batch size != 1 and with seq len == 1 #3002
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
github-actions
bot
added
NNCF PT
ljaljushkin
force-pushed
the
nl/3d_activations_weight_compression
branch
from
91773fd
to
9add286
Compare
ljaljushkin
changed the title
ljaljushkin
force-pushed
the
nl/3d_activations_weight_compression
branch
from
9add286
to
1214ce8
Compare
|
This PR resolves 155538 |
andreyanufr
approved these changes
Oct 22, 2024
ljaljushkin
changed the title
ljaljushkin
commented
Oct 23, 2024
| @@ -264,7 +264,7 @@ def _quantize_weights( | |||
| scales.append(scale) | |||
| else: | |||
| if self._scale_estimation and block_compression_config.num_bits == 4: | |||
| activations = [inp.squeeze()[:, (i1 + i) : (i1 + i + group_size)] for inp in inputs] | |||
| activations = [inp[..., (i1 + i) : (i1 + i + group_size)] for inp in inputs] | |||
There was a problem hiding this comment.
Slice by last dimension, which is supposed to be hidden one.
It's aligned with processing statistics in activations_to_wc_statistics when reduction axes are all dimensions except the last one @nikita-savelyevv
| [ | ||
| LMLinearModel.INPUT_SHAPE, | ||
| [3, 5, 16], | ||
| [1, 1, 16], |
There was a problem hiding this comment.
Added test case for the bug 155538 with tiny-llama example. The root cause is not filtered data with sequence length == 1.
@kshpv
| "input_shape", | ||
| [ | ||
| LMLinearModel.INPUT_SHAPE, | ||
| [3, 5, 16], |
There was a problem hiding this comment.
test case to cover SD case with batch size != 1
nikita-savelyevv
approved these changes
Oct 23, 2024
alexsu52
approved these changes
Oct 24, 2024
andreyanufr
approved these changes
Oct 24, 2024
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.
Changes
GPTQ correctly processes inputs with batch size != 1 and with batch size and sequence length equal 1.
Also changed the errors we are raising in NNCF from built-in Python errors to NNCF-specific ones.
Reason for changes
Stable-diffusion models, e.g.
runwayml/stable-diffusion-v1-5has as an input for linear layers with the following shapes:[2*num_images_in_prompt, text_embedding_size, hidden_dimension].https://github.com/openvinotoolkit/nncf/blob/develop/examples/llm_compression/openvino/tiny_llama/main.py
uses not filtered data from
wikitextthat leads to the corner case with sequence length == 1.Related tickets
150851, 155538
Tests
CI