[Quantization] Add support for Comfy-Kitchen / Comfy-Quants - #14747
Open
PrakshaaleJain wants to merge 16 commits into
Open
PrakshaaleJain wants to merge 16 commits into
PrakshaaleJain wants to merge 16 commits into
Conversation
github-actions
Bot
added
documentation
PrakshaaleJain
force-pushed
the
feat/comfy-quant-support
branch
from
September 13, 2026 09:37
a99b253 to
e9bca91
Compare
Contributor
Author
|
Hi @DN6, please have a review for the PR. |
DN6
reviewed
Sep 14, 2026
| **kwargs, | ||
| ): | ||
| self.quant_method = QuantizationMethod.COMFY_QUANT | ||
| self.quant_format = quant_format |
Collaborator
There was a problem hiding this comment.
We need to raise when an unsupported format is passed here.
| ): | ||
| module, tensor_name = get_module_from_name(model, param_name) | ||
|
|
||
| import comfy_kitchen.tensor as ck_tensor |
Collaborator
There was a problem hiding this comment.
Can run a is_comfy_kitchen_available check and import at top level.
|
|
||
| @require_torch | ||
| @pytest.mark.skipif(not is_comfy_kitchen_available(), reason="comfy-kitchen is not available") | ||
| class TestComfyQuantizer: |
Collaborator
There was a problem hiding this comment.
Let's setup testing in a similar manner to GGUF:
| if isinstance(param_value, ck_tensor.QuantizedTensor): | ||
| quantized_weight = param_value | ||
| else: | ||
| layout = layout_map.get(self.quant_format.lower()) |
Collaborator
There was a problem hiding this comment.
Isn't this is a constant for a given quant format? Can't we set it once on init?
|
|
||
| import comfy_kitchen.tensor as ck_tensor | ||
|
|
||
| layout_map = { |
Collaborator
There was a problem hiding this comment.
I think a few of these are not real layout objects. Please check.
| } | ||
|
|
||
| # Check if it's already a QuantizedTensor (e.g., if loaded directly from a custom loader) | ||
| if isinstance(param_value, ck_tensor.QuantizedTensor): |
Collaborator
There was a problem hiding this comment.
This wouldn't work since the state dict would contain just torch.Tensor objects. We would need to construct the QuantizedTensor manually from the state dict.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
What does this PR do?
This PR adds support for the Comfy Quant quantization toolkit by integrating its inference engine, Comfy-Kitchen, directly into Diffusers library.
Key additions:
ComfyQuantConfigandComfyQuantizerto handle wrapping of tensors.create_quantized_paramand wraps them viaQuantizedTensor.from_floatto dispatch to Comfy-Kitchen's optimized GPU kernels.fp8,int8,mxfp8,nvfp4,int4_svd,int4_awq).tests/quantization/comfy_quant/.Fixes #14705
Before submitting
self-reviewskill on the diff?AI Agent Self-Review Notes
Agent: Antigravity (Google DeepMind)
Review Focus: Diffusers Quantization Standards & Guidelines
ComfyQuantizerrelies onis_comfy_kitchen_available(). If the backend is not installed, Diffusers will safely fall back to standard behavior or fail gracefully ifComfyQuantConfigis passed explicitly without the toolkit.make style && make fix-copies.utils/check_dummies.pysuccessfully generateddummy_comfy_kitchen_objects.pyand correctly registered the imports in `diffusers/__inittests/quantization/comfy_quant/test_comfy_quant.py. Tests cover both valid format wrapping (FP8mapping toTensorCoreFP8Layout) and error handling for unsupported formats. Tests ran locally and passed with 100% success.