dequantization offload accounting (fixes Flux2 OOMs - incl TEs) #11171
+22
−8
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
rattus128
requested review from
Kosinkadink,
comfyanonymous and
guill
as code owners
Contributor
|
Confirmed that it resolves #10891 (comment). |
comfy/sd.py
Outdated
|
|
||
| self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) | ||
| self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device) | ||
| self.patcher.set_model_compute_dtype(dtype) |
Owner
There was a problem hiding this comment.
This is slightly wrong because the text encoder is always upcasted to fp32.
Contributor
Author
There was a problem hiding this comment.
Fixed to match:
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -127,7 +127,8 @@ class CLIP:
self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
- self.patcher.set_model_compute_dtype(dtype)
+ #Match torch.float32 hardcode upcast in TE implemention
+ self.patcher.set_model_compute_dtype(torch.float32)
self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
self.patcher.is_clip = True
self.apply_hooks_to_conds = None
Handle the case where the attribute doesnt exist by returning a static sentinel (distinct from None). If the sentinel is passed in as the set value, del the attr.
When measuring the cost of offload, identify weights that need a type change or dequantization and add the size of the conversion result to the offload cost. This is mutually exclusive with lowvram patches which already has a large conservative estimate and wont overlap the dequant cost so\ dont double count.
So that the loader can know the size of weights for dequant accounting.
rattus128
force-pushed
the
prs/dequant-reserve
branch
from
a03d98e to
9663183
Compare
Owner
I get this when running the basic hidream dev workflow on: https://comfyanonymous.github.io/ComfyUI_examples/hidream/ with simulated 16GB vram. |
Owner
|
Nevermind I think it was my fault: #11201 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
This is the hopefully full root cause fix on:
#10891
Primary commit message:
Example Test case:
RTX3060 Flux2 workflow with ModelComputeDtype node set to fp32
Before:
After: