Conversation
- Cleaned up unnecessary whitespace and comments in model_management_mgpu.py, nodes.py, wanvideo.py, and wrappers.py for better code clarity. - Replaced list comprehensions with direct list conversions in nodes.py for efficiency. - Updated memory logging format in model_management_mgpu.py to streamline data capture. - Enhanced device management in wanvideo.py by ensuring consistent device setting and loading. - Added linting configurations in pyproject.toml to enforce code quality standards. - Removed unused imports and optimized existing ones across multiple files.
…ons, fix checkpoint loading.
…rove accessibility
Contributor
There was a problem hiding this comment.
Pull request overview
Release 2.6.0 focused on restoring runtime correctness and compatibility with current ComfyUI, especially for non-primary CUDA devices and basic DynamicVRAM scenarios, plus updating loader/integration glue to match current upstream behavior.
Changes:
- Add CUDA current-device/runtime guards and patch Comfy stream + sampling entrypoints to keep execution aligned with the intended GPU.
- Refresh advanced checkpoint loader behavior (incl.
disable_dynamic) and update several integration/compat nodes (WanVideo alias input, restored device selector, safer optional-node registration). - Sync tooling/lint config and add a starter example workflow.
Reviewed changes
Copilot reviewed 13 out of 14 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
__init__.py |
Core runtime/device-guard additions; patches Comfy stream/sampling; DynamicVRAM multi-device init; hardened optional node registration. |
wrappers.py |
Adds CUDA current-device guarding for standard MultiGPU wrappers (UNet/VAE paths). |
checkpoint_multigpu.py |
Updates patched checkpoint loading path for current ComfyUI loader semantics and forwards disable_dynamic. |
distorch_2.py |
Refactors/updates DisTorch2 patching and allocation logic for current ComfyUI behavior. |
device_utils.py |
Improves device enumeration and cache-clearing behavior (incl. CUDA synchronize before empty_cache). |
nodes.py |
Restores DeviceSelectorMultiGPU; adjusts UNet LP loader call pattern; minor robustness tweaks. |
wanvideo.py |
Adds vace_model backward-compatible alias for extra_model; trims unused imports/formatting. |
model_management_mgpu.py |
Minor cleanup in memory snapshot logging loop. |
pyproject.toml |
Bumps version to 2.6.0 and adds Ruff/Pylint configuration. |
ci/run_workflows.py |
Switches prints to explicit stdout/stderr writers with flush. |
ci/summarize_log.py |
Switches prints to explicit stdout writer with flush. |
ci/extract_allocation.py |
Switches prints to explicit stdout writer with flush. |
example_workflows/ComfyUI-starter_multigpu.json |
Adds starter workflow demonstrating basic MultiGPU usage. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
You can also share your feedback on Copilot code review. Take the survey.
Comment on lines
33
to
+35
| def patched_load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, | ||
| embedding_directory=None, output_model=True, model_options={}, | ||
| te_model_options={}, metadata=None): | ||
| te_model_options={}, metadata=None, disable_dynamic=False): |
There was a problem hiding this comment.
patched_load_state_dict_guess_config uses mutable default arguments (model_options={}, te_model_options={}), which can leak state across invocations if ComfyUI mutates these dicts during loading. Use None defaults and create new dicts inside the function (e.g., model_options = model_options or {}) to avoid cross-call contamination.
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
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.
Summary
This PR ships
ComfyUI-MultiGPU2.6.0.The focus of this release is compatibility and runtime correctness on current ComfyUI:
cuda:1This is not a DistTorch + DynamicVRAM integration PR. That work remains separate.
What changed
Runtime and DynamicVRAM compatibility
comfy.model_management.current_stream(device)usage so stream lookup stays aligned with the requested CUDA devicecomfy_aimdoon all visible CUDA devices when DynamicVRAM is enabled, instead of assuming only the startup device mattersPractical result: standard MultiGPU workflows that place UNet execution on a non-primary GPU such as
cuda:1now work with DynamicVRAM both on and off.Checkpoint loader compatibility
disable_dynamicPractical result: the advanced checkpoint loader path is no longer a stale fork of older ComfyUI loader semantics.
Compatibility and helper fixes
custom_nodesroots with normalized module-name matchingcomfy_kitchenDLPack export compatibilitytorch.cuda.synchronize()beforeempty_cache()to be consistent with the addition to the Comfy Core routine on Feb 3DeviceSelectorMultiGPUvace_modelcompatibility by treating it as a backward-compatible alias forextra_modelTooling and examples
User-visible fixes
comfy_kitchencu130 cross-device DLPack export failure reported in On cu130: Can't export tensors on a different CUDA device index. cu128 works in light testing #167vace_modelcompatibility reported in WanVideoModelLoaderMultiGPU doesn't have a vace_model input. #172Validation
Validated in local workflow/harness testing and static checks:
cuda:0andcuda:1ruff check .passespy_compilepassespylintpassesCloses
Closes #167
Closes #169
Closes #172
Supersedes #168
Not in scope
access violation 0xC0000005when… #141Follow-up
The next major follow-up is DistTorch + DynamicVRAM-on integration, with DistTorch retaining placement control and DynamicVRAM handling compute-side transient residency rather than replacing DistTorch's loading policy wholesale.