ComfyUI Dynamic VRAM compatibility by Pizzawookiee · Pull Request #572 · numz/ComfyUI-SeedVR2_VideoUpscaler

Pizzawookiee · 2026-04-13T16:47:00Z

Patches the model with ComfyUI's operations layers (i.e. replacing nn.Linear with operations.Linear) before loading the weights. Also uses ComfyUI's safetensors loader. In theory this should be all that is necessary to implement ComfyUI's recent Dynamic VRAM features.

thehhmdb · 2026-04-13T19:01:58Z

I had a quick test and it to errors with:

[ERROR] Error during processing: module 'comfy' has no attribute 'ops'

Pizzawookiee · 2026-04-13T19:20:05Z

Is it failing at this line specifically?

src/common/config.py, line 219 --> default_ops = comfy.ops.manual_cast

I am on comfy stable release 0.18.1

thehhmdb · 2026-04-13T19:31:34Z

Yes, here are more log entries:

[20:26:15.005] ❌ [ERROR] Error during processing: module 'comfy' has no attribute 'ops'
Traceback (most recent call last):
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\inference_cli.py", line 1643, in main
    frames = process_single_file(file_path, args, device_list, output_path,
                                format_auto_detected=format_auto_detected,
                                runner_cache=runner_cache)
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\inference_cli.py", line 537, in process_single_file
    for result in _stream_video_chunks(
                  ~~~~~~~~~~~~~~~~~~~~^
        cap=cap,
        ^^^^^^^^
    ...<9 lines>...
        cleanup_timer_name="chunk_cleanup"
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    ):
    ^
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\inference_cli.py", line 696, in _stream_video_chunks
    result = _process_frames_core(
        frames_tensor=frames.to(torch.float16),
    ...<3 lines>...
        runner_cache=runner_cache
    )
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\inference_cli.py", line 916, in _process_frames_core
    runner, cache_context = prepare_runner(
                            ~~~~~~~~~~~~~~^
        dit_model=args.dit_model,
        ^^^^^^^^^^^^^^^^^^^^^^^^^
    ...<22 lines>...
        torch_compile_args_vae=torch_compile_args_vae
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\core\generation_utils.py", line 491, in prepare_runner
    runner, cache_context = configure_runner(
                            ~~~~~~~~~~~~~~~~^
        dit_model=dit_model,
        ^^^^^^^^^^^^^^^^^^^^
    ...<18 lines>...
        torch_compile_args_vae=torch_compile_args_vae
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\core\model_configuration.py", line 829, in configure_runner
    _setup_models(
    ~~~~~~~~~~~~~^
        runner, cache_context, dit_model, vae_model,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        base_cache_dir, block_swap_config, debug
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\core\model_configuration.py", line 949, in _setup_models
    dit_created = _setup_dit_model(runner, cache_context, dit_model, base_cache_dir,
                                   block_swap_config, debug)
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\core\model_configuration.py", line 1055, in _setup_dit_model
    runner = prepare_model_structure(runner, "dit", dit_checkpoint_path,
                                    runner.config, debug, block_swap_config)
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\core\model_loader.py", line 453, in prepare_model_structure
    model = create_object(model_config)
  File "Y:\ComfyUI\custom_nodes\ComfyUI-SeedVR2_VideoUpscaler\src\common\config.py", line 219, in create_object
    default_ops = comfy.ops.manual_cast
                  ^^^^^^^^^
AttributeError: module 'comfy' has no attribute 'ops'

I have upgraded Comfyui

Pizzawookiee · 2026-04-13T20:05:51Z

Just to clarify, are you using the standalone CLI? And if that is true, are you able to test the node itself within ComfyUI?

thehhmdb · 2026-04-13T20:38:47Z

Yes, I was using the CLI.

I have also tested the node within confyui, and Phase 2 DIT upscaling gives an OOM error (Error in Phase 2 (Upscaling): Allocation on device) without block swap, despite lots of RAM being available.

Pizzawookiee · 2026-04-13T20:45:34Z

Okay then I'll mark this as draft while I look at these issues.

wywywywy · 2026-05-21T11:36:53Z

Just tested this MR with the latest Comfy using 7b Q4 on a 5090 batch size 41 on a 2.7MP video, no crash at all. Upscaling processed successfully.

I tested with both blocks_to_swap=36 and blocks_to_swap=0, both worked. Not sure if this parameter actually does anything any more.

EDIT: Btw this is through the UI not CLI

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8a2308f770

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-21T15:57:10Z

                    debug.log("Using CPU intermediate loading for MPS compatibility", 
                            category="info", indent_level=1)
-                state = load_safetensors_file(checkpoint_path, device="cpu")
+                state = comfy.utils.load_torch_file(checkpoint_path, device=device_str)


Load MPS fallback tensors on CPU

In the MPS allocator-error fallback path, the retry still calls comfy.utils.load_torch_file(..., device=device_str), where device_str is still "mps". That re-runs the same failing load path instead of doing the intended CPU intermediate load, so users on MPS who hit allocation/watermark errors will continue to fail rather than recovering.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-05-21T15:57:10Z


+        # 2b. Handle InflatedCausalConv3d
+        elif isinstance (child, InflatedCausalConv3d) and hasattr(target_ops, "Conv3d"):
+            child.__bases__ = (target_ops.Conv3d,)


Convert InflatedCausalConv3d by replacing module instance

Setting child.__bases__ = (target_ops.Conv3d,) mutates an instance attribute, not the class inheritance used for dispatch, so this does not actually convert InflatedCausalConv3d layers to Comfy ops. As a result, models that rely on InflatedCausalConv3d keep the original layer type and miss the dynamic-VRAM operation patching this commit intends to add.

Useful? React with 👍 / 👎.

Pizzawookiee · 2026-05-21T15:58:07Z

Just tested this MR with the latest Comfy using 7b Q4 on a 5090 batch size 41 on a 2.7MP video, no crash at all. Upscaling processed successfully.

I tested with both blocks_to_swap=36 and blocks_to_swap=0, both worked. Not sure if this parameter actually does anything any more.

EDIT: Btw this is through the UI not CLI

To clarify, with the base repository (without this pull request) were you able to run the model without crashes at both blocks to swap = 0 and blocks to swap = 36?

wywywywy · 2026-05-21T16:03:11Z

To clarify, with the base repository (without this pull request) were you able to run the model without crashes at both blocks to swap = 0 and blocks to swap = 36?

On the official main branch yes I can run blocks to swap = 36 no problems as well. I haven't tested 0 in a long time so I'm not sure.

wywywywy · 2026-05-21T20:39:11Z

Let me know if there's anything you want me to test

Pizzawookiee · 2026-05-21T21:00:55Z

Let me know if there's anything you want me to test

In theory, I could use some testing with block swap turned completely off. That said, on my device which is less powerful, disabling block swap while using my pull request currently still results in OOM so I don't consider this PR up to par yet.

wywywywy · 2026-05-22T08:21:05Z

Just realised that swap_io_components needs to be off as well.

So I pulled the latest commit and tested it again. Still works with batch size of 41.

When I use batch size 81, the VRAM goes to 100%, but I guess at that point it's all taken up by the latents not the layers, so it won't be a good test for offloading.

Pizzawookiee added 2 commits April 13, 2026 12:35

Update config.py to patch SeedVR2 model with comfy operations

765cac6

Update model_loader.py to use comfy's model loader for safetensor files

8a2308f

Pizzawookiee marked this pull request as draft April 13, 2026 20:44

Pizzawookiee marked this pull request as ready for review May 21, 2026 15:54

chatgpt-codex-connector Bot reviewed May 21, 2026

View reviewed changes

Fix inheritance error for conv3d layers in VAE

b5f0186

Pizzawookiee marked this pull request as draft May 21, 2026 20:56

Uh oh!

Conversation

Pizzawookiee commented Apr 13, 2026

Uh oh!

thehhmdb commented Apr 13, 2026

Uh oh!

Pizzawookiee commented Apr 13, 2026

Uh oh!

thehhmdb commented Apr 13, 2026

Uh oh!

Pizzawookiee commented Apr 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

thehhmdb commented Apr 13, 2026

Uh oh!

Pizzawookiee commented Apr 13, 2026

Uh oh!

wywywywy commented May 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 21, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 21, 2026

Choose a reason for hiding this comment

Uh oh!

Pizzawookiee commented May 21, 2026

Uh oh!

wywywywy commented May 21, 2026

Uh oh!

wywywywy commented May 21, 2026

Uh oh!

Pizzawookiee commented May 21, 2026

Uh oh!

wywywywy commented May 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Pizzawookiee commented Apr 13, 2026 •

edited

Loading

wywywywy commented May 21, 2026 •

edited

Loading