Fixed comfyui_patches_ll to newest ComfyUI

ParaAttentionNode Fix Summary (Detailed Version)

1. Error That Occurred

• Affected Node: `KSamplerAdvanced` (when using MultiGPU patch internally)

• Stack Trace Origin

  • When `WrapperExecutor.execute()` calls `fb_cache_prepare_wrapper` via CFG guidance processing in `comfy/samplers.py`, the keyword argument `latent_shapes` is passed

  • `fb_cache_prepare_wrapper` retains the old function signature and does not expect `latent_shapes`, causing a TypeError and stopping execution

2. Root Cause Analysis

1. ComfyUI Core Specification Change

   • `WrapperExecutor.execute()` now passes new keywords like `latent_shapes` to CFG wrappers

2. comfyui_patches_ll Wrapper Has Outdated Signature

   • `fb_cache_prepare_wrapper(wrapper_executor, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None)` stops at this definition and rejects new keywords

3. Result

   • When running KSampler on models with First Block Cache (FBC) applied, a TypeError halts execution

   • Specifically, the FBC wrapper in ParaAttentionNode (`comfyui_patches_ll`) is the culprit

3. Fix Implementation

• File Modified: `ComfyUI/custom_nodes/comfyui_patches_ll/nodes/ParaAttentionNode.py`

• Function Modified: `fb_cache_prepare_wrapper`

def fb_cache_prepare_wrapper(wrapper_executor, noise, latent_image, sampler, sigmas,
                             denoise_mask=None, callback=None, disable_pbar=False,
                             seed=None, latent_shapes=None, **kwargs):
    cfg_guider = wrapper_executor.class_obj

    try:
        out = wrapper_executor(
            noise,
            latent_image,
            sampler,
            sigmas,
            denoise_mask=denoise_mask,
            callback=callback,
            disable_pbar=disable_pbar,
            seed=seed,
            latent_shapes=latent_shapes,
            **kwargs
        )
    finally:
        diffusion_model = cfg_guider.model_patcher.model.diffusion_model
        if hasattr(diffusion_model, fb_cache_model_temp):
            delattr(diffusion_model, fb_cache_model_temp)

    return out

Key Changes

1. Extended Function Signature

   • Now accepts the new keyword argument `latent_shapes`, and additionally captures `**kwargs` for future extensibility

2. Pass Through to Lower Layer

   • During wrapper execution, forward received arguments as `wrapper_executor(..., latent_shapes=latent_shapes, **kwargs)` to downstream handlers

3. Existing Cleanup Logic Unchanged

   • The `finally:` block that clears temporary attributes from the diffusion model remains unmodified, so no side effects occur

4. Results Achieved

• TypeError Resolved
  When `latent_shapes` and other additional arguments are passed, the wrapper can now handle them, allowing KSampler to complete successfully

• Backward Compatibility Maintained

  • Existing FBC operations (cache registration and cleanup) work as before

  • No impact on other wrappers or MultiGPU nodes

• Future-Proof

  • Because `**kwargs` is captured, even if more parameters are added in the future, TypeError will not occur

5. Supplementary Notes

• The problem originated from parameter additions accompanying ComfyUI core updates

• If ParaAttentionNode is not updated, a TypeError will halt execution as seen here

• `latent_shapes` is a parameter added when samplers like DPM++ handle multiple latents; this fix ensures the wrapper adapts to that change

Above is the detailed explanation of the ParaAttentionNode (First Block Cache wrapper) fix and its implementation.