Fix MPS upscale/VAE failures: enforce contiguous tiled input and safer split-attention chunking

[fritzprix]

Summary

This PR fixes two Apple Silicon/MPS runtime failures seen in upscale + VAE encode workflows.

Problem

On MPS, some workflows fail with:

• view size is not compatible with input tensor's size and stride
• MPSGraph does not support tensor dims larger than INT_MAX

Root cause

1. Tiled upscale inputs can become non-contiguous after slicing/narrowing, which breaks models that assume view-compatible memory layout.
2. VAE split-attention chunking could still produce oversized matmul shapes on MPS for large token counts.

Changes

• In tiled scaling, pass contiguous tiles into model calls (s.contiguous(), s_in.contiguous()).
• In VAE slice_attention:
  • use robust ceil-based chunk sizing,
  • cap per-slice size for MPS INT_MAX safety,
  • retry with increased steps on MPS INT_MAX/MPSGraph runtime errors.

Impact / compatibility

• No intended functional change in outputs.
• Main tradeoff: potentially slower attention on very large MPS workloads due to smaller chunks.
• CUDA/CPU behavior is unchanged except for safer chunk sizing logic.

Validation

• Reproduced failing upscale + VAE path on macOS/MPS before patch.
• Confirmed workflow completes after patch.

#11851

[Copilot]

Pull request overview

This PR fixes Apple Silicon/MPS runtime failures in upscale and VAE encode workflows by addressing tensor contiguity issues and attention computation limits.

Changes:

• Added .contiguous() calls to tiled upscaling to ensure memory layout compatibility with model view operations
• Improved VAE attention slice sizing with proper ceil-based chunking (fixes incorrect logic for non-evenly-divisible cases)
• Added MPS-specific safeguards to prevent INT_MAX tensor size errors with automatic retry logic

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
comfy/utils.py	Added contiguous() calls before passing tiled tensors to model functions, fixing "view size is not compatible" errors on MPS
comfy/ldm/modules/diffusionmodules/model.py	Improved slice_attention chunking with ceil-based sizing, MPS INT_MAX safeguards, and retry logic for MPS-specific errors

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

There appears to be a typo in the error string check: "MPSGaph" should likely be "MPSGraph". This typo means the error handling won't catch errors containing "MPSGraph" misspelled in this way. Consider removing this misspelled variant unless it's intentionally included to handle a known typo in MPS error messages.

Suggested change

	if q.device.type == "mps" and ("INT_MAX" in str(e) or "MPSGraph" in str(e) or "MPSGaph" in str(e)):
	if q.device.type == "mps" and ("INT_MAX" in str(e) or "MPSGraph" in str(e)):

[nick-morhun]

Did you consider moving this complex logic to a new function?