[ROCm] revert VAE restrictions on RDNA4 and CDNA4

[Apophis3158]

Better support for AMD latest GPU arches: RDNA4 (gfx1200, gfx1201) and CDNA4 (gfx950), determined based on fp8 support.

Reference: __hip_fp8_e4m3 and __hip_fp8_e5m2 supports in HIP C++ type implementation support table at AMD data types and precision support

[coderabbitai]

📝 Walkthrough

Walkthrough

The PR makes AMD-specific behavior conditional on SUPPORT_FP8_OPS. In comfy/model_management.py, pytorch_attention_enabled_vae() now returns False for AMD only when SUPPORT_FP8_OPS is false; if SUPPORT_FP8_OPS is true it defers to pytorch_attention_enabled(). In comfy/sd.py, VAE.__init__ sets VAE_KL_MEM_RATIO to 2.73 only when is_amd() is true and SUPPORT_FP8_OPS is false; otherwise it uses the non-AMD value. No public signatures changed.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description check	✅ Passed	The description clearly explains the architectural improvements and references AMD's official documentation on FP8 support, directly relating to the code changes in model_management.py and sd.py.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title references reverting VAE restrictions on RDNA4 and CDNA4 AMD architectures, which aligns with the PR's core objective of improving support for these architectures by conditionally enabling PyTorch VAE attention and adjusting memory ratios based on FP8 support.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[Apophis3158]

• Updated inline comments.

These two improvements has already been tested by ROCm users on Windows and Linux long time ago:

• Revert "Disable pytorch attention in VAE for AMD." #8289
• AMD/ROCm - Fix VAE_KL_MEM_RATIO overestimation for modern ROCm #12685

and more.

Most feedback is from gfx120x, so it's better to use SUPPORT_FP8_OPS for the restriction.

ComfyUI/comfy/model_management.py

Lines 409 to 411 in 1de83f9

	if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
	if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]): # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0
	SUPPORT_FP8_OPS = True