🩹 FIXED: ComfyUI VRAM Spike Issue on Python 3.12 - Auto-Patched in v4.10.9

[diodiogod]

TL;DR

ComfyUI v0.3.57+ causes severe VRAM spikes on Python 3.12. We have built-in an automatic fix to TTS Audio Suite v4.10.9 and submitted the patch to ComfyUI (PR #10026).

What Happened

Since I introduced IndexTTS-2, I noticed massive VRAM spikes on Python 3.12. This drove me absolutely crazy - I initially thought it was my code, then dependency conflicts, then just Python version differences...

Plot twist: The spikes affect ALL TTS engines (even classic ChatterBox), and only happen on Python 3.12. After extensive testing with clean environments and old code versions, I confirmed it is a ComfyUI regression.

Root Cause Found ✅

• Culprit: ComfyUI v0.3.57+ enables CUDNN benchmarking that conflicts with Python 3.12s garbage collection
• Affected: All TTS workflows on Python 3.12 using --fast flag
• Symptoms: Multi-GB VRAM spikes before/after inference (not during generation)
• Unaffected: Python 3.13 works fine

✅ ALREADY FIXED IN TTS AUDIO SUITE v4.10.9!

Good news: The fix is now automatically built into TTS Audio Suite v4.10.9! Just update to the latest version and the VRAM spikes will be automatically resolved.

The automatic fix:

• 🔍 Automatically detects Python 3.12 + problematic ComfyUI versions
• 🛠️ Disables the problematic CUDNN benchmarking before TTS generation
• 🩹 Shows friendly message: TTS AUDIO SUITE CUDNN FIX APPLIED
• ✅ Zero user action required - just update and run!

Alternative Solutions (Not Needed)

Option 1: Don not Use --fast Flag

Remove --fast from your ComfyUI launch arguments (reduces performance slightly)

Option 2: Downgrade ComfyUI (Not Recommended)

• ComfyUI: v0.3.56 (last working version)
• ComfyUI-frontend: v1.25.11

Status

• ✅ Root cause identified (CUDNN benchmark + Python 3.12 GC conflict)
• ✅ Automatic fix built into TTS Audio Suite v4.10.9
• ✅ PR submitted to ComfyUI (#10026) for upstream fix
• ✅ Users just need to update TTS Audio Suite

Python 3.13 users: You are unaffected and can continue using latest ComfyUI.

How to Update

1. Git users: git pull in your TTS Audio Suite folder
2. Download users: Re-download from the releases page
3. Restart ComfyUI and run any TTS workflow
4. Look for the fix message: You will see 🩹 TTS AUDIO SUITE CUDNN FIX APPLIED when it activates

That is it! Your VRAM spikes are now eliminated automatically. 🎉

---

Technical Details: This is a targeted compatibility patch that only disables CUDNN benchmarking on Python 3.12 when the problematic --fast autotune flag is used. Other Python versions continue to benefit from the performance optimization. The patch applies automatically before any TTS generation, ensuring zero VRAM spikes while preserving all functionality.