LTX 2.5 silent videos traced to PyTorch attention backend bug on AMD GPUs
Captain_Doobie · reddit · 2026-09-09
An RX 7900 XTX user fought LTX 2.5's prompt-ignoring and missing-speech issues for days before finding the root cause: not the workflow or prompt, but PyTorch's scaled-dot-product attention backend computing wrong results on AMD/HIP gfx1100.
- Symptoms: video generated fine but dialogue became music/gibberish, subject and framing drifted, unwanted subtitles appeared; the same prompt worked in LTX 2.3.
- Diagnosis: seeds, prompt enhancer removal, quantized encoders, and official BF16 weights all failed; layer-by-layer comparison against the Hugging Face Gemma 4 reference showed the first 5 blocks agreed, then the first global-attention block diverged dramatically — Q/K projections and positional encoding matched, so the bug sat inside the selected SDPA backend.
- Fix: patch ComfyUI's gemma4.py to force SDPBackend.MATH only when AMD/HIP gfx1100 + head dim 512 + BF16/FP32 + attention mask are all true; all other calls keep the original path. After the patch, encoder output matched the reference within 0.42% relative L2, cosine similarity 0.99999.
More from Multimodal
- Butter.video wins over skeptical creator with hundreds of tasteful timeline 'blocks' for video effects — beechinour · 2026-10-27
- $50, 2-hour Kurzgesagt-style AI video reignites debate over intent in AI creation — motionbynick · 2026-10-27
- Fake BTS of a dam disaster made with Dreamina Seedance for 29 cents — Standard_Spell5873 · 2026-10-27
- Redditor struggles with outfit swap: Qwen Image 2.1 breaks body proportions — Existing-Box-7866 · 2026-10-27
- Kling 4.0 lands on Deevid AI: 30-second videos, 10 keyframes, 15 multimodal references — nikola_mr64990 · 2026-10-27
- Consistent characters in ComfyUI: RealVisXL V5 + img2img face reference workflow — MiaLarsonJL · 2026-10-27