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
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- LichtFeld Studio v0.5.4.3 ships community-built SteadySplats, POPSpa and Voxelization — janusch_patas · 2026-10-26
- One prompt to a 2-3 min narrated short: hands-on test of Antigravity video generation — AI_Andrew · 2026-10-26
- Creator makes 10,564-frame animated short entirely with Gemini toolchain — AI_Andrew · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26