24h GitHub PR Update

GitHub Tracked repos · Mar 20, 2026, 10:03 p.m.

24h pull request roundup with 9 new PRs across 3 active repos in the tracked set.

#13082 Auto-regressive video generation by @Talmaj: Proposes auto-regressive video generation support, likely enabling iterative frame extension workflows.
#13081 fix: follow RFC 2183 for Content-Disposition header by @chulinhcql-art: Makes Content-Disposition handling standards-compliant to improve file naming behavior across clients.
#13080 fix: RFC 2183 compliant Content-Disposition header in /view endpoint by @TJUEZ: Aligns /view response headers with RFC 2183 to reduce attachment/rendering edge cases.
#13079 Make EmptyImage node follow intermediate device/dtype. by @comfyanonymous: Ensures EmptyImage honors intermediate compute settings, reducing dtype/device mismatch surprises.
#13074 ltx: vae: Fix missing init variable by @rattus128: Fixes missing LTX VAE init state that could fail graph execution in video jobs.
#13073 Fix VRAM leak in tiler fallback in video VAEs by @rattus128: Addresses a VRAM leak in video VAE fallback logic for better stability in long or repeated runs.

#742 Fix CPU/CUDA device mismatch when training Klein edit with control_path by @HuangYuChuh: Fixes mixed device placement during Klein edit training so control_path runs don’t crash on tensor-device mismatches.

#170 Add @torch.inference_mode() to pipeline call methods by @garrick99: Wraps pipeline calls in inference mode to cut autograd overhead and reduce runtime memory usage.
#169 Fix —quantization fp8-scaled-mm CLI crash by @garrick99: Fixes a CLI crash path for fp8-scaled-mm quantization, improving reliability for low-memory runs.