Repository navigation
Open
Labels
performancePerformance and memory issuesPerformance and memory issuesscope: attentionAttention backends (VSA, STA, Flash, etc.)Attention backends (VSA, STA, Flash, etc.)scope: dataData preprocessing, datasetsData preprocessing, datasetsscope: docsDocumentationDocumentationscope: inferenceInference pipeline, serving, CLIInference pipeline, serving, CLIscope: kernelCUDA kernels, fastvideo-kernelCUDA kernels, fastvideo-kernelscope: modelModel architecture (DiTs, encoders, VAEs)Model architecture (DiTs, encoders, VAEs)scope: trainingTraining pipeline, methods, configsTraining pipeline, methods, configs
Description
Updated 2026-11-07T15:05Z at main a01c50a.
Ready for a maintainer (0)
None.
One step away (15)
- [perf] fused RoPE Triton kernel + FP32LayerNorm fp32 cache #1245 blocked on red lane: [perf] fused RoPE Triton kernel + FP32LayerNorm fp32 cache (Godmook) — approved by alexzms; buildkite/ci failed on head and the fastcheck lanes never ran
- [bugfix]: build_eval_kwargs returns (T, C, H, W) per metric contract #1412 blocked on review: [bugfix]: build_eval_kwargs returns (T, C, H, W) per metric contract (klhhhhh) — CI green on head f8d1ac8, no human approval yet
- [perf] LTX2 FP4-linear on DGX Spark (GB10/sm_121): ~24% faster denoise at 1080p #1594 blocked on review: [perf] LTX2 FP4-linear on DGX Spark (GB10/sm_121): ~24% faster denoise at 1080p (Mister-Raggs) — CI green on head 2a88bbf, no human approval yet
- [feat] Add opt-in FlashInfer attention backend for Wan #1799 blocked on review: [feat] Add opt-in FlashInfer attention backend for Wan (klhhhhh) — CI green on head f6b71c5, no human approval yet
- [perf]FlashInfer cuDNN batched prefill attention #1827 blocked on review: [perf]FlashInfer cuDNN batched prefill attention (klhhhhh) — CI green on head a087519, no human approval yet
- [perf] Shard tokens across SP ranks in causal Wan training (opt-in) #1951 blocked on review: [perf] Shard tokens across SP ranks in causal Wan training (opt-in) (zmy1260222041) — CI green on head 10e0797, no human approval yet
- [bugfix] Resolve Studio attention settings and isolate job execution #1952 blocked on review: [bugfix] Resolve Studio attention settings and isolate job execution (FlappyBob) — CI green on head 97a5754, no human approval yet
- [kernel] sm_103a VSA: take the softmax row max from tcgen05.ld.red #1982 blocked on review: [kernel] sm_103a VSA: take the softmax row max from tcgen05.ld.red (lpc0220) — CI green on head aa2c332, no human approval yet
- [perf]: training-free token reduction for MiniMax H3 #2004 blocked on review: [perf]: training-free token reduction for MiniMax H3 (KyleNeverGivesUp) — CI green on head abd7a97, no human approval yet
- [perf] Wan: bit-exact Triton residual/modulate kernels as the default… #2020 blocked on review: [perf] Wan: bit-exact Triton residual/modulate kernels as the default… (incisors) — CI green on head f34cd9c, no human approval yet
- [perf] Kandinsky-5 kernel fusion: Triton-fused residual+LayerNorm+modulate inference path #2021 blocked on review: [perf] Kandinsky-5 kernel fusion: Triton-fused residual+LayerNorm+modulate inference path (arcusbuilds) — CI green on head a6573a6, no human approval yet
- [new-model] Re-add PAN2 text-to-video and image-to-video, with inference optimizations and guardrails #2029 blocked on review: [new-model] Re-add PAN2 text-to-video and image-to-video, with inference optimizations and guardrails (SolitaryThinker) — CI green on head 6b5eef4, no human approval yet; changes requested by wu-qing-157 on an earlier head
- [perf] Wan: fused-norm tier at the norm1 site, exact tier on causal Wan #2033 blocked on review: [perf] Wan: fused-norm tier at the norm1 site, exact tier on causal Wan (incisors) — CI green on head d080930, no human approval yet
- [docs] Slim README, add FastH3 Omni news, add a docs news archive #2066 blocked on review: [docs] Slim README, add FastH3 Omni news, add a docs news archive (fv-superagent) — CI green on head 5754bce, no human approval yet
- [docs] List the missing registered model IDs in the support matrix #2071 blocked on review: [docs] List the missing registered model IDs in the support matrix (fv-superagent) — CI green on head 1523d89, no human approval yet
Anyone with a question about a PR can ask by @-mentioning fv-superagent.
— fv-superagent, automated maintainer run by aryan5v
Metadata
Metadata
Assignees
Labels
performancePerformance and memory issuesPerformance and memory issuesscope: attentionAttention backends (VSA, STA, Flash, etc.)Attention backends (VSA, STA, Flash, etc.)scope: dataData preprocessing, datasetsData preprocessing, datasetsscope: docsDocumentationDocumentationscope: inferenceInference pipeline, serving, CLIInference pipeline, serving, CLIscope: kernelCUDA kernels, fastvideo-kernelCUDA kernels, fastvideo-kernelscope: modelModel architecture (DiTs, encoders, VAEs)Model architecture (DiTs, encoders, VAEs)scope: trainingTraining pipeline, methods, configsTraining pipeline, methods, configs
Activity