Repository navigation

[perf] Batch H3 vision frames with PyTorch varlen attention - #2050

Open

leo2ultimate wants to merge 11 commits into

hao-ai-lab:mainfrom

leo2ultimate:codex/h3-vision-attention-batching

Open

[perf] Batch H3 vision frames with PyTorch varlen attention#2050
leo2ultimate wants to merge 11 commits into
hao-ai-lab:mainfrom
leo2ultimate:codex/h3-vision-attention-batching

Commits on Nov 10, 2026
- [perf] Batch independent H3 vision attention segments
  
  leo
  authored and
  leo
  committed
  7a7323c
- [docs] Design direct packed H3 VAE loading
  
  leo
  authored and
  leo
  committed
  9dd7e94
- [perf] Pack H3 vision segments with FlashAttention
  
  leo
  authored and
  leo
  committed
  7f93fdc
- [refactor] Use existing backend policy for H3 vision attention
  
  leo
  authored and
  leo
  committed
  d90afea
- [cleanup] Remove unreleased H3 vision env deprecations
  
  leo
  authored and
  leo
  committed
  befec88
- [perf] Batch H3 vision SDPA only on fused backends
  
  leo
  authored and
  leo
  committed
  6f71078
- [perf] Use native PyTorch varlen for H3 vision attention
  
  leo
  authored and
  leo
  committed
  40ceb38
- [test] Trim H3 varlen regression coverage
  
  leo
  authored and
  leo
  committed
  0ced630
- [cleanup] Trim H3 vision attention comments
  
  leo
  authored and
  leo
  committed
  d73b597
- [cleanup] Remove redundant H3 test and benchmark comments
  
  leo
  authored and
  leo
  committed
  67a23ee
- [bugfix] Preserve H3 vision SDPA compatibility fallback
  
  leo
  authored and
  leo
  committed
  95c6c1c