Repository navigation

[dev] Make fused CSA Top-K deterministic across no-grad and grad forwards - #7169

Merged

hxbai merged 11 commits into

bobboli:fix/csa-fused-topk-order

Sep 17, 2026

Merged

[dev] Make fused CSA Top-K deterministic across no-grad and grad forwards#7169
hxbai merged 11 commits into
NVIDIA:devfrom
bobboli:fix/csa-fused-topk-order

Commits on Sep 9, 2026
- [dev] Make fused CSA Top-K deterministic across no-grad and grad forwards
  
  bobboli
  and
  claude
  committed
  20cbd65
Commits on Sep 11, 2026
- Name the grad-enabled lowering the no-grad CSA paths mirror
  
  bobboli
  and
  claude
  committed
  6e5d9ce
- Describe both build_attention_indices layouts at the THD no-grad site
  
  bobboli
  and
  claude
  committed
  4edec16
- Drop the real-kernel Top-K repeatability test
  
  bobboli
  and
  claude
  committed
  92473a3
Commits on Sep 15, 2026
- Sort the deterministic Top-K fallback in row slabs
  bobboli
  committed
  c78fef6
- Check no-grad and grad CSA forwards agree on the real kernels
  bobboli
  committed
  f6802bc
- Merge branch 'dev' into fix/csa-fused-topk-order
  bobboli
  committed
  e79f16f
- Use supported shapes for fused CSA parity tests
  bobboli
  committed
  6fc209b
- Mask raw THD padding before sparse attention forward
  bobboli
  committed
  003e334
Commits on Sep 16, 2026
- Reuse negative infinity constant in stable CSA Top-K
  bobboli
  committed
  da81780
Commits on Sep 17, 2026
- Mask CP padding before fused sparse attention forward
  bobboli
  committed
  b57ad25