Skip to content

Repository navigation

docs(cuda): name DsaIndexSelect in the coverage matrix, which main is red without #2271

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
justinchuby wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

docs(cuda): name DsaIndexSelect in the coverage matrix, which main is red without #2271

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a6ab8c5
docs(cuda): name DsaIndexSelect in the coverage matrix, which main is…
justinchuby Aug 27, 2026
385602a
Merge branch 'main' into gaff1-cuda-coverage-dsaindexselect
justinchuby Aug 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions docs/execution/CUDA_COVERAGE.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -202,6 +202,7 @@ not yet wired) · **🔬 custom** (needs a fused NVRTC/CUTLASS kernel).
| `GroupQueryAttention`, `PagedAttention` | `com.microsoft` | ✅ | **shared tiled attention cores** | Registered attention/KV variants with f32/f16/bf16 compute and integer sequence metadata; each variant's shape, cache, and capture bounds are enforced by its claim gate and dedicated GPU suite. |
| `VarlenAttention`, `PackedVarlenAttention` | `pkg.nxrt` | ✅ | **shared tiled attention cores** | Runtime-owned variable-length attention variants over the shared tiled cores. |
| `CompressedSparseAttention`, `SparseKvGather`, `IndexShare`, `KvCacheCapacityAppend` | `pkg.nxrt` | ✅ | **NVRTC sparse/KV kernels** | Runtime-owned sparse-attention and stable-KV primitives. `KvCacheCapacityAppend` and `IndexShare` have explicit capture-safe device-error paths; see their dedicated GPU suites. |
| `DsaIndexSelect` | `pkg.nxrt` | ✅ | **NVRTC index-selection kernel** | Device-resident query-dependent sparse index selection for the GLM-5.2 DSA / `IndexShare` indexer. Reproduces the frozen CPU reference (`onnx-runtime-ep-cpu`'s `dsa_index_select.rs`) on the device: head-weighted ReLU'd scaled dot-product scoring, additive f32 causal bias, a `bias > -1e30` allow mask, then the top `min(#allowed, top_k)` positions by `(score desc, index asc)`, emitted ascending and right-padded with the `-1` sentinel. Final NaNs are canonicalized to positive quiet NaN on both backends, so selection is bit-parity with the CPU oracle. |
| `LinearAttention`, `CausalConvWithState` | `` | ✅ | **NVRTC recurrent kernels** | Standard-domain f32/f16/bf16 hybrid-decoder state kernels. `LinearAttention` supports the four update rules and fails closed for unsupported geometry; `CausalConvWithState` implements depthwise causal state carry. |
| `LinearAttention`, `CausalConvWithState` | `com.microsoft` | ✅ | **NVRTC recurrent kernels** | Contrib-domain registrations of the same recurrent kernels. |
| `FusedAttention` | `com.microsoft` | 🔬 | **fusion rewrite to `Attention`** | The fused kernel exists behind `AttentionKernel`; registering/lowering this op name remains. |
Expand Down
Loading