[ROCm][Kernel] Add HybridW4A16LinearKernel: Triton prefill + HIP skinny decode (#40977)

Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Matthias Gehre
2026-07-13 20:22:00 -07:00
committed by GitHub
co-authored by Andreas Karatzas
parent b50ef9c6ed
commit 382bbd5144
11 changed files with 2165 additions and 5 deletions
+8
View File
@@ -26,6 +26,14 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
"Tensor");
rocm_ops.impl("wvSplitK", torch::kCUDA, &wvSplitK);
// W4A16 grouped skinny GEMM: packed int4 weights, per-group scales,
// optional zero points for asymmetric quantization
rocm_ops.def(
"wvSplitK_int4_g(Tensor in_a, Tensor in_b, Tensor in_scale, "
"Tensor? in_zero_points, Tensor? in_bias, int CuCount, "
"int group_size) -> Tensor");
rocm_ops.impl("wvSplitK_int4_g", torch::kCUDA, &wvSplitK_int4_g);
// Custom gemm op for skinny matrix-matrix multiplication
rocm_ops.def(
"wvSplitKrc(Tensor in_a, Tensor in_b, Tensor? in_bias, int CuCount) -> "