[Kernel][ROCm] Native W4A16 kernel for AMD RDNA3 (gfx1100) — fp16 + bf16 (#41394)
Signed-off-by: JartX <sagformas@epdcenter.es>
This commit is contained in:
@@ -39,6 +39,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
|
||||
" Tensor scale_b, int CuCount) -> ()");
|
||||
rocm_ops.impl("wvSplitKQ", torch::kCUDA, &wvSplitKQ);
|
||||
|
||||
#ifdef VLLM_ROCM_GFX1100
|
||||
// W4A16 GPTQ kernels for AMD RDNA3 (gfx1100).
|
||||
rocm_ops.def(
|
||||
"gptq_gemm_rdna3(Tensor a, Tensor b_q_weight, Tensor b_qzeros, "
|
||||
"Tensor b_scales, Tensor b_g_idx, bool use_v2_format) -> Tensor");
|
||||
rocm_ops.impl("gptq_gemm_rdna3", torch::kCUDA, &gptq_gemm_rdna3);
|
||||
|
||||
rocm_ops.def(
|
||||
"gptq_gemm_rdna3_wmma(Tensor a, Tensor b_q_weight, Tensor b_qzeros, "
|
||||
"Tensor b_scales, Tensor b_g_idx, bool use_v2_format) -> Tensor");
|
||||
rocm_ops.impl("gptq_gemm_rdna3_wmma", torch::kCUDA, &gptq_gemm_rdna3_wmma);
|
||||
#endif
|
||||
|
||||
// Custom attention op
|
||||
// Compute the attention between an input query and the cached
|
||||
// keys/values using PagedAttention.
|
||||
|
||||
Reference in New Issue
Block a user