[ROCm][Perf] Fused MoE W4A16 HIP kernel for AMD RDNA3 (gfx1100) (#44075)

Signed-off-by: JartX <sagformas@epdcenter.es>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
JartX
2026-06-06 15:30:39 -05:00
committed by GitHub
co-authored by TJian
parent fa27d4e9cf
commit 062b05ff3a
11 changed files with 1906 additions and 5 deletions
+9
View File
@@ -50,6 +50,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
"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);
rocm_ops.def(
"moe_gptq_gemm_rdna3(Tensor a, Tensor! c, Tensor b_q_weight, "
"Tensor b_scales, Tensor b_qzeros, Tensor topk_weights, "
"Tensor sorted_token_ids, Tensor expert_ids, "
"Tensor num_tokens_post_padded, "
"int top_k, int block_size_m, bool mul_topk_weight, "
"int output_topk) -> ()");
rocm_ops.impl("moe_gptq_gemm_rdna3", torch::kCUDA, &moe_gptq_gemm_rdna3);
#endif
// Custom attention op