From 60d56fa2d064d6b7b7c96940fb2ba44832b8f826 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Sun, 2 Nov 2025 10:00:26 +0200 Subject: [PATCH] WIP --- ggml/src/ggml-cuda.cu | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index c78a0f39..bed99058 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -3097,7 +3097,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg ggml_are_same_shape(dst, cgraph->nodes[i+1]->src[1]) && cgraph->nodes[i+1] == cgraph->nodes[i+2]->src[0] && ops_are_same_device(cgraph, i, i+2)) { - //printf("Fusing add->add->fused_rms of %s, %s, %s\n", dst->name, cgraph->nodes[i+1]->name, cgraph->nodes[i+2]->name); + printf("Fusing add->add->fused_rms of %s, %s, %s\n", dst->name, cgraph->nodes[i+1]->name, cgraph->nodes[i+2]->name); ggml_cuda_op_fused_add_add_rms_norm(ctx, dst, cgraph->nodes[i+1], cgraph->nodes[i+2]); i += 2; } @@ -3249,6 +3249,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg // printf("=== Fused rms_norm(%s)\n", dst->name); // for (int j = 1; j <= 6; ++j) printf(" %s(%s)\n", ggml_op_name(cgraph->nodes[i+j]->op), cgraph->nodes[i+j]->name); //} + // Disabled because currently there is something wrong in the fused kernel implementation if (false && ENABLE_FUSION && i + 4 < cgraph->n_nodes && cgraph->nodes[i+1]->op == GGML_OP_VIEW && cgraph->nodes[i+2]->op == GGML_OP_FUSED_RMS_NORM &&