metal: template GLU kernels to support f16/f32 (#23882)

Drops the hardcoded f32 GLU kernels in favor of a single template. We now load/store in the native tensor type (half or float) to save memory bandwidth, but keep the actual ALU compute in float to avoid exploding math in geglu/swiglu. Also opened up the dispatch gate to allow f16 inputs.
2026-06-01 15:40:28 +03:00
parent 55ac0909e5
commit 95b8b8ec1a
2 changed files with 67 additions and 31 deletions
@@ -1107,7 +1107,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                case GGML_GLU_OP_SWIGLU_OAI:
                case GGML_GLU_OP_GEGLU_ERF:
                case GGML_GLU_OP_GEGLU_QUICK:
-                    return ggml_is_contiguous_1(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                    return ggml_is_contiguous_1(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
               default:
                    return false;
            }