metal : avoid too much shared memory usage with mul_mat_id

ggml-ci

Author: ggerganov

ggml/src/ggml-metal/ggml-metal.m

@@ -2856,17 +2856,19 @@ static void ggml_metal_encode_node(
                 // ne21 = n_rows
                 const int dst_rows = ne20*ne21;
                 const int dst_rows_min = n_as;
-                const int dst_rows_max = (device.maxThreadgroupMemoryLength - 32 - 8192)/4;
+                const int dst_rows_max = (device.maxThreadgroupMemoryLength/2 - 8192)/4;
 
                 // max size of the rowids array in the kernel shared buffer
-                GGML_ASSERT(dst_rows <= dst_rows_max);
+                //GGML_ASSERT(dst_rows <= dst_rows_max);
 
                 // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                 // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
                 if ([device supportsFamily:MTLGPUFamilyApple7] &&
                         ne00 % 32 == 0 && ne00 >= 64 &&
-                        ne01 / ne02 >= 512 &&    // NOTE: this is based on Mixtral shapes, might need adjustments
-                        dst_rows > dst_rows_min) {
+                        //ne01 / ne02 >= 512 &&    // NOTE: this is based on Mixtral shapes, might need adjustments
+                        dst_rows >  dst_rows_min &&
+                        dst_rows <= dst_rows_max) {
+
                     // some Metal matrix data types require aligned pointers
                     // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
                     switch (src0->type) {