[GPU] Size LoRA opt SLM buffers to actual rank instead of hardcoded 256

Author: Aleksandr Nesterov

src/plugins/intel_gpu/src/graph/impls/ocl_v2/lora.cpp

@@ -58,11 +58,22 @@ class LoraRefBase : public KernelGenerator {
         jit_constants.add(make_type_jit_constants("ACCUMULATOR", params.get_input_layout(0).data_type));
         jit_constants.add(make_type_jit_constants("STATE", params.get_input_layout(2).data_type));
 
-        jit_constants.make("MAX_LORA_RANK", 256);
-
         LayoutJitter alpha_state_dims(params.input_layouts[3], params.in_port_to_shape_info_offset.at(3));
         jit_constants.make("LORA_RANK", alpha_state_dims.dim(ChannelName::FEATURE));
 
+        // MAX_LORA_RANK is the compile-time size/stride of the kernels' SLM buffers
+        // (fma_buff/reduce/tmp_buf, and MAX_GEMMA_N = MAX_LORA_RANK * LORA_COUNT). The LoRA
+        // rank is a static model property and every access stays within [0, LORA_RANK), so
+        // size the buffers to the actual rank (rounded up to the subgroup size, capped at the
+        // historical 256 upper bound) instead of always reserving 256. This keeps the SLM
+        // footprint proportional to the rank: it avoids local-memory overflow on small-SLM
+        // iGPUs without a matrix engine (clBuildProgram CL_OUT_OF_RESOURCES) and improves
+        // occupancy on all devices, while leaving MAX_GEMMA_SGK / dispatch unchanged.
+        size_t actual_lora_rank = extract_channel(ChannelName::FEATURE, params.input_layouts[3]);
+        size_t mlr_subgroup = params.get_input_layout(0).data_type == ov::element::f16 ? 16ul : 8ul;
+        size_t max_lora_rank = std::min<size_t>(256ul, round_up_to(std::max<size_t>(actual_lora_rank, 1ul), mlr_subgroup));
+        jit_constants.make("MAX_LORA_RANK", max_lora_rank);
+
         jit_constants.make("LORA_COUNT", get_lora_count(params));
 
         return jit_constants;