hyperray 0.0.1__py3-none-any.whl

hyperdrone/_vendor/rl-tools/include/rl_tools/dyn/operations_generic.h

@@ -0,0 +1,769 @@
+#include "../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_DYN_OPERATIONS_GENERIC_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_DYN_OPERATIONS_GENERIC_H
+
+#include "tensor_operations_generic.h"
+
+#include "../rl_tools.h"
+
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools{
+    namespace dyn{
+
+        // --- Activation ---
+        template <typename MATH_DEVICE>
+        RL_TOOLS_FUNCTION_PLACEMENT inline float apply_activation(MATH_DEVICE& math_dev, ActivationFunction af, float x){
+            switch(af){
+                case ActivationFunction::IDENTITY: return x;
+                case ActivationFunction::RELU: return math::max(math_dev, x, 0.0f);
+                case ActivationFunction::GELU: { float a = math::FRAC_2_SQRTPI<float> * math::SQRT1_2<float> * 0.5f; return 0.5f * (x + x * math::tanh(math_dev, a * (0.044715f * x * x * x + x))); }
+                case ActivationFunction::TANH: return math::tanh(math_dev, x);
+                case ActivationFunction::FAST_TANH: return math::fast_tanh(math_dev, x);
+                case ActivationFunction::SIGMOID: return 1.0f / (1.0f + math::exp(math_dev, -x));
+                default: return x;
+            }
+        }
+
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT TI infer_flat_input_dim(const Layer<TI>& layer){
+            switch(layer.type){
+                case LayerType::DENSE: return layer.template as<const layers::Dense<TI>>().input_dim;
+                case LayerType::STANDARDIZE: return layer.template as<const layers::Standardize<TI>>().dim;
+                case LayerType::SEQUENTIAL: case LayerType::MLP:
+                    for(TI i = 0; i < layer.num_children; i++){
+                        TI dim = infer_flat_input_dim(layer.children[i]);
+                        if(dim > 0) return dim;
+                    }
+                    return 0;
+                default: return 0;
+            }
+        }
+
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT TI product(const TI* shape, TI rank){
+            TI s = 1;
+            for(TI i = 0; i < rank; i++) s *= shape[i];
+            return s;
+        }
+
+        // --- Shape propagation ---
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void compute_leaf_output_shape(Layer<TI>& layer, const TI* in_shape, TI in_rank){
+            TI in_size = product(in_shape, in_rank);
+            auto replace_last_dim = [&](TI new_dim){
+                layer.output_rank = in_rank; layer.output_size = 1;
+                for(TI i = 0; i < in_rank - 1; i++){ layer.output_shape[i] = in_shape[i]; layer.output_size *= in_shape[i]; }
+                layer.output_shape[in_rank - 1] = new_dim; layer.output_size *= new_dim;
+            };
+            // Mirror real nn_models shape propagation: preserve all leading batch/time
+            // dims, only transform the last 3 (HWC for Conv/Pool) or the last N dims.
+            auto replace_last_three = [&](TI new_h, TI new_w, TI new_c){
+                layer.output_rank = in_rank;
+                for(TI i = 0; i < in_rank - 3; i++) layer.output_shape[i] = in_shape[i];
+                layer.output_shape[in_rank - 3] = new_h;
+                layer.output_shape[in_rank - 2] = new_w;
+                layer.output_shape[in_rank - 1] = new_c;
+                TI leading = 1; for(TI i = 0; i < in_rank - 3; i++) leading *= in_shape[i];
+                layer.output_size = leading * new_h * new_w * new_c;
+            };
+            auto collapse_last_three_to_one = [&](TI flat_dim){
+                layer.output_rank = in_rank - 2;
+                for(TI i = 0; i < in_rank - 3; i++) layer.output_shape[i] = in_shape[i];
+                layer.output_shape[in_rank - 3] = flat_dim;
+                TI leading = 1; for(TI i = 0; i < in_rank - 3; i++) leading *= in_shape[i];
+                layer.output_size = leading * flat_dim;
+            };
+            auto insert_last_three = [&](TI h, TI w, TI c){
+                layer.output_rank = in_rank + 2;
+                for(TI i = 0; i < in_rank - 1; i++) layer.output_shape[i] = in_shape[i];
+                layer.output_shape[in_rank - 1] = h;
+                layer.output_shape[in_rank] = w;
+                layer.output_shape[in_rank + 1] = c;
+                layer.output_size = in_size;
+            };
+            switch(layer.type){
+                case LayerType::DENSE: replace_last_dim(layer.template as<layers::Dense<TI>>().output_dim); break;
+                case LayerType::GRU: replace_last_dim(layer.template as<layers::GRU<TI>>().hidden_dim); break;
+                case LayerType::CONV2D: {
+                    if(in_rank < 3){ layer.output_size = 0; break; }
+                    auto& c = layer.template as<layers::Conv2d<TI>>();
+                    TI ih = in_shape[in_rank-3], iw = in_shape[in_rank-2];
+                    TI oh = (ih+2*c.padding_h-c.kernel_height)/c.stride_h+1;
+                    TI ow = (iw+2*c.padding_w-c.kernel_width)/c.stride_w+1;
+                    replace_last_three(oh, ow, c.output_channels);
+                    break;
+                }
+                case LayerType::MAX_POOL2D: {
+                    if(in_rank < 3){ layer.output_size = 0; break; }
+                    auto& mp = layer.template as<layers::MaxPool2d<TI>>();
+                    TI ih = in_shape[in_rank-3], iw = in_shape[in_rank-2], ch = in_shape[in_rank-1];
+                    TI oh = (ih+2*mp.padding_h-mp.kernel_height)/mp.stride_h+1;
+                    TI ow = (iw+2*mp.padding_w-mp.kernel_width)/mp.stride_w+1;
+                    replace_last_three(oh, ow, ch);
+                    break;
+                }
+                case LayerType::AVG_POOL2D: {
+                    if(in_rank < 3){ layer.output_size = 0; break; }
+                    TI ch = in_shape[in_rank-1];
+                    // Global avg pool collapses H,W → one value per channel; preserve leading dims, last dim becomes C.
+                    layer.output_rank = in_rank - 2;
+                    for(TI i = 0; i < in_rank - 3; i++) layer.output_shape[i] = in_shape[i];
+                    layer.output_shape[in_rank - 3] = ch;
+                    TI leading = 1; for(TI i = 0; i < in_rank - 3; i++) leading *= in_shape[i];
+                    layer.output_size = leading * ch;
+                    break;
+                }
+                case LayerType::FLATTEN: {
+                    if(in_rank >= 3){
+                        TI flat = 1;
+                        for(TI i = in_rank-3; i < in_rank; i++) flat *= in_shape[i];
+                        collapse_last_three_to_one(flat);
+                    }
+                    else{ layer.output_rank = in_rank; for(TI i = 0; i < in_rank; i++) layer.output_shape[i] = in_shape[i]; layer.output_size = in_size; }
+                    break;
+                }
+                case LayerType::UNFLATTEN: {
+                    auto& u = layer.template as<layers::Unflatten<TI>>();
+                    if(u.height > 0 && u.width > 0 && u.channels > 0){
+                        insert_last_three(u.height, u.width, u.channels);
+                    }
+                    else{
+                        // Unflatten requires its target (height, width, channels) to be known — typically loaded
+                        // from the checkpoint's saved attributes. If they're missing we cannot recover them:
+                        // spatial = height * width is underdetermined and channels has no independent source here.
+                        // Signal failure explicitly via output_size=0 so the composite propagates the failure.
+                        layer.output_rank = 0;
+                        layer.output_size = 0;
+                    }
+                    break;
+                }
+                case LayerType::SAMPLE_AND_SQUASH: {
+                    replace_last_dim(in_shape[in_rank-1] / 2);
+                    break;
+                }
+                default: {
+                    layer.output_rank = in_rank; layer.output_size = in_size;
+                    for(TI i = 0; i < in_rank; i++) layer.output_shape[i] = in_shape[i];
+                    break;
+                }
+            }
+        }
+
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void propagate_parallel_output(Layer<TI>& layer){
+            auto* p = layer.data ? &layer.template as<const layers::Parallel<TI>>() : nullptr;
+            if(p == nullptr || p->num_branches == 0){ layer.output_size = 0; return; }
+            TI num_branches = p->num_branches;
+            TI rank = layer.children[0].output_rank;
+            TI concat_last = 0;
+            TI total_size = 0;
+            for(TI i = 0; i < num_branches; i++){
+                concat_last += layer.children[i].output_shape[rank - 1];
+                total_size += layer.children[i].output_size;
+            }
+            layer.output_rank = rank;
+            layer.output_size = total_size;
+            for(TI i = 0; i < rank - 1; i++) layer.output_shape[i] = layer.children[0].output_shape[i];
+            layer.output_shape[rank - 1] = concat_last;
+            if(p->has_head){
+                TI head_idx = num_branches;
+                propagate_shapes(layer.children[head_idx], layer.output_shape, layer.output_rank);
+                layer.output_rank = layer.children[head_idx].output_rank;
+                layer.output_size = layer.children[head_idx].output_size;
+                for(TI i = 0; i < layer.output_rank; i++) layer.output_shape[i] = layer.children[head_idx].output_shape[i];
+            }
+        }
+
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void propagate_shapes(Layer<TI>& layer, const TensorTuple<TI>& inputs){
+            if(layer.type == LayerType::PARALLEL && layer.num_children >= 2){
+                auto* p = layer.data ? &layer.template as<const layers::Parallel<TI>>() : nullptr;
+                TI num_branches = p ? p->num_branches : 0;
+                for(TI i = 0; i < num_branches && i < inputs.num_tensors; i++){
+                    propagate_shapes(layer.children[i], inputs.tensors[i].shape, inputs.tensors[i].rank);
+                }
+                propagate_parallel_output(layer);
+            } else if(inputs.num_tensors > 0){
+                propagate_shapes(layer, inputs.tensors[0].shape, inputs.tensors[0].rank);
+            }
+        }
+
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void propagate_shapes(Layer<TI>& layer, const TI* in_shape, TI in_rank){
+            if(layer.num_children == 0){
+                compute_leaf_output_shape(layer, in_shape, in_rank);
+            }
+            else if(layer.type == LayerType::PARALLEL){
+                layer.output_size = 0;
+            }
+            else{
+                const TI* cur_shape = in_shape; TI cur_rank = in_rank;
+                for(TI i = 0; i < layer.num_children; i++){
+                    propagate_shapes(layer.children[i], cur_shape, cur_rank);
+                    if(layer.children[i].output_size == 0){ layer.output_size = 0; return; }
+                    cur_shape = layer.children[i].output_shape;
+                    cur_rank = layer.children[i].output_rank;
+                }
+                if(layer.type == LayerType::RESNET_BLOCK && layer.num_children == 3){
+                    propagate_shapes(layer.children[2], in_shape, in_rank);
+                }
+                TI last = (layer.type == LayerType::RESNET_BLOCK) ? 1 : layer.num_children - 1;
+                layer.output_rank = layer.children[last].output_rank;
+                layer.output_size = layer.children[last].output_size;
+                for(TI i = 0; i < layer.output_rank; i++) layer.output_shape[i] = layer.children[last].output_shape[i];
+            }
+        }
+
+        // --- Buffer sizing (no switch needed for composites — just recurse children) ---
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT TI max_output_size(const Layer<TI>& layer){
+            TI m = layer.output_size;
+            if(layer.type == LayerType::PARALLEL){
+                auto* p = layer.data ? &layer.template as<const layers::Parallel<TI>>() : nullptr;
+                TI num_branches = p ? p->num_branches : 0;
+                TI concat = 0;
+                for(TI i = 0; i < num_branches; i++) concat += layer.children[i].output_size;
+                if(concat > m) m = concat;
+            }
+            for(TI i = 0; i < layer.num_children; i++){ TI s = max_output_size(layer.children[i]); if(s > m) m = s; }
+            return m;
+        }
+        template <typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT TI max_scratch_size(const Layer<TI>& layer){
+            TI m = 0;
+            // Leaf-specific scratch needs
+            switch(layer.type){
+                case LayerType::GRU: {
+                    auto& g = layer.template as<const layers::GRU<TI>>();
+                    TI batch = (layer.output_rank == 3) ? layer.output_shape[1] : (layer.output_size / g.hidden_dim);
+                    m = batch * 3 * g.hidden_dim;
+                    break;
+                }
+                case LayerType::RESNET_BLOCK: {
+                    m = layer.children[0].output_size; // conv1 intermediate
+                    TI child_max = 0;
+                    for(TI i = 0; i < layer.num_children; i++){ TI s = max_scratch_size(layer.children[i]); if(s > child_max) child_max = s; }
+                    m += child_max;
+                    return m;
+                }
+                case LayerType::PARALLEL: {
+                    auto* p = layer.data ? &layer.template as<const layers::Parallel<TI>>() : nullptr;
+                    TI num_branches = p ? p->num_branches : 0;
+                    TI inter_size = 0;
+                    for(TI i = 0; i < num_branches; i++) inter_size += layer.children[i].output_size;
+                    if(inter_size > m) m = inter_size;
+                    break;
+                }
+                default: break;
+            }
+            // Recurse into children
+            for(TI i = 0; i < layer.num_children; i++){ TI s = max_scratch_size(layer.children[i]); if(s > m) m = s; }
+            return m;
+        }
+
+        // --- Leaf evaluate helpers ---
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void evaluate_dense(DEVICE& device, const layers::Dense<TI>& layer, const Tensor<TensorSpecification<TI>>& input, Tensor<TensorSpecification<TI>>& output){
+            TI batch_size = input.size() / layer.input_dim;
+            if(input.type == Type::FLOAT32 && layer.weights.type == Type::FLOAT32 && layer.biases.type == Type::FLOAT32 && output.type == Type::FLOAT32){
+                const float* w = reinterpret_cast<const float*>(layer.weights.data);
+                const float* bias = reinterpret_cast<const float*>(layer.biases.data);
+                const float* in = reinterpret_cast<const float*>(input.data);
+                float* out = reinterpret_cast<float*>(output.data);
+                for(TI b = 0; b < batch_size; b++)
+                    for(TI o = 0; o < layer.output_dim; o++){
+                        float acc = bias[o];
+                        for(TI i = 0; i < layer.input_dim; i++) acc += w[o * layer.input_dim + i] * in[b * layer.input_dim + i];
+                        out[b * layer.output_dim + o] = apply_activation(device.math, layer.activation_function, acc);
+                    }
+            } else {
+                for(TI b = 0; b < batch_size; b++)
+                    for(TI o = 0; o < layer.output_dim; o++){
+                        float acc = get(device, layer.biases, o);
+                        for(TI i = 0; i < layer.input_dim; i++) acc += get(device, layer.weights, o * layer.input_dim + i) * get(device, input, b * layer.input_dim + i);
+                        set(device, output, apply_activation(device.math, layer.activation_function, acc), b * layer.output_dim + o);
+                    }
+            }
+        }
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void evaluate_step_gru(DEVICE& device, const layers::GRU<TI>& layer, const Tensor<TensorSpecification<TI>>& input, state::GRU<TI>& gru_state, Tensor<TensorSpecification<TI>>& scratch){
+            TI batch_size = input.size() / layer.input_dim, hidden_dim = layer.hidden_dim;
+            if(input.type == Type::FLOAT32 && layer.weights_input.type == Type::FLOAT32 && layer.biases_input.type == Type::FLOAT32 && layer.weights_hidden.type == Type::FLOAT32 && layer.biases_hidden.type == Type::FLOAT32 && layer.initial_hidden_state.type == Type::FLOAT32 && gru_state.hidden.type == Type::FLOAT32 && scratch.type == Type::FLOAT32){
+                const float* wi_ptr = reinterpret_cast<const float*>(layer.weights_input.data);
+                const float* bi_ptr = reinterpret_cast<const float*>(layer.biases_input.data);
+                const float* wh_ptr = reinterpret_cast<const float*>(layer.weights_hidden.data);
+                const float* bh_ptr = reinterpret_cast<const float*>(layer.biases_hidden.data);
+                const float* init_h = reinterpret_cast<const float*>(layer.initial_hidden_state.data);
+                const float* in = reinterpret_cast<const float*>(input.data);
+                float* hid = reinterpret_cast<float*>(gru_state.hidden.data);
+                float* sc = reinterpret_cast<float*>(scratch.data);
+                if(!gru_state.initialized){
+                    for(TI b = 0; b < batch_size; b++) for(TI h = 0; h < hidden_dim; h++) hid[b*hidden_dim+h] = init_h[h];
+                    gru_state.initialized = true;
+                }
+                TI input_dim = layer.input_dim;
+                for(TI b = 0; b < batch_size; b++){
+                    for(TI g = 0; g < 2*hidden_dim; g++){
+                        float wh = bh_ptr[g];
+                        for(TI h = 0; h < hidden_dim; h++) wh += wh_ptr[g*hidden_dim+h] * hid[b*hidden_dim+h];
+                        float wi = bi_ptr[g];
+                        for(TI i = 0; i < input_dim; i++) wi += wi_ptr[g*input_dim+i] * in[b*input_dim+i];
+                        sc[b*2*hidden_dim+g] = 1.0f / (1.0f + math::exp(device.math, -(wh+wi)));
+                    }
+                    for(TI h = 0; h < hidden_dim; h++){
+                        TI g = 2*hidden_dim+h;
+                        float wh_n = bh_ptr[g];
+                        for(TI hh = 0; hh < hidden_dim; hh++) wh_n += wh_ptr[g*hidden_dim+hh] * hid[b*hidden_dim+hh];
+                        float wi_n = bi_ptr[g];
+                        for(TI i = 0; i < input_dim; i++) wi_n += wi_ptr[g*input_dim+i] * in[b*input_dim+i];
+                        float r = sc[b*2*hidden_dim+h];
+                        float n_pre = r * wh_n + wi_n;
+                        float n = math::tanh(device.math, n_pre);
+                        float z = sc[b*2*hidden_dim+hidden_dim+h];
+                        sc[b*2*hidden_dim+h] = z * hid[b*hidden_dim+h] + (1.0f-z) * n;
+                    }
+                    for(TI h = 0; h < hidden_dim; h++) hid[b*hidden_dim+h] = sc[b*2*hidden_dim+h];
+                }
+            } else {
+                if(!gru_state.initialized){
+                    for(TI b = 0; b < batch_size; b++) for(TI h = 0; h < hidden_dim; h++) set(device, gru_state.hidden, get(device, layer.initial_hidden_state, h), b*hidden_dim+h);
+                    gru_state.initialized = true;
+                }
+                for(TI b = 0; b < batch_size; b++){
+                    for(TI g = 0; g < 2*hidden_dim; g++){
+                        float wh = get(device, layer.biases_hidden, g);
+                        for(TI h = 0; h < hidden_dim; h++) wh += get(device, layer.weights_hidden, g*hidden_dim+h) * get(device, gru_state.hidden, b*hidden_dim+h);
+                        float wi = get(device, layer.biases_input, g);
+                        for(TI i = 0; i < layer.input_dim; i++) wi += get(device, layer.weights_input, g*layer.input_dim+i) * get(device, input, b*layer.input_dim+i);
+                        set(device, scratch, 1.0f / (1.0f + math::exp(device.math, -(wh+wi))), b*2*hidden_dim+g);
+                    }
+                    for(TI h = 0; h < hidden_dim; h++){
+                        TI g = 2*hidden_dim+h;
+                        float wh_n = get(device, layer.biases_hidden, g);
+                        for(TI hh = 0; hh < hidden_dim; hh++) wh_n += get(device, layer.weights_hidden, g*hidden_dim+hh) * get(device, gru_state.hidden, b*hidden_dim+hh);
+                        float wi_n = get(device, layer.biases_input, g);
+                        for(TI i = 0; i < layer.input_dim; i++) wi_n += get(device, layer.weights_input, g*layer.input_dim+i) * get(device, input, b*layer.input_dim+i);
+                        float r = get(device, scratch, b*2*hidden_dim+h);
+                        float n_pre = r * wh_n + wi_n;
+                        float n = math::tanh(device.math, n_pre);
+                        float z = get(device, scratch, b*2*hidden_dim+hidden_dim+h);
+                        set(device, scratch, z * get(device, gru_state.hidden, b*hidden_dim+h) + (1.0f-z) * n, b*2*hidden_dim+h);
+                    }
+                    for(TI h = 0; h < hidden_dim; h++) set(device, gru_state.hidden, get(device, scratch, b*2*hidden_dim+h), b*hidden_dim+h);
+                }
+            }
+        }
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void evaluate_conv2d(DEVICE& device, const layers::Conv2d<TI>& layer, const Tensor<TensorSpecification<TI>>& input, Tensor<TensorSpecification<TI>>& output){
+            TI ih = input.shape[input.rank-3], iw = input.shape[input.rank-2];
+            TI batch = input.size() / (ih * iw * layer.input_channels);
+            TI oh = (ih + 2*layer.padding_h - layer.kernel_height) / layer.stride_h + 1;
+            TI ow = (iw + 2*layer.padding_w - layer.kernel_width) / layer.stride_w + 1;
+            bool all_f32 = input.type == Type::FLOAT32 && layer.weights.type == Type::FLOAT32 && layer.biases.type == Type::FLOAT32 && output.type == Type::FLOAT32;
+            if(all_f32){
+                const float* w = reinterpret_cast<const float*>(layer.weights.data);
+                const float* bias = reinterpret_cast<const float*>(layer.biases.data);
+                const float* in = reinterpret_cast<const float*>(input.data);
+                float* out = reinterpret_cast<float*>(output.data);
+                for(TI bi = 0; bi < batch; bi++)
+                    for(TI ohi = 0; ohi < oh; ohi++)
+                        for(TI owi = 0; owi < ow; owi++)
+                            for(TI oc = 0; oc < layer.output_channels; oc++){
+                                float acc = bias[oc];
+                                for(TI kh = 0; kh < layer.kernel_height; kh++){
+                                    TI ihi = ohi*layer.stride_h+kh-layer.padding_h; if(ihi >= ih) continue;
+                                    for(TI kw = 0; kw < layer.kernel_width; kw++){
+                                        TI iwi = owi*layer.stride_w+kw-layer.padding_w; if(iwi >= iw) continue;
+                                        for(TI ic = 0; ic < layer.input_channels; ic++)
+                                            acc += w[((oc*layer.kernel_height+kh)*layer.kernel_width+kw)*layer.input_channels+ic]
+                                                 * in[((bi*ih+ihi)*iw+iwi)*layer.input_channels+ic];
+                                    }
+                                }
+                                out[((bi*oh+ohi)*ow+owi)*layer.output_channels+oc] = acc;
+                            }
+                if(layer.normalization == layers::Conv2d<TI>::Normalization::BATCH_NORM){
+                    const float* gamma = reinterpret_cast<const float*>(layer.gamma.data);
+                    const float* beta = reinterpret_cast<const float*>(layer.beta.data);
+                    const float* running_mean = reinterpret_cast<const float*>(layer.running_mean.data);
+                    const float* running_var = reinterpret_cast<const float*>(layer.running_var.data);
+                    float eps = 1e-5f;
+                    TI total = output.size();
+                    for(TI i = 0; i < total; i++){
+                        TI oc = i % layer.output_channels;
+                        out[i] = gamma[oc] * (out[i] - running_mean[oc]) / math::sqrt(device.math, running_var[oc] + eps) + beta[oc];
+                    }
+                }
+                if(layer.activation_function != ActivationFunction::IDENTITY){
+                    TI total = output.size();
+                    for(TI i = 0; i < total; i++) out[i] = apply_activation(device.math, layer.activation_function, out[i]);
+                }
+            } else {
+                for(TI bi = 0; bi < batch; bi++)
+                    for(TI ohi = 0; ohi < oh; ohi++)
+                        for(TI owi = 0; owi < ow; owi++)
+                            for(TI oc = 0; oc < layer.output_channels; oc++){
+                                float acc = get(device, layer.biases, oc);
+                                for(TI kh = 0; kh < layer.kernel_height; kh++){
+                                    TI ihi = ohi*layer.stride_h+kh-layer.padding_h; if(ihi >= ih) continue;
+                                    for(TI kw = 0; kw < layer.kernel_width; kw++){
+                                        TI iwi = owi*layer.stride_w+kw-layer.padding_w; if(iwi >= iw) continue;
+                                        for(TI ic = 0; ic < layer.input_channels; ic++)
+                                            acc += get(device, layer.weights, ((oc*layer.kernel_height+kh)*layer.kernel_width+kw)*layer.input_channels+ic)
+                                                 * get(device, input, ((bi*ih+ihi)*iw+iwi)*layer.input_channels+ic);
+                                    }
+                                }
+                                set(device, output, acc, ((bi*oh+ohi)*ow+owi)*layer.output_channels+oc);
+                            }
+                if(layer.normalization == layers::Conv2d<TI>::Normalization::BATCH_NORM){
+                    float eps = 1e-5f;
+                    for(TI i = 0; i < output.size(); i++){
+                        TI oc = i % layer.output_channels;
+                        set(device, output, get(device, layer.gamma, oc) * (get(device, output, i) - get(device, layer.running_mean, oc)) / math::sqrt(device.math, get(device, layer.running_var, oc) + eps) + get(device, layer.beta, oc), i);
+                    }
+                }
+                if(layer.activation_function != ActivationFunction::IDENTITY)
+                    for(TI i = 0; i < output.size(); i++) set(device, output, apply_activation(device.math, layer.activation_function, get(device, output, i)), i);
+            }
+        }
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void evaluate_max_pool2d(DEVICE& device, const layers::MaxPool2d<TI>& layer, const Tensor<TensorSpecification<TI>>& input, Tensor<TensorSpecification<TI>>& output){
+            TI ih = input.shape[input.rank-3], iw = input.shape[input.rank-2], ch = input.shape[input.rank-1];
+            TI batch = input.size() / (ih*iw*ch);
+            TI oh = (ih + 2*layer.padding_h - layer.kernel_height) / layer.stride_h + 1;
+            TI ow = (iw + 2*layer.padding_w - layer.kernel_width) / layer.stride_w + 1;
+            for(TI bi = 0; bi < batch; bi++)
+                for(TI ohi = 0; ohi < oh; ohi++)
+                    for(TI owi = 0; owi < ow; owi++)
+                        for(TI c = 0; c < ch; c++){
+                            float mx = -1e30f;
+                            for(TI kh = 0; kh < layer.kernel_height; kh++){
+                                TI ihi = ohi*layer.stride_h+kh-layer.padding_h; if(ihi >= ih) continue;
+                                for(TI kw = 0; kw < layer.kernel_width; kw++){
+                                    TI iwi = owi*layer.stride_w+kw-layer.padding_w; if(iwi >= iw) continue;
+                                    float v = get(device, input, ((bi*ih+ihi)*iw+iwi)*ch+c); if(v > mx) mx = v;
+                                }
+                            }
+                            set(device, output, mx, ((bi*oh+ohi)*ow+owi)*ch+c);
+                        }
+        }
+    }
+
+    // Forward declarations
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT bool evaluate(DEVICE& device, const dyn::Layer<TI>& layer, const dyn::Tensor<dyn::TensorSpecification<TI>>& input, dyn::Tensor<dyn::TensorSpecification<TI>>& output, dyn::Buffer<TI>& buffer);
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT bool evaluate_step(DEVICE& device, const dyn::Layer<TI>& layer, const dyn::Tensor<dyn::TensorSpecification<TI>>& input, dyn::State<TI>& state, dyn::Tensor<dyn::TensorSpecification<TI>>& output, dyn::Buffer<TI>& buffer);
+
+    namespace dyn{
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT bool evaluate_chain(DEVICE& device, const Layer<TI>& layer, const Tensor<TensorSpecification<TI>>& input, Tensor<TensorSpecification<TI>>& output, Buffer<TI>& buffer){
+            const Tensor<TensorSpecification<TI>>* current_input = &input;
+            for(TI i = 0; i < layer.num_children; i++){
+                Tensor<TensorSpecification<TI>>* current_output;
+                if(i == layer.num_children - 1) current_output = &output;
+                else current_output = (current_input == &buffer.tick) ? &buffer.tock : &buffer.tick;
+                if(!rl_tools::evaluate(device, layer.children[i], *current_input, *current_output, buffer)) return false;
+                current_input = current_output;
+            }
+            return true;
+        }
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT bool evaluate_step_chain(DEVICE& device, const Layer<TI>& layer, state::Composite<TI>& comp_state, const Tensor<TensorSpecification<TI>>& input, Tensor<TensorSpecification<TI>>& output, Buffer<TI>& buffer){
+            const Tensor<TensorSpecification<TI>>* current_input = &input;
+            for(TI i = 0; i < layer.num_children; i++){
+                Tensor<TensorSpecification<TI>>* current_output;
+                if(i == layer.num_children - 1) current_output = &output;
+                else current_output = (current_input == &buffer.tick) ? &buffer.tock : &buffer.tick;
+                if(!rl_tools::evaluate_step(device, layer.children[i], *current_input, comp_state.child_states[i], *current_output, buffer)) return false;
+                current_input = current_output;
+            }
+            return true;
+        }
+        template <typename DEVICE, typename TI>
+        RL_TOOLS_FUNCTION_PLACEMENT void evaluate_gru(DEVICE& device, const layers::GRU<TI>& gru, const Layer<TI>& layer, const Tensor<TensorSpecification<TI>>& input, Tensor<TensorSpecification<TI>>& output, Buffer<TI>& buffer){
+            TI hidden_dim = gru.hidden_dim;
+            TI batch = (input.rank == 3) ? input.shape[1] : (input.size() / gru.input_dim);
+            state::GRU<TI> tmp_state;
+            TI state_shape[] = {batch, hidden_dim};
+            set_shape(tmp_state.hidden, (TI)2, state_shape);
+            tmp_state.hidden.type = Type::FLOAT32;
+            tmp_state.hidden.data = buffer.scratch.data;
+            tmp_state.initialized = false;
+            Tensor<TensorSpecification<TI>> gate_scratch;
+            TI gate_shape[] = {batch * 2 * hidden_dim};
+            set_shape(gate_scratch, (TI)1, gate_shape);
+            gate_scratch.type = Type::FLOAT32;
+            gate_scratch.data = reinterpret_cast<char*>(buffer.scratch.data) + batch * hidden_dim * sizeof(float);
+            if(input.rank == 2){
+                evaluate_step_gru(device, gru, input, tmp_state, gate_scratch);
+                for(TI b = 0; b < batch; b++) for(TI h = 0; h < hidden_dim; h++) set(device, output, get(device, tmp_state.hidden, b*hidden_dim+h), b*hidden_dim+h);
+            }
+            else{
+                TI seq_len = input.shape[0];
+                Tensor<TensorSpecification<TI>> step_input;
+                TI step_shape[] = {batch, gru.input_dim};
+                set_shape(step_input, (TI)2, step_shape);
+                step_input.type = input.type;
+                for(TI t = 0; t < seq_len; t++){
+                    step_input.data = reinterpret_cast<char*>(const_cast<void*>(input.data)) + t * batch * gru.input_dim * size_of<TI>(input.type);
+                    evaluate_step_gru(device, gru, step_input, tmp_state, gate_scratch);
+                    for(TI b = 0; b < batch; b++) for(TI h = 0; h < hidden_dim; h++) set(device, output, get(device, tmp_state.hidden, b*hidden_dim+h), (t*batch+b)*hidden_dim+h);
+                }
+            }
+        }
+    }
+
+    // --- Buffer malloc/free ---
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, dyn::Buffer<TI>& buffer){
+        TI tt_size = dyn::max_output_size(*buffer.layer);
+        TI sc_size = dyn::max_scratch_size(*buffer.layer);
+        TI tt[] = {tt_size}; dyn::set_shape(buffer.tick, (TI)1, tt); buffer.tick.type = dyn::Type::FLOAT32; rl_tools::malloc(device, buffer.tick);
+        dyn::set_shape(buffer.tock, (TI)1, tt); buffer.tock.type = dyn::Type::FLOAT32; rl_tools::malloc(device, buffer.tock);
+        if(sc_size > 0){ TI sc[] = {sc_size}; dyn::set_shape(buffer.scratch, (TI)1, sc); buffer.scratch.type = dyn::Type::FLOAT32; rl_tools::malloc(device, buffer.scratch); }
+    }
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, dyn::Buffer<TI>& buffer){
+        rl_tools::free(device, buffer.tick); rl_tools::free(device, buffer.tock); rl_tools::free(device, buffer.scratch);
+    }
+
+    // --- State malloc/free/reset ---
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, dyn::State<TI>& state){
+        state.type = state.layer->type;
+        if(state.type == dyn::LayerType::GRU){
+            auto& g = state.layer->template as<const dyn::layers::GRU<TI>>();
+            auto* gs = new dyn::state::GRU<TI>();
+            TI shape[] = {state.batch_size, g.hidden_dim};
+            dyn::set_shape(gs->hidden, (TI)2, shape); gs->hidden.type = dyn::Type::FLOAT32;
+            rl_tools::malloc(device, gs->hidden); gs->initialized = false;
+            state.data = gs;
+        }
+        else if(state.layer->num_children > 0){
+            auto* cs = new dyn::state::Composite<TI>();
+            cs->num_children = state.layer->num_children;
+            cs->child_states = new dyn::State<TI>[cs->num_children];
+            for(TI i = 0; i < cs->num_children; i++){
+                cs->child_states[i].batch_size = state.batch_size;
+                cs->child_states[i].layer = &state.layer->children[i];
+                rl_tools::malloc(device, cs->child_states[i]);
+            }
+            state.data = cs;
+        }
+        else{ state.data = nullptr; }
+    }
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, dyn::State<TI>& state){
+        if(state.type == dyn::LayerType::GRU){
+            auto* gs = reinterpret_cast<dyn::state::GRU<TI>*>(state.data);
+            rl_tools::free(device, gs->hidden); delete gs;
+        }
+        else if(state.data){
+            auto* cs = reinterpret_cast<dyn::state::Composite<TI>*>(state.data);
+            for(TI i = 0; i < cs->num_children; i++) rl_tools::free(device, cs->child_states[i]);
+            delete[] cs->child_states; delete cs;
+        }
+        state.data = nullptr;
+    }
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset(DEVICE& device, const dyn::Layer<TI>& layer, dyn::State<TI>& state){
+        if(state.type == dyn::LayerType::GRU) reinterpret_cast<dyn::state::GRU<TI>*>(state.data)->initialized = false;
+        else if(state.data){
+            auto* cs = reinterpret_cast<dyn::state::Composite<TI>*>(state.data);
+            for(TI i = 0; i < cs->num_children; i++) reset(device, layer.children[i], cs->child_states[i]);
+        }
+    }
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(DEVICE& device, DEVICE&, const dyn::State<TI>& src, dyn::State<TI>& dst){
+        if(src.type == dyn::LayerType::GRU){
+            auto* src_gs = reinterpret_cast<const dyn::state::GRU<TI>*>(src.data);
+            auto* dst_gs = reinterpret_cast<dyn::state::GRU<TI>*>(dst.data);
+            TI bytes = src_gs->hidden.size() * dyn::size_of<TI>(src_gs->hidden.type);
+            const char* s = reinterpret_cast<const char*>(src_gs->hidden.data);
+            char* d = reinterpret_cast<char*>(dst_gs->hidden.data);
+            for(TI i = 0; i < bytes; i++) d[i] = s[i];
+            dst_gs->initialized = src_gs->initialized;
+        }
+        else if(src.data && dst.data){
+            auto* src_cs = reinterpret_cast<const dyn::state::Composite<TI>*>(src.data);
+            auto* dst_cs = reinterpret_cast<dyn::state::Composite<TI>*>(dst.data);
+            for(TI i = 0; i < src_cs->num_children; i++) copy(device, device, src_cs->child_states[i], dst_cs->child_states[i]);
+        }
+    }
+
+    // --- Layer free ---
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, dyn::Layer<TI>& layer){
+        // Free children first
+        for(TI i = 0; i < layer.num_children; i++) rl_tools::free(device, layer.children[i]);
+        if(layer.children){ delete[] layer.children; layer.children = nullptr; }
+        // Free leaf data
+        switch(layer.type){
+            case dyn::LayerType::DENSE: { auto& d = layer.template as<dyn::layers::Dense<TI>>(); rl_tools::free(device, d.weights); rl_tools::free(device, d.biases); delete &d; break; }
+            case dyn::LayerType::GRU: { auto& g = layer.template as<dyn::layers::GRU<TI>>(); rl_tools::free(device, g.weights_input); rl_tools::free(device, g.biases_input); rl_tools::free(device, g.weights_hidden); rl_tools::free(device, g.biases_hidden); rl_tools::free(device, g.initial_hidden_state); delete &g; break; }
+            case dyn::LayerType::CONV2D: { auto& c = layer.template as<dyn::layers::Conv2d<TI>>(); rl_tools::free(device, c.weights); rl_tools::free(device, c.biases); if(c.normalization != dyn::layers::Conv2d<TI>::Normalization::NONE){ rl_tools::free(device, c.gamma); rl_tools::free(device, c.beta); if(c.normalization == dyn::layers::Conv2d<TI>::Normalization::BATCH_NORM){ rl_tools::free(device, c.running_mean); rl_tools::free(device, c.running_var); } } delete &c; break; }
+            case dyn::LayerType::MAX_POOL2D: delete reinterpret_cast<dyn::layers::MaxPool2d<TI>*>(layer.data); break;
+            case dyn::LayerType::STANDARDIZE: { auto& s = layer.template as<dyn::layers::Standardize<TI>>(); rl_tools::free(device, s.mean); rl_tools::free(device, s.precision); delete &s; break; }
+            case dyn::LayerType::EMBEDDING: { auto& e = layer.template as<dyn::layers::Embedding<TI>>(); rl_tools::free(device, e.weights); delete &e; break; }
+            case dyn::LayerType::PARALLEL: delete reinterpret_cast<dyn::layers::Parallel<TI>*>(layer.data); break;
+            default: break;
+        }
+        layer.data = nullptr;
+    }
+
+    // --- Top-level evaluate ---
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT bool evaluate(DEVICE& device, const dyn::Layer<TI>& layer, const dyn::Tensor<dyn::TensorSpecification<TI>>& input, dyn::Tensor<dyn::TensorSpecification<TI>>& output, dyn::Buffer<TI>& buffer){
+        if(layer.output_size == 0) return false;
+        if(output.capacity < layer.output_size) return false;
+        if(buffer.tick.data == nullptr) return false;
+        if(input.data == nullptr) return false;
+        dyn::set_shape(output, layer.output_rank, layer.output_shape);
+        output.type = dyn::Type::FLOAT32;
+        switch(layer.type){
+            case dyn::LayerType::DENSE: dyn::evaluate_dense(device, layer.template as<const dyn::layers::Dense<TI>>(), input, output); return true;
+            case dyn::LayerType::GRU: dyn::evaluate_gru(device, layer.template as<const dyn::layers::GRU<TI>>(), layer, input, output, buffer); return true;
+            case dyn::LayerType::CONV2D: dyn::evaluate_conv2d(device, layer.template as<const dyn::layers::Conv2d<TI>>(), input, output); return true;
+            case dyn::LayerType::MAX_POOL2D: dyn::evaluate_max_pool2d(device, layer.template as<const dyn::layers::MaxPool2d<TI>>(), input, output); return true;
+            case dyn::LayerType::SAMPLE_AND_SQUASH: {
+                TI last = input.shape[input.rank-1], half = last/2, batch = input.size()/last;
+                for(TI b = 0; b < batch; b++) for(TI i = 0; i < half; i++) rl_tools::set(device, output, math::tanh(device.math, rl_tools::get(device, input, b*last+i)), b*half+i);
+                return true;
+            }
+            case dyn::LayerType::STANDARDIZE: {
+                auto& s = layer.template as<const dyn::layers::Standardize<TI>>();
+                TI batch = input.size() / s.dim;
+                if(input.type == dyn::Type::FLOAT32 && s.mean.type == dyn::Type::FLOAT32 && s.precision.type == dyn::Type::FLOAT32 && output.type == dyn::Type::FLOAT32){
+                    const float* in = reinterpret_cast<const float*>(input.data);
+                    const float* mean = reinterpret_cast<const float*>(s.mean.data);
+                    const float* prec = reinterpret_cast<const float*>(s.precision.data);
+                    float* out = reinterpret_cast<float*>(output.data);
+                    for(TI b = 0; b < batch; b++) for(TI i = 0; i < s.dim; i++) out[b*s.dim+i] = (in[b*s.dim+i] - mean[i]) * prec[i];
+                } else {
+                    for(TI b = 0; b < batch; b++) for(TI i = 0; i < s.dim; i++) rl_tools::set(device, output, (rl_tools::get(device, input, b*s.dim+i) - rl_tools::get(device, s.mean, i)) * rl_tools::get(device, s.precision, i), b*s.dim+i);
+                }
+                return true;
+            }
+            case dyn::LayerType::FLATTEN: case dyn::LayerType::UNFLATTEN: {
+                output.type = input.type;
+                TI bytes = input.size() * dyn::size_of<TI>(input.type);
+                const char* src = reinterpret_cast<const char*>(input.data); char* dst = reinterpret_cast<char*>(output.data);
+                for(TI i = 0; i < bytes; i++) dst[i] = src[i];
+                return true;
+            }
+            case dyn::LayerType::AVG_POOL2D: {
+                TI ch = input.shape[input.rank-1], w = input.shape[input.rank-2], h = input.shape[input.rank-3], batch = input.size()/(h*w*ch);
+                float scale = 1.0f / (float)(h*w);
+                for(TI b = 0; b < batch; b++) for(TI c = 0; c < ch; c++){
+                    float sum = 0; for(TI hi = 0; hi < h; hi++) for(TI wi = 0; wi < w; wi++) sum += rl_tools::get(device, input, ((b*h+hi)*w+wi)*ch+c);
+                    rl_tools::set(device, output, sum*scale, b*ch+c);
+                }
+                return true;
+            }
+            case dyn::LayerType::SEQUENTIAL: case dyn::LayerType::MLP:
+                return dyn::evaluate_chain(device, layer, input, output, buffer);
+            case dyn::LayerType::PARALLEL:
+                return false;
+            case dyn::LayerType::RESNET_BLOCK: {
+                if(!rl_tools::evaluate(device, layer.children[0], input, buffer.scratch, buffer)) return false;
+                if(!rl_tools::evaluate(device, layer.children[1], buffer.scratch, output, buffer)) return false;
+                if(layer.num_children == 3){
+                    dyn::Buffer<TI> sub_buffer = buffer;
+                    TI conv1_bytes = layer.children[0].output_size * sizeof(float);
+                    sub_buffer.scratch.data = reinterpret_cast<char*>(buffer.scratch.data) + conv1_bytes;
+                    sub_buffer.scratch.capacity = buffer.scratch.capacity - layer.children[0].output_size;
+                    if(!rl_tools::evaluate(device, layer.children[2], input, buffer.scratch, sub_buffer)) return false;
+                }
+                const dyn::Tensor<dyn::TensorSpecification<TI>>& shortcut = (layer.num_children == 3) ? buffer.scratch : input;
+                for(TI i = 0; i < output.size(); i++){
+                    float val = rl_tools::get(device, output, i) + rl_tools::get(device, shortcut, i);
+                    rl_tools::set(device, output, val > 0 ? val : 0, i);
+                }
+                return true;
+            }
+            default: return false;
+        }
+    }
+
+    // --- Top-level evaluate (TensorTuple) ---
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT bool evaluate(DEVICE& device, const dyn::Layer<TI>& layer, const dyn::TensorTuple<TI>& inputs, dyn::Tensor<dyn::TensorSpecification<TI>>& output, dyn::Buffer<TI>& buffer){
+        if(layer.type != dyn::LayerType::PARALLEL){
+            if(inputs.num_tensors > 0) return rl_tools::evaluate(device, layer, inputs.tensors[0], output, buffer);
+            return false;
+        }
+        auto* p = layer.data ? &layer.template as<const dyn::layers::Parallel<TI>>() : nullptr;
+        if(p == nullptr || p->num_branches == 0) return false;
+        TI num_branches = p->num_branches;
+        if(layer.output_size == 0) return false;
+        if(output.capacity < layer.output_size) return false;
+        dyn::set_shape(output, layer.output_rank, layer.output_shape);
+        output.type = dyn::Type::FLOAT32;
+        if(inputs.num_tensors < num_branches) return false;
+        dyn::Tensor<dyn::TensorSpecification<TI>> intermediates[dyn::TensorTuple<TI>::MAX_TENSORS];
+        TI offset = 0;
+        for(TI i = 0; i < num_branches; i++){
+            auto& child = layer.children[i];
+            dyn::set_shape(intermediates[i], child.output_rank, child.output_shape);
+            intermediates[i].type = dyn::Type::FLOAT32;
+            intermediates[i].data = reinterpret_cast<char*>(buffer.scratch.data) + offset * sizeof(float);
+            intermediates[i].capacity = child.output_size;
+            if(!rl_tools::evaluate(device, child, inputs.tensors[i], intermediates[i], buffer)) return false;
+            offset += child.output_size;
+        }
+        TI rank = layer.children[0].output_rank;
+        TI last_dims[dyn::TensorTuple<TI>::MAX_TENSORS];
+        TI last_out = 0;
+        for(TI i = 0; i < num_branches; i++){ last_dims[i] = layer.children[i].output_shape[rank - 1]; last_out += last_dims[i]; }
+        TI batch = layer.children[0].output_size / last_dims[0];
+        dyn::Tensor<dyn::TensorSpecification<TI>>& concat_target = p->has_head ? buffer.tick : output;
+        TI concat_shape[dyn::TensorSpecification<TI>::MAX_RANK];
+        for(TI i = 0; i < rank - 1; i++) concat_shape[i] = layer.children[0].output_shape[i];
+        concat_shape[rank - 1] = last_out;
+        dyn::set_shape(concat_target, rank, concat_shape);
+        concat_target.type = dyn::Type::FLOAT32;
+        for(TI b = 0; b < batch; b++){
+            TI col = 0;
+            for(TI i = 0; i < num_branches; i++){
+                for(TI j = 0; j < last_dims[i]; j++){
+                    rl_tools::set(device, concat_target, rl_tools::get(device, intermediates[i], b * last_dims[i] + j), b * last_out + col + j);
+                }
+                col += last_dims[i];
+            }
+        }
+        if(p->has_head) return rl_tools::evaluate(device, layer.children[num_branches], buffer.tick, output, buffer);
+        return true;
+    }
+
+    // --- Top-level evaluate_step ---
+    template <typename DEVICE, typename TI>
+    RL_TOOLS_FUNCTION_PLACEMENT bool evaluate_step(DEVICE& device, const dyn::Layer<TI>& layer, const dyn::Tensor<dyn::TensorSpecification<TI>>& input, dyn::State<TI>& state, dyn::Tensor<dyn::TensorSpecification<TI>>& output, dyn::Buffer<TI>& buffer){
+        if(layer.output_size == 0) return false;
+        if(output.capacity < layer.output_size) return false;
+        if(buffer.tick.data == nullptr) return false;
+        switch(layer.type){
+            case dyn::LayerType::GRU: {
+                auto& g = layer.template as<const dyn::layers::GRU<TI>>();
+                auto* gs = reinterpret_cast<dyn::state::GRU<TI>*>(state.data);
+                if(!gs) return false;
+                dyn::set_shape(output, layer.output_rank, layer.output_shape); output.type = dyn::Type::FLOAT32;
+                dyn::Tensor<dyn::TensorSpecification<TI>> gate_scratch;
+                TI batch = input.size() / g.input_dim;
+                TI gs_shape[] = {batch * 2 * g.hidden_dim};
+                dyn::set_shape(gate_scratch, (TI)1, gs_shape); gate_scratch.type = dyn::Type::FLOAT32; gate_scratch.data = buffer.scratch.data;
+                dyn::evaluate_step_gru(device, g, input, *gs, gate_scratch);
+                for(TI b = 0; b < batch; b++) for(TI h = 0; h < g.hidden_dim; h++) rl_tools::set(device, output, rl_tools::get(device, gs->hidden, b*g.hidden_dim+h), b*g.hidden_dim+h);
+                return true;
+            }
+            case dyn::LayerType::SEQUENTIAL: case dyn::LayerType::MLP: {
+                auto* cs = reinterpret_cast<dyn::state::Composite<TI>*>(state.data);
+                if(!cs) return false;
+                return dyn::evaluate_step_chain(device, layer, *cs, input, output, buffer);
+            }
+            default: return evaluate(device, layer, input, output, buffer);
+        }
+    }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif