OpenVINO Backend: Upgrade graph cache to use buffer pointers instead of names

.github/workflows/build-openvino.yml

@@ -78,18 +78,16 @@ jobs:
 
       - name: Test (CPU)
         id: cmake_test_cpu
-        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main --verbose --timeout 2000
 
       - name: Test (GPU)
         id: cmake_test_gpu
-        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
           export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main --verbose --timeout 3000
 
   openvino-windows-2022:
     runs-on: windows-2022
@@ -159,11 +157,10 @@ jobs:
       - name: Test (CPU)
         id: cmake_test_cpu
         shell: cmd
-        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           REM Find extracted OpenVINO folder dynamically
           for /d %%i in (openvino_toolkit\*) do set OPENVINO_ROOT=%%i
           call "%OPENVINO_ROOT%\setupvars.bat"
 
           cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -C Release --verbose --timeout 3000

docs/backend/OPENVINO.md

@@ -206,7 +206,7 @@ cmake -B build/ReleaseOV -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_OPENVINO=ON
 cmake --build build/ReleaseOV --parallel
 ```
 
-- **Windows:** Open a **Developer Command Prompt for VS 2022** (so the MSVC toolchain is on `PATH`), then run:
+- **Windows:** Open **x64 Native Tools Command Prompt for VS** (so the MSVC toolchain is on `PATH`), then run:
 
 ```cmd
 C:\Intel\openvino\setupvars.bat

ggml/src/ggml-openvino/ggml-decoder.h

@@ -20,6 +20,7 @@ struct ModelParams {
     int n_seq = 1;
     int n_heads_kv = -1;
     int head_size = -1;
+    int state_size = -1;  // for SSM molels, eg qwen35
     int32_t rope_params[15];
     bool mixed_rope_params = false;
     std::vector<int> swa_layers;
@@ -48,6 +49,37 @@ struct ComputeParams {
     int token_len_per_seq = -1;
     int past_kv_len = -1;
     int output_len = 1;
+
+    int cache_rs_reset_idx = -1;
+    int cache_rs_reset_len = -1;
+    // SSM/DeltaNet models otionally clear cache_r and cache_s of certain slots in the cgraph
+    // 3: [ 18432,     4,     1,     1] RESHAPE              cache_r_l0 (reshaped)
+    //    [ 18432,     4,     1,     1]            0: NONE        cache_r_l0
+    // 4: [ 18432,     1,     1,     1] VIEW                 cache_r_l0 (reshaped) (view)
+    //    [ 18432,     4,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)
+    // 5: [ 18432,     1,     1,     1] SCALE                cache_r_l0 (reshaped) (view) (view)
+    //    [ 18432,     1,     1,     1]            0: VIEW        cache_r_l0 (reshaped) (view)
+
+    int s_copy_active_slot_idx = -1;
+    int s_copy_active_slot_len = -1;
+    // SSM/DeltaNet models otionally reorder slots of state cache, to make the active slots contiguous
+    // leaf_5 is the inp->s_copy in llama-graph.cpp, eg if there are 8 slots in total and slot 3 and 7
+    // are active in the current batch, leaf_5 will be [3, 7, 5, 6, 4]
+    //  6: [     2,     1,     1,     1] VIEW                  (view)
+    //      [     2,     1,     1,     1]            0: NONE        leaf_5
+    //  7: [ 18432,     2,     1,     1] GET_ROWS             conv_states-0
+    //      [ 18432,     4,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)
+    //      [     2,     1,     1,     1]            1: VIEW         (view)
+    //  8: [     0,     1,     1,     1] VIEW                  (view)
+    //      [     2,     1,     1,     1]            0: NONE        leaf_5
+    //  9: [ 18432,     0,     1,     1] GET_ROWS             node_9
+    //      [ 18432,     4,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)
+    //      [     0,     1,     1,     1]            1: VIEW         (view)
+    // 10: [ 18432,     0,     1,     1] VIEW                 cache_r_l0 (view)
+    //      [ 18432,     4,     1,     1]            0: NONE        cache_r_l0
+    // 11: [ 18432,     0,     1,     1] CPY                  cache_r_l0 (view) (copy of )
+    //      [ 18432,     0,     1,     1]            0: GET_ROWS    node_9
+    //      [ 18432,     0,     1,     1]            1: VIEW        cache_r_l0 (view)
 };
 
 class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
@@ -59,8 +91,6 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
         std::map<std::string, ggml_tensor *> node_inputs;
         std::map<std::string, std::vector<std::pair<std::string, ggml_tensor *>>> node_inputs_views;
         std::vector<std::string> node_inputs_names;
-        ggml_tensor * node_output;
-        std::string node_output_name;
         int node_op_case = 0;
         void * data_addr;
     };
@@ -156,6 +186,10 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
 
     virtual std::vector<std::string> get_output_names(int node_idx) const override;
 
+    virtual std::string get_inplace_op_src(int node_idx) const override;
+
+    virtual bool is_view_like_alias_of(int node_idx, const std::string & view_src_name) const override;
+
     virtual const std::string & get_op_type() const override;
 
     virtual const std::string & get_op_type(int node_idx) const override;
@@ -173,23 +207,19 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
 
     virtual int get_op_case(int node_idx) const override { return m_node_info_list[node_idx].node_op_case; }
 
-    virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_inputs() const override {
+    virtual const std::map<std::string, ov::frontend::ggml::ModelInputInfo> & get_model_inputs() const override {
         return m_model_inputs;
     }
 
-    virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_extra_inputs() const override {
+    virtual const std::map<std::string, ov::frontend::ggml::ModelExtraInputInfo> & get_model_extra_inputs() const override {
         return m_model_extra_inputs;
     }
 
-    virtual const std::map<std::string, std::shared_ptr<ov::Tensor>> & get_model_extra_input_values() const {
-        return m_model_extra_input_values;
-    }
-
     virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_weights() const override {
         return m_model_weights;
     }
 
-    virtual std::vector<std::string> get_model_output_names() const override { return m_model_output_names; }
+    virtual std::set<std::string> get_model_output_names() const override { return m_model_output_names; }
 
     const std::map<std::string, ggml_tensor *> & get_model_outputs() const { return m_model_outputs; }
 
@@ -214,6 +244,8 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
 
     virtual bool has_mixed_rope_params() const override { return m_model_params.mixed_rope_params; }
 
+    virtual int get_ssm_state_size() const override { return m_model_params.state_size; }
+
     virtual std::map<std::string, std::string> get_kv_param_res_names() const override;
 
     virtual bool is_static() const override { return m_is_static; }
@@ -287,8 +319,12 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
         return op->op == GGML_OP_ROPE && tensor == op->src[2];
     }
 
+    // also returns true for cache_s and cache_r in SSM/DeltaNet models
     inline static bool is_kvcache(const ggml_tensor * tensor, const ggml_tensor * op) {
-        return tensor->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY ||
+        if (tensor == nullptr) {
+            return false;
+        }
+        return (tensor->buffer != nullptr && tensor->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY) ||
                (op != nullptr && op->op == GGML_OP_SET_ROWS && op->src[2] == tensor);
     }
 
@@ -301,7 +337,13 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
                op->src[1]->op == GGML_OP_NONE;
     }
 
-    std::string get_graph_input_ov_name(const ggml_tensor * tensor, const ggml_tensor * op) {
+    // the state permutation index input used in SSM/DeltaNet models (inp->s_copy in llama-graph.cpp)
+    inline static bool is_inp_s_copy(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] &&
+               op->src[0]->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY;
+    }
+
+    std::string get_graph_input_ov_name(const ggml_tensor * tensor, const ggml_tensor * op) const {
         if (is_inp_pos(tensor, op)) {
             return "inp_pos";
         }
@@ -321,6 +363,10 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
     void compute_model_inputs();
     void compute_model_outputs();
 
+    // True if tensor is the inp->s_copy index leaf gathered by a recurrent state cache GET_ROWS
+    // (possibly through a VIEW), so it gets a dynamic [1,1,1,-1] graph-input shape.
+    bool is_s_copy_leaf(const ggml_tensor * tensor) const;
+
     // Infer and propagate dynamic-dimension indices for all tensors in the GGML graph.
     void compute_node_dynamic_dims();
 
@@ -329,12 +375,11 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
     ggml_cgraph * m_cgraph = nullptr;
     std::map<std::string, ggml_tensor *> m_inputs;
 
-    std::map<std::string, std::shared_ptr<ov::Node>> m_model_inputs;
-    std::map<std::string, std::shared_ptr<ov::Node>> m_model_extra_inputs;
-    std::map<std::string, std::shared_ptr<ov::Tensor>> m_model_extra_input_values;
+    std::map<std::string, ov::frontend::ggml::ModelInputInfo> m_model_inputs;
+    std::map<std::string, ov::frontend::ggml::ModelExtraInputInfo> m_model_extra_inputs;
     std::map<std::string, std::shared_ptr<ov::Node>> m_model_weights;
     std::map<std::string, ggml_tensor *> m_model_outputs;
-    std::vector<std::string> m_model_output_names;
+    std::set<std::string> m_model_output_names;
     std::vector<NodeInfo> m_node_info_list;
     std::map<ggml_tensor *, int> m_node_dynamic_dims;
 

ggml/src/ggml-openvino/ggml-openvino-extra.cpp

@@ -252,14 +252,24 @@ ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_ten
         return layout;
     }
 
-    // Only handle 2D weight tensors
-    if (tensor->ne[2] != 1 || tensor->ne[3] != 1) {
+    // Most quantized weights use the existing 2D extraction path. MXFP4 also
+    // appears as 3D expert weights for MUL_MAT_ID, so allow that type through.
+    if (tensor->type != GGML_TYPE_MXFP4 && (tensor->ne[2] != 1 || tensor->ne[3] != 1)) {
         return layout;
     }
 
     int64_t n_elements = ggml_nelements(tensor);
     const size_t alignment = 64;  // Good for SIMD
 
+    if (tensor->type == GGML_TYPE_MXFP4 && (tensor->ne[2] > 1 || tensor->ne[3] > 1)) {
+        layout.weights_per_block = 32;
+        layout.is_symmetric = true;
+        layout.weights_size = ggml_nbytes(tensor);
+        layout.weights_offset = 0;
+        layout.total_size = layout.weights_size;
+        return layout;
+    }
+
     // Check if requantization is needed (NPU-specific)
     auto requant_type = ggml_openvino_get_requant_type(tensor, use_bias);
     if (requant_type.has_value()) {
@@ -334,6 +344,11 @@ ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_ten
     layout.is_symmetric = false;
 
     switch (tensor->type) {
+    case GGML_TYPE_MXFP4:
+        layout.is_u4 = true;
+        layout.is_symmetric = true;
+        break;
+
     case GGML_TYPE_Q4_0:
         layout.is_u4 = true;
         layout.is_symmetric = true;
@@ -369,9 +384,9 @@ ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_ten
     // Weights: U4 = n_elements/2 bytes, U8 = n_elements bytes
     layout.weights_size = layout.is_u4 ? (n_elements / 2) : n_elements;
 
-    // Scales: F16 per block
+    // Scales: F16 per block, except MXFP4 which stores one E8M0 byte per block.
     int64_t n_blocks = n_elements / layout.weights_per_block;
-    layout.scales_size = n_blocks * sizeof(uint16_t);  // F16 = 2 bytes
+    layout.scales_size = n_blocks * (tensor->type == GGML_TYPE_MXFP4 ? sizeof(uint8_t) : sizeof(uint16_t));
     // For symmetric quantization, no zp needed (weights stored as signed)
     if (layout.is_symmetric) {
         layout.zp_size = 0;