OpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback

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
@@ -710,11 +710,15 @@ Boolean flags follow a uniform convention: set to a **positive integer** (e.g. `
 |-----------------------------------|-----------|------------|-------------------------------------------------------------------------------------------------------------|
 | `GGML_OPENVINO_DEVICE`            | String    | `CPU`      | Specify the target device (CPU, GPU, NPU). On systems with multiple GPUs, use `GPU.0` or `GPU.1` to explicitly target specific GPU. See [OpenVINO GPU Device](https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/gpu-device.html). When set to **NPU**, static compilation mode is enabled for optimal performance. |
 | `GGML_OPENVINO_CACHE_DIR`         | String    | `not set`  | Directory for OpenVINO model caching (recommended: `/tmp/ov_cache`). Enables model caching when set. **Not supported on NPU devices.** |
+| `GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR` | String | `not set` | Directory for the frontend compiled-model cache. When set, OpenVINO compiled models are exported as blobs and imported on later runs to skip weight requantization, graph conversion, and compilation for matching single-graph models. |
 | `GGML_OPENVINO_PREFILL_CHUNK_SIZE`| Integer   | `256`      | Token chunk size for **NPU** prefill (NPU-only; ignored on CPU/GPU). Must be a positive integer; otherwise the default is used. |
 | `GGML_OPENVINO_STATEFUL_EXECUTION`| Boolean   | `0`        | Enable stateful KV cache for better performance. Recommended on CPU, GPU.                                   |
 | `GGML_OPENVINO_DISABLE_CACHE`     | Boolean   | `0`        | Disable the in-process compiled-model / decoder cache (cache is on by default). Set to `1` to disable.      |
 | `GGML_OPENVINO_DISABLE_KV_SLICE`  | Boolean   | `0`        | Disable the KV-cache input-tensor slicing optimization (slicing is on by default on CPU/GPU). Set to `1` to disable. |
 | `GGML_OPENVINO_MANUAL_GQA_ATTN`   | Boolean   | device-based | Tri-state. When **unset**, manual GQA attention is enabled by default on `GPU` and disabled on other devices. Set to a positive integer to force-enable, or `0` to force-disable. |
+| `GGML_OPENVINO_MEMORY_OPTIMIZE`   | Boolean   | `0`        | Umbrella switch for compile-time memory reductions. Enables `GGML_OPENVINO_REDUCE_COMPILE_MEM` and, on GPU, `GGML_OPENVINO_RELEASE_WEIGHTS` unless those fine-grained variables are explicitly set. |
+| `GGML_OPENVINO_REDUCE_COMPILE_MEM`| Boolean   | inherits from `GGML_OPENVINO_MEMORY_OPTIMIZE` | Reduce compile-time host memory use by streaming weight requantization and avoiding extra weight-node materialization where possible. Set explicitly to override the umbrella switch. |
+| `GGML_OPENVINO_RELEASE_WEIGHTS`   | Boolean   | inherits from `GGML_OPENVINO_MEMORY_OPTIMIZE` on GPU | GPU-only. Release host weight buffers after the compiled model cache can reuse the device/plugin copy. Requires stable graph shapes; dynamic workloads that need recompilation should leave this disabled. |
 | `GGML_OPENVINO_PROFILING`         | Boolean   | `0`        | Enable execution-time profiling.                                                                            |
 | `GGML_OPENVINO_DUMP_CGRAPH`       | Boolean   | `0`        | Dump the GGML compute graph to `cgraph_ov.txt`.                                                             |
 | `GGML_OPENVINO_DUMP_IR`           | Boolean   | `0`        | Serialize OpenVINO IR files with timestamps.                                                                |

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

@@ -11,6 +11,8 @@
 #include <memory>
 #include <openvino/core/partial_shape.hpp>
 #include <optional>
+#include <set>
+#include <string>
 #include <vector>
 
 struct ModelParams {
@@ -20,6 +22,7 @@ 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 +51,47 @@ 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_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)
+
+    struct RsWriteback {
+        int slot_begin = 0;  // first cache slot written by the CPY
+        int src_begin = 0;   // where the copied data starts in the source tensor (in rows of it)
+    };
+
+    std::map<std::string, RsWriteback> rs_writebacks;
+    // Offsets of the state cache writeback CPY nodes, keyed by node name. They change with the
+    // batch (kv head, active sequence count, token count) and, with rollback enabled
+    // (cparams.n_rs_seq > 0), the conv state is written back once per snapshot slot, each snapshot
+    // taking a different conv_input window. Passed to the cached model as runtime inputs.
 };
 
 class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
@@ -59,8 +103,6 @@ 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 +198,10 @@ 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 +219,19 @@ 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 +256,8 @@ 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; }
@@ -235,6 +279,11 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
     static std::map<std::string, std::shared_ptr<ov::Node>> create_weight_nodes(ggml_cgraph * cgraph,
                                                                                 bool naive = false);
 
+    // Collect just the set of weight-tensor names referenced by the graph, without
+    // building (or requantizing) any OV weight nodes. Used by topology checks like
+    // is_model_splitted that only need name membership.
+    static std::set<std::string> collect_weight_names(ggml_cgraph * cgraph);
+
     const ggml_tensor * get_tensor_used_op(const ggml_tensor * tensor) const;
 
     const ggml_tensor * get_tensor_from_name(const std::string & name) const;
@@ -274,6 +323,12 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
         return op->op == GGML_OP_ROPE && tensor == op->src[1];
     }
 
+    // IMROPE packs 4 stacked position planes (t/h/w/e) into inp_pos, each of length
+    // n_tokens; other modes carry a single position per token.
+    inline static int get_inp_pos_n_planes(const ggml_tensor * op) {
+        return op->op_params[2] == GGML_ROPE_TYPE_IMROPE ? 4 : 1;
+    }
+
     inline static bool is_inp_emb(const ggml_tensor * tensor, const ggml_tensor * op) {
         return tensor->op == GGML_OP_GET_ROWS && op->op == GGML_OP_RMS_NORM;
     }
@@ -287,8 +342,12 @@ 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 +360,13 @@ 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 +386,10 @@ 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 +398,11 @@ 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

@@ -31,6 +31,7 @@ void ggml_openvino_device_config::init() {
         // String values (use ggml_openvino_getenv_str)
         "GGML_OPENVINO_DEVICE",
         "GGML_OPENVINO_CACHE_DIR",
+        "GGML_OPENVINO_DEBUG_NODE",
         // Integer values (use ggml_openvino_getenv_int)
         "GGML_OPENVINO_PREFILL_CHUNK_SIZE",
         // Boolean toggles (treated as int flags via ggml_openvino_getenv_int)
@@ -44,7 +45,12 @@ void ggml_openvino_device_config::init() {
         "GGML_OPENVINO_ENABLE_CACHE",
         "GGML_OPENVINO_DISABLE_CACHE",
         "GGML_OPENVINO_DISABLE_KV_SLICE",
+        "GGML_OPENVINO_ENABLE_FALLBACK",
         "GGML_OPENVINO_MANUAL_GQA_ATTN",
+        "GGML_OPENVINO_MEMORY_OPTIMIZE",
+        "GGML_OPENVINO_RELEASE_WEIGHTS",
+        "GGML_OPENVINO_REDUCE_COMPILE_MEM",
+        "GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR",
     };
 
     for (const char * const & env_var : env_var_names) {
@@ -168,6 +174,22 @@ int ggml_openvino_getenv_int(const char * var, int default_value) {
     return v ? std::atoi(v) : default_value;
 }
 
+bool ggml_openvino_reduce_compile_mem_enabled() {
+    const char * reduce_compile_mem = ggml_openvino_getenv_str("GGML_OPENVINO_REDUCE_COMPILE_MEM");
+    if (reduce_compile_mem != nullptr) {
+        return ggml_openvino_getenv_int("GGML_OPENVINO_REDUCE_COMPILE_MEM") != 0;
+    }
+    return ggml_openvino_getenv_int("GGML_OPENVINO_MEMORY_OPTIMIZE") != 0;
+}
+
+bool ggml_openvino_release_weights_enabled(const std::string & device) {
+    const char * release_weights = ggml_openvino_getenv_str("GGML_OPENVINO_RELEASE_WEIGHTS");
+    if (release_weights != nullptr) {
+        return device == "GPU" && ggml_openvino_getenv_int("GGML_OPENVINO_RELEASE_WEIGHTS") != 0;
+    }
+    return device == "GPU" && ggml_openvino_getenv_int("GGML_OPENVINO_MEMORY_OPTIMIZE") != 0;
+}
+
 // Check if running on NPU
 bool ggml_openvino_is_npu() {
     return ggml_openvino_get_device_config().is_npu;
@@ -252,14 +274,31 @@ 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. 3D expert weights for
+    // MUL_MAT_ID (MoE) are also supported, either as MXFP4 (packed, dedicated branch below) or via the
+    // generic sizing math below, which is shape-agnostic (based on total element count). Only reject 4D.
+    if (tensor->ne[3] != 1) {
         return layout;
     }
 
+    // 3D MoE expert weights that are not requantized (see below) always use the exact f16
+    // zero-point extraction (see extract_quantized_weights), which needs a wider zp slot than
+    // the packed integer zero point -- must be kept in sync with that function so the buffer
+    // sizing here matches what process_weight_tensor actually writes.
+    const bool for_gather_matmul = tensor->ne[2] > 1;
+
     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 +373,11 @@ 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;
@@ -369,12 +413,17 @@ 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
-    // For symmetric quantization, no zp needed (weights stored as signed)
+    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). Asymmetric
+    // for_gather_matmul (3D MoE expert) weights use an exact f16 zero point (see
+    // extract_quantized_weights/make_int8_weights/make_int4_weights), which needs one f16 per
+    // block instead of a packed u4/u8 integer zero point.
     if (layout.is_symmetric) {
         layout.zp_size = 0;
+    } else if (use_bias || for_gather_matmul) {
+        layout.zp_size = n_blocks * sizeof(uint16_t);
     } else {
         layout.zp_size = layout.is_u4 ? ((n_blocks + 1) / 2) : n_blocks;
     }