whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp

@@ -0,0 +1,65 @@
+#pragma once
+
+#include "openvino/op/op.hpp"
+
+namespace ov::op::internal {
+/// \note GatedDeltaNet op class is under development and subject to change
+///
+/// \brief Operator performing Gated Delta Net computation
+/// \ingroup ov_ops_cpp_api
+class OPENVINO_API GatedDeltaNet : public ov::op::Op {
+public:
+    OPENVINO_OP("GatedDeltaNet")
+
+    GatedDeltaNet() = default;
+    /// \brief Constructs a GatedDeltaNet operation.
+    ///
+    /// \param query Query tensor input.
+    /// \param key Key tensor input.
+    /// \param value Value tensor input.
+    /// \param recurrent_state Initial recurrent state tensor.
+    /// \param gate Gate tensor controlling state decay/update.
+    /// \param beta Beta tensor scaling the delta update.
+    /// \param fuse_qk_l2norm Enables fusing q/k L2-normalization into this op.
+    /// \param q_l2_norm_eps Epsilon used for query L2-normalization when fusion is enabled.
+    /// \param k_l2_norm_eps Epsilon used for key L2-normalization when fusion is enabled.
+    GatedDeltaNet(const Output<Node>& query,
+                  const Output<Node>& key,
+                  const Output<Node>& value,
+                  const Output<Node>& recurrent_state,
+                  const Output<Node>& gate,
+                  const Output<Node>& beta,
+                  const bool fuse_qk_l2norm = false,
+                  const float q_l2_norm_eps = 1e-6F,
+                  const float k_l2_norm_eps = 1e-6F);
+
+    /// \brief Constructs a GatedDeltaNet operation from input vector.
+    ///
+    /// \param args Input tensor vector in order: query, key, value, recurrent_state, gate, beta.
+    /// \param fuse_qk_l2norm Enables fusing q/k L2-normalization into this op.
+    /// \param q_l2_norm_eps Epsilon used for query L2-normalization when fusion is enabled.
+    /// \param k_l2_norm_eps Epsilon used for key L2-normalization when fusion is enabled.
+    GatedDeltaNet(const ov::OutputVector& args,
+                  const bool fuse_qk_l2norm = false,
+                  const float q_l2_norm_eps = 1e-6F,
+                  const float k_l2_norm_eps = 1e-6F);
+    void validate_and_infer_types() override;
+    bool visit_attributes(AttributeVisitor& visitor) override;
+    std::shared_ptr<ov::Node> clone_with_new_inputs(const ov::OutputVector& new_args) const override;
+    bool get_fuse_qk_l2norm() const {
+        return m_fuse_qk_l2norm;
+    }
+    float get_q_l2_norm_eps() const {
+        return m_q_l2_norm_eps;
+    }
+    float get_k_l2_norm_eps() const {
+        return m_k_l2_norm_eps;
+    }
+
+private:
+    bool m_fuse_qk_l2norm = false;
+    float m_q_l2_norm_eps = 1e-6F;
+    float m_k_l2_norm_eps = 1e-6F;
+};
+
+}  // namespace ov::op::internal

data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp

@@ -18,16 +18,9 @@ namespace op {
 OutputVector translate_get_rows(const NodeContext & context) {
     num_inputs_check(context, 2, 2);
 
-    int op_case = context.get_op_case();
-
     Output<Node> res;
-    auto data = context.get_input(0);
-    auto indices = context.get_input(1);
-
-    if (op_case == 2) {
-        // The input comes from a VIEW
-        indices = process_view_input(context, 1);
-    }
+    auto data = process_view_input_new(context, 0);
+    auto indices = process_view_input_new(context, 1);
 
     // data[1,b,x,y] ind[1,1,b,x'] test-backend-ops case
     // data[x,y] ind[1,1,1,x'] normal case

data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp

@@ -4,6 +4,7 @@
 
 #include <memory>
 #include <openvino/core/node_output.hpp>
+#include <openvino/op/clamp.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/gelu.hpp>
 #include <openvino/op/multiply.hpp>
@@ -21,23 +22,26 @@ OutputVector translate_glu_geglu(const NodeContext & context) {
     ov::Output<ov::Node> src0;
     ov::Output<ov::Node> src1;
     if (context.get_input_size() == 2) {
-        src0 = context.get_input(0);
-        src1 = context.get_input(1);
+        // Inputs may be VIEW slices of a combined gate_up tensor (MoE experts):
+        // resolve them so each half has its real sliced shape, not the base tensor.
+        src0 = process_view_input_new(context, 0);
+        src1 = process_view_input_new(context, 1);
     } else {
         // GGML splits along ne[0] (OV last axis) using floor division: nc = ne[0] / 2.
         // Both halves are nc elements; if the dimension is odd, the last element is dropped.
         // Use Slice instead of Split to handle odd dimensions correctly.
-        auto combined = context.get_input(0);
+        // Resolve a VIEW input (e.g. non-contiguous slice) to its real shape first.
+        auto combined = process_view_input_new(context, 0);
         auto combined_shape = combined.get_partial_shape();
         int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
         int64_t nc = last_dim_val / 2;
 
-        auto axis   = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
-        auto step   = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
         auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
-        auto stop0  = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto stop0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
         auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
-        auto stop1  = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
+        auto stop1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
 
         src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
         src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
@@ -49,6 +53,16 @@ OutputVector translate_glu_geglu(const NodeContext & context) {
         std::swap(src0, src1);
     }
 
+    if (context.is_static()) {
+        // TODO: Temporary solution for NPU accuracy issue due to fp16 overflow
+        // To be removed once permanent solution is implemented
+        // Justification:
+        // For |x| > 5, GELU(x) ≈ max(x, 0)  (behaves like ReLU)
+        // So Clamp(-10, 10) only affects values where GELU would return ≈ x anyway.
+        // The only loss: values > 10 get mapped to 10 instead of x.
+        // In practice, FFN intermediates rarely exceed 10 after GEGLU gating.
+        src0 = std::make_shared<ov::op::v0::Clamp>(src0, -10.0, 10.0);
+    }
     auto gelu = std::make_shared<ov::op::v7::Gelu>(src0);
     auto res = std::make_shared<ov::op::v1::Multiply>(gelu, src1);
 

data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp

@@ -3,8 +3,11 @@
 #include "../utils.h"
 
 #include <cstdint>
+#include <limits>
 #include <memory>
 #include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/clamp.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/multiply.hpp>
 #include <openvino/op/sigmoid.hpp>
@@ -15,29 +18,32 @@ namespace frontend {
 namespace ggml {
 namespace op {
 
-OutputVector translate_glu_swiglu(const NodeContext & context) {
+static std::pair<ov::Output<ov::Node>, ov::Output<ov::Node>> get_glu_inputs(const NodeContext & context) {
     num_inputs_check(context, 1, 2);
 
     ov::Output<ov::Node> src0;
     ov::Output<ov::Node> src1;
     if (context.get_input_size() == 2) {
-        src0 = context.get_input(0);
-        src1 = context.get_input(1);
+        // Inputs may be VIEW slices of a combined gate_up tensor (MoE experts):
+        // resolve them so each half has its real sliced shape, not the base tensor.
+        src0 = process_view_input_new(context, 0);
+        src1 = process_view_input_new(context, 1);
     } else {
         // GGML splits along ne[0] (OV last axis) using floor division: nc = ne[0] / 2.
         // Both halves are nc elements; if the dimension is odd, the last element is dropped.
         // Use Slice instead of Split to handle odd dimensions correctly.
-        auto combined = context.get_input(0);
+        // Resolve a VIEW input (e.g. non-contiguous slice) to its real shape first.
+        auto combined = process_view_input_new(context, 0);
         auto combined_shape = combined.get_partial_shape();
         int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
         int64_t nc = last_dim_val / 2;
 
-        auto axis   = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
-        auto step   = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
         auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
-        auto stop0  = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto stop0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
         auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
-        auto stop1  = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
+        auto stop1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
 
         src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
         src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
@@ -49,6 +55,12 @@ OutputVector translate_glu_swiglu(const NodeContext & context) {
         std::swap(src0, src1);
     }
 
+    return {src0, src1};
+}
+
+OutputVector translate_glu_swiglu(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
     auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(src0);
     auto silu = std::make_shared<ov::op::v1::Multiply>(src0, sigmoid);
     auto res = std::make_shared<ov::op::v1::Multiply>(silu, src1);
@@ -56,6 +68,27 @@ OutputVector translate_glu_swiglu(const NodeContext & context) {
     return rename_outputs_with_suffix({res}, context.get_name());
 }
 
+OutputVector translate_glu_swiglu_oai(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
+    const int32_t * params = context.get_output_op_params();
+    const float alpha = reinterpret_cast<const float *>(params)[2];
+    const float limit = reinterpret_cast<const float *>(params)[3];
+
+    auto gate = std::make_shared<ov::op::v0::Clamp>(src0, -std::numeric_limits<float>::infinity(), limit);
+    auto alpha_const = ov::op::v0::Constant::create(ov::element::f32, {}, {alpha});
+    auto scaled_gate = std::make_shared<ov::op::v1::Multiply>(gate, alpha_const);
+    auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(scaled_gate);
+    auto out_glu = std::make_shared<ov::op::v1::Multiply>(gate, sigmoid);
+
+    auto up = std::make_shared<ov::op::v0::Clamp>(src1, -limit, limit);
+    auto one = ov::op::v0::Constant::create(ov::element::f32, {}, {1.0f});
+    auto up_plus_one = std::make_shared<ov::op::v1::Add>(up, one);
+    auto res = std::make_shared<ov::op::v1::Multiply>(out_glu, up_plus_one);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
 }  // namespace op
 }  // namespace ggml
 }  // namespace frontend

data/ext/sources/ggml/src/ggml-openvino/openvino/op/im2col.cpp

@@ -0,0 +1,120 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml-impl.h"
+
+#include <cstddef>
+#include <memory>
+#include <openvino/core/shape.hpp>
+#include <openvino/core/strides.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/extractimagepatches.hpp>
+#include <openvino/op/pad.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/util/attr_types.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_im2col(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+    const int32_t * params = context.get_output_op_params();
+    int32_t s0 = params[0];
+    int32_t s1 = params[1];
+    int32_t p0 = params[2];
+    int32_t p1 = params[3];
+    int32_t d0 = params[4];
+    int32_t d1 = params[5];
+    bool is_2D = params[6] == 1;
+    ov::Output<Node> res;
+
+    ov::Output<Node> image = context.get_input(1);
+    const ov::Shape kernel_shape = context.get_input(0).get_shape();
+
+    const size_t IC = is_2D ? kernel_shape[1] : kernel_shape[2];
+    const size_t KH = is_2D ? kernel_shape[2] : 1;
+    const size_t KW = kernel_shape[3];
+
+    int32_t stride_w = s0;
+    int32_t stride_h = is_2D ? s1 : 1;
+    int32_t pad_w = p0;
+    int32_t pad_h = is_2D ? p1 : 0;
+    int32_t dil_w = d0;
+    int32_t dil_h = is_2D ? d1 : 1;
+
+    if (!is_2D) {
+        // GGML input shape: [IW, IC, N, 1]
+        // OpenVINO input shape: [1, N, IC, IW]
+        // Reshape image to: [N, IC, 1, IW]
+        const ov::Shape image_shape = image.get_shape();
+        const size_t N = image_shape[1];
+        const size_t IW = image_shape[3];
+        auto image_reshape_shape = ov::op::v0::Constant::create(
+            ov::element::i64, ov::Shape{4},
+            std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(IC), 1, static_cast<int64_t>(IW)});
+        image = std::make_shared<ov::op::v1::Reshape>(image, image_reshape_shape, false);
+    }
+
+    const ov::Shape patch_sizes = {KH, KW};
+    const ov::Strides strides = {static_cast<size_t>(stride_h), static_cast<size_t>(stride_w)};
+    const ov::Shape rates = {static_cast<size_t>(dil_h), static_cast<size_t>(dil_w)};
+
+    auto pads_begin =
+        ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 0, pad_h, pad_w});
+    auto pads_end =
+        ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 0, pad_h, pad_w});
+
+    auto pad = std::make_shared<ov::op::v1::Pad>(image, pads_begin, pads_end, ov::op::PadMode::CONSTANT);
+    auto patches =
+        std::make_shared<ov::op::v3::ExtractImagePatches>(pad, patch_sizes, strides, rates, ov::op::PadType::VALID);
+
+    // [N, KH*KW*IC, OH, OW] → [N, OH, OW, KH*KW*IC]
+    auto perm1 = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 2, 3, 1});
+    auto t1 = std::make_shared<ov::op::v1::Transpose>(patches, perm1);
+
+    // [N, OH, OW, KH*KW*IC] → [N, OH, OW, KH*KW, IC]
+    const ov::Shape out_shape = t1->get_output_shape(0);
+    const size_t N = out_shape[0];
+    const size_t OH = out_shape[1];
+    const size_t OW = out_shape[2];
+    auto reshape1_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{5},
+        std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(KH * KW), static_cast<int64_t>(IC)});
+    auto r1 = std::make_shared<ov::op::v1::Reshape>(t1, reshape1_shape, false);
+
+    // [N, OH, OW, KH*KW, IC] → [N, OH, OW, IC, KH*KW]
+    auto perm2 = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{5}, std::vector<int64_t>{0, 1, 2, 4, 3});
+    auto t2 = std::make_shared<ov::op::v1::Transpose>(r1, perm2);
+
+    // flatten back to [N, OH, OW, IC*KH*KW]
+    auto r2_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{4},
+        std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(IC * KH * KW)});
+    res = std::make_shared<ov::op::v1::Reshape>(t2, r2_shape, false);
+
+    if (!is_2D) {
+        // [N, 1, OW, IC * KW] -> [1, N, OW, IC * KW]
+        auto final_reshape_shape = ov::op::v0::Constant::create(
+            ov::element::i64, ov::Shape{4},
+            std::vector<int64_t>{1, static_cast<int64_t>(N), static_cast<int64_t>(OW), static_cast<int64_t>(IC * KW)});
+        res = std::make_shared<ov::op::v1::Reshape>(res, final_reshape_shape, false);
+    }
+
+    auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

data/ext/sources/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp

@@ -0,0 +1,44 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/maximum.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reduce_sum.hpp>
+#include <openvino/op/sqrt.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_l2_norm(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input_node = process_view_input_new(context, 0);
+
+    auto squared = std::make_shared<ov::op::v1::Multiply>(input_node, input_node);
+
+    auto sum_squared = std::make_shared<ov::op::v1::ReduceSum>(
+        squared, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}), true);
+
+    auto l2_norm = std::make_shared<ov::op::v0::Sqrt>(sum_squared);
+
+    float eps;
+    memcpy(&eps, context.get_output_op_params(), sizeof(float));
+
+    auto eps_const = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{1}, {eps});
+    auto clamped_norm = std::make_shared<ov::op::v1::Maximum>(l2_norm, eps_const);
+
+    auto res = std::make_shared<ov::op::v1::Divide>(input_node, clamped_norm);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

data/ext/sources/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp

@@ -0,0 +1,226 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <memory>
+#include <openvino/op/bitwise_and.hpp>
+#include <openvino/op/bitwise_right_shift.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+namespace {
+
+std::shared_ptr<ov::op::v0::Constant> const_i64(const std::vector<int64_t> & values) {
+    return ov::op::v0::Constant::create(ov::element::i64, ov::Shape{values.size()}, values);
+}
+
+ov::Output<ov::Node> slice_axis(const ov::Output<ov::Node> & input, int64_t axis, int64_t begin, int64_t end) {
+    return std::make_shared<ov::op::v8::Slice>(input, const_i64({begin}), const_i64({end}), const_i64({1}),
+                                              const_i64({axis}));
+}
+
+ov::Output<ov::Node> translate_mul_mat_id_mxfp4_packed(const NodeContext & context,
+                                                       ov::Output<ov::Node> expert_weights,
+                                                       ov::Output<ov::Node> activations,
+                                                       ov::Output<ov::Node> ids) {
+    auto packed_shape = expert_weights.get_partial_shape().to_shape();
+    FRONT_END_OP_CONVERSION_CHECK(packed_shape.size() == 5 && packed_shape[4] == 17,
+                                  "Expected packed MXFP4 expert weights with shape [1, n_expert, m, k_blocks, 17]");
+
+    const int64_t n_expert = static_cast<int64_t>(packed_shape[1]);
+    const int64_t rows = static_cast<int64_t>(packed_shape[2]);
+    const int64_t k_blocks = static_cast<int64_t>(packed_shape[3]);
+    const int64_t qk = 32;
+    const int64_t cols = k_blocks * qk;
+
+    auto packed_shape_4d = const_i64({n_expert, rows, k_blocks, 17});
+    expert_weights = std::make_shared<ov::op::v1::Reshape>(expert_weights, packed_shape_4d, false);
+
+    auto activations_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto activations_shape_3d = get_dimensions(activations_shape_4d, {1, 2, 3});
+    auto ids_shape_2d = get_dimensions(ids_shape_4d, {2, 3});
+
+    activations = std::make_shared<ov::op::v1::Reshape>(activations, activations_shape_3d, false);
+    ids = std::make_shared<ov::op::v1::Reshape>(ids, ids_shape_2d, false);
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+
+    static const std::vector<float> f4e2m1_lut = {0.0f,  0.5f,  1.0f,  1.5f,  2.0f,  3.0f,  4.0f,  6.0f,
+                                                  -0.0f, -0.5f, -1.0f, -1.5f, -2.0f, -3.0f, -4.0f, -6.0f};
+    std::vector<float> e8m0_lut(256);
+    for (size_t i = 0; i < e8m0_lut.size(); ++i) {
+        uint32_t bits = static_cast<uint32_t>(i) << 23;
+        memcpy(&e8m0_lut[i], &bits, sizeof(float));
+    }
+    e8m0_lut[0] = std::numeric_limits<float>::min() / 2.0f;
+    e8m0_lut[255] = std::numeric_limits<float>::quiet_NaN();
+
+    auto f4_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{f4e2m1_lut.size()}, f4e2m1_lut);
+    auto scale_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{e8m0_lut.size()}, e8m0_lut);
+
+    auto selected_packed_weights = std::make_shared<ov::op::v8::Gather>(expert_weights, ids, gather_axis);
+    auto scale_byte = slice_axis(selected_packed_weights, 4, 0, 1);
+    auto qs = slice_axis(selected_packed_weights, 4, 1, 17);
+    auto low = std::make_shared<ov::op::v13::BitwiseAnd>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {0x0F}), ov::op::AutoBroadcastType::NUMPY);
+    auto high_shift = std::make_shared<ov::op::v15::BitwiseRightShift>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {4}), ov::op::AutoBroadcastType::NUMPY);
+    auto nibbles = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{low, high_shift}, 4);
+    auto nibble_indices = std::make_shared<ov::op::v0::Convert>(nibbles, ov::element::i32);
+    auto weights_f32 = std::make_shared<ov::op::v8::Gather>(f4_lut, nibble_indices, gather_axis);
+
+    auto scale_indices = std::make_shared<ov::op::v0::Convert>(scale_byte, ov::element::i32);
+    auto scales_f32 = std::make_shared<ov::op::v8::Gather>(scale_lut, scale_indices, gather_axis);
+    ov::Output<ov::Node> selected_weights = std::make_shared<ov::op::v1::Multiply>(weights_f32, scales_f32,
+                                                                                  ov::op::AutoBroadcastType::NUMPY);
+
+    auto ids_shape = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto selected_weights_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{get_dimensions(ids_shape, {0, 1}), const_i64({rows, cols})}, 0);
+    selected_weights = std::make_shared<ov::op::v1::Reshape>(selected_weights, selected_weights_target_dims, false);
+
+    auto activations_shape = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    ov::Output<ov::Node> acts_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{
+            get_dimensions(activations_shape, {0}),
+            get_dimensions(ids_shape, {1}),
+            get_dimensions(activations_shape, {2}),
+        },
+        0);
+    ov::Output<ov::Node> acts_broadcasted =
+        std::make_shared<ov::op::v3::Broadcast>(activations, acts_target_dims, ov::op::BroadcastType::BIDIRECTIONAL);
+
+    auto activations_expanded = std::make_shared<ov::op::v0::Unsqueeze>(acts_broadcasted, const_i64({2}));
+    ov::Output<ov::Node> result =
+        std::make_shared<ov::op::v0::MatMul>(activations_expanded, selected_weights, false, true);
+
+    auto batch_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto row_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {rows});
+    auto result_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{batch_dim, get_dimensions(ids_shape, {0, 1}), row_dim}, 0);
+    result = std::make_shared<ov::op::v1::Reshape>(result, result_target_dims, false);
+
+    const auto output_type = context.get_output_type();
+    if (result.get_element_type() != output_type) {
+        result = std::make_shared<ov::op::v0::Convert>(result, output_type);
+    }
+    return result;
+}
+
+}  // namespace
+
+OutputVector translate_mul_mat_id(const NodeContext & context) {
+    num_inputs_check(context, 3, 3);
+
+    auto expert_weights = process_view_input_new(context, 0);
+    auto activations = process_view_input_new(context, 1);
+    auto ids = process_view_input_new(context, 2);
+
+    if (expert_weights.get_element_type() == ov::element::u8 && expert_weights.get_partial_shape().rank().is_static() &&
+        expert_weights.get_partial_shape().rank().get_length() == 5) {
+        return rename_outputs_with_suffix({translate_mul_mat_id_mxfp4_packed(context, expert_weights, activations, ids)},
+                                          context.get_name());
+    }
+
+    // OpenVINO sees GGML tensors in reversed dimension order:
+    //   weights: [1, n_expert, m, k]
+    //   activations: [1, n_tokens, n_used_or_1, k]
+    //   ids: [1, 1, n_tokens, n_used]
+    // Rebuild the logical ranks explicitly from the 4D inputs instead of relying
+    // on fixed squeeze axes: real graphs can arrive through VIEW/RESHAPE chains
+    // where singleton axes are still represented differently at this point.
+    auto expert_weights_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(expert_weights, ov::element::i64);
+    auto activations_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+
+    auto expert_weights_shape_3d = get_dimensions(expert_weights_shape_4d, {1, 2, 3});
+    auto activations_shape_3d = get_dimensions(activations_shape_4d, {1, 2, 3});
+    auto ids_shape_2d = get_dimensions(ids_shape_4d, {2, 3});
+
+    expert_weights = std::make_shared<ov::op::v1::Reshape>(expert_weights, expert_weights_shape_3d, false);
+    activations = std::make_shared<ov::op::v1::Reshape>(activations, activations_shape_3d, false);
+    ids = std::make_shared<ov::op::v1::Reshape>(ids, ids_shape_2d, false);
+
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+    ov::Output<ov::Node> selected_weights = std::make_shared<ov::op::v8::Gather>(expert_weights, ids, gather_axis);
+
+    const auto output_type = context.get_output_type();
+    if (selected_weights.get_element_type() != ov::element::f32) {
+        selected_weights = std::make_shared<ov::op::v0::Convert>(selected_weights, ov::element::f32);
+    }
+    if (activations.get_element_type() != ov::element::f32) {
+        activations = std::make_shared<ov::op::v0::Convert>(activations, ov::element::f32);
+    }
+
+    auto activations_shape = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    auto ids_shape = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    ov::Output<ov::Node> acts_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{
+            get_dimensions(activations_shape, {0}),
+            get_dimensions(ids_shape, {1}),
+            get_dimensions(activations_shape, {2}),
+        },
+        0);
+    ov::Output<ov::Node> acts_broadcasted =
+        std::make_shared<ov::op::v3::Broadcast>(activations, acts_target_dims, ov::op::BroadcastType::BIDIRECTIONAL);
+
+    auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
+    auto activations_expanded = std::make_shared<ov::op::v0::Unsqueeze>(acts_broadcasted, unsqueeze_axes);
+
+    auto batch_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto output_shape = context.get_output_shape();
+    FRONT_END_OP_CONVERSION_CHECK(output_shape.rank().is_static() && output_shape.rank().get_length() == 4,
+                                  "Unexpected MUL_MAT_ID output rank");
+    FRONT_END_OP_CONVERSION_CHECK(output_shape[3].is_static(), "Expected static row dimension for MUL_MAT_ID output");
+    const auto row_dim_value = output_shape[3].get_length();
+    auto row_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {row_dim_value});
+
+    ov::Output<ov::Node> result =
+        std::make_shared<ov::op::v0::MatMul>(activations_expanded, selected_weights, false, true);
+
+    auto result_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{
+            batch_dim,
+            get_dimensions(ids_shape, {0, 1}),
+            row_dim,
+        },
+        0);
+    result = std::make_shared<ov::op::v1::Reshape>(result, result_target_dims, false);
+
+    if (result.get_element_type() != output_type) {
+        result = std::make_shared<ov::op::v0::Convert>(result, output_type);
+    }
+
+    return rename_outputs_with_suffix({result}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov