whispercpp 1.3.6 → 1.3.8

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

@@ -0,0 +1,146 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/sigmoid.hpp>
+#include <openvino/op/tile.hpp>
+#include <openvino/op/util/precision_sensitive_attribute.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+namespace {
+
+bool is_silu_div_pattern(const ov::Output<ov::Node> & numerator,
+                         const ov::Output<ov::Node> & denominator,
+                         const NodeContext & context) {
+    if (context.get_input_size() != 2) {
+        return false;
+    }
+
+    const auto * unary_op = reinterpret_cast<const ggml_unary_op *>(context.get_input_op_params(0));
+    if (unary_op == nullptr || *unary_op != GGML_UNARY_OP_SILU) {
+        return false;
+    }
+
+    auto mul = std::dynamic_pointer_cast<ov::op::v1::Multiply>(numerator.get_node_shared_ptr());
+    if (!mul) {
+        return false;
+    }
+
+    const auto denom_node = denominator.get_node_shared_ptr();
+    const auto mul_input_0 = mul->input_value(0).get_node_shared_ptr();
+    const auto mul_input_1 = mul->input_value(1).get_node_shared_ptr();
+
+    auto sigmoid = std::dynamic_pointer_cast<ov::op::v0::Sigmoid>(mul_input_1);
+    if (mul_input_0 == denom_node && sigmoid && sigmoid->input_value(0).get_node_shared_ptr() == denom_node) {
+        return true;
+    }
+
+    sigmoid = std::dynamic_pointer_cast<ov::op::v0::Sigmoid>(mul_input_0);
+    return mul_input_1 == denom_node && sigmoid && sigmoid->input_value(0).get_node_shared_ptr() == denom_node;
+}
+
+ov::Output<ov::Node> repeat_input_to_match(const NodeContext & context,
+                                           const ov::Output<ov::Node> & input,
+                                           const ov::Output<ov::Node> & target,
+                                           size_t input_index) {
+    const auto input_shape = context.get_input_shape(input_index);
+    const auto target_shape = context.get_input_shape(0);
+
+    if (input_shape == target_shape) {
+        return input;
+    }
+
+    if (input_shape.rank().is_static() && target_shape.rank().is_static()) {
+        const auto rank = static_cast<size_t>(input_shape.rank().get_length());
+        std::vector<int64_t> repeats(rank, 1);
+        bool needs_repeat = false;
+
+        for (size_t axis = 0; axis < rank; ++axis) {
+            FRONT_END_OP_CONVERSION_CHECK(input_shape[axis].is_static() && target_shape[axis].is_static(),
+                                          "DIV repeat requires static dimensions on both inputs");
+
+            const int64_t input_dim = input_shape[axis].get_length();
+            const int64_t target_dim = target_shape[axis].get_length();
+
+            FRONT_END_OP_CONVERSION_CHECK(input_dim > 0 && target_dim > 0 && target_dim % input_dim == 0,
+                                          "DIV input shape ", input_shape, " cannot repeat to match ", target_shape);
+
+            repeats[axis] = target_dim / input_dim;
+            needs_repeat = needs_repeat || repeats[axis] != 1;
+        }
+
+        if (!needs_repeat) {
+            return input;
+        }
+
+        auto repeats_node = ov::op::v0::Constant::create(ov::element::i64, {repeats.size()}, repeats);
+        return std::make_shared<ov::op::v0::Tile>(input, repeats_node);
+    }
+
+    auto input_shape_node = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
+    auto target_shape_node = std::make_shared<ov::op::v3::ShapeOf>(target, ov::element::i64);
+    auto repeats_node = std::make_shared<ov::op::v1::Divide>(target_shape_node, input_shape_node);
+    return std::make_shared<ov::op::v0::Tile>(input, repeats_node);
+}
+
+}  // namespace
+
+OutputVector translate_div(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    auto input_0 = process_view_input_new(context, 0);
+    auto input_1 = process_view_input_new(context, 1);
+
+    if (is_silu_div_pattern(input_0, input_1, context)) {
+        ov::Output<ov::Node> res = std::make_shared<ov::op::v0::Sigmoid>(input_1);
+        if (res.get_element_type() != context.get_output_type()) {
+            res = std::make_shared<ov::op::v0::Convert>(res, context.get_output_type());
+        }
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    input_1 = repeat_input_to_match(context, input_1, input_0, 1);
+
+    const auto output_type = context.get_output_type();
+    const bool use_f32_compute = input_0.get_element_type() != ov::element::f32 ||
+                                 input_1.get_element_type() != ov::element::f32 || output_type != ov::element::f32;
+
+    if (use_f32_compute) {
+        input_0 = std::make_shared<ov::op::v0::Convert>(input_0, ov::element::f32);
+        input_1 = std::make_shared<ov::op::v0::Convert>(input_1, ov::element::f32);
+    }
+
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Divide>(input_0, input_1);
+    if (use_f32_compute) {
+        // Keep the reciprocal/divide path in FP32. Without this hint, the GPU
+        // plugin can still compress the subgraph back to FP16 and overflow on
+        // small shexp gate values (e.g. silu(x) / x in qwen2moe).
+        ov::mark_as_precision_sensitive(res.get_node_shared_ptr()->input(0));
+        ov::mark_as_precision_sensitive(res.get_node_shared_ptr()->input(1));
+    }
+    if (res.get_element_type() != output_type) {
+        auto output_convert = std::make_shared<ov::op::v0::Convert>(res, output_type);
+        if (use_f32_compute) {
+            ov::mark_as_precision_sensitive(output_convert->input(0));
+        }
+        res = output_convert;
+    }
+    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/flash_attn_ext.cpp

@@ -1,15 +1,21 @@
 #include "../node_context.h"
 #include "../op_table.h"
 #include "../utils.h"
+#include "ggml-openvino/ggml-openvino-extra.h"
 
 #include <cstdint>
+#include <cstdlib>
 #include <memory>
+#include <openvino/op/add.hpp>
 #include <openvino/op/broadcast.hpp>
 #include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/scaled_dot_product_attention.hpp>
+#include <openvino/op/softmax.hpp>
 #include <openvino/op/transpose.hpp>
 #include <openvino/op/unsqueeze.hpp>
 #include <string>
@@ -34,36 +40,115 @@ OutputVector translate_flash_attn_ext(const NodeContext & context) {
     auto q = std::make_shared<ov::op::v0::Convert>(q_f32, ov::element::f16);
     auto scale_node = std::make_shared<ov::op::v0::Constant>(ov::element::f16, ov::Shape{}, std::vector<float>{scale});
 
-    ov::Output<ov::Node> mask_sliced, res;
+    ov::Output<ov::Node> res;
+
+    // For stateful
     std::string mask_name = "KQ_mask_sliced";
     if (context.get_input_names()[3].find("swa") != std::string::npos) {
         mask_name = "KQ_mask_swa_sliced";
     }
     if (context.has_input(mask_name)) {
-        mask_sliced = context.get_input(mask_name);
-    } else {
-        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
-        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
-        auto two = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
-        auto token_len = get_dimensions(q, {2});
-        mask_sliced = std::make_shared<ov::op::v8::Slice>(mask, zero, token_len, one, two);
+        mask = context.get_input(mask_name);
     }
 
-    if (mask_sliced.get_element_type() != ov::element::f16) {
-        mask_sliced = std::make_shared<ov::op::v0::Convert>(mask_sliced, ov::element::f16);
+    if (mask.get_element_type() != ov::element::f16) {
+        mask = std::make_shared<ov::op::v0::Convert>(mask, ov::element::f16);
     }
 
-    auto tile_kv = [&](int64_t num_heads, int64_t num_heads_kv, int64_t head_size, ov::Output<Node> kv) {
-        int64_t factor = num_heads / num_heads_kv;
-        if (factor > 1 && num_heads_kv > 1) {
+    //auto tile_kv = [&](int64_t num_heads, int64_t num_heads_kv, int64_t head_size, ov::Output<Node> kv) {
+    //    int64_t factor = num_heads / num_heads_kv;
+    //    if (factor > 1 && num_heads_kv > 1) {
+    auto q_shape = context.get_input_shape(0).to_shape();
+    auto k_shape = context.get_input_shape(1).to_shape();
+    const int64_t num_heads = q_shape[1];
+    const int64_t num_heads_kv = k_shape[1];
+    const int64_t head_size = q_shape[3];
+    const int64_t factor = num_heads / num_heads_kv;
+
+    // Manual GQA attention: enabled by default on GPU in stateless mode.
+    // Set GGML_OPENVINO_MANUAL_GQA_ATTN to a positive value (e.g. 1) to force-enable,
+    // or to 0 to force-disable. Unset falls back to the device-based default.
+    static const bool manual_gqa_enabled = []() {
+        const char * env = ggml_openvino_getenv_str("GGML_OPENVINO_MANUAL_GQA_ATTN");
+        if (env != nullptr) {
+            return ggml_openvino_getenv_int("GGML_OPENVINO_MANUAL_GQA_ATTN") > 0;
+        }
+        const char * dev = ggml_openvino_getenv_str("GGML_OPENVINO_DEVICE");
+        return dev != nullptr && std::string(dev) == "GPU";
+    }();
+    const bool use_manual_gqa_attention =
+        manual_gqa_enabled && factor > 1 && num_heads_kv > 1 && !context.is_stateful();
+
+    if (use_manual_gqa_attention) {
+        // Q, K, V arrive as [B, n_heads(_kv), S, head_size], where B is the active
+        // batch (n_seq_active) and may be > 1 (llama-perplexity, llama-server -np > 1)
+        // or dynamic. Reshape to
+        //   K_r: [B, num_heads_kv, 1, S, head_size]
+        //   Q_r: [B, num_heads_kv, factor, S_q, head_size]
+        // and let MatMul broadcast across the factor dim without materialising
+        // an expanded K/V. The leading 0 + special_zero=true copies B at runtime,
+        // so this is correct for B == 1, B > 1, and dynamic B alike. Only the head
+        // dims and head_size are baked in as literals; the sequence dim stays -1.
+        auto k_5d_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                       std::vector<int64_t>{0, num_heads_kv, 1, -1, head_size});
+        auto v_5d_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                       std::vector<int64_t>{0, num_heads_kv, 1, -1, head_size});
+        auto q_5d_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                       std::vector<int64_t>{0, num_heads_kv, factor, -1, head_size});
+
+        auto k_r = std::make_shared<ov::op::v1::Reshape>(k, k_5d_shape, true);
+        auto v_r = std::make_shared<ov::op::v1::Reshape>(v, v_5d_shape, true);
+        auto q_r = std::make_shared<ov::op::v1::Reshape>(q, q_5d_shape, true);
+
+        // QK^T → [B, num_heads_kv, factor, S_q, S_k]
+        auto qk = std::make_shared<ov::op::v0::MatMul>(q_r, k_r, /*tA=*/false, /*tB=*/true);
+        auto qk_scaled = std::make_shared<ov::op::v1::Multiply>(qk, scale_node);
+
+        // Mask arrives as [B, 1, S_q, S_k]. Unsqueeze a factor axis at position 2 to
+        // get [B, 1, 1, S_q, S_k], which NUMPY-broadcasts cleanly against the
+        // [B, num_heads_kv, factor, S_q, S_k] scores: B==B, then 1→num_heads_kv and
+        // 1→factor on the head dims.
+        auto mask_unsq1 =
+            std::make_shared<ov::op::v0::Unsqueeze>(mask, ov::op::v0::Constant::create(ov::element::i64, {1}, {2}));
+        // mask_unsq1: [B, 1, 1, S_q, S_k] (rank 5)
+        ov::Output<ov::Node> qk_masked = std::make_shared<ov::op::v1::Add>(qk_scaled, mask_unsq1);
+
+        auto softmax = std::make_shared<ov::op::v8::Softmax>(qk_masked, /*axis=*/-1);
+
+        // softmax @ V → [B, num_heads_kv, factor, S_q, head_size]
+        auto attn = std::make_shared<ov::op::v0::MatMul>(softmax, v_r);
+
+        // Reshape back to [B, num_heads, S_q, head_size] (combine num_heads_kv * factor).
+        // Leading 0 + special_zero=true copies B at runtime.
+        auto out_4d_shape =
+            ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, num_heads, -1, head_size});
+        auto out_4d = std::make_shared<ov::op::v1::Reshape>(attn, out_4d_shape, true);
+
+        // The standard SDPA path's downstream is Transpose(0,2,1,3) → Convert(f32).
+        // Replicate it here so callers see the same output layout/dtype.
+        res = std::make_shared<ov::op::v1::Transpose>(
+            out_4d, ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3}));
+        res = std::make_shared<ov::op::v0::Convert>(res, ov::element::f32);
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    // Default path: explicit Broadcast → SDPA. Kept as the fallback because
+    // (a) it goes through the GPU plugin's micro-SDPA fast path (FlashAttention
+    // tiles via DPAS), and (b) the manual path above is still being validated.
+    auto tile_kv = [&](int64_t n_heads, int64_t n_heads_kv, int64_t hs, ov::Output<Node> kv) {
+        int64_t f = n_heads / n_heads_kv;
+        if (f > 1 && n_heads_kv > 1) {
             ov::Output<ov::Node> kv_broadcast_shape, kv_unsqueezed, new_kv_shape;
             auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, Shape{}, {2});
             kv_unsqueezed = std::make_shared<ov::op::v0::Unsqueeze>(kv, unsqueeze_axes);
 
-            kv_broadcast_shape = ov::op::v0::Constant::create(
-                ov::element::i64, {5}, {(int64_t) 1, (int64_t) 1, factor, (int64_t) 1, (int64_t) 1});
+            kv_broadcast_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                              {(int64_t) 1, (int64_t) 1, f, (int64_t) 1, (int64_t) 1});
             new_kv_shape =
-                ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 0, num_heads, (int64_t) -1, head_size});
+                ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 0, n_heads, (int64_t) -1, hs});
+            //    ov::element::i64, {5}, {(int64_t) 1, (int64_t) 1, factor, (int64_t) 1, (int64_t) 1});
+            //new_kv_shape =
+            //    ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 0, num_heads, (int64_t) -1, head_size});
 
             kv = std::make_shared<ov::op::v3::Broadcast>(kv_unsqueezed, kv_broadcast_shape,
                                                          ov::op::BroadcastType::BIDIRECTIONAL);
@@ -72,12 +157,14 @@ OutputVector translate_flash_attn_ext(const NodeContext & context) {
         return kv;
     };
 
-    auto q_shape = context.get_input_shape(0).to_shape();
-    auto k_shape = context.get_input_shape(1).to_shape();
-    k = tile_kv(q_shape[1], k_shape[1], q_shape[3], k);
-    v = tile_kv(q_shape[1], k_shape[1], q_shape[3], v);
+    //auto q_shape = context.get_input_shape(0).to_shape();
+    //auto k_shape = context.get_input_shape(1).to_shape();
+    //k = tile_kv(q_shape[1], k_shape[1], q_shape[3], k);
+    //v = tile_kv(q_shape[1], k_shape[1], q_shape[3], v);
+    k = tile_kv(num_heads, num_heads_kv, head_size, k);
+    v = tile_kv(num_heads, num_heads_kv, head_size, v);
 
-    auto sdpa = std::make_shared<ov::op::v13::ScaledDotProductAttention>(q, k, v, mask_sliced, scale_node, false);
+    auto sdpa = std::make_shared<ov::op::v13::ScaledDotProductAttention>(q, k, v, mask, scale_node, false);
     res = std::make_shared<ov::op::v1::Transpose>(sdpa,
                                                   ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3}));
     res = std::make_shared<ov::op::v0::Convert>(res, ov::element::f32);

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

@@ -0,0 +1,282 @@
+#include "gated_delta_net.hpp"
+
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cmath>
+#include <cstdint>
+#include <memory>
+#include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/exp.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/loop.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+static OutputVector translate_gated_delta_net_ref(const NodeContext & context);
+
+OutputVector translate_gated_delta_net(const NodeContext & context) {
+    // auto v_shape = context.get_input_shape(2).to_shape();  // [B, T, H_v, S_v]
+    // auto q_shape = context.get_input_shape(0).to_shape();  // [B, T, H_k, S_k]
+
+    // // Fused GatedDeltaNet op only supports scalar gate (kda=0).
+    // // Fall back to reference implementation for per-key-dimension gating.
+    // // if (kda) {
+    // //     return translate_gated_delta_net_ref(context);
+    // // }
+
+    // auto q = context.get_input(0);
+    // auto k = context.get_input(1);
+    // auto v = context.get_input(2);
+    // auto g = context.get_input(3);
+    // auto beta = context.get_input(4);
+    // auto state = context.get_input(5);
+
+    // const int64_t B = v_shape[0];
+    // const int64_t T = v_shape[1];
+    // const int64_t H_v = v_shape[2];
+    // const int64_t S_v = v_shape[3];
+    // const int64_t S_k = q_shape[3];
+
+    // // ggml state layout (OV notation): [B, H_v, value_dim, key_dim]
+    // // GatedDeltaNet op expects: [B, H_v, key_dim, value_dim]
+    // auto state_reshape_shape =
+    //     ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, S_v, S_k});
+    // state = std::make_shared<ov::op::v1::Reshape>(state, state_reshape_shape, false);
+    // auto state_perm = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, 1, 3, 2});
+    // state = std::make_shared<ov::op::v1::Transpose>(state, state_perm);
+
+    // g = std::make_shared<ov::op::v0::Squeeze>(g, ov::op::v0::Constant::create(ov::element::i64, {1}, {3}));
+    // beta = std::make_shared<ov::op::v0::Squeeze>(beta, ov::op::v0::Constant::create(ov::element::i64, {1}, {3}));
+
+    // auto gdn = std::make_shared<ov::op::internal::GatedDeltaNet>(q, k, v, state, g, beta);
+
+    // auto attn_4d = gdn->output(0);
+    // auto state_4d = gdn->output(1);  // [B, H_v, key_dim, value_dim]
+    // // Transpose output state back to ggml layout [B, H_v, value_dim, key_dim]
+    // auto state_transposed = std::make_shared<ov::op::v1::Transpose>(state_4d, state_perm);
+    // auto flat_shape_1d = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+    // auto attn = std::make_shared<ov::op::v1::Reshape>(attn_4d, flat_shape_1d, false);
+    // auto new_state = std::make_shared<ov::op::v1::Reshape>(state_transposed, flat_shape_1d, false);
+    // auto packed = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{attn, new_state}, 0);
+    // auto out_shape =
+    //     ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, T * B + S_v * B, S_v * H_v});
+    // auto res = std::make_shared<ov::op::v1::Reshape>(packed, out_shape, false);
+
+    // return rename_outputs_with_suffix({res}, context.get_name());
+
+    // The OV version in CI does not have the GatedDeltaNet op, so use reference implementation for now.
+    return translate_gated_delta_net_ref(context);
+}
+
+static OutputVector translate_gated_delta_net_ref(const NodeContext & context) {
+    num_inputs_check(context, 6, 6);
+
+    // Inputs (OV shapes are reversed from ggml):
+    // ggml: q[S_k, H_k, T, B], k[S_k, H_k, T, B], v[S_v, H_v, T, B]
+    // OV:   q[B, T, H_k, S_k], k[B, T, H_k, S_k], v[B, T, H_v, S_v]
+    // ggml: g[1 or S_v, H_v, T, B], beta[1, H_v, T, B]
+    // OV:   g[B, T, H_v, 1 or S_v], beta[B, T, H_v, 1]
+    // ggml: state[S_v, S_v, H_v, B]
+    // OV:   state[B, H_v, S_v, S_v]
+    auto q = process_view_input_new(context, 0);
+    auto k = process_view_input_new(context, 1);
+    auto v = process_view_input_new(context, 2);
+    auto g = process_view_input_new(context, 3);
+    auto beta = process_view_input_new(context, 4);
+    auto state = process_view_input_new(context, 5);
+
+    auto v_shape = context.get_input_shape(2).to_shape();  // [B, T, H_v, S_v]
+    auto q_shape = context.get_input_shape(0).to_shape();  // [B, T, H_k, S_k]
+    auto g_shape = context.get_input_shape(3).to_shape();  // [B, T, H_v, 1 or S_v]
+
+    const int64_t B = v_shape[0];
+    const int64_t T = v_shape[1];
+    const int64_t H_v = v_shape[2];
+    const int64_t S_v = v_shape[3];
+    const int64_t H_k = q_shape[2];
+    const bool kda = (g_shape[3] == (size_t) S_v);
+
+    const int64_t rq1 = H_v / H_k;  // head repeat factor
+    const float scale = 1.0f / std::sqrt((float) S_v);
+
+    auto axis_1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto axis_2 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
+
+    // Transpose inputs from [B, T, H, S] to [B, H, T, S] for easier per-head processing
+    auto perm_0213 = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, 2, 1, 3});
+    auto q_t = std::make_shared<ov::op::v1::Transpose>(q, perm_0213);        // [B, H_k, T, S_k]
+    auto k_t = std::make_shared<ov::op::v1::Transpose>(k, perm_0213);        // [B, H_k, T, S_k]
+    auto v_t = std::make_shared<ov::op::v1::Transpose>(v, perm_0213);        // [B, H_v, T, S_v]
+    auto g_t = std::make_shared<ov::op::v1::Transpose>(g, perm_0213);        // [B, H_v, T, 1 or S_v]
+    auto beta_t = std::make_shared<ov::op::v1::Transpose>(beta, perm_0213);  // [B, H_v, T, 1]
+
+    // Broadcast Q, K heads to match V heads if GQA is used (H_v > H_k)
+    ov::Output<ov::Node> q_bh = q_t;
+    ov::Output<ov::Node> k_bh = k_t;
+    if (rq1 > 1) {
+        auto q_unsq = std::make_shared<ov::op::v0::Unsqueeze>(q_t, axis_2);  // [B, H_k, 1, T, S]
+        auto k_unsq = std::make_shared<ov::op::v0::Unsqueeze>(k_t, axis_2);  // [B, H_k, 1, T, S]
+
+        auto bcast_shape = ov::op::v0::Constant::create(ov::element::i64, {5}, std::vector<int64_t>{1, 1, rq1, 1, 1});
+        auto q_bcast =
+            std::make_shared<ov::op::v3::Broadcast>(q_unsq, bcast_shape, ov::op::BroadcastType::BIDIRECTIONAL);
+        auto k_bcast =
+            std::make_shared<ov::op::v3::Broadcast>(k_unsq, bcast_shape, ov::op::BroadcastType::BIDIRECTIONAL);
+
+        // Transpose [B, H_k, rq1, T, S] -> [B, rq1, H_k, T, S] so that reshape merges
+        // as [rq1, H_k] giving repeat-blocks pattern matching CPU: iq1 = iv1 % H_k
+        auto perm_5d = ov::op::v0::Constant::create(ov::element::i64, {5}, std::vector<int64_t>{0, 2, 1, 3, 4});
+        auto q_transposed = std::make_shared<ov::op::v1::Transpose>(q_bcast, perm_5d);
+        auto k_transposed = std::make_shared<ov::op::v1::Transpose>(k_bcast, perm_5d);
+
+        auto new_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, T, S_v});
+        q_bh = std::make_shared<ov::op::v1::Reshape>(q_transposed, new_shape, false);
+        k_bh = std::make_shared<ov::op::v1::Reshape>(k_transposed, new_shape, false);
+    }
+
+    // Merge batch and head dims: [B*H_v, T, S_v]
+    auto merge_bh = [&](ov::Output<ov::Node> x, int64_t last_dim) {
+        auto shape = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{B * H_v, T, last_dim});
+        return std::make_shared<ov::op::v1::Reshape>(x, shape, false);
+    };
+
+    auto q_m = merge_bh(q_bh, S_v);           // [B*H_v, T, S_v]
+    auto k_m = merge_bh(k_bh, S_v);           // [B*H_v, T, S_v]
+    auto v_m = merge_bh(v_t, S_v);            // [B*H_v, T, S_v]
+    auto g_m = merge_bh(g_t, kda ? S_v : 1);  // [B*H_v, T, 1 or S_v]
+    auto beta_m = merge_bh(beta_t, 1);        // [B*H_v, T, 1]
+
+    // State: [B, H_v, S_v, S_v] -> [B*H_v, S_v, S_v]
+    auto state_shape = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{B * H_v, S_v, S_v});
+    auto state_m = std::make_shared<ov::op::v1::Reshape>(state, state_shape, false);
+
+    auto scale_const = ov::op::v0::Constant::create(ov::element::f32, {}, std::vector<float>{scale});
+
+    // --- Build Loop body ---
+    // Body parameters (no iteration counter needed, use -1 in special ports)
+    auto body_state = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_q = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_k = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_v = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_g = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_beta = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_iter = std::make_shared<ov::op::v0::Parameter>(ov::element::i64, ov::Shape{1});
+
+    // Condition output (always true - we rely on trip_count for termination)
+    auto body_cond_out = ov::op::v0::Constant::create(ov::element::boolean, ov::Shape{1}, std::vector<bool>{true});
+
+    // Gather current token from invariant inputs using iteration counter
+    auto q_t_cur = std::make_shared<ov::op::v8::Gather>(body_q, body_iter, axis_1);     // [B*H_v, 1, S_v]
+    auto k_t_cur = std::make_shared<ov::op::v8::Gather>(body_k, body_iter, axis_1);     // [B*H_v, 1, S_v]
+    auto v_t_cur = std::make_shared<ov::op::v8::Gather>(body_v, body_iter, axis_1);     // [B*H_v, 1, S_v]
+    auto g_t_cur = std::make_shared<ov::op::v8::Gather>(body_g, body_iter, axis_1);     // [B*H_v, 1, 1 or S_v]
+    auto b_t_cur = std::make_shared<ov::op::v8::Gather>(body_beta, body_iter, axis_1);  // [B*H_v, 1, 1]
+
+    // Squeeze token dim
+    auto q_cur = std::make_shared<ov::op::v0::Squeeze>(q_t_cur, axis_1);  // [B*H_v, S_v]
+    auto k_cur = std::make_shared<ov::op::v0::Squeeze>(k_t_cur, axis_1);  // [B*H_v, S_v]
+    auto v_cur = std::make_shared<ov::op::v0::Squeeze>(v_t_cur, axis_1);  // [B*H_v, S_v]
+    auto g_cur = std::make_shared<ov::op::v0::Squeeze>(g_t_cur, axis_1);  // [B*H_v, 1 or S_v]
+    auto b_cur = std::make_shared<ov::op::v0::Squeeze>(b_t_cur, axis_1);  // [B*H_v, 1]
+
+    // Step 1: Apply decay gate to state
+    auto exp_g = std::make_shared<ov::op::v0::Exp>(g_cur);                                // [B*H_v, 1 or S_v]
+    auto exp_g_unsq = std::make_shared<ov::op::v0::Unsqueeze>(exp_g, axis_1);             // [B*H_v, 1, 1 or S_v]
+    auto state_decayed = std::make_shared<ov::op::v1::Multiply>(body_state, exp_g_unsq);  // [B*H_v, S_v, S_v]
+
+    // Step 2: delta = (v - S @ k) * beta
+    auto k_col = std::make_shared<ov::op::v0::Unsqueeze>(k_cur, axis_2);                 // [B*H_v, S_v, 1]
+    auto sk = std::make_shared<ov::op::v0::MatMul>(state_decayed, k_col, false, false);  // [B*H_v, S_v, 1]
+    auto sk_sq = std::make_shared<ov::op::v0::Squeeze>(sk, axis_2);                      // [B*H_v, S_v]
+    auto v_minus_sk = std::make_shared<ov::op::v1::Subtract>(v_cur, sk_sq);              // [B*H_v, S_v]
+    auto delta = std::make_shared<ov::op::v1::Multiply>(v_minus_sk, b_cur);              // [B*H_v, S_v]
+
+    // Step 3: state += outer(delta, k)
+    auto delta_col = std::make_shared<ov::op::v0::Unsqueeze>(delta, axis_2);                 // [B*H_v, S_v, 1]
+    auto k_row = std::make_shared<ov::op::v0::Unsqueeze>(k_cur, axis_1);                     // [B*H_v, 1, S_v]
+    auto outer_prod = std::make_shared<ov::op::v0::MatMul>(delta_col, k_row, false, false);  // [B*H_v, S_v, S_v]
+    auto state_updated = std::make_shared<ov::op::v1::Add>(state_decayed, outer_prod);       // [B*H_v, S_v, S_v]
+
+    // Step 4: attn_out = S @ q * scale
+    auto q_col = std::make_shared<ov::op::v0::Unsqueeze>(q_cur, axis_2);                 // [B*H_v, S_v, 1]
+    auto sq = std::make_shared<ov::op::v0::MatMul>(state_updated, q_col, false, false);  // [B*H_v, S_v, 1]
+    auto sq_squeezed = std::make_shared<ov::op::v0::Squeeze>(sq, axis_2);                // [B*H_v, S_v]
+    auto attn_out = std::make_shared<ov::op::v1::Multiply>(sq_squeezed, scale_const);    // [B*H_v, S_v]
+
+    // Unsqueeze attn_out to [B*H_v, 1, S_v] for scan output concatenation
+    auto attn_out_unsq = std::make_shared<ov::op::v0::Unsqueeze>(attn_out, axis_1);  // [B*H_v, 1, S_v]
+
+    // --- Assemble Loop ---
+    // Body: results = [condition, state_updated, attn_out_unsq]
+    auto body = std::make_shared<ov::Model>(
+        ov::OutputVector{body_cond_out, state_updated, attn_out_unsq},
+        ov::ParameterVector{body_iter, body_state, body_q, body_k, body_v, body_g, body_beta});
+
+    auto trip_count = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, std::vector<int64_t>{T});
+    auto exec_cond = ov::op::v0::Constant::create(ov::element::boolean, ov::Shape{1}, std::vector<bool>{true});
+
+    auto loop = std::make_shared<ov::op::v5::Loop>(trip_count, exec_cond);
+    loop->set_function(body);
+    loop->set_special_body_ports(ov::op::v5::Loop::SpecialBodyPorts{0, 0});
+
+    // Carried state: feeds back from body output 1 to body_state param
+    loop->set_merged_input(body_state, state_m, state_updated);
+    // Invariant inputs: passed through unchanged each iteration
+    loop->set_invariant_input(body_q, q_m);
+    loop->set_invariant_input(body_k, k_m);
+    loop->set_invariant_input(body_v, v_m);
+    loop->set_invariant_input(body_g, g_m);
+    loop->set_invariant_input(body_beta, beta_m);
+
+    // Loop outputs:
+    // 1) Final state (last iteration value of state_updated)
+    auto final_state_out = loop->get_iter_value(state_updated, -1);  // [B*H_v, S_v, S_v]
+    // 2) Concatenated attention outputs across all iterations along axis 1
+    auto attn_concat_out = loop->get_concatenated_slices(attn_out_unsq, 0, 1, 1, -1, 1);  // [B*H_v, T, S_v]
+
+    // --- Pack outputs to match ggml layout ---
+    // ggml output ne = {S_v*H, T*B + S_v*B, 1, 1} -> OV [1, 1, T*B+S_v*B, S_v*H_v]
+    // attn: [B, T, H_v, S_v] row-major, state: [B, H_v, S_v, S_v] row-major
+
+    // attn: [B*H_v, T, S_v] -> [B, H_v, T, S_v] -> transpose to [B, T, H_v, S_v] -> flatten
+    auto attn_4d_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, T, S_v});
+    auto attn_4d = std::make_shared<ov::op::v1::Reshape>(attn_concat_out, attn_4d_shape, false);
+    auto attn_perm = std::make_shared<ov::op::v1::Transpose>(attn_4d, perm_0213);  // [B, T, H_v, S_v]
+
+    auto flat_shape_1d = ov::op::v0::Constant::create(ov::element::i64, {1}, std::vector<int64_t>{-1});
+    auto attn_1d = std::make_shared<ov::op::v1::Reshape>(attn_perm, flat_shape_1d, false);
+
+    // state: [B*H_v, S_v, S_v] -> [B, H_v, S_v, S_v] -> flatten
+    auto state_4d_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, S_v, S_v});
+    auto state_4d = std::make_shared<ov::op::v1::Reshape>(final_state_out, state_4d_shape, false);
+    auto state_1d = std::make_shared<ov::op::v1::Reshape>(state_4d, flat_shape_1d, false);
+
+    // Concat [attn | state] and reshape to final output
+    auto packed = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{attn_1d, state_1d}, 0);
+    auto out_shape =
+        ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, T * B + S_v * B, S_v * H_v});
+    auto res = std::make_shared<ov::op::v1::Reshape>(packed, out_shape, false);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov