ravi9 · mostafafaheem · Jul 16, 2026
@@ -366,6 +366,32 @@ int GgmlOvDecoder::compute_op_case(const ggml_tensor * node) const {
                    is_kvcache(node->src[1]->view_src, nullptr)) {
             // s_copy defrag remainder writeback: gathered extra state rows copied back into the cache
             op_case = 3;
+        } else if (node->src[1] != nullptr && node->src[1]->op == GGML_OP_VIEW &&
+                node->src[1]->view_src != nullptr) {
+            // op_case 5: KV write for decoder self-attention (dynamic write offset)
+            // op_case 6: KV write for encoder self-attn or cross-attn (static offset)
+            const ggml_tensor * kv_buf = node->src[1]->view_src;
+            if (kv_buf->ne[1] == 1 && kv_buf->ne[2] == 1 && kv_buf->ne[3] == 1) {
+                op_case = 6;
+                // Forward-scan the graph for a FLASH_ATTN_EXT that reads from
+                // the same buffer. Having a mask (src[3] != nullptr) implies
+                // decoder self-attention and the write offset is dynamic.
+                for (int i = 0; i < m_cgraph->n_nodes; i++) {
+                    const ggml_tensor * n = m_cgraph->nodes[i];
+                    if (n->op != GGML_OP_FLASH_ATTN_EXT) {
+                        continue;
+                    }
+                    // K (src[1]) and V (src[2]) are 3-D views whose view_src is
+                    // the flat KV buffer we are writing to.
+                    if ((n->src[1] != nullptr && n->src[1]->view_src == kv_buf) ||
+                            (n->src[2] != nullptr && n->src[2]->view_src == kv_buf)) {
+                        if (n->src[3] != nullptr) {
+                            op_case = 5;  // decoder self-attention: mask present
+                        }
+                        break;
+                    }
+                }
+            }
         }
         break;
     }
@@ -380,6 +406,16 @@ int GgmlOvDecoder::compute_op_case(const ggml_tensor * node) const {
             op_case = 1;
         }
         break;
+    }
+        case GGML_OP_FLASH_ATTN_EXT: {
+        if (node->src[1] != nullptr && node->src[1]->op == GGML_OP_VIEW &&
+                node->src[1]->view_src != nullptr) {
+            const ggml_tensor * kv_buf = node->src[1]->view_src;
+            if (kv_buf->ne[1] == 1 && kv_buf->ne[2] == 1 && kv_buf->ne[3] == 1) {
+                op_case = (node->src[3] != nullptr) ? 1 : 2;
+            }
+        }
+        break;
     }
     default:
         break;
@@ -412,23 +448,36 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr
 
         switch (node->op) {
         case GGML_OP_FLASH_ATTN_EXT:
-            if (node->src[0] == nullptr || node->src[1] == nullptr || node->src[3] == nullptr) {
+            if (node->src[0] == nullptr || node->src[1] == nullptr) {
                 return -1;
             }
             switch (node->src[1]->op) {
             case GGML_OP_PERMUTE:
-                // case 0: node op is FLASH_ATTN_EXT, src 1 not null & op is PERMUTE & the permuted tensor src is the view of cache k
-                if (node->src[1]->src[0] != nullptr && node->src[1]->src[0]->op == GGML_OP_VIEW) {
+                // case 0: src[1] is PERMUTE of a cache VIEW, mask required
+                if (node->src[3] != nullptr && node->src[1]->src[0] != nullptr &&
+                    node->src[1]->src[0]->op == GGML_OP_VIEW) {
                     return 0;
                 }
                 break;
             case GGML_OP_CPY:
-                // case 1: node op is FLASH_ATTN_EXT, src 1 not null & op is CPY & the copied tensor src is PERMUTE & the permuted tensor src is the view of cache k
-                if (node->src[1]->src[0] != nullptr && node->src[1]->src[0]->op == GGML_OP_PERMUTE &&
-                    node->src[1]->src[0]->src[0] != nullptr && node->src[1]->src[0]->src[0]->op == GGML_OP_VIEW) {
+                // case 1: src[1] is CPY of a PERMUTE(VIEW), mask required
+                if (node->src[3] != nullptr && node->src[1]->src[0] != nullptr &&
+                    node->src[1]->src[0]->op == GGML_OP_PERMUTE &&
+                    node->src[1]->src[0]->src[0] != nullptr &&
+                    node->src[1]->src[0]->src[0]->op == GGML_OP_VIEW) {
                     return 1;
                 }
                 break;
+            case GGML_OP_VIEW:
+                // cases 4/5/6: whisper - K is a direct non-contiguous VIEW_3D of a KV cache
+                if (node->src[1]->view_src != nullptr) {
+                    if (node->src[3] != nullptr) {
+                        return 4; // decoder self-attention
+                    } else {
+                        return 5; // cross-attention or encoder self-attention
+                    };
+                }
+                break;
             default:
                 break;
             }
@@ -481,6 +530,18 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr
                 cache_k_permute = node->src[0]->src[0]->src[0];
                 mask = node->src[1];
                 break;
+            case 4:
+            case 5: {
+                // whisper: K is a direct VIEW_3D of the KV buffer, no PERMUTE node
+                auto * cache_k_view = node->src[1];  // VIEW_3D of kv_self.k or kv_cross.k`
+                compute_params.token_len_per_seq = node->src[0]->ne[1];
+                if (attention_pattern_case == 4) {
+                    compute_params.attention_size = cache_k_view->ne[1];
+                } else {
+                    compute_params.attention_size_static = cache_k_view->ne[1];
+                }
+                continue;
+            }
             default:
                 break;
             }
@@ -634,11 +695,15 @@ ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op,
     } else if (is_kvcache(input, op)) {
         // kvcache
         input_shape = ov::PartialShape{get_shape(input)};
-        if (!m_is_static) {
+        // Whisper.cpp uses a fixed size 1D KV buffer [N, 1, 1, 1] (GGML) or [1, 1, 1, N] (OV).
+        // the token fill level is handled by token_len_per_seq + dynamic mask input.
+        // skip dynamic dim and stateful reshape for this layout.
+        const bool is_flat_kv = (input->ne[1] == 1 && input->ne[2] == 1 && input->ne[3] == 1);
+        if (!m_is_static && !is_flat_kv) {
             // do not fix ctx size to make llama-bench work across test params
             input_shape[2] = -1;
         }
-        if (is_stateful()) {
+        if (is_stateful() && !is_flat_kv) {
             // Convert stateless KV cache layout [1, 1, seq, n_heads_kv * head_size]
             // to stateful layout [1, seq, n_heads_kv, head_size].
             assert(input_shape.size() == 4 && input_shape[0] == 1 && input_shape[1] == 1 &&
@@ -713,6 +778,9 @@ void GgmlOvDecoder::add_extra_inputs() {
     if (m_compute_params.attention_size != -1) {
         create_1d_input("attention_size", m_compute_params.attention_size);
     }
+    if (m_compute_params.attention_size_static != -1) {
+        create_1d_input("attention_size_static", m_compute_params.attention_size_static);
+    }
     if (m_compute_params.attention_size_swa != -1) {
         create_1d_input("attention_size_swa", m_compute_params.attention_size_swa);
     }

@@ -45,6 +45,7 @@ struct ComputeParams {
     int seq_active_start = 0;
     int attention_size = -1;
     int attention_size_swa = -1;
+    int attention_size_static = -1;  // encoder/cross-attn KV fill level (whisper)
     int input_len = -1;
     int token_len_per_seq = -1;
     int past_kv_len = -1;

@@ -809,11 +809,27 @@ static bool has_non_contiguous_view_input(const ggml_tensor * op) {
 }
 
 static bool is_supported_flash_attn_pattern(const ggml_tensor * op) {
-    // pattern of q,k,v should be q->op==PERMUTE, q->src[0]->op==VIEW, q->src[0]->src[0]->view_src==nullptr
+    // Each Q/K/V input must follow one of:
+    //   PERMUTE -> VIEW  -> base (view_src==nullptr)   (llama KV-cache path)
+    //   PERMUTE -> RESHAPE -> base (view_src==nullptr)  (whisper Q)
+    //   VIEW -> base (view_src==nullptr)                (whisper K/V from kv_pad)
     for (int i = 0; i < 3; i++) {
         const ggml_tensor * src = op->src[i];
-        if (src->op != GGML_OP_PERMUTE || src->src[0] == nullptr || src->src[0]->op != GGML_OP_VIEW ||
-            src->src[0]->src[0] == nullptr || src->src[0]->src[0]->view_src != nullptr) {
+        if (src->op == GGML_OP_PERMUTE) {
+            if (src->src[0] == nullptr) {
+                return false;
+            }
+            if (src->src[0]->op != GGML_OP_VIEW && src->src[0]->op != GGML_OP_RESHAPE) {
+                return false;
+            }
+            if (src->src[0]->src[0] == nullptr || src->src[0]->src[0]->view_src != nullptr) {
+                return false;
+            }
+        } else if (src->op == GGML_OP_VIEW) {
+            if (src->src[0] == nullptr || src->src[0]->view_src != nullptr) {
+                return false;
+            }
+        } else {
             return false;
         }
     }

@@ -3,15 +3,22 @@
 #include "../utils.h"
 
 #include <climits>
+#include <cstdint>
+#include <cstdio>
 #include <memory>
+#include <openvino/frontend/exception.hpp>
 #include <openvino/op/add.hpp>
 #include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
-#include <openvino/op/gather.hpp>
 #include <openvino/op/multiply.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/gather.hpp>
 #include <openvino/op/negative.hpp>
 #include <openvino/op/reshape.hpp>
+#include <openvino/op/scatter_update.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/squeeze.hpp>
 #include <openvino/op/shape_of.hpp>
 #include <openvino/op/slice.hpp>
 
@@ -22,6 +29,7 @@ namespace op {
 
 OutputVector translate_cpy(const NodeContext & context) {
     auto op_case = context.get_op_case();
+    auto input = process_view_input_new(context, 0);
     auto input_shape = context.get_input_shape(0);
     auto output_shape = context.get_input_shape(1);
 
@@ -121,7 +129,64 @@ OutputVector translate_cpy(const NodeContext & context) {
         return rename_outputs_with_suffix({res}, context.get_name());
     }
 
-    auto input = process_view_input_new(context, 0);
+    if (op_case == 5 || op_case == 6) {
+        auto input_shape = context.get_input_shape(0);
+        auto output_shape = context.get_output_shape();
+        auto dst_ggml_shape = context.get_view_input_ggml_shape(1, 0);
+        auto dst_stride = context.get_view_input_stride(1, 0);
+        size_t offset_bytes = context.get_view_input_offset(1, 0);
+        auto n_state = (int64_t)context.get_input_shape(0)[3].get_length();
+        auto n_state_c  = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_state});
+        auto kv_buf = context.get_input(1);  // shape {1,1,1,N}
+
+        Output<Node> token_len_per_seq;
+        Output<Node> n_write_dyn;
+        if (context.has_input("token_len_per_seq")){
+            token_len_per_seq = context.get_input("token_len_per_seq");
+            n_write_dyn = std::make_shared<ov::op::v1::Multiply>(token_len_per_seq, n_state_c);
+        } else {
+            n_write_dyn = ov::op::v0::Constant::create(ov::element::i64, {1}, {(int64_t)dst_ggml_shape[3]});
+        }
+        size_t elem_size = dst_stride[3];
+        FRONT_END_OP_CONVERSION_CHECK(elem_size > 0, "CPY KV cache view update has invalid element size");
+        int64_t start_elem = (int64_t)(offset_bytes / elem_size);
+        // op_case 5: decoder self-attention – write offset advances each step.
+        // op_case 6: encoder self-attn or cross-attn – offset fixed at compile time.
+        const bool is_decoder_self_attn = (op_case == 5);
+        auto ones_c   = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{1, 1, 1});
+        auto new_shape = std::make_shared<ov::op::v0::Concat>(
+            ov::OutputVector{ones_c, n_write_dyn}, 0);
+
+        auto reshaped = std::make_shared<ov::op::v1::Reshape>(input, new_shape, false);
+        auto data = std::make_shared<ov::op::v0::Convert>(
+            reshaped,
+            context.get_output_type());
+        // Indices [start_elem .. start_elem + n_write) on axis 3 of {1,1,1,N}
+        // For decoder self-attention the write offset advances each step, so compute it
+        // dynamically from the model inputs: start = (attention_size - token_len_per_seq) * n_state.
+        // For encoder self-attn and cross-attn the offset is fixed at graph-compile time.
+        ov::Output<ov::Node> start;
+        if (is_decoder_self_attn &&
+                context.has_input("attention_size") && context.has_input("token_len_per_seq")) {
+            auto attention_size_in = context.get_input("attention_size");
+            auto token_len_in = context.get_input("token_len_per_seq");
+            auto past_tokens = std::make_shared<ov::op::v1::Subtract>(attention_size_in, token_len_in);
+            auto new_start = std::make_shared<ov::op::v1::Multiply>(past_tokens, n_state_c);
+            start = std::make_shared<ov::op::v1::Add>(new_start, ov::op::v0::Constant::create(ov::element::i64, {1}, {start_elem}));
+        } else {
+            start = ov::op::v0::Constant::create(ov::element::i64, {1}, {start_elem});
+        }
+        auto start_squeezed = std::make_shared<ov::op::v0::Squeeze>(start);
+        auto end = std::make_shared<ov::op::v1::Add>(start_squeezed, n_write_dyn);
+        auto end_squeezed = std::make_shared<ov::op::v0::Squeeze>(end);
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto step_squeezed = std::make_shared<ov::op::v0::Squeeze>(step);
+        auto indices = std::make_shared<ov::op::v4::Range>(start_squeezed, end_squeezed, step_squeezed, ov::element::i64);
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+
+        auto kv_updated = std::make_shared<ov::op::v3::ScatterUpdate>(kv_buf, indices, data, axis);
+        return rename_outputs_with_suffix({kv_updated}, context.get_name());
+    }
 
     if (input_shape != output_shape) {
         auto new_shape = ov::op::v0::Constant::create(