openvino : enable qwen35 on NPU

ggml/src/ggml-openvino/ggml-decoder.cpp

@@ -357,6 +357,18 @@ int GgmlOvDecoder::compute_op_case(const ggml_tensor * node) const {
         break;
     }
     case GGML_OP_VIEW: {
+        if (m_is_static && node->src[0] != nullptr &&
+            (node->src[0]->op == GGML_OP_GATED_DELTA_NET || node->src[0]->op == GGML_OP_CONCAT)) {
+            // VIEW slicing a GATED_DELTA_NET combined [attn|state] output, or the conv_input
+            // CONCAT. The consuming CPY/RMS_NORM op recovers the true window at runtime via
+            // ssm_state_size / the fixed conv kernel width, so this VIEW must stay an identity
+            // pass-through of the full source here too (it already is on the dynamic path);
+            // otherwise the generic static-mode Slice below would bake in the *captured*
+            // cgraph's token count, which is wrong once the compiled static model runs with a
+            // different token count (prefill chunk size or 1).
+            op_case = 1;
+            break;
+        }
         if (node->src[0]->op == GGML_OP_VIEW) {
             auto * src = node->src[0];
             if (ggml_nelements(node) != ggml_nelements(src)) {
@@ -712,10 +724,8 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr
                 ComputeParams::RsWriteback writeback;
                 writeback.slot_begin = (int) (dest_view->view_offs / row_bytes);
                 if (is_conv) {
-                    // conv_input column the copied window starts at
                     writeback.src_begin = (int) (node->src[0]->view_offs / node->src[0]->view_src->nb[0]);
                 } else if (is_gdn) {
-                    // first row of the state part of the gated-delta-net output
                     writeback.src_begin = (int) (node->src[0]->view_offs / node->src[0]->view_src->nb[1]);
                 }
                 compute_params.rs_writebacks[get_tensor_ov_name(cgraph, node)] = writeback;
@@ -800,7 +810,9 @@ ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op,
         input_shape = ov::PartialShape{1, 1, 1, len};
 
     } else if (is_inp_s_copy(input, op) || is_s_copy_leaf(input)) {
-        input_shape = ov::PartialShape{1, 1, 1, -1};
+        // On NPU the total slot count (n_seq_max) is fixed at translation time, so the s_copy
+        // index list has a static length; on CPU/GPU it may change across compiles (defrag).
+        input_shape = m_is_static ? ov::PartialShape{get_shape(input)} : ov::PartialShape{1, 1, 1, -1};
 
     } else {
         input_shape = ov::PartialShape{get_shape(input)};
@@ -852,8 +864,8 @@ void GgmlOvDecoder::add_extra_inputs() {
     //     see llama_kv_cache_unified::get_n_kv and llama_kv_cache_unified::get_padding.
     // 2. `n_seq_active` and `seq_active_start`, used in FLASH_ATTN_EXT to indicate the active sequences in the batch
 
-    auto create_1d_input = [this](const std::string & name, int64_t value) {
-        m_model_extra_inputs[name] = {ov::element::i64, ov::Shape{1}, value, !m_is_static};
+    auto create_1d_input = [this](const std::string & name, int64_t value, bool force_parameter = false) {
+        m_model_extra_inputs[name] = {ov::element::i64, ov::Shape{1}, value, force_parameter || !m_is_static};
     };
 
     if (m_compute_params.attention_size != -1) {
@@ -874,17 +886,32 @@ void GgmlOvDecoder::add_extra_inputs() {
     // create_1d_input("token_len", m_compute_params.token_len_per_seq * m_compute_params.n_seq_active);
 
     if (m_compute_params.cache_rs_reset_idx != -1) {
-        create_1d_input("cache_rs_reset_idx", m_compute_params.cache_rs_reset_idx);
-        create_1d_input("cache_rs_reset_len", m_compute_params.cache_rs_reset_len);
+        // Whether/which cache slot to reset varies per compute call (e.g. a new sequence starting
+        // vs. continued decoding). can_reuse_statically() does not invalidate the cached static
+        // model on ComputeParams changes, so these must stay runtime Parameters even when static
+        // (scale.cpp op_case 1 only uses them in value comparisons, never as Slice bounds, so this
+        // does not reintroduce dynamic shapes).
+        create_1d_input("cache_rs_reset_idx", m_compute_params.cache_rs_reset_idx, /*force_parameter=*/true);
+        create_1d_input("cache_rs_reset_len", m_compute_params.cache_rs_reset_len, /*force_parameter=*/true);
     }
 
     if (m_compute_params.s_copy_active_slot_len != -1) {
         create_1d_input("s_copy_active_slot_len", m_compute_params.s_copy_active_slot_len);
+        if (m_is_static) {
+            // Number of real tokens in the current prefill chunk. The last chunk is padded with
+            // fabricated token ids; attention masks them out, but the recurrent (GDN/conv) path
+            // would otherwise fold them into cache_r/cache_s permanently. Varies per chunk, so it
+            // must stay a runtime Parameter; it is only compared against a Range or used as Gather
+            // indices, so it does not make any shape dynamic.
+            create_1d_input("chunk_valid_len", get_static_n_tokens(), /*force_parameter=*/true);
+        }
     }
 
     for (const auto & [node_name, writeback] : m_compute_params.rs_writebacks) {
         create_1d_input("rs_slot_begin_" + node_name, writeback.slot_begin);
-        create_1d_input("rs_src_begin_" + node_name, writeback.src_begin);
+        if (!m_is_static) {
+            create_1d_input("rs_src_begin_" + node_name, writeback.src_begin);
+        }
     }
 }
 
@@ -1850,13 +1877,23 @@ void GgmlOvDecoder::compute_node_dynamic_dims() {
                     auto dynamic_dim_stride = src_logical_nb[dynamic_dim_idx] / ggml_type_size(node->src[0]->type) *
                                               ggml_type_size(node->type);
                     int matched_dim_count = 0;
+                    int first_matched_dim = -1;
                     for (int i = 0; i < GGML_MAX_DIMS; i++) {
                         if (node->nb[i] == dynamic_dim_stride && node->ne[i] == node->src[0]->ne[dynamic_dim_idx]) {
+                            if (first_matched_dim == -1) {
+                                first_matched_dim = i;
+                            }
                             m_node_dynamic_dims[node] = i;
                             matched_dim_count++;
                         }
                     }
-                    if (matched_dim_count != 1) {
+                    if (matched_dim_count > 1 && node->src[0]->ne[dynamic_dim_idx] == 1) {
+                        // Single-token capture: every trailing dim is size 1 with the same stride, so
+                        // the match is ambiguous. The lowest index is the real axis; the rest are
+                        // ggml's size-1 padding. Bailing out here would bake the captured token count
+                        // into the static prefill model, which then runs with a different one.
+                        m_node_dynamic_dims[node] = first_matched_dim;
+                    } else if (matched_dim_count != 1) {
                         m_node_dynamic_dims[node] = -1;
                         GGML_LOG_WARN("ggml-openvino: cannot determine dynamic dim for CONT node '%s', src[0]: '%s'\n",
                                       node->name, node->src[0]->name);

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

@@ -85,14 +85,15 @@ struct ComputeParams {
 
     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)
+        int src_begin = 0;   // first source row or column copied by the CPY
     };
 
     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.
+    // Destination slot offset of each state cache writeback CPY node, keyed by node name. It
+    // changes with the batch (kv head, active sequence count) and, with rollback enabled
+    // (cparams.n_rs_seq > 0), the conv state is written back once per snapshot slot. Passed to the
+    // cached model as a runtime input. Dynamic models also receive the source-side offset; static
+    // models use a fixed end-anchored offset in the translator.
 };
 
 class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {

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

@@ -42,6 +42,9 @@ void ggml_openvino_device_config::init() {
         "GGML_OPENVINO_DUMP_IR",
         "GGML_OPENVINO_DEBUG_INPUT",
         "GGML_OPENVINO_DEBUG_OUTPUT",
+        // Force the static (NPU-shape) compute path on any device, e.g. GGML_OPENVINO_DEVICE=CPU,
+        // to test the static-shape translation without NPUW/real NPU hardware in the loop.
+        "GGML_OPENVINO_FORCE_STATIC",
         "GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS",
         "GGML_OPENVINO_ENABLE_CACHE",
         "GGML_OPENVINO_DISABLE_CACHE",

ggml/src/ggml-openvino/openvino/op/cpy.cpp

@@ -6,6 +6,7 @@
 #include <cstdint>
 #include <cstdio>
 #include <memory>
+#include <numeric>
 #include <openvino/frontend/exception.hpp>
 #include <openvino/op/add.hpp>
 #include <openvino/op/concat.hpp>
@@ -30,7 +31,6 @@ 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);
 
@@ -69,10 +69,27 @@ OutputVector translate_cpy(const NodeContext & context) {
         return rename_outputs_with_suffix({res}, context.get_name());
     }
 
-    // Recurrent state cache writeback into a slot block of the cache. Where the block starts and
-    // where the copied data starts in the source are runtime inputs, so the cached model works for
-    // any kv head, active sequence count and token count. The result is the full updated cache.
+    // Recurrent state cache writeback into a slot block of the cache. Where the block starts is a
+    // runtime input, so the cached model works for any kv head and active sequence count. The
+    // result is the full updated cache.
     // op_case 1: gated-delta-net state, op_case 2: conv state, op_case 3: defrag remainder.
+    if (op_case == 3) {
+        // With -np 1 (and generally whenever there is no defrag remainder) this GET_ROWS gathers
+        // zero rows: nothing to write back, and the cache is unchanged. NPU rejects zero-size
+        // tensors, so short-circuit instead of building a degenerate Slice/Concat chain.
+        bool is_empty = false;
+        if (input_shape.rank().is_static()) {
+            for (const auto & d : input_shape) {
+                if (d.is_static() && d.get_length() == 0) {
+                    is_empty = true;
+                    break;
+                }
+            }
+        }
+        if (is_empty) {
+            return {context.get_input(1)};
+        }
+    }
     const std::string slot_begin_name = "rs_slot_begin_" + context.get_name();
     const bool slice_assign =
         context.has_input(slot_begin_name) && !context.is_stateful() && (op_case >= 1 && op_case <= 3);
@@ -89,19 +106,49 @@ OutputVector translate_cpy(const NodeContext & context) {
         ov::Output<ov::Node> begin = context.get_input(slot_begin_name);
         auto base = context.get_input(1);
         if (op_case == 1) {
-            // GDN packs [attn | state snapshots]; the state part runs from src_begin to the end.
-            auto src_begin = context.get_input("rs_src_begin_" + context.get_name());
-            auto state_part = std::make_shared<ov::op::v8::Slice>(context.get_input(0), src_begin, int_max, one, axis);
+            ov::Output<ov::Node> state_begin;
+            const std::string src_begin_name = "rs_src_begin_" + context.get_name();
+            if (context.has_input(src_begin_name)) {
+                state_begin = context.get_input(src_begin_name);
+            } else {
+                auto ssm_state_size = context.get_ssm_state_size();
+                if (context.has_input("s_copy_active_slot_len")) {
+                    auto len = context.get_input("s_copy_active_slot_len");
+                    auto state_rows = std::make_shared<ov::op::v1::Multiply>(
+                        ov::op::v0::Constant::create(ov::element::i64, {1}, {ssm_state_size}), len);
+                    state_begin = std::make_shared<ov::op::v0::Negative>(state_rows);
+                } else {
+                    state_begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {-ssm_state_size});
+                }
+            }
+            auto state_part =
+                std::make_shared<ov::op::v8::Slice>(context.get_input(0), state_begin, int_max, one, axis);
             src = std::make_shared<ov::op::v1::Reshape>(state_part, feature, false);
         } else if (op_case == 2) {
-            // conv_input is [previous conv state | new tokens]; copy the conv_kernel_size - 1 wide
-            // window starting at src_begin, which is the snapshot this writeback corresponds to.
+            // conv_input is [previous conv state | new tokens]; the snapshot is the conv_kernel_size - 1
+            // columns ending at the last *valid* token. Gather (rather than Slice) keeps the output
+            // shape static even though the window start is a runtime value.
             auto window_size = (int64_t) input_shape[3].get_length();
-            auto src_begin = context.get_input("rs_src_begin_" + context.get_name());
-            auto src_end = std::make_shared<ov::op::v1::Add>(
-                src_begin, ov::op::v0::Constant::create(ov::element::i64, {1}, {window_size}));
-            auto window = std::make_shared<ov::op::v8::Slice>(context.get_input(0), src_begin, src_end, one,
-                                                              ov::op::v0::Constant::create(ov::element::i64, {1}, {3}));
+            ov::Output<ov::Node> window;
+            auto col_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+            const std::string src_begin_name = "rs_src_begin_" + context.get_name();
+            if (context.has_input(src_begin_name)) {
+                auto src_begin = context.get_input(src_begin_name);
+                auto src_end = std::make_shared<ov::op::v1::Add>(
+                    src_begin, ov::op::v0::Constant::create(ov::element::i64, {1}, {window_size}));
+                window = std::make_shared<ov::op::v8::Slice>(context.get_input(0), src_begin, src_end, one, col_axis);
+            } else if (context.has_input("chunk_valid_len")) {
+                std::vector<int64_t> offsets(window_size);
+                std::iota(offsets.begin(), offsets.end(), 0);
+                auto indices = std::make_shared<ov::op::v1::Add>(
+                    ov::op::v0::Constant::create(ov::element::i64, {(size_t) window_size}, offsets),
+                    context.get_input("chunk_valid_len"));
+                window = std::make_shared<ov::op::v8::Gather>(context.get_input(0), indices, col_axis);
+            } else {
+                auto window_begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {-window_size});
+                window =
+                    std::make_shared<ov::op::v8::Slice>(context.get_input(0), window_begin, int_max, one, col_axis);
+            }
             const auto base_shape = base.get_partial_shape();
             FRONT_END_OP_CONVERSION_CHECK(base_shape.rank().is_static() && base_shape.rank().get_length() == 4,
                                           "CPY conv state cache update requires rank-4 base cache");
@@ -163,6 +210,8 @@ 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();

ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp

@@ -7,12 +7,15 @@
 #include <cmath>
 #include <cstdint>
 #include <memory>
+#include <numeric>
 #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/exp.hpp>
 #include <openvino/op/gather.hpp>
+#include <openvino/op/less.hpp>
 #include <openvino/op/loop.hpp>
 #include <openvino/op/matmul.hpp>
 #include <openvino/op/multiply.hpp>
@@ -80,6 +83,28 @@ OutputVector translate_gated_delta_net(const NodeContext & context) {
     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}));
 
+    if (context.has_input("chunk_valid_len")) {
+        // The last prefill chunk is padded with fabricated tokens. The recurrence is
+        //   S_t = S_{t-1} * exp(g_t) + k_t (x) ((v_t - S_{t-1}^T k_t) * beta_t)
+        // so forcing g = 0 and beta = 0 makes a padded step an exact identity and keeps the final
+        // state equal to the state after the last real token. Attention output at those positions
+        // is garbage but never read.
+        const auto & g_ps = g.get_partial_shape();
+        FRONT_END_OP_CONVERSION_CHECK(g_ps.rank().is_static() && g_ps.rank().get_length() == 3 && g_ps[1].is_static(),
+                                      "GATED_DELTA_NET pad masking requires a static token dimension");
+        const int64_t n_tokens = g_ps[1].get_length();
+        std::vector<int64_t> positions(n_tokens);
+        std::iota(positions.begin(), positions.end(), 0);
+        auto valid = std::make_shared<ov::op::v1::Less>(
+            ov::op::v0::Constant::create(ov::element::i64, {(size_t) n_tokens}, positions),
+            context.get_input("chunk_valid_len"));
+        auto mask = std::make_shared<ov::op::v0::Unsqueeze>(
+            std::make_shared<ov::op::v0::Convert>(valid, g.get_element_type()),
+            ov::op::v0::Constant::create(ov::element::i64, {2}, std::vector<int64_t>{0, 2}));
+        g = std::make_shared<ov::op::v1::Multiply>(g, mask);
+        beta = std::make_shared<ov::op::v1::Multiply>(beta, mask);
+    }
+
     // std::cout << "GatedDeltaNet input shapes: q=" << q.get_partial_shape() << ", k=" << k.get_partial_shape()
     //           << ", v=" << v.get_partial_shape() << ", g=" << g.get_partial_shape()
     //           << ", beta=" << beta.get_partial_shape() << ", state=" << state.get_partial_shape() << std::endl;

ggml/src/ggml-openvino/openvino/op/view.cpp

@@ -17,6 +17,13 @@ namespace op {
 OutputVector translate_view(const NodeContext & context) {
     num_inputs_check(context, 1, 1);
 
+    if (context.get_op_case() == 1) {
+        // Static-mode identity pass-through for VIEWs over a GATED_DELTA_NET combined output or
+        // the conv_input CONCAT; the consuming op (CPY/RMS_NORM) does its own runtime-correct
+        // slicing on the full tensor (see ggml-decoder.cpp compute_op_case, GGML_OP_VIEW).
+        return {context.get_input(0)};
+    }
+
     if (!context.is_static()) {
         // On the stateless/non-static path VIEW is normally a no-op (consumers re-slice).
         // EXCEPTION: the MoE expert aggregation slices each expert plane out of

ggml/src/ggml-openvino/openvino/utils.cpp

@@ -72,6 +72,7 @@ OutputVector rename_outputs_with_suffix(const OutputVector & outputs, const std:
         name += "_";
         name += suffix;
         node->set_friendly_name(name);
+        // Uncomment to dump every node's inferred shape (used to hunt down dynamic dims on NPU).
         // std::cout << name << "  " << output.get_partial_shape() << std::endl;
     }
     return outputs;