Merge commit '213c4a0b81788e058c30479842954fb0815be61a' into concedo_experimental

# Conflicts:
#	CODEOWNERS
#	common/CMakeLists.txt
#	common/chat-peg-parser.cpp
#	common/chat.cpp
#	docs/backend/SYCL.md
#	docs/development/parsing.md
#	docs/ops.md
#	docs/ops/SYCL.csv
#	embd_res/templates/Apriel-1.6-15b-Thinker-fixed.jinja
#	embd_res/templates/Bielik-11B-v3.0-Instruct.jinja
#	embd_res/templates/GLM-4.7-Flash.jinja
#	embd_res/templates/LFM2-8B-A1B.jinja
#	embd_res/templates/StepFun3.5-Flash.jinja
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-sycl/CMakeLists.txt
#	ggml/src/ggml-sycl/backend.hpp
#	ggml/src/ggml-sycl/common.hpp
#	ggml/src/ggml-sycl/convert.cpp
#	ggml/src/ggml-sycl/convert.hpp
#	ggml/src/ggml-sycl/count-equal.cpp
#	ggml/src/ggml-sycl/dpct/helper.hpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-sycl/presets.hpp
#	ggml/src/ggml-sycl/softmax.cpp
#	ggml/src/ggml-sycl/vecdotq.hpp
#	models/templates/Apertus-8B-Instruct.jinja
#	models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja
#	models/templates/Qwen-QwQ-32B.jinja
#	models/templates/Qwen3-Coder.jinja
#	models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja
#	models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja
#	models/templates/deepseek-ai-DeepSeek-V3.1.jinja
#	models/templates/fireworks-ai-llama-3-firefunction-v2.jinja
#	models/templates/moonshotai-Kimi-K2.jinja
#	models/templates/unsloth-Apriel-1.5.jinja
#	tests/CMakeLists.txt
#	tests/peg-parser/test-basic.cpp
#	tests/peg-parser/tests.h
#	tests/test-backend-ops.cpp
#	tests/test-chat-peg-parser.cpp
#	tests/test-chat-template.cpp
#	tests/test-chat.cpp
#	tests/test-json-schema-to-grammar.cpp
#	tests/test-peg-parser.cpp
#	tools/CMakeLists.txt
#	tools/cli/cli.cpp

src/llama-context.cpp

@@ -153,6 +153,9 @@ llama_context::llama_context(
     cparams.flash_attn = params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_DISABLED;
     cparams.auto_fa    = params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO;
 
+    cparams.fused_gdn_ar = true;
+    cparams.fused_gdn_ch = false; // TODO: implement
+
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
 
@@ -431,7 +434,7 @@ void llama_context::sched_reserve() {
     if (cparams.auto_fa) {
         auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
         if (!gf) {
-            throw std::runtime_error("failed to split graph for Flash Attention check");
+            throw std::runtime_error("failed to reserve graph for Flash Attention check");
         }
 
         const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
@@ -441,8 +444,7 @@ void llama_context::sched_reserve() {
             if (n->op != GGML_OP_FLASH_ATTN_EXT) {
                 continue;
             }
-            ggml_backend_dev_t device_fa = ggml_backend_get_device(
-                    ggml_backend_sched_get_tensor_backend(sched.get(), n));
+            ggml_backend_dev_t device_fa = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
 
             // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
             GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
@@ -457,6 +459,7 @@ void llama_context::sched_reserve() {
                 break;
             }
         }
+
         if (fa_device_mismatch) {
             cparams.flash_attn = false;
             LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
@@ -468,6 +471,39 @@ void llama_context::sched_reserve() {
         cparams.auto_fa = false;
     }
 
+    if (cparams.fused_gdn_ar) {
+        auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
+        if (!gf) {
+            throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check");
+        }
+
+        const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDNAR) + 1;
+        bool gdn_device_mismatch = false;
+        for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+            ggml_tensor * n = ggml_graph_node(gf, i);
+            if (n->op != GGML_OP_GATED_DELTA_NET) {
+                continue;
+            }
+            ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
+
+            GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDNAR "-", prefix_len) == 0);
+            const int il = std::stoi(n->name + prefix_len);
+            ggml_backend_dev_t device_kv = model.dev_layer(il);
+            if (device_gdn != device_kv) {
+                LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
+                        "is assigned to device %s (usually due to missing support)\n",
+                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
+                gdn_device_mismatch = true;
+                break;
+            }
+        }
+
+        if (gdn_device_mismatch) {
+            cparams.fused_gdn_ar = false;
+            LLAMA_LOG_WARN("%s: fused Gated Delta Net not supported, set to disabled\n", __func__);
+        }
+    }
+
     // reserve worst-case graph
     int n_splits_pp = -1;
     int n_nodes_pp  = -1;

src/llama-cparams.h

@@ -31,6 +31,8 @@ struct llama_cparams {
     bool offload_kqv;
     bool flash_attn;
     bool auto_fa;
+    bool fused_gdn_ar;       // use fused gated delta net (autoregressive)
+    bool fused_gdn_ch;       // use fused gated delta net (chunked)
     bool no_perf;
     bool warmup;
     bool op_offload;

src/llama-impl.h

@@ -70,4 +70,6 @@ std::string llama_format_tensor_shape(const struct ggml_tensor * t);
 
 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
 
-#define LLAMA_TENSOR_NAME_FATTN "__fattn__"
+#define LLAMA_TENSOR_NAME_FATTN  "__fattn__"
+#define LLAMA_TENSOR_NAME_FGDNAR "__fgdnar__"
+#define LLAMA_TENSOR_NAME_FGDNCH "__fgdnch__"

src/models/delta-net-base.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include "llama-impl.h"
+
 // utility to get one slice from the third dimension
 // input dim:  [x, y, c, b]
 // output dim: [x, y, 1, b]
@@ -39,6 +41,13 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
     GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
 
+    if (cparams.fused_gdn_ch) {
+        //ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
+        //cb(result, LLAMA_TENSOR_NAME_FGDNCH, il);
+
+        GGML_ABORT("not implemented yet");
+    }
+
     const float scale = 1.0f / sqrtf(S_k);
 
     q = ggml_scale(ctx0, q, scale);
@@ -316,6 +325,26 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
     GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
 
+    if (cparams.fused_gdn_ar) {
+        ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
+        cb(result, LLAMA_TENSOR_NAME_FGDNAR, il);
+
+        ggml_tensor * output = ggml_view_4d(ctx0, result,
+            S_v, H_v, n_tokens, n_seqs,
+            ggml_row_size(result->type, S_v),
+            ggml_row_size(result->type, S_v * H_v),
+            ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
+
+        ggml_tensor * new_state = ggml_view_4d(ctx0, result,
+            S_v, S_v, H_v, n_seqs,
+            ggml_row_size(result->type, S_v),
+            ggml_row_size(result->type, S_v * S_v),
+            ggml_row_size(result->type, S_v * S_v * H_v),
+            ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
+
+        return {output, new_state};
+    }
+
     const float scale = 1.0f / sqrtf(S_k);
 
     q = ggml_scale(ctx0, q, scale);

src/models/qwen35.cpp

@@ -332,8 +332,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
-    std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
+    std::pair<ggml_tensor *, ggml_tensor *> attn_out;
     if (n_seq_tokens == 1) {
         attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
     } else {

src/models/qwen35moe.cpp

@@ -332,8 +332,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
-    std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
+    std::pair<ggml_tensor *, ggml_tensor *> attn_out;
     if (n_seq_tokens == 1) {
         attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
     } else {