Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	docs/backend/SYCL.md
#	ggml/src/ggml-cpu/ggml-cpu.c
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/cvt.cl
#	ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl
#	ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl
#	ggml/src/ggml-sycl/backend.hpp
#	ggml/src/ggml-sycl/common.hpp
#	ggml/src/ggml-sycl/dmmv.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	scripts/sync_vendor.py
#	tests/CMakeLists.txt
#	tests/test-alloc.cpp
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
#	tests/test-gguf.cpp
#	tests/test-save-load-state.cpp

src/llama-context.cpp

@@ -58,6 +58,12 @@ static const llm_fused_op_probe llm_fused_op_gdn_ch_probe = {
     /*.n_tokens_per_seq =*/ 16,
 };
 
+static const llm_fused_op_probe llm_fused_op_lid_probe = {
+    /*.op               =*/ LLM_FUSED_OP_LIGHTNING_INDEXER,
+    /*.name             =*/ "Lightning Indexer",
+    /*.n_tokens_per_seq =*/ 1,
+};
+
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
@@ -229,6 +235,9 @@ llama_context::llama_context(
     cparams.fused_gdn_ch = true;
     cparams.auto_fgdn    = true;
 
+    cparams.fused_lid    = true;
+    cparams.auto_flid    = true;
+
     // 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;
 
@@ -530,6 +539,12 @@ void llama_context::resolve_fused_ops(const llama_memory_context_i * mctx, uint3
         resolve(llm_fused_op_gdn_ch_probe, cparams.fused_gdn_ch);
         cparams.auto_fgdn = false;
     }
+
+    if (cparams.auto_flid) {
+        LLAMA_LOG_INFO("%s: resolving fused Lightning Indexer support:\n", func);
+        resolve(llm_fused_op_lid_probe, cparams.fused_lid);
+        cparams.auto_flid = false;
+    }
 }
 
 void llama_context::sched_reserve() {

src/llama-cparams.h

@@ -41,6 +41,8 @@ struct llama_cparams {
     bool fused_gdn_ar;       // use fused gated delta net (autoregressive)
     bool fused_gdn_ch;       // use fused gated delta net (chunked)
     bool auto_fgdn;
+    bool fused_lid;          // use fused lightning indexer
+    bool auto_flid;
     bool no_perf;
     bool warmup;             // TODO: remove [TAG_LLAMA_GRAPH_NO_WARMUP]
     bool op_offload;

src/llama-graph.cpp

@@ -843,7 +843,7 @@ static void dsv4_build_comp_inputs(
         GGML_ASSERT(n_stream > 0);
         GGML_ASSERT(n_tokens%n_stream == 0);
 
-        inp.kq_mask = ggml_new_tensor_4d(ctx, cparams.flash_attn && strcmp(name, "lid") != 0 ? GGML_TYPE_F16 : GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
+        inp.kq_mask = ggml_new_tensor_4d(ctx, (strcmp(name, "lid") != 0 && cparams.flash_attn) || (strcmp(name, "lid") == 0 && cparams.fused_lid) ? GGML_TYPE_F16 : GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp.kq_mask);
         ggml_set_name(inp.kq_mask, (std::string("dsv4_") + name + "_kq_mask").c_str());
     }
@@ -3026,9 +3026,9 @@ llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
     {
         inp->self_k_idxs_lid = mctx_cur->get_lid()->build_input_k_idxs(ctx0, ubatch);
 
-        // ensure F32 mask
+        // ensure that mask type matches fused lightning indexer use (requires f16 mask)
         auto cparams_copy = cparams;
-        cparams_copy.flash_attn = false;
+        cparams_copy.flash_attn = cparams.fused_lid;
 
         inp->self_kq_mask_lid = build_attn_inp_kq_mask(ctx0, mctx_cur->get_lid(), ubatch, cparams_copy);
         inp->self_kq_mask_lid_cnv = inp->self_kq_mask_lid;

src/llama-graph.h

@@ -42,6 +42,7 @@ enum llm_fused_op {
     LLM_FUSED_OP_FLASH_ATTN,
     LLM_FUSED_OP_GDN_AR,
     LLM_FUSED_OP_GDN_CH,
+    LLM_FUSED_OP_LIGHTNING_INDEXER,
 };
 
 enum llm_ffn_op_type : int {

src/llama-kv-cache-dsv4.cpp

@@ -29,6 +29,15 @@ static uint32_t dsv4_comp_size(uint32_t kv_size, uint32_t ratio) {
     return std::max<uint32_t>(1, (kv_size + ratio - 1)/ratio);
 }
 
+static void dsv4_clear_tensor_stream(ggml_tensor * tensor, uint32_t stream) {
+    GGML_ASSERT(ggml_is_contiguous(tensor));
+    GGML_ASSERT(tensor->ne[3] == 1);
+    GGML_ASSERT(stream < (uint32_t) tensor->ne[2]);
+
+    const size_t stream_size = tensor->nb[2];
+    ggml_backend_tensor_memset(tensor, 0, stream*stream_size, stream_size);
+}
+
 static int64_t dsv4_stream_offset(uint32_t n_stream, llama_seq_id seq_id, uint32_t size) {
     if (n_stream <= 1) {
         return 0;
@@ -781,11 +790,20 @@ llama_dsv4_comp_state::llama_dsv4_comp_state(
             __func__, name, ratio, state_size, n_embd_state, n_stream, layers.size(), total_size()/1024.0/1024.0);
 }
 
-void llama_dsv4_comp_state::clear(bool data) {
+void llama_dsv4_comp_state::clear(llama_seq_id seq_id, bool data) {
     if (!data) {
         return;
     }
 
+    if (seq_id >= 0) {
+        GGML_ASSERT((uint32_t) seq_id < n_stream);
+        for (const auto & layer : layers) {
+            dsv4_clear_tensor_stream(layer.kv,    (uint32_t) seq_id);
+            dsv4_clear_tensor_stream(layer.score, (uint32_t) seq_id);
+        }
+        return;
+    }
+
     for (auto & [_, buf] : ctxs_bufs) {
         ggml_backend_buffer_clear(buf.get(), 0);
     }
@@ -1034,7 +1052,7 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4(
     // graph does not necessarily overwrite; uninitialized buffer contents would
     // otherwise leak in (instance-specific garbage) and corrupt recall. Zero all
     // compressed buffers up front so reads of un-written rows are deterministic.
-    clear_compressed(true);
+    clear_compressed(-1, true);
 }
 
 llama_memory_context_ptr llama_kv_cache_dsv4::init_batch(
@@ -1147,7 +1165,7 @@ bool llama_kv_cache_dsv4::get_can_shift() const {
 
 void llama_kv_cache_dsv4::clear(bool data) {
     kv_raw->clear(data);
-    clear_compressed(true); // DSV4 compressed buffers must never expose stale/uninit rows
+    clear_compressed(-1, true); // DSV4 compressed buffers must never expose stale/uninit rows
 }
 
 bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
@@ -1169,7 +1187,7 @@ bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1
     const bool res = kv_raw->seq_rm(seq_id, p0, p1);
 
     if (res) {
-        clear_compressed(true);
+        clear_compressed(seq_id, true);
     }
 
     return res;
@@ -1177,22 +1195,29 @@ bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1
 
 void llama_kv_cache_dsv4::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     kv_raw->seq_cp(seq_id_src, seq_id_dst, p0, p1);
-    clear_compressed(true);
 }
 
 void llama_kv_cache_dsv4::seq_keep(llama_seq_id seq_id) {
+    GGML_ASSERT(seq_id >= 0 && (uint32_t) seq_id < n_seq_max);
+
     kv_raw->seq_keep(seq_id);
-    clear_compressed(true);
+
+    for (llama_seq_id id = 0; id < (llama_seq_id) n_seq_max; ++id) {
+        if (id == seq_id) {
+            continue;
+        }
+
+        kv_raw->seq_rm(id, -1, -1);
+        clear_compressed(id, true);
+    }
 }
 
 void llama_kv_cache_dsv4::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     kv_raw->seq_add(seq_id, p0, p1, shift);
-    clear_compressed(true);
 }
 
 void llama_kv_cache_dsv4::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     kv_raw->seq_div(seq_id, p0, p1, d);
-    clear_compressed(true);
 }
 
 llama_pos llama_kv_cache_dsv4::seq_pos_min(llama_seq_id seq_id) const {
@@ -1328,13 +1353,32 @@ llama_dsv4_comp_state * llama_kv_cache_dsv4::get_lid_state() const {
     return lid_state.get();
 }
 
-void llama_kv_cache_dsv4::clear_compressed(bool data) {
-    kv_csa->clear(data);
-    kv_hca->clear(data);
-    kv_lid->clear(data);
-    csa_state->clear(data);
-    hca_state->clear(data);
-    lid_state->clear(data);
+void llama_kv_cache_dsv4::clear_compressed(llama_seq_id seq_id, bool data) {
+    if (seq_id < 0) {
+        kv_csa->clear(data);
+        kv_hca->clear(data);
+        kv_lid->clear(data);
+    } else {
+        GGML_ASSERT((uint32_t) seq_id < n_seq_max);
+
+        const auto clear_seq = [seq_id, data](llama_kv_cache * kv) {
+            kv->seq_rm(seq_id, -1, -1);
+
+            if (data) {
+                for (uint32_t il : kv->get_layer_ids()) {
+                    dsv4_clear_tensor_stream(kv->get_k_storage(il), (uint32_t) seq_id);
+                }
+            }
+        };
+
+        clear_seq(kv_csa.get());
+        clear_seq(kv_hca.get());
+        clear_seq(kv_lid.get());
+    }
+
+    csa_state->clear(seq_id, data);
+    hca_state->clear(seq_id, data);
+    lid_state->clear(seq_id, data);
 }
 
 //

src/llama-kv-cache-dsv4.h

@@ -21,7 +21,7 @@ public:
             const char    * name,
         const llama_memory_i::layer_filter_cb & filter);
 
-    void clear(bool data);
+    void clear(llama_seq_id seq_id, bool data);
 
     uint32_t get_ratio()    const;
     uint32_t get_state_size() const;
@@ -67,6 +67,8 @@ private:
 // DSV4 uses a normal raw/SWA token cache plus compressed K-only block caches.
 // The compressed caches are storage only; DSV4-specific visibility and block
 // planning are handled by llama_kv_cache_dsv4_context / llm_graph_input_dsv4.
+// FIXME: currently the cache only supports non-unified mode even if unified flag is passed
+// FIXME: we currently conflate token_pos and buffer contents. See https://github.com/ggml-org/llama.cpp/pull/25521#discussion_r3558173819
 
 class llama_kv_cache_dsv4 : public llama_memory_i {
 public:
@@ -146,7 +148,7 @@ private:
     std::unique_ptr<llama_dsv4_comp_state> hca_state;
     std::unique_ptr<llama_dsv4_comp_state> lid_state;
 
-    void clear_compressed(bool data);
+    void clear_compressed(llama_seq_id seq_id, bool data);
 };
 
 // DSV4 raw attention only uses the SWA half of kv_raw. The base half is kept

src/llama-model.cpp

@@ -451,8 +451,7 @@ llama_model * llama_model_create(llm_arch arch, const llama_model_params & param
 
     if (model != nullptr) {
         model->arch = arch;
-        auto & devices = model->devices;
-        if (!devices.empty() && devices[0].is_meta && !llm_arch_supports_sm_tensor(arch)) {
+        if (params.split_mode == LLAMA_SPLIT_MODE_TENSOR && !llm_arch_supports_sm_tensor(arch)) {
             throw std::runtime_error(std::string("LLAMA_SPLIT_MODE_TENSOR not implemented for architecture '") + llm_arch_name(arch) + "'");
         }
     }

src/models/deepseek32.cpp

@@ -301,43 +301,50 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
                 indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream, indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
                 indexer_weights = ggml_view_4d(ctx0, indexer_weights, indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream, indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
 
-                // calculate indexer kq
-                indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
-                cb(indexer_q, "indexer_q", il);
-                indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
-                cb(indexer_k, "indexer_k", il);
-
-                ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
-                cb(indexer_kq, "indexer_kq", il);
-
-                // ReLU requires contiguous tensors
-                indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
-                cb(indexer_kq, "indexer_kq", il);
-
-                // apply ReLU
-                ggml_tensor * indexer_score = ggml_relu(ctx0, indexer_kq);
-                cb(indexer_score, "indexer_score", il);
-
                 // pre-scale weights to avoid scaling operations on huge indexer_score tensor
                 indexer_weights = ggml_scale(ctx0, indexer_weights, 1.0f / sqrtf(float(n_embd_indexer_head * n_indexer_head)));
                 cb(indexer_weights, "indexer_weights", il);
 
-                // multiply scores by indexer weights
-                indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
-                cb(indexer_score, "indexer_score", il);
+                ggml_tensor * indexer_score = nullptr;
+                if (cparams.fused_lid) {
+                    indexer_score = ggml_lightning_indexer(ctx0, indexer_q, indexer_k, indexer_weights, inp_attn_dsa->get_kq_mask_lid());
+                    cb(indexer_score, "indexer_score", il);
+                    res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, indexer_score, il});
+                } else {
+                    // calculate indexer kq
+                    indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
+                    cb(indexer_q, "indexer_q", il);
+                    indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
+                    cb(indexer_k, "indexer_k", il);
 
-                // sum by q n_indexer_head dimension
-                indexer_score = ggml_sum_rows(ctx0, indexer_score);
-                cb(indexer_score, "indexer_score", il);
+                    ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
+                    cb(indexer_kq, "indexer_kq", il);
 
-                // permute result to match KQ mask
-                indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
-                cb(indexer_score, "indexer_score", il);
+                    // ReLU requires contiguous tensors
+                    indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
+                    cb(indexer_kq, "indexer_kq", il);
 
-                // mask indexer scores
-                ggml_tensor * indexer_kq_mask = inp_attn_dsa->get_kq_mask_lid();
-                indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
-                cb(indexer_score, "indexer_score", il);
+                    // apply ReLU
+                    indexer_score = ggml_relu(ctx0, indexer_kq);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // multiply scores by indexer weights
+                    indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // sum by q n_indexer_head dimension
+                    indexer_score = ggml_sum_rows(ctx0, indexer_score);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // permute result to match KQ mask
+                    indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
+                    cb(indexer_score, "indexer_score", il);
+
+                    // mask indexer scores
+                    ggml_tensor * indexer_kq_mask = inp_attn_dsa->get_kq_mask_lid();
+                    indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
+                    cb(indexer_score, "indexer_score", il);
+                }
 
                 // get indices of top k indexer scores
                 uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k;

src/models/deepseek4.cpp

@@ -556,25 +556,32 @@ ggml_tensor * llama_model_deepseek4::graph::build_lid_top_k(
             indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream,
             indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
 
-    indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
-    cb(indexer_q, "lid_q", il);
-    indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
-    cb(indexer_k, "lid_k", il);
+    ggml_tensor * indexer_score = nullptr;
+    if (cparams.fused_lid) {
+        indexer_score = ggml_lightning_indexer(ctx0, indexer_q, indexer_k, indexer_weights, inp_lid.kq_mask);
+        cb(indexer_score, "lid_score_masked", il);
+        res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, indexer_score, il});
+    } else {
+        indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
+        cb(indexer_q, "lid_q", il);
+        indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
+        cb(indexer_k, "lid_k", il);
 
-    ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
-    cb(indexer_kq, "lid_kq", il);
+        ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
+        cb(indexer_kq, "lid_kq", il);
 
-    indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
-    cb(indexer_kq, "lid_kq", il);
+        indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
+        cb(indexer_kq, "lid_kq", il);
 
-    ggml_tensor * indexer_score = ggml_relu(ctx0, indexer_kq);
-    indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
-    indexer_score = ggml_sum_rows(ctx0, indexer_score);
-    indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
-    cb(indexer_score, "lid_score", il);
+        indexer_score = ggml_relu(ctx0, indexer_kq);
+        indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
+        indexer_score = ggml_sum_rows(ctx0, indexer_score);
+        indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
+        cb(indexer_score, "lid_score", il);
 
-    indexer_score = ggml_add(ctx0, indexer_score, inp_lid.kq_mask);
-    cb(indexer_score, "lid_score_masked", il);
+        indexer_score = ggml_add(ctx0, indexer_score, inp_lid.kq_mask);
+        cb(indexer_score, "lid_score_masked", il);
+    }
 
     const uint32_t n_top_k = indexer_score->ne[0] < hparams.indexer_top_k ? indexer_score->ne[0] : hparams.indexer_top_k;
     ggml_tensor * top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k));