@fugood/llama.node 1.1.11 → 1.2.0-rc.0

package/src/llama.cpp/src/llama-context.cpp

@@ -270,19 +270,7 @@ llama_context::llama_context(
         }
     }
 
-    // resolve automatic Flash Attention use and reserve worst-case graph
     if (!hparams.vocab_only) {
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
-        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-
-        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
-
-        int n_splits_pp = -1;
-        int n_nodes_pp  = -1;
-
-        int n_splits_tg = -1;
-        int n_nodes_tg  = -1;
-
         llama_memory_context_ptr mctx;
         if (memory) {
             LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -294,54 +282,69 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        // reserve pp (prompt processing) graph first so that buffers are only allocated once
-        {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+
+        // avoid reserving graphs with zero outputs
+        n_outputs = 1;
+
+        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
+
+        // resolve automatic Flash Attention use
+        if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
+            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
             if (!gf) {
-                throw std::runtime_error("failed to allocate compute pp buffers");
+                throw std::runtime_error("failed to split graph for Flash Attention check");
             }
 
-            if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
-                ggml_backend_sched_alloc_graph(sched.get(), gf);
-
-                const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
-                bool fa_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_FLASH_ATTN_EXT) {
-                        continue;
-                    }
-                    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);
-                    const int il = std::stoi(n->name + prefix_len);
-                    ggml_backend_dev_t device_kv = model.dev_layer(il);
-                    if (device_fa != device_kv) {
-                        LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention 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_fa));
-                        // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
-                        fa_device_mismatch = true;
-                        break;
-                    }
+            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
+            bool fa_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_FLASH_ATTN_EXT) {
+                    continue;
                 }
-                if (fa_device_mismatch) {
-                    cparams.flash_attn = false;
-                    LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
-                    if (ggml_is_quantized(params.type_v)) {
-                        throw std::runtime_error("quantized V cache was requested, but this requires Flash Attention");
-                    }
-                    auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
-                    if (!gf) {
-                        throw std::runtime_error("failed to allocate compute pp buffers");
-                    }
-                } else {
-                    cparams.flash_attn = true;
-                    LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
+                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);
+                const int il = std::stoi(n->name + prefix_len);
+                ggml_backend_dev_t device_kv = model.dev_layer(il);
+                if (device_fa != device_kv) {
+                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention 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_fa));
+                    // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
+                    fa_device_mismatch = true;
+                    break;
                 }
             }
+            if (fa_device_mismatch) {
+                cparams.flash_attn = false;
+                LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
+                if (ggml_is_quantized(params.type_v)) {
+                    throw std::runtime_error("quantized V cache was requested, but this requires Flash Attention");
+                }
+            } else {
+                cparams.flash_attn = true;
+                LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
+            }
+        }
+
+        // reserve worst-case graph
+        int n_splits_pp = -1;
+        int n_nodes_pp  = -1;
+
+        int n_splits_tg = -1;
+        int n_nodes_tg  = -1;
+
+        // reserve pp (prompt processing) graph first so that buffers are only allocated once
+        {
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            if (!gf) {
+                throw std::runtime_error("failed to allocate compute pp buffers");
+            }
 
             n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
             n_nodes_pp  = ggml_graph_n_nodes(gf);
@@ -1366,8 +1369,9 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
+ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
+    GGML_ASSERT(n_outputs >= 1);
 
     if (n_tokens % n_seqs != 0) {
         n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
@@ -1401,7 +1405,9 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     this->n_outputs = save_n_outputs;
 
     // initialize scheduler with the specified graph
-    if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+    if (split_only) {
+        ggml_backend_sched_split_graph(sched.get(), gf);
+    } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }

package/src/llama.cpp/src/llama-context.h

@@ -196,7 +196,7 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
 
 private:
     llm_graph_params graph_params(

package/src/llama.cpp/src/llama-graph.cpp

@@ -258,6 +258,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
+    LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
+    const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
+                          (swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
+                          (swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
+                          (swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
+    LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
+    LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
+    LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
+
+    LLAMA_LOG_DEBUG("    ");
+    for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
+        LLAMA_LOG_DEBUG("%2d", j);
+    }
+    LLAMA_LOG_DEBUG("\n");
+
+    for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
+        LLAMA_LOG_DEBUG(" %2d ", i);
+        for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
+            float val = data[i * n_kv + j];
+            if (val == -INFINITY) {
+                LLAMA_LOG_DEBUG(" ∞");
+            } else {
+                LLAMA_LOG_DEBUG(" 0");
+            }
+        }
+        LLAMA_LOG_DEBUG("\n");
+    }
+}
+
 void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
     const int64_t n_kv     = ubatch->n_tokens;
     const int64_t n_tokens = ubatch->n_tokens;
@@ -267,6 +297,9 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 
     float * data = (float *) kq_mask->data;
 
+    // [TAG_NO_CACHE_ISWA]
+    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");
+
     for (int h = 0; h < 1; ++h) {
         for (int i1 = 0; i1 < n_tokens; ++i1) {
             const llama_seq_id s1 = ubatch->seq_id[i1][0];
@@ -277,21 +310,33 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                 for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
                     const llama_seq_id s0 = ubatch->seq_id[i0][0];
 
-                    // TODO: reimplement this like in llama_kv_cache
-                    if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
-                        if (hparams.use_alibi) {
-                            f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
-                        } else {
-                            f = 0.0f;
-                        }
-                        break;
+                    if (s0 != s1) {
+                        continue; // skip different sequences
                     }
-                }
 
+                    if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
+                        continue; // skip future tokens for causal attention
+                    }
+
+                    // TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
+                    //if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
+                    //    continue; // skip masked tokens for SWA
+                    //}
+
+                    // TODO: reimplement this like in llama_kv_cache_unified
+                    if (hparams.use_alibi) {
+                        f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
+                    } else {
+                        f = 0.0f;
+                    }
+                }
                 data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
             }
         }
     }
+    if (debug) {
+        print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
+    }
 }
 
 void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {

package/src/llama.cpp/src/llama-graph.h

@@ -78,6 +78,11 @@ struct llm_graph_params;
 
 class llm_graph_input_i {
 public:
+    llm_graph_input_i() {
+        const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");
+        debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;
+    }
+
     virtual ~llm_graph_input_i() = default;
 
     virtual void set_input(const llama_ubatch * ubatch) = 0;
@@ -90,6 +95,9 @@ public:
         GGML_UNUSED(params);
         return false;
     }
+protected:
+    // env: LLAMA_GRAPH_INPUT_DEBUG
+    int debug = 0;
 };
 
 using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;

package/src/llama.cpp/src/llama-hparams.cpp

@@ -1,6 +1,7 @@
 #include "llama-hparams.h"
 
 #include "ggml.h"
+#include <cassert>
 
 void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
     if (dense_first) {
@@ -178,3 +179,39 @@ uint32_t llama_hparams::n_layer_kv() const {
 
     return res;
 }
+
+bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
+    assert(p0 >= 0 && p1 >= 0);
+
+    switch (swa_type) {
+        case LLAMA_SWA_TYPE_NONE:
+            {
+            } break;
+        case LLAMA_SWA_TYPE_STANDARD:
+            {
+                if (p1 - p0 >= (int32_t) n_swa) {
+                    return true;
+                }
+            } break;
+        case LLAMA_SWA_TYPE_CHUNKED:
+            {
+                const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
+
+                if (p0 < pos_chunk_start) {
+                    return true;
+                }
+            } break;
+        case LLAMA_SWA_TYPE_SYMMETRIC:
+            {
+                const int32_t half_n_swa = (int32_t) n_swa / 2;
+                const int32_t pos_diff = p1 - p0;
+
+                // Mask if outside the symmetric window
+                if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
+                    return true;
+                }
+            } break;
+    }
+
+    return false;
+}

package/src/llama.cpp/src/llama-hparams.h

@@ -16,9 +16,10 @@ enum llama_expert_gating_func_type {
 };
 
 enum llama_swa_type {
-    LLAMA_SWA_TYPE_NONE     = 0,
-    LLAMA_SWA_TYPE_STANDARD = 1,
-    LLAMA_SWA_TYPE_CHUNKED  = 2,
+    LLAMA_SWA_TYPE_NONE      = 0,
+    LLAMA_SWA_TYPE_STANDARD  = 1,
+    LLAMA_SWA_TYPE_CHUNKED   = 2,
+    LLAMA_SWA_TYPE_SYMMETRIC = 3,
 };
 
 struct llama_hparams_posnet {
@@ -227,6 +228,11 @@ struct llama_hparams {
 
     // number of layers for which has_kv() returns true
     uint32_t n_layer_kv() const;
+
+    // note that this function uses different SWA parameters from those in the hparams
+    // TODO: think of a better place for this function
+    // TODO: pack the SWA params in a struct?
+    static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");

package/src/llama.cpp/src/llama-kv-cache.cpp

@@ -1393,29 +1393,7 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
 }
 
 bool llama_kv_cache::is_masked_swa(llama_pos p0, llama_pos p1) const {
-    assert(p0 >= 0 && p1 >= 0);
-
-    switch (swa_type) {
-        case LLAMA_SWA_TYPE_NONE:
-            {
-            } break;
-        case LLAMA_SWA_TYPE_STANDARD:
-            {
-                if (p1 - p0 >= (int32_t) n_swa) {
-                    return true;
-                }
-            } break;
-        case LLAMA_SWA_TYPE_CHUNKED:
-            {
-                const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
-
-                if (p0 < pos_chunk_start) {
-                    return true;
-                }
-            } break;
-    }
-
-    return false;
+    return llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1);
 }
 
 void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {

package/src/llama.cpp/src/llama-kv-cache.h

@@ -212,6 +212,7 @@ private:
     // env: LLAMA_KV_CACHE_DEBUG
     int debug = 0;
 
+    // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;

package/src/llama.cpp/src/llama-model.cpp

@@ -1110,7 +1110,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
-                    case 18: type = LLM_TYPE_537M; break;
+                    case 18: type = LLM_TYPE_270M; break;
                     case 26: type = LLM_TYPE_1B; break;
                     case 34: type = LLM_TYPE_4B; break;
                     case 48: type = LLM_TYPE_12B; break;
@@ -1142,6 +1142,26 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GEMMA_EMBEDDING:
+            {
+                hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;
+                hparams.set_swa_pattern(6);
+
+                hparams.causal_attn = false; // embeddings do not use causal attention
+                hparams.rope_freq_base_train_swa  = 10000.0f;
+                hparams.rope_freq_scale_train_swa = 1.0f;
+
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_POOLING_TYPE,                hparams.pooling_type);
+
+                switch (hparams.n_layer) {
+                    case 24: type = LLM_TYPE_0_3B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+                hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -3484,6 +3504,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_GEMMA3:
+            case LLM_ARCH_GEMMA_EMBEDDING:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -11045,6 +11066,137 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
     }
 };
 
+struct llm_build_gemma_embedding_iswa : public llm_graph_context {
+    llm_build_gemma_embedding_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_k;
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+        if (ubatch.token) {
+            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+            cb(inpL, "inp_scaled", -1);
+        }
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        // TODO: support cacheless iSWA embeddings [TAG_NO_CACHE_ISWA]
+        auto * inp_attn = build_attn_inp_kv_iswa();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+            // norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_post_norm", il);
+
+            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = build_norm(sa_out,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = build_norm(cur,
+                    model.layers[il].ffn_post_norm, NULL,
+                    LLM_NORM_RMS, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 // TODO: move up next to build_starcoder
 struct llm_build_starcoder2 : public llm_graph_context {
     llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
@@ -18481,6 +18633,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
+        //case LLM_ARCH_GEMMA_EMBEDDING: // TODO: disabled until the cacheless SWA logic is fixed [TAG_NO_CACHE_ISWA]
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
             {
@@ -18761,6 +18914,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_gemma3n_iswa>(*this, params);
             } break;
+        case LLM_ARCH_GEMMA_EMBEDDING:
+            {
+                llm = std::make_unique<llm_build_gemma_embedding_iswa>(*this, params);
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 llm = std::make_unique<llm_build_starcoder2>(*this, params);
@@ -19161,6 +19318,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GEMMA2:
         case LLM_ARCH_GEMMA3:
         case LLM_ARCH_GEMMA3N:
+        case LLM_ARCH_GEMMA_EMBEDDING:
         case LLM_ARCH_STARCODER2:
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:

package/src/llama.cpp/src/llama-model.h

@@ -39,7 +39,6 @@ enum llm_type {
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
-    LLM_TYPE_537M,
     LLM_TYPE_558M,
     LLM_TYPE_700M,
     LLM_TYPE_770M,