@fugood/llama.node 1.4.11 → 1.4.13

package/src/llama.cpp/src/llama.cpp

@@ -140,6 +140,10 @@ enum layer_fraction_t {
 };
 // this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
 
+class llama_params_fit_exception : public std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
 static void llama_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
@@ -181,12 +185,11 @@ static void llama_params_fit_impl(
         }
     }
 
-    int64_t sum_total           = 0;
+    int64_t sum_free            = 0;
     int64_t sum_projected_free  = 0;
     int64_t min_projected_free  = INT64_MAX;
     int64_t sum_projected_used  = 0;
     int64_t sum_projected_model = 0;
-    int64_t sum_projected_ctx   = 0;
 
     if (nd > 1) {
         LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
@@ -197,12 +200,11 @@ static void llama_params_fit_impl(
         const int64_t projected_used = dmd.mb.total();
         const int64_t projected_free = dmd.free - projected_used;
 
-        sum_total           += dmd.total;
+        sum_free            += dmd.free;
         sum_projected_used  += projected_used;
         sum_projected_free  += projected_free;
         min_projected_free   = std::min(min_projected_free, projected_free);
         sum_projected_model += dmd.mb.model;
-        sum_projected_ctx   += dmd.mb.context;
 
         if (nd > 1) {
             LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
@@ -210,10 +212,9 @@ static void llama_params_fit_impl(
                 projected_free >= 0 ? "surplus" : "deficit");
         }
     }
-    assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
-    assert(sum_projected_used >= sum_projected_ctx);
+    assert(sum_free >= 0 && sum_projected_used >= 0);
     LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
-        __func__, sum_projected_used/MiB, sum_total/MiB);
+        __func__, sum_projected_used/MiB, sum_free/MiB);
     if (min_projected_free >= margin) {
         if (nd == 1) {
             LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
@@ -236,9 +237,7 @@ static void llama_params_fit_impl(
                 __func__, margin/MiB, -global_surplus/MiB);
             if (cparams->n_ctx == 0) {
                 if (hp_nct > n_ctx_min) {
-                    const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
-
-                    int64_t memory_reduction = -global_surplus;
+                    int64_t sum_used_target = sum_free - nd*margin_s;
                     if (nd > 1) {
                         // for multiple devices we need to be more conservative in terms of how much context we think can fit:
                         //   - for dense models only whole layers can be assigned to devices
@@ -246,24 +245,34 @@ static void llama_params_fit_impl(
                         //   - on average we expect a waste of 0.5 layers/tensors per device
                         //   - use slightly more than the expected average for nd devices to be safe
                         const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
-                        memory_reduction += (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
+                        sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
                     }
 
-                    uint32_t ctx_reduction = std::min(uint32_t((memory_reduction + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
-                    cparams->n_ctx = hp_nct - ctx_reduction;
-                    cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
-
-                    ctx_reduction = hp_nct - cparams->n_ctx;
-                    memory_reduction = ctx_reduction * bytes_per_ctx;
-                    global_surplus += memory_reduction;
-                    LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                        __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
-                    if (global_surplus >= 0) {
+                    int64_t sum_projected_used_min_ctx = 0;
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    for (const auto & dmd : dmds_min_ctx) {
+                        sum_projected_used_min_ctx += dmd.mb.total();
+                    }
+                    if (sum_used_target > sum_projected_used_min_ctx) {
+                        // linear interpolation between minimum and maximum context size:
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
+                            / (sum_projected_used - sum_projected_used_min_ctx);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
+
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
+                        LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                         if (nd == 1) {
                             LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__);
                             return;
                         }
                         LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__);
+                    } else {
+                        const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
+                        LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
@@ -276,28 +285,28 @@ static void llama_params_fit_impl(
     }
 
     if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
-        throw std::runtime_error("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+        throw llama_params_fit_exception("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
     }
     if (nd > 1) {
         if (!tensor_split) {
-            throw std::runtime_error("did not provide a buffer to write the tensor_split to, abort");
+            throw llama_params_fit_exception("did not provide a buffer to write the tensor_split to, abort");
         }
         if (mparams->tensor_split) {
             for (size_t id = 0; id < nd; id++) {
                 if (mparams->tensor_split[id] != 0.0f) {
-                    throw std::runtime_error("model_params::tensor_split already set by user, abort");
+                    throw llama_params_fit_exception("model_params::tensor_split already set by user, abort");
                 }
             }
         }
         if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
-            throw std::runtime_error("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
+            throw llama_params_fit_exception("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
         }
     }
     if (!tensor_buft_overrides) {
-        throw std::runtime_error("did not provide buffer to set tensor_buft_overrides, abort");
+        throw llama_params_fit_exception("did not provide buffer to set tensor_buft_overrides, abort");
     }
     if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
-        throw std::runtime_error("model_params::tensor_buft_overrides already set by user, abort");
+        throw llama_params_fit_exception("model_params::tensor_buft_overrides already set by user, abort");
     }
 
     // step 3: iteratively fill the back to front with "dense" layers
@@ -350,6 +359,11 @@ static void llama_params_fit_impl(
 
         // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
         layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+
+        uint32_t n_full() const {
+            assert(n_layer >= n_part);
+            return n_layer - n_part;
+        }
     };
 
     const size_t ntbo = llama_max_tensor_buft_overrides();
@@ -373,18 +387,18 @@ static void llama_params_fit_impl(
 
         size_t itbo = 0;
         for (size_t id = 0; id < nd; id++) {
-            il0 += ngl_per_device[id].n_layer - ngl_per_device[id].n_part;
+            il0 += ngl_per_device[id].n_full();
             for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
                 if (itbo + 1 >= ntbo) {
                     tensor_buft_overrides[itbo].pattern = nullptr;
                     tensor_buft_overrides[itbo].buft    = nullptr;
                     itbo++;
                     mparams.tensor_buft_overrides = tensor_buft_overrides;
-                    throw std::runtime_error("llama_params_fit_n_tensor_buft_overrides() == "
-                        + std::to_string(ntbo) + " is insufficient for model\n");
+                    throw llama_params_fit_exception("llama_max_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model");
                 }
                 tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
-                tensor_buft_overrides[itbo].buft = overflow_bufts[id];
+                tensor_buft_overrides[itbo].buft = il == il0 ? overflow_bufts[id] : ggml_backend_cpu_buffer_type();
                 itbo++;
             }
             il0 += ngl_per_device[id].n_part;
@@ -459,20 +473,14 @@ static void llama_params_fit_impl(
         LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
     }
 
-    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the partial layers of a device overflow to:
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the first partial layer of a device overflows to:
     overflow_bufts.reserve(nd);
-    for (size_t id = 0; id < nd - 1; ++id) {
-        overflow_bufts.push_back(ggml_backend_dev_buffer_type(devs[id + 1]));
+    for (size_t id = 0; id < nd; id++) {
+        overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
     }
-    overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
 
     std::vector<ngl_t> ngl_per_device(nd);
     std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts);
-    if (hp_nex > 0) {
-        for (size_t id = 0; id < nd; id++) {
-            ngl_per_device[id].overflow_type = LAYER_FRACTION_MOE;
-        }
-    }
 
     // optimize the number of layers per device using the method of false position:
     //   - ngl_per_device has 0 layers for each device, lower bound
@@ -512,7 +520,8 @@ static void llama_params_fit_impl(
                     std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
                     ngl_per_device_test[id].n_layer += step_size;
                     if (hp_nex) {
-                        ngl_per_device_test[id].n_part += step_size;
+                        ngl_per_device_test[id].n_part += size_t(id) == nd - 1 && ngl_per_device_test[id].n_part == 0 ?
+                            step_size - 1 : step_size; // the first layer is the output layer which must always be full
                     }
                     const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
 
@@ -561,7 +570,7 @@ static void llama_params_fit_impl(
     assert(id_dense_start < nd);
 
     LLAMA_LOG_INFO("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
-    for (size_t id = 0; id <= id_dense_start; id++) {
+    for (size_t id = 0; id <= id_dense_start && id_dense_start < nd; id++) {
         std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
         for (size_t jd = id_dense_start; jd < nd; jd++) {
             const uint32_t n_layer_move = jd < nd - 1 ? ngl_per_device_high[jd].n_layer : ngl_per_device_high[jd].n_layer - 1;
@@ -573,12 +582,8 @@ static void llama_params_fit_impl(
         std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
 
         if (mem_high[id] > targets[id]) {
-            assert(ngl_per_device_high[id].n_layer >= ngl_per_device_high[id].n_part);
-            assert(ngl_per_device[id].n_layer >= ngl_per_device[id].n_part);
-            assert((ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
-                   >= ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
-            uint32_t delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
-                - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+            uint32_t delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
             while (delta > 1) {
                 uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
                 step_size = std::max(step_size, uint32_t(1));
@@ -594,7 +599,7 @@ static void llama_params_fit_impl(
                     ngl_per_device_test[id].n_layer += n_convert_jd;
                     n_converted_test += n_convert_jd;
 
-                    if (ngl_per_device_test[id_dense_start_test].n_layer > 0) {
+                    if (ngl_per_device_test[id_dense_start_test].n_part > 0) {
                         break;
                     }
                 }
@@ -613,8 +618,8 @@ static void llama_params_fit_impl(
                     LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
                         __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
                 }
-                delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
-                    - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+                assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+                delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
             }
         } else {
             ngl_per_device = ngl_per_device_high;
@@ -632,14 +637,19 @@ static void llama_params_fit_impl(
             ngl_per_device_test[id_dense_start_test].n_part--;
             ngl_per_device_test[id].n_layer++;
             ngl_per_device_test[id].n_part++;
-            if (ngl_per_device_test[id_dense_start_test].n_layer == 0) {
+            if (ngl_per_device_test[id_dense_start_test].n_part == 0) {
                 id_dense_start_test++;
             }
             ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            std::vector<ggml_backend_buffer_type_t> overflow_bufts_test = overflow_bufts;
+            if (id < nd - 1) {
+                overflow_bufts_test[id] = ggml_backend_dev_buffer_type(devs[id + 1]);
+            }
             LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
-            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
-            if (mem_test[id] < targets[id]) {
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+            if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
                 ngl_per_device = ngl_per_device_test;
+                overflow_bufts = overflow_bufts_test;
                 mem            = mem_test;
                 id_dense_start = id_dense_start_test;
                 LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
@@ -647,9 +657,10 @@ static void llama_params_fit_impl(
 
                 ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
                 LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
-                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
-                if (mem_test[id] < targets[id]) {
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
                     ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
                     mem            = mem_test;
                     id_dense_start = id_dense_start_test;
                     LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
@@ -658,9 +669,10 @@ static void llama_params_fit_impl(
             } else {
                 ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
                 LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
-                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
-                if (mem_test[id] < targets[id]) {
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
                     ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
                     mem            = mem_test;
                     id_dense_start = id_dense_start_test;
                     LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
@@ -675,30 +687,41 @@ static void llama_params_fit_impl(
             __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
     }
 
+    // print info for devices that were not changed during the conversion from dense only to full layers:
+    for (size_t id = id_dense_start + 1; id < nd; id++) {
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
     set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
 }
 
-bool llama_params_fit(
+enum llama_params_fit_status llama_params_fit(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
         size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
-    bool ok = true;
+    llama_params_fit_status status = LLAMA_PARAMS_FIT_STATUS_SUCCESS;
     try {
         llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
         LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
-    } catch (const std::runtime_error & e) {
+    } catch (const llama_params_fit_exception & e) {
         LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
-        ok = false;
+        status = LLAMA_PARAMS_FIT_STATUS_FAILURE;
+    } catch (const std::runtime_error & e) {
+        LLAMA_LOG_ERROR("%s: encountered an error while trying to fit params to free device memory: %s\n", __func__, e.what());
+        status = LLAMA_PARAMS_FIT_STATUS_ERROR;
     }
     const int64_t t1_us = llama_time_us();
     LLAMA_LOG_INFO("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
-    return ok;
+    return status;
 }
 
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
-        /*.no_perf                     =*/ true,
+        /*.no_perf =*/ true,
     };
 
     return result;

package/src/llama.cpp/src/models/afmoe.cpp

@@ -22,8 +22,15 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
     const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
 
     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);
+
         ggml_tensor * inpSA = inpL;
 
+        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
+        const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                              (il + 1) % hparams.n_no_rope_layer_step != 0;
+
         // dual attention normalization (pre)
         cur = build_norm(inpL,
                 model.layers[il].attn_norm, NULL,
@@ -56,19 +63,16 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
             cb(Qcur, "Qcur_normed", il);
             cb(Kcur, "Kcur_normed", il);
 
-            // RoPE only for sliding_attention layers
-            const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
-                                ((il + 1) % hparams.n_no_rope_layer_step) != 0;
             if (use_rope) {
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur_rope", il);
 
                 Kcur = ggml_rope_ext(
                         ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur_rope", il);
             }

package/src/llama.cpp/src/models/bert.cpp

@@ -142,11 +142,13 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
                     LLM_FFN_GELU, LLM_FFN_SEQ, il);
             cb(cur, "ffn_out", il);
         } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+            const bool up_contains_gate = !model.layers[il].ffn_gate && model.layers[il].ffn_up->ne[1] != hparams.n_ff();
+            auto type_op = up_contains_gate ? LLM_FFN_GEGLU : LLM_FFN_GELU;
             cur = build_ffn(cur,
-                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
                     model.layers[il].ffn_gate, NULL, NULL,
                     model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
-                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
+                    type_op, LLM_FFN_PAR, il);
             cb(cur, "ffn_out", il);
         } else {
             cur = build_ffn(cur,

package/src/llama.cpp/src/models/cogvlm.cpp

@@ -3,12 +3,14 @@
 llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
-    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+    const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
     GGML_ASSERT(n_embd_head == hparams.n_rot);
 
-    ggml_tensor *inpL, *cur;
+    ggml_tensor * inpL;
+    ggml_tensor * cur;
+
     inpL = build_inp_embd(model.tok_embd);
 
     ggml_tensor * inp_pos = build_inp_pos();
@@ -44,7 +46,7 @@ llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_pa
         }
 
         ggml_tensor * inpSA = inpL;
-        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cur = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 
         // build self attention
         {

package/src/llama.cpp/src/models/cohere2-iswa.cpp

@@ -21,6 +21,9 @@ llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const
 
     for (int il = 0; il < n_layer; ++il) {
         const bool is_swa = hparams.is_swa(il);
+        // UNUSED:
+        // 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, il);

package/src/llama.cpp/src/models/deepseek2.cpp

@@ -215,7 +215,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 model.layers[il].ffn_exp_probs_b,
                 n_expert, n_expert_used,
                 LLM_FFN_SILU, hparams.expert_weights_norm,
-                true, hparams.expert_weights_scale,
+                hparams.expert_weights_scale, hparams.expert_weights_scale,
                 (llama_expert_gating_func_type) hparams.expert_gating_func,
                 il);
             cb(moe_out, "ffn_moe_out", il);

package/src/llama.cpp/src/models/gemma-embedding.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k;
@@ -12,10 +10,8 @@ llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model,
     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);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
 
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();

package/src/llama.cpp/src/models/gemma2-iswa.cpp

@@ -19,6 +19,9 @@ llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const ll
     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,
@@ -43,12 +46,12 @@ llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const ll
 
             Qcur = ggml_rope_ext(
                     ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                     ext_factor, attn_factor, beta_fast, beta_slow);
 
             Kcur = ggml_rope_ext(
                     ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    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);

package/src/llama.cpp/src/models/gemma3.cpp

@@ -10,10 +10,9 @@ llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_gr
     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);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 

package/src/llama.cpp/src/models/gemma3n-iswa.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model),
@@ -15,10 +13,9 @@ llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const
     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);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -248,7 +245,7 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_altup, n_layer, n_tokens]
 ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
-    auto          inp = std::make_unique<llm_graph_input_embd>();
+    auto inp = std::make_unique<llm_graph_input_embd>();
     ggml_tensor * inp_per_layer;
     if (ubatch.token) {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);

package/src/llama.cpp/src/models/llama-iswa.cpp

@@ -25,8 +25,12 @@ llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_
     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);
+
         ggml_tensor * inpSA = inpL;
 
+        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
         const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
                               (il + 1) % hparams.n_no_rope_layer_step != 0;
 
@@ -67,13 +71,13 @@ llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_
             if (use_rope) {
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         ext_factor, attn_factor, beta_fast, beta_slow
                         );
 
                 Kcur = ggml_rope_ext(
                         ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         ext_factor, attn_factor, beta_fast, beta_slow
                         );
             } else if (inp_attn_scale) {

package/src/llama.cpp/src/models/llama.cpp

@@ -1,6 +1,7 @@
 #include "models.h"
 
-llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+template <bool embed>
+llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -14,7 +15,14 @@ llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_para
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
-    auto * inp_attn = build_attn_inp_kv();
+    using inp_attn_type = std::conditional_t<embed, llm_graph_input_attn_no_cache, llm_graph_input_attn_kv>;
+
+    inp_attn_type * inp_attn = nullptr;
+    if constexpr (embed) {
+        inp_attn = build_attn_inp_no_cache();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
 
     const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -145,11 +153,16 @@ llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_para
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
-    // lm_head
-    cur = build_lora_mm(model.output, cur);
+    if constexpr (!embed) {
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
 
-    cb(cur, "result_output", -1);
-    res->t_logits = cur;
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+    }
 
     ggml_build_forward_expand(gf, cur);
 }
+
+template struct llm_build_llama<false>;
+template struct llm_build_llama<true>;