whispercpp 1.3.5 → 1.3.6

data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp

@@ -2,11 +2,12 @@
 
 llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = hparams.n_embd;
-    const int64_t n_embd_head = hparams.n_embd_head_v;
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;
@@ -16,17 +17,6 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
-    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
-
-    if (ubatch.embd) {
-        // Image input: split main embd and deepstack embds
-        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
-        for (size_t i = 0; i < n_deepstack_layers; i++) {
-            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
-        }
-        inpL = inpL_main;
-    }
-
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -109,7 +99,7 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
                     nullptr,
                     n_expert, n_expert_used,
                     LLM_FFN_SILU, true,
-                    false, 0.0,
+                    hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                     il);
         cb(moe_out, "ffn_moe_out", il);
@@ -120,8 +110,9 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
         cur = build_cvec(cur, il);
         cb(cur, "l_out", il);
 
-        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
-            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+        if (il < (int) n_deepstack_layers) {
+            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
+            cur = ggml_add(ctx0, cur, ds);
             cb(cur, "deepstack_out", il);
         }
 

data/ext/sources/examples/talk-llama/models/qwen3vl.cpp

@@ -2,11 +2,12 @@
 
 llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = hparams.n_embd;
-    const int64_t n_embd_head = hparams.n_embd_head_v;
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;
@@ -16,17 +17,6 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
-    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
-
-    if (ubatch.embd) {
-        // Image input: split main embd and deepstack embds
-        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
-        for (size_t i = 0; i < n_deepstack_layers; i++) {
-            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
-        }
-        inpL = inpL_main;
-    }
-
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -113,8 +103,9 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
         cur = build_cvec(cur, il);
         cb(cur, "l_out", il);
 
-        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
-            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+        if (il < (int) n_deepstack_layers) {
+            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
+            cur = ggml_add(ctx0, cur, ds);
             cb(cur, "deepstack_out", il);
         }
 

data/ext/sources/examples/talk-llama/models/refact.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_refact::llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/rnd1.cpp

@@ -2,10 +2,10 @@
 
 // RND1 is a Qwen3Moe AR model converted to diffusion model.
 llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;
@@ -93,7 +93,7 @@ llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params
                     nullptr,
                     n_expert, n_expert_used,
                     LLM_FFN_SILU, true,
-                    false, 0.0,
+                    hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                     il);
         cb(moe_out, "ffn_moe_out", il);

data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 llm_build_rwkv6_base::llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model) {}

data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include "llama-memory-recurrent.h"
+
 llm_build_rwkv7_base::llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model) {}

data/ext/sources/examples/talk-llama/models/seed-oss.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_seed_oss::llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/smallthinker.cpp

@@ -2,10 +2,10 @@
 
 template <bool iswa>
 llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;
@@ -93,7 +93,7 @@ llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model,
                     nullptr,
                     n_expert, n_expert_used,
                     LLM_FFN_RELU, true,
-                    false, 0.0,
+                    hparams.expert_weights_scale,
                     static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
                     il, probs);
 

data/ext/sources/examples/talk-llama/models/smollm3.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_smollm3::llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/stablelm.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_stablelm::llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/starcoder.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_starcoder::llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/starcoder2.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_starcoder2::llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/step35-iswa.cpp

@@ -0,0 +1,165 @@
+#include "models.h"
+
+llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    ggml_tensor * inp_pos     = build_inp_pos();
+    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) {
+        ggml_tensor * inpSA = inpL;
+
+        const uint32_t n_head_l    = hparams.n_head(il);
+        const uint32_t n_head_kv_l = hparams.n_head_kv(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);
+
+        cur = inpL;
+
+        // dump pre-attn RMSNorm input to pinpoint layer boundary issues
+        cb(cur, "attn_norm_in", il);
+
+        // self-attention
+        {
+            cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
+
+            // Q/K per-head RMSNorm (Step35 q_norm / k_norm)
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+            }
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+            }
+
+            // RoPE (partial rotary factors per layer)
+            const bool is_swa = hparams.is_swa(il);
+            ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il);
+            const int64_t n_rot_l = hparams.n_rot(il);
+            Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, rope_factors,
+                n_rot_l, 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_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow
+            );
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+
+            const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));
+            ggml_tensor * attn_out = build_attn(inp_attn,
+                    nullptr, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(attn_out, "attn_out", il);
+            // head-wise attention gate: sigmoid(g_proj(x)) in torch
+            if (model.layers[il].wqkv_gate) {
+                ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, cur); // [n_head_l, n_tokens]
+                cb(gate, "attn_gate", il);
+
+                gate = ggml_sigmoid(ctx0, gate);
+                cb(gate, "attn_gate_sigmoid", il);
+
+                // reshape + broadcast to [n_embd_head_v, n_head_l, n_tokens]
+                ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens);
+                ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate,       1,          n_head_l, n_tokens);
+                cb(gate_3d, "attn_gate_3d", il);
+
+                attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
+                cb(attn_3d, "attn_gated_3d", il);
+
+                attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens);
+                cb(attn_out, "attn_gated", il);
+            }
+
+            // output projection
+            cur = build_lora_mm(model.layers[il].wo, attn_out);
+            cb(cur, "attn_proj", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            // dense MLP
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   nullptr,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, nullptr,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, nullptr,
+                    nullptr,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE routed experts
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // shared expert MLP (always added on MoE layers in Step35)
+            ggml_tensor * sh_out = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   nullptr, nullptr,
+                    model.layers[il].ffn_gate_shexp, nullptr, nullptr,
+                    model.layers[il].ffn_down_shexp, nullptr, nullptr,
+                    nullptr,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(sh_out, "ffn_shared_out", il);
+
+            cur = ggml_add(ctx0, moe_out, sh_out);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

data/ext/sources/examples/talk-llama/models/t5-dec.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_t5_dec::llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     //const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/t5-enc.cpp

@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_t5_enc::llm_build_t5_enc(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/models/xverse.cpp

@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_xverse::llm_build_xverse(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

data/ext/sources/examples/talk-llama/unicode.cpp

@@ -1,16 +1,10 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "unicode.h"
 #include "unicode-data.h"
 
 #include <algorithm>
 #include <cassert>
-#include <codecvt>
 #include <cstddef>
 #include <cstdint>
-#include <locale>
 #include <map>
 #include <regex>
 #include <stdexcept>
@@ -199,27 +193,6 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
     return map;
 }
 
-static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
-#if defined(__clang__)
-    // disable C++17 deprecation warning for std::codecvt_utf8
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-    return conv.from_bytes(s);
-}
-
 static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
     std::vector<std::string> bpe_encoded_words;
     for (const auto & word : bpe_words) {
@@ -497,49 +470,26 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
     return bpe_offsets;
 }
 
-// use std::wregex to split the text
-static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {
-    std::wregex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
-    std::vector<size_t> bpe_offsets; // store the offset of each word
-    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
-    size_t start = 0;
-    for (auto offset : offsets) {
-        std::wcregex_iterator it(wtext.data() + start, wtext.data() + start + offset, expr);
-        std::wcregex_iterator end;
-
-        int64_t start_idx = 0;
-        while (it != end) {
-            std::wcmatch match = *it;
-            if (match.position() > start_idx) {
-                bpe_offsets.emplace_back(match.position() - start_idx);
-            }
-            bpe_offsets.emplace_back(match.length());
-            start_idx = match.position() + match.length();
-            ++it;
-        }
-
-        if (start_idx < (int64_t) offset) {
-            bpe_offsets.emplace_back(offset - start_idx);
-        }
-        start += offset;
-    }
-
-    return bpe_offsets;
-}
-
-// use std::regex to split the text
-static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
-    std::regex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
+template <typename CharT>
+static std::vector<size_t> unicode_regex_split_stl(const std::basic_string<CharT> & text, const std::basic_string<CharT> & regex, const std::vector<size_t> & offsets) {
+    using BidirIt = typename std::basic_string<CharT>::const_iterator;
+#ifdef _MSC_VER
+    // Bypass bug in MSVC: https://github.com/ggml-org/llama.cpp/issues/17830
+    constexpr auto regex_flags = std::regex_constants::ECMAScript;
+#else
+    constexpr auto regex_flags = std::regex_constants::optimize | std::regex_constants::nosubs;
+#endif
+    std::basic_regex<CharT> expr(regex, regex_flags);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;
     for (auto offset : offsets) {
-        std::cregex_iterator it(text.data() + start, text.data() + start + offset, expr);
-        std::cregex_iterator end;
+        std::regex_iterator<BidirIt> it(text.begin() + start, text.begin() + start + offset, expr);
+        std::regex_iterator<BidirIt> end;
 
         int64_t start_idx = 0;
         while (it != end) {
-            std::cmatch match = *it;
+            std::match_results<BidirIt> match = *it;
             if (match.position() > start_idx) {
                 bpe_offsets.emplace_back(match.position() - start_idx);
             }
@@ -819,6 +769,12 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
     } else if (regex_expr == "\\p{AFMoE_digits}") {
         // AFMOE digit pattern - use custom implementation for proper splitting
         bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
+    } else if (regex_expr == "\\d{1,3}(?=(?:\\d{3})*\\b)") {
+        // tiny_aya digit grouping pattern from tokenizer.json:
+        //   {"type": "Split", "pattern": {"Regex": "\\d{1,3}(?=(?:\\d{3})*\\b)"}, "behavior": "Isolated"}
+        // Splits digits into groups of 3 from the right (e.g., 1234567 -> 1, 234, 567)
+        // TODO: Revisit this regex, in case there are any subtle tokenization differences with the original regex.
+        bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
     }
 
     return bpe_offsets;
@@ -1051,10 +1007,10 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                     break;
                 }
             }
+            const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
 
             if (use_collapsed) {
                 // sanity-check that the original regex does not contain any non-ASCII characters
-                const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
                 for (size_t i = 0; i < cpts_regex.size(); ++i) {
                     if (cpts_regex[i] >= 128) {
                         throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported");
@@ -1110,7 +1066,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
             } else {
                 // no unicode category used, we can use std::wregex directly
-                const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
+                std::wstring wregex_expr(cpts_regex.begin(), cpts_regex.end());
 
                 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
                 std::wstring wtext(cpts.begin(), cpts.end());

data/ext/sources/ggml/CMakeLists.txt

@@ -1,10 +1,10 @@
-cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 5)
+set(GGML_VERSION_PATCH 8)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -228,6 +228,8 @@ option(GGML_WEBGPU_CPU_PROFILE              "ggml: enable WebGPU profiling (CPU)
 option(GGML_WEBGPU_GPU_PROFILE              "ggml: enable WebGPU profiling (GPU)"             OFF)
 option(GGML_WEBGPU_JSPI                     "ggml: use JSPI for WebGPU"                       ON)
 option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
+option(GGML_VIRTGPU                         "ggml: use the VirtGPU/Virglrenderer API Remoting frontend"     OFF)
+option(GGML_VIRTGPU_BACKEND                 "ggml: build the VirtGPU/Virglrenderer API Remoting backend"    OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
 option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
@@ -246,12 +248,14 @@ set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
 set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
                                             "ggml: sycl device architecture")
 
+option(GGML_OPENVINO                        "ggml: use OPENVINO"                              OFF)
+
 option(GGML_OPENCL                          "ggml: use OpenCL"                                OFF)
 option(GGML_OPENCL_PROFILING                "ggml: use OpenCL profiling (increases overhead)" OFF)
 option(GGML_OPENCL_EMBED_KERNELS            "ggml: embed kernels"                             ON)
 option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adreno"          ON)
 set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
-                                            "gmml: OpenCL API version to target")
+                                            "ggml: OpenCL API version to target")
 
 option(GGML_HEXAGON                         "ggml: enable Hexagon backend"                    OFF)
 set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml: quantize group size (32, 64, or 128)")
@@ -320,10 +324,12 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
+    include/ggml-virtgpu.h
     include/ggml-sycl.h
     include/ggml-vulkan.h
     include/ggml-webgpu.h
     include/ggml-zendnn.h
+    include/ggml-openvino.h
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")