whispercpp 1.3.5 → 1.3.6

data/ext/sources/examples/talk-llama/llama-arch.h

@@ -4,6 +4,7 @@
 
 #include <string>
 #include <set>
+#include <vector>
 
 //
 // gguf constants (sync with gguf.py)
@@ -30,6 +31,7 @@ enum llm_arch {
     LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_JINA_BERT_V3,
+    LLM_ARCH_EUROBERT,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
@@ -41,6 +43,8 @@ enum llm_arch {
     LLM_ARCH_QWEN3NEXT,
     LLM_ARCH_QWEN3VL,
     LLM_ARCH_QWEN3VLMOE,
+    LLM_ARCH_QWEN35,
+    LLM_ARCH_QWEN35MOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -76,15 +80,18 @@ enum llm_arch {
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
     LLM_ARCH_GLM4_MOE,
+    LLM_ARCH_GLM_DSA,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
+    LLM_ARCH_JAIS2,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
     LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
+    LLM_ARCH_EXAONE_MOE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
     LLM_ARCH_RWKV7,
@@ -120,9 +127,12 @@ enum llm_arch {
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_MISTRAL3,
+    LLM_ARCH_PADDLEOCR,
     LLM_ARCH_MIMO2,
+    LLM_ARCH_STEP35,
     LLM_ARCH_LLAMA_EMBED,
     LLM_ARCH_MAINCODER,
+    LLM_ARCH_KIMI_LINEAR,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -164,6 +174,8 @@ enum llm_kv {
     LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,
+    LLM_KV_SWIGLU_CLAMP_EXP,
+    LLM_KV_SWIGLU_CLAMP_SHEXP,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
     LLM_KV_EXPERT_COUNT,
@@ -177,6 +189,7 @@ enum llm_kv {
     LLM_KV_EXPERT_GROUP_SCALE,
     LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
+    LLM_KV_MOE_LATENT_SIZE,
     LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_NUM_DEEPSTACK_LAYERS,
     LLM_KV_POOLING_TYPE,
@@ -194,6 +207,7 @@ enum llm_kv {
     LLM_KV_EMBEDDING_SCALE,
     LLM_KV_TOKEN_SHIFT_COUNT,
     LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
+    LLM_KV_FULL_ATTENTION_INTERVAL,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -221,8 +235,14 @@ enum llm_kv {
     LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_KEY_LENGTH_SWA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
+    LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
+    LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
+    LLM_KV_ATTENTION_INDEXER_TOP_K,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_COUNT_SWA,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
     LLM_KV_ROPE_FREQ_BASE_SWA,
@@ -249,6 +269,8 @@ enum llm_kv {
     LLM_KV_SSM_GROUP_COUNT,
     LLM_KV_SSM_DT_B_C_RMS,
 
+    LLM_KV_KDA_HEAD_DIM,
+
     LLM_KV_WKV_HEAD_SIZE,
 
     LLM_KV_TOKENIZER_MODEL,
@@ -356,6 +378,7 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
     LLM_TENSOR_FFN_GATE_EXPS,
     LLM_TENSOR_FFN_UP_EXPS,
+    LLM_TENSOR_FFN_GATE_UP_EXPS,
     LLM_TENSOR_FFN_DOWN_SHEXP,
     LLM_TENSOR_FFN_GATE_SHEXP,
     LLM_TENSOR_FFN_UP_SHEXP,
@@ -363,6 +386,8 @@ enum llm_tensor {
     LLM_TENSOR_FFN_GATE_CHEXPS,
     LLM_TENSOR_FFN_UP_CHEXPS,
     LLM_TENSOR_FFN_EXP_PROBS_B,
+    LLM_TENSOR_FFN_LATENT_DOWN,
+    LLM_TENSOR_FFN_LATENT_UP,
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,
@@ -397,6 +422,16 @@ enum llm_tensor {
     LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
     LLM_TENSOR_SSM_BETA_ALPHA,      // qwen3next
+    LLM_TENSOR_SSM_ALPHA,           // qwen3.5
+    // Kimi Linear KDA (using SSM_ prefix for consistency)
+    LLM_TENSOR_SSM_CONV1D_Q,        // kimi: Q conv1d weight
+    LLM_TENSOR_SSM_CONV1D_K,        // kimi: K conv1d weight
+    LLM_TENSOR_SSM_CONV1D_V,        // kimi: V conv1d weight
+    LLM_TENSOR_SSM_F_A,             // kimi: forget gate projection A
+    LLM_TENSOR_SSM_F_B,             // kimi: forget gate projection B
+    LLM_TENSOR_SSM_BETA,            // kimi: beta mixing coefficient and qwen3.5
+    LLM_TENSOR_SSM_G_A,             // kimi: output gate projection A
+    LLM_TENSOR_SSM_G_B,             // kimi: output gate projection B
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
@@ -473,6 +508,7 @@ enum llm_tensor {
     LLM_TENSOR_ENC_OUTPUT_NORM,
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_CLS_NORM,
     LLM_TENSOR_CONV1D,
     LLM_TENSOR_CONVNEXT_DW,
     LLM_TENSOR_CONVNEXT_NORM,
@@ -497,6 +533,10 @@ enum llm_tensor {
     LLM_TENSOR_VISEXP_FFN_GATE,
     LLM_TENSOR_VISEXP_FFN_DOWN,
     LLM_TENSOR_VISEXP_FFN_UP,
+    LLM_TENSOR_INDEXER_K_NORM,
+    LLM_TENSOR_INDEXER_PROJ,
+    LLM_TENSOR_INDEXER_ATTN_K,
+    LLM_TENSOR_INDEXER_ATTN_Q_B,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,
@@ -575,6 +615,8 @@ struct llm_tensor_info {
     ggml_op op;
 };
 
+std::vector<llm_arch> llm_arch_all();
+
 const char * llm_arch_name(llm_arch arch);
 
 llm_arch llm_arch_from_string(const std::string & name);

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

@@ -394,11 +394,13 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
     clear();
     split_reset();
 
+    const int64_t n_pos_all = (int64_t) n_tokens*n_pos_per_embd;
+
     auto udata = std::make_shared<llama_ubatch::data_t>();
 
     udata->token     .resize(n_tokens);
     udata->embd      .clear();
-    udata->pos       .resize(n_tokens);
+    udata->pos       .resize(n_pos_all);
     udata->n_seq_id  .resize(n_tokens);
     udata->seq_id    .resize(n_tokens);
     udata->seq_id_unq.resize(0);

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

@@ -57,6 +57,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "exaone4",           LLM_CHAT_TEMPLATE_EXAONE_4          },
+    { "exaone-moe",        LLM_CHAT_TEMPLATE_EXAONE_MOE        },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
@@ -137,6 +138,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains("[gMASK]<sop>")) {
         return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
+        if (tmpl_contains("<|tool_declare|>")) {
+            return LLM_CHAT_TEMPLATE_EXAONE_MOE;
+        }
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
         return LLM_CHAT_TEMPLATE_GLMEDGE;
@@ -229,7 +233,7 @@ int32_t llm_chat_apply_template(
     llm_chat_template tmpl,
     const std::vector<const llama_chat_message *> & chat,
     std::string & dest, bool add_ass) {
-    // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
+    // Taken from the research: https://github.com/ggml-org/llama.cpp/issues/5527
     std::stringstream ss;
     if (tmpl == LLM_CHAT_TEMPLATE_CHATML) {
         // chatml template
@@ -576,6 +580,22 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[|assistant|]";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_MOE) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|system|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "user") {
+                ss << "<|user|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "assistant") {
+                ss << "<|assistant|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "tool") {
+                ss << "<|tool|>\n" << trim(message->content) << "<|endofturn|>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
         // this template requires the model to have "\n\n" as EOT token
         for (size_t i = 0; i < chat.size(); i++) {

data/ext/sources/examples/talk-llama/llama-chat.h

@@ -36,6 +36,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_EXAONE_4,
+    LLM_CHAT_TEMPLATE_EXAONE_MOE,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,