cui-llama.rn 1.6.0 → 1.6.1

package/cpp/ggml.h

@@ -394,8 +394,8 @@ extern "C" {
 
     // precision
     enum lm_ggml_prec {
-        LM_GGML_PREC_DEFAULT,
-        LM_GGML_PREC_F32,
+        LM_GGML_PREC_DEFAULT =  0, // stored as lm_ggml_tensor.op_params, 0 by default
+        LM_GGML_PREC_F32     = 10,
     };
 
     // model file types
@@ -482,6 +482,7 @@ extern "C" {
         LM_GGML_OP_CONV_TRANSPOSE_1D,
         LM_GGML_OP_IM2COL,
         LM_GGML_OP_IM2COL_BACK,
+        LM_GGML_OP_CONV_2D_DW,
         LM_GGML_OP_CONV_TRANSPOSE_2D,
         LM_GGML_OP_POOL_1D,
         LM_GGML_OP_POOL_2D,
@@ -508,17 +509,12 @@ extern "C" {
 
         LM_GGML_OP_UNARY,
 
-        LM_GGML_OP_MAP_UNARY,
-        LM_GGML_OP_MAP_BINARY,
-
-        LM_GGML_OP_MAP_CUSTOM1_F32,
-        LM_GGML_OP_MAP_CUSTOM2_F32,
-        LM_GGML_OP_MAP_CUSTOM3_F32,
-
         LM_GGML_OP_MAP_CUSTOM1,
         LM_GGML_OP_MAP_CUSTOM2,
         LM_GGML_OP_MAP_CUSTOM3,
 
+        LM_GGML_OP_CUSTOM,
+
         LM_GGML_OP_CROSS_ENTROPY_LOSS,
         LM_GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         LM_GGML_OP_OPT_STEP_ADAMW,
@@ -683,6 +679,9 @@ extern "C" {
     LM_GGML_API bool lm_ggml_is_contiguous_1(const struct lm_ggml_tensor * tensor); // contiguous for dims >= 1
     LM_GGML_API bool lm_ggml_is_contiguous_2(const struct lm_ggml_tensor * tensor); // contiguous for dims >= 2
 
+    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
+    LM_GGML_API bool lm_ggml_is_contiguous_channels(const struct lm_ggml_tensor * tensor);
+
     LM_GGML_API bool lm_ggml_are_same_shape (const struct lm_ggml_tensor * t0, const struct lm_ggml_tensor * t1);
     LM_GGML_API bool lm_ggml_are_same_stride(const struct lm_ggml_tensor * t0, const struct lm_ggml_tensor * t1);
 
@@ -1666,7 +1665,7 @@ extern "C" {
             struct lm_ggml_tensor  * a,
             struct lm_ggml_tensor  * b);
 
-    // depthwise
+    // depthwise (via im2col and mul_mat)
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_2d_dw(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,  // convolution kernel
@@ -1678,6 +1677,22 @@ extern "C" {
             int                  d0,  // dilation dimension 0
             int                  d1); // dilation dimension 1
 
+    // Depthwise 2D convolution
+    // may be faster than lm_ggml_conv_2d_dw, but not available in all backends
+    // a:   KW    KH    1    C    convolution kernel
+    // b:   W     H     C    N    input data
+    // res: W_out H_out C    N
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_2d_dw_direct(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,
+            struct lm_ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_transpose_2d_p0(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
@@ -1723,24 +1738,29 @@ extern "C" {
             float                 p0,
             float                 p1);
 
-    // nearest interpolate
+    enum lm_ggml_scale_mode {
+        LM_GGML_SCALE_MODE_NEAREST  = 0,
+        LM_GGML_SCALE_MODE_BILINEAR = 1,
+    };
+
+    // interpolate
     // multiplies ne0 and ne1 by scale factor
-    // used in stable-diffusion
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_upscale(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
-            int                   scale_factor);
+            int                   scale_factor,
+            enum lm_ggml_scale_mode  mode);
 
-    // nearest interpolate
-    // nearest interpolate to specified dimensions
-    // used in tortoise.cpp
+    // interpolate
+    // interpolate scale to specified dimensions
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_upscale_ext(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
-            int                   ne3);
+            int                   ne3,
+            enum lm_ggml_scale_mode  mode);
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_pad(
@@ -1917,83 +1937,6 @@ extern "C" {
 
     // custom operators
 
-    typedef void (*lm_ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*lm_ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*lm_ggml_custom1_op_f32_t)(struct lm_ggml_tensor *, const struct lm_ggml_tensor *);
-    typedef void (*lm_ggml_custom2_op_f32_t)(struct lm_ggml_tensor *, const struct lm_ggml_tensor *, const struct lm_ggml_tensor *);
-    typedef void (*lm_ggml_custom3_op_f32_t)(struct lm_ggml_tensor *, const struct lm_ggml_tensor *, const struct lm_ggml_tensor *, const struct lm_ggml_tensor *);
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_unary_f32(
-            struct lm_ggml_context        * ctx,
-            struct lm_ggml_tensor         * a,
-                   lm_ggml_unary_op_f32_t   fun),
-        "use lm_ggml_map_custom1 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_unary_inplace_f32(
-            struct lm_ggml_context        * ctx,
-            struct lm_ggml_tensor         * a,
-                   lm_ggml_unary_op_f32_t   fun),
-        "use lm_ggml_map_custom1_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_binary_f32(
-            struct lm_ggml_context         * ctx,
-            struct lm_ggml_tensor          * a,
-            struct lm_ggml_tensor          * b,
-                   lm_ggml_binary_op_f32_t   fun),
-        "use lm_ggml_map_custom2 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_binary_inplace_f32(
-            struct lm_ggml_context         * ctx,
-            struct lm_ggml_tensor          * a,
-            struct lm_ggml_tensor          * b,
-                   lm_ggml_binary_op_f32_t   fun),
-        "use lm_ggml_map_custom2_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom1_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-                   lm_ggml_custom1_op_f32_t   fun),
-        "use lm_ggml_map_custom1 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom1_inplace_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-                   lm_ggml_custom1_op_f32_t   fun),
-        "use lm_ggml_map_custom1_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom2_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-                   lm_ggml_custom2_op_f32_t   fun),
-        "use lm_ggml_map_custom2 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom2_inplace_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-                   lm_ggml_custom2_op_f32_t   fun),
-        "use lm_ggml_map_custom2_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom3_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-            struct lm_ggml_tensor           * c,
-                   lm_ggml_custom3_op_f32_t   fun),
-        "use lm_ggml_map_custom3 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom3_inplace_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-            struct lm_ggml_tensor           * c,
-                   lm_ggml_custom3_op_f32_t   fun),
-        "use lm_ggml_map_custom3_inplace instead");
-
-    // custom operators v2
-
     typedef void (*lm_ggml_custom1_op_t)(struct lm_ggml_tensor * dst , const struct lm_ggml_tensor * a, int ith, int nth, void * userdata);
     typedef void (*lm_ggml_custom2_op_t)(struct lm_ggml_tensor * dst , const struct lm_ggml_tensor * a, const struct lm_ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*lm_ggml_custom3_op_t)(struct lm_ggml_tensor * dst , const struct lm_ggml_tensor * a, const struct lm_ggml_tensor * b, const struct lm_ggml_tensor * c, int ith, int nth, void * userdata);
@@ -2049,6 +1992,30 @@ extern "C" {
             int                     n_tasks,
             void                  * userdata);
 
+    typedef void (*lm_ggml_custom_op_t)(struct lm_ggml_tensor * dst , int ith, int nth, void * userdata);
+
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_custom_4d(
+            struct lm_ggml_context * ctx,
+            enum lm_ggml_type        type,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            struct lm_ggml_tensor ** args,
+            int                   n_args,
+            lm_ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_custom_inplace(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,
+            struct lm_ggml_tensor ** args,
+            int                   n_args,
+            lm_ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
     // loss function
 
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_cross_entropy_loss(

package/cpp/json-schema-to-grammar.cpp

@@ -16,6 +16,9 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
+    if (max_items == 0) {
+        return "";
+    }
     if (min_items == 0 && max_items == 1) {
         return item_rule + "?";
     }

package/cpp/llama-arch.cpp

@@ -19,6 +19,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
+    { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -54,6 +55,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK,         "deepseek"         },
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
+    { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -105,6 +107,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -139,6 +142,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
     { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
+    { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -469,6 +474,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NOMIC_BERT_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {
@@ -1102,6 +1125,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_Q_B,           "blk.%d.attn_q_b" },
             { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
             { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_K_B,           "blk.%d.attn_k_b" },
+            { LLM_TENSOR_ATTN_V_B,           "blk.%d.attn_v_b" },
             { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
@@ -1152,6 +1177,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
         },
     },
+    {
+        LLM_ARCH_GLM4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
@@ -1543,23 +1587,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q,                     {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_K,                     {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP_SHEXP,               {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_A,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, LM_GGML_OP_MUL_MAT}},

package/cpp/llama-arch.h

@@ -23,6 +23,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -58,6 +59,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
+    LLM_ARCH_GLM4,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -109,6 +111,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -143,6 +146,8 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -256,6 +261,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_POST_ATTN_NORM,
+    LLM_TENSOR_POST_MLP_NORM,
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -303,6 +310,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_K_B,
+    LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,

package/cpp/llama-batch.cpp

@@ -189,7 +189,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
     LM_GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
@@ -203,6 +203,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
     for (size_t i = 0; i < n_tokens; ++i) {
         ids[i] = i;
     }
+
     if (simple_split) {
         seq.resize(1);
         llama_sbatch_seq & s = seq[0];
@@ -212,6 +213,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         s.length = n_tokens;
         return;
     }
+
     std::sort(ids.begin(), ids.end(),
             [&batch](size_t a, size_t b) {
                 int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
@@ -239,6 +241,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
                 return n_seq_a > n_seq_b;
             }
     );
+
     // init seq
     llama_sbatch_seq * last_seq = nullptr;
 
@@ -262,6 +265,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         seq.push_back(new_seq);
         last_seq = &seq.back();
     }
+
     // keep shared prompts first at the end, then sort by length descending.
     std::sort(seq.begin(), seq.end(),
             [](llama_sbatch_seq & a, llama_sbatch_seq & b) {

package/cpp/llama-batch.h

@@ -70,7 +70,8 @@ struct llama_sbatch {
     // sequence-wise split
     llama_ubatch split_seq(size_t n_ubatch);
 
-    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
 };
 
 // temporary allocate memory for the input batch if needed

package/cpp/llama-chat.cpp

@@ -50,8 +50,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "deepseek3",         LLM_CHAT_TEMPLATE_DEEPSEEK_3        },
     { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
-    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
-    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGLM_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGLM_4         },
     { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
@@ -62,6 +62,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
+    { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -81,7 +82,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     if (tmpl_contains("<|im_start|>")) {
         return tmpl_contains("<|im_sep|>")
             ? LLM_CHAT_TEMPLATE_PHI_4
-            : LLM_CHAT_TEMPLATE_CHATML;
+            : tmpl_contains("<end_of_utterance>")
+                ? LLM_CHAT_TEMPLATE_SMOLVLM // SmolVLM uses <|im_start|> as BOS, but it is NOT chatml
+                : LLM_CHAT_TEMPLATE_CHATML;
     } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
         if (tmpl_contains("[SYSTEM_PROMPT]")) {
             return LLM_CHAT_TEMPLATE_MISTRAL_V7;
@@ -119,8 +122,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         }
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
         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;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -149,9 +156,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_LLAMA_3;
     } else if (tmpl_contains("[gMASK]sop")) {
         // chatglm3-6b
-        return LLM_CHAT_TEMPLATE_CHATGML_3;
-    } else if (tmpl_contains("[gMASK]<sop>")) {
-        return LLM_CHAT_TEMPLATE_CHATGML_4;
+        return LLM_CHAT_TEMPLATE_CHATGLM_3;
     } else if (tmpl_contains(LU8("<用户>"))) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         return LLM_CHAT_TEMPLATE_MINICPM;
@@ -432,7 +437,7 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_3) {
         // chatglm3-6b
         ss << "[gMASK]" << "sop";
         for (auto message : chat) {
@@ -442,14 +447,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
         for (auto message : chat) {
             std::string role(message->role);
             ss << "<|" << role << "|>" << "\n" << message->content;
         }
         if (add_ass) {
-            ss << "<|assistant|>";
+            ss << "<|assistant|>\n";
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         for (auto message : chat) {
@@ -620,7 +625,23 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|header_start|>assistant<|header_end|>\n\n";
         }
-    }  else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SMOLVLM) {
+        // SmolVLM
+        ss << "<|im_start|>"; // uses <|im_start|> as BOS, but the actual content is NOT chatml
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "<end_of_utterance>\n";
+            } else {
+                ss << "Assistant: " << message->content << "<end_of_utterance>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
+    } else {
         // template not supported
         return -1;
     }

package/cpp/llama-chat.h

@@ -29,8 +29,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_DEEPSEEK_3,
     LLM_CHAT_TEMPLATE_COMMAND_R,
     LLM_CHAT_TEMPLATE_LLAMA_3,
-    LLM_CHAT_TEMPLATE_CHATGML_3,
-    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_CHATGLM_3,
+    LLM_CHAT_TEMPLATE_CHATGLM_4,
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
@@ -41,6 +41,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
+    LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };