llama: update to commit 2016f07b (#10352)

Makefile.sync

 UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=71e90e8813f90097701e62f7fce137d96ddf41e2
+FETCH_HEAD=2016f07bd106c73699ecbaace80f55db5ed95dac
 
 .PHONY: help
 help:

llama/build-info.cpp

 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "71e90e8813f90097701e62f7fce137d96ddf41e2";
+char const *LLAMA_COMMIT = "2016f07bd106c73699ecbaace80f55db5ed95dac";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/examples/llava/clip-impl.h

@@ -50,7 +50,6 @@
 // tensor name constants
 //
 
-#define TN_TOKEN_EMBD      "%s.token_embd.weight"
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
 #define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
@@ -66,8 +65,6 @@
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
-#define TN_TEXT_PROJ       "text_projection.weight"
-#define TN_VIS_PROJ        "visual_projection.weight"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"

llama/llama.cpp/src/llama-arch.cpp

@@ -145,6 +145,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
     { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,       "%s.attention.cross_attention_layers"       },
+    { 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"              },
@@ -1142,6 +1144,8 @@ 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" },
@@ -1636,23 +1640,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_K,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP_SHEXP,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_A,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},

llama/llama.cpp/src/llama-arch.h

@@ -149,6 +149,8 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
     LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -311,6 +313,8 @@ 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,

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

@@ -10,6 +10,7 @@
 #include <cstring>
 #include <stdexcept>
 #include <cinttypes>
+#include <cmath>
 
 //
 // llama_context
@@ -473,7 +474,6 @@ ggml_tensor * llama_context::build_rope_shift(
     const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
 
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
-    const auto & yarn_attn_factor = cparams.yarn_attn_factor;
     const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
 
@@ -482,6 +482,10 @@ ggml_tensor * llama_context::build_rope_shift(
     const auto & n_rot     = hparams.n_rot;
     const auto & rope_type = hparams.rope_type;
 
+    // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
+    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+    const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;
+
     ggml_tensor * tmp;
 
     if (ggml_is_quantized(cur->type)) {

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

@@ -1194,6 +1194,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_tensor * v,
          ggml_tensor * kq_b,
          ggml_tensor * kq_mask,
+         ggml_tensor * v_mla,
              bool      v_trans,
              float     kq_scale) const {
   //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
@@ -1205,8 +1206,6 @@ ggml_tensor * llm_graph_context::build_attn_mha(
   //const auto & n_embd_head_k = hparams.n_embd_head_k;
   //const auto & n_embd_head_v = hparams.n_embd_head_v;
 
-    const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0];
-
     const auto n_tokens = q->ne[1];
     const auto n_head   = q->ne[2];
     const auto n_kv     = k->ne[1];
@@ -1235,7 +1234,12 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
-        cur = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens);
+        if (v_mla) {
+            cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
+            cur = ggml_mul_mat(ctx0, v_mla, cur);
+        }
+
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
 
@@ -1273,9 +1277,14 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
 
-        ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+        // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
+        if (v_mla) {
+            kqv = ggml_mul_mat(ctx0, v_mla, kqv);
+        }
+
+        cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
 
-        cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
 
         if (!cparams.offload_kqv) {
             // all nodes between the KV store and the attention output are run on the CPU
@@ -1310,6 +1319,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     GGML_UNUSED(n_tokens);
@@ -1331,7 +1341,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
     //cb(k, "v", il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
 
     cb(cur, "kqv_out", il);
 
@@ -1385,6 +1395,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
@@ -1470,7 +1481,7 @@ ggml_tensor * llm_graph_context::build_attn(
                 ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v,
                 0);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, v_trans, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1529,6 +1540,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
@@ -1548,7 +1560,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
     //cb(k, "v", il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
 
     cb(cur, "kqv_out", il);
 
@@ -1717,4 +1729,3 @@ void llm_graph_context::build_pooling(
 
     ggml_build_forward_expand(gf, cur);
 }
-

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

@@ -522,6 +522,7 @@ struct llm_graph_context {
              ggml_tensor * v,     // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
              ggml_tensor * kq_b,
              ggml_tensor * kq_mask,
+             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                     bool   v_trans,
                    float   kq_scale) const;
 
@@ -536,6 +537,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
@@ -550,6 +552,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
@@ -564,6 +567,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 

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

@@ -46,6 +46,10 @@ struct llama_hparams {
     uint32_t n_rel_attn_bkts = 0;
     uint32_t n_vocab = 0;
 
+    // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+    uint32_t n_embd_head_k_mla = 0;
+    uint32_t n_embd_head_v_mla = 0;
+
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;

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

@@ -27,7 +27,7 @@ bool llama_kv_cache_unified::init(
 
     recurrent = llama_model_is_recurrent(&model);
     v_trans   = !recurrent && !cparams.flash_attn;
-    can_shift = !recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+    can_shift = !recurrent;
 
     LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n",
             __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift);

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

@@ -174,6 +174,8 @@ struct llama_layer {
     struct ggml_tensor * wq_b      = nullptr;
     struct ggml_tensor * wkv_a_mqa = nullptr;
     struct ggml_tensor * wkv_b     = nullptr;
+    struct ggml_tensor * wk_b      = nullptr;
+    struct ggml_tensor * wv_b      = nullptr;
     struct ggml_tensor * wq_cross  = nullptr;
     struct ggml_tensor * wk_cross  = nullptr;
     struct ggml_tensor * wv_cross  = nullptr;

llama/llama.cpp/src/llama-vocab.cpp

@@ -1833,6 +1833,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_prefix|>"  // Qwen
                         || t.first == "<fim-prefix>"
+                        || t.first == "<fim_prefix>"    // Granite
                         || t.first == "<｜fim▁begin｜>" // DeepSeek
                         || t.first == "<PRE>"
                         || t.first == "▁<PRE>"          // CodeLlama
@@ -1851,6 +1852,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_suffix|>" // Qwen
                         || t.first == "<fim-suffix>"
+                        || t.first == "<fim_suffix>"   // Granite
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == "<SUF>"
                         || t.first == "▁<SUF>"         // CodeLlama
@@ -1869,6 +1871,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_middle|>" // Qwen
                         || t.first == "<fim-middle>"
+                        || t.first == "<fim_middle>"   // Granite
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == "<MID>"
                         || t.first == "▁<MID>"         // CodeLlama
@@ -1887,6 +1890,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_pad|>" // Qwen
                         || t.first == "<fim-pad>"
+                        || t.first == "<fim_pad>"   // Granite
                         || t.first == "<PAD>"
                         ) {
                     special_fim_pad_id = t.second;
@@ -1905,6 +1909,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|repo_name|>"
                         || t.first == "<fim-repo>"
                         || t.first == "<REPO>"
+                        || t.first == "<reponame>"    // Granite
                         ) {
                     special_fim_rep_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch

@@ -65,10 +65,10 @@ index 273075f4..dd11f304 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index cec36b36..4b057973 100644
+index e2617b06..242e50a7 100644
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
-@@ -530,6 +530,7 @@ static void ggml_backend_cann_buffer_free_buffer(
+@@ -800,6 +800,7 @@ static void ggml_backend_cann_buffer_free_buffer(
      ggml_backend_cann_buffer_context* ctx =
          (ggml_backend_cann_buffer_context*)buffer->context;
      delete ctx;
@@ -76,7 +76,7 @@ index cec36b36..4b057973 100644
  }
  
  /**
-@@ -1199,6 +1200,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
+@@ -1472,6 +1473,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
   */
  static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
      ACL_CHECK(aclrtFreeHost(buffer->context));
@@ -85,10 +85,10 @@ index cec36b36..4b057973 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index fafe9633..59a49560 100644
+index a7febef7..31750b6f 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -533,6 +533,7 @@ struct ggml_backend_cuda_buffer_context {
+@@ -534,6 +534,7 @@ struct ggml_backend_cuda_buffer_context {
  static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
      delete ctx;
@@ -96,7 +96,7 @@ index fafe9633..59a49560 100644
  }
  
  static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-@@ -788,6 +789,7 @@ struct ggml_backend_cuda_split_buffer_context {
+@@ -789,6 +790,7 @@ struct ggml_backend_cuda_split_buffer_context {
  static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
      delete ctx;
@@ -104,7 +104,7 @@ index fafe9633..59a49560 100644
  }
  
  static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1061,6 +1063,7 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
+@@ -1062,6 +1064,7 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
  
  static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      CUDA_CHECK(cudaFreeHost(buffer->context));
@@ -125,10 +125,10 @@ index 50579227..2799a0a5 100644
  
  static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 9f1c6c6c..310afe8a 100644
+index 266d8af4..12886cd3 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -4641,6 +4641,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -4759,6 +4759,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
      }
  
      free(ctx);
@@ -137,10 +137,10 @@ index 9f1c6c6c..310afe8a 100644
  
  static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
-index b8b5cbd3..14d4561b 100644
+index 05a2f4e6..392cc18d 100644
 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp
 +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
-@@ -1443,6 +1443,7 @@ struct ggml_backend_opencl_buffer_context {
+@@ -1940,6 +1940,7 @@ struct ggml_backend_opencl_buffer_context {
  static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
      delete ctx;
@@ -149,10 +149,10 @@ index b8b5cbd3..14d4561b 100644
  
  static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
-index 862b9b66..34536681 100644
+index a0667b7d..bd83adc5 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -443,6 +443,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -468,6 +468,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
      GGML_ASSERT(status);
      delete ctx;
@@ -161,7 +161,7 @@ index 862b9b66..34536681 100644
  
  static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index 3e48a924..a3d182fc 100644
+index 1de34c96..4600f61e 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
 @@ -316,6 +316,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
@@ -189,10 +189,10 @@ index 3e48a924..a3d182fc 100644
  
  static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 783a0ff8..8ac1e07e 100644
+index 39f3cd34..c569a8a5 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -8639,6 +8639,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -8653,6 +8653,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
      ggml_vk_destroy_buffer(ctx->dev_buffer);
      delete ctx;
@@ -200,7 +200,7 @@ index 783a0ff8..8ac1e07e 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -8782,6 +8783,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -8796,6 +8797,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
  static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
      ggml_vk_host_free(vk_instance.devices[0], buffer->context);

llama/patches/0002-pretokenizer.patch

@@ -10,7 +10,7 @@ logs instead of throwing an error
  1 file changed, 3 insertions(+), 11 deletions(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 464ff01e..0125ee53 100644
+index 48060517..a35b498c 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1491,16 +1491,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {

llama/patches/0003-embeddings.patch

@@ -11,10 +11,10 @@ instead of forcing one or the error
  1 file changed, 3 insertions(+), 3 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 4735e98e..65135172 100644
+index 983385f8..32f59819 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -1232,7 +1232,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1236,7 +1236,7 @@ int llama_context::decode(llama_batch & inp_batch) {
      int64_t n_outputs_all = 0;
  
      // count outputs
@@ -23,7 +23,7 @@ index 4735e98e..65135172 100644
          for (uint32_t i = 0; i < n_tokens_all; ++i) {
              n_outputs_all += batch.logits[i] != 0;
          }
-@@ -1344,7 +1344,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1348,7 +1348,7 @@ int llama_context::decode(llama_batch & inp_batch) {
          //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
          //}
  
@@ -32,7 +32,7 @@ index 4735e98e..65135172 100644
          auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
  
          if (t_embd && res->get_embd_pooled()) {
-@@ -1488,7 +1488,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
+@@ -1492,7 +1492,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead

llama/patches/0004-clip-unicode.patch

@@ -10,12 +10,12 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 49c90b75..4b72ea9f 100644
+index 75970615..d57b4bd6 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -28,6 +28,19 @@
- #include <cinttypes>
+@@ -29,6 +29,19 @@
  #include <limits>
+ #include <array>
  
 +#if defined(_WIN32)
 +#define WIN32_LEAN_AND_MEAN
@@ -33,7 +33,7 @@ index 49c90b75..4b72ea9f 100644
  struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
  
  //#define CLIP_DEBUG_FUNCTIONS
-@@ -1429,7 +1442,29 @@ struct clip_model_loader {
+@@ -1430,7 +1443,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 49c90b75..4b72ea9f 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -1456,7 +1491,11 @@ struct clip_model_loader {
+@@ -1457,7 +1492,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0005-solar-pro.patch

 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
 From: jmorganca <jmorganca@gmail.com>
-Date: Tue, 8 Apr 2025 16:03:51 -0700
+Date: Sun, 20 Apr 2025 16:11:09 -0700
 Subject: [PATCH] solar-pro
 
 adds support for the Solar Pro architecture
@@ -15,7 +15,7 @@ adds support for the Solar Pro architecture
  7 files changed, 248 insertions(+)
 
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index a6fddc7f..0b0fedcd 100644
+index 62e1480b..f754bc8f 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
 @@ -68,6 +68,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
@@ -31,10 +31,10 @@ index a6fddc7f..0b0fedcd 100644
      { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
 +    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
+     { 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"              },
-@@ -1478,6 +1480,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -1482,6 +1484,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
          },
      },
@@ -59,7 +59,7 @@ index a6fddc7f..0b0fedcd 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -1671,6 +1691,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -1660,6 +1680,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
@@ -68,7 +68,7 @@ index a6fddc7f..0b0fedcd 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 2c2099b3..74aa3dd0 100644
+index 98ca00a1..439aaeab 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
 @@ -72,6 +72,7 @@ enum llm_arch {
@@ -84,10 +84,10 @@ index 2c2099b3..74aa3dd0 100644
      LLM_KV_ATTENTION_SLIDING_WINDOW,
      LLM_KV_ATTENTION_SCALE,
 +    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-     LLM_KV_ROPE_DIMENSION_COUNT,
-     LLM_KV_ROPE_DIMENSION_SECTIONS,
-@@ -340,6 +342,7 @@ enum llm_tensor {
+@@ -344,6 +346,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -115,10 +115,10 @@ index 90dfe7a7..8a667960 100644
      if (il < n_layer) {
          return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index 4e0b5719..c3147cbc 100644
+index 80fcd65d..6e278945 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
-@@ -51,6 +51,8 @@ struct llama_hparams {
+@@ -55,6 +55,8 @@ struct llama_hparams {
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
  
@@ -127,7 +127,7 @@ index 4e0b5719..c3147cbc 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -149,6 +151,9 @@ struct llama_hparams {
+@@ -153,6 +155,9 @@ struct llama_hparams {
      // dimension of the recurrent state embeddings
      uint32_t n_embd_v_s() const;
  
@@ -150,10 +150,10 @@ index ea73a8a7..a012aeae 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index b74dd72c..5fbd0055 100644
+index 6b7bfecf..aba42819 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1372,6 +1372,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1374,6 +1374,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -175,7 +175,7 @@ index b74dd72c..5fbd0055 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -3701,6 +3716,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -3717,6 +3732,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
  
                          layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
  
@@ -210,7 +210,7 @@ index b74dd72c..5fbd0055 100644
                          layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                          layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
                          layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-@@ -12244,6 +12287,165 @@ struct llm_build_chameleon : public llm_graph_context {
+@@ -12296,6 +12339,165 @@ struct llm_build_chameleon : public llm_graph_context {
      }
  };
  
@@ -316,7 +316,7 @@ index b74dd72c..5fbd0055 100644
 +
 +                cur = build_attn(inp_attn, gf,
 +                        model.layers[il].wo, model.layers[il].bo,
-+                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
++                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
 +                cb(cur, "attn_out", il);
 +            }
 +
@@ -376,7 +376,7 @@ index b74dd72c..5fbd0055 100644
  struct llm_build_wavtokenizer_dec : public llm_graph_context {
      llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
          ggml_tensor * cur;
-@@ -12993,6 +13195,10 @@ llm_graph_result_ptr llama_model::build_graph(
+@@ -13045,6 +13247,10 @@ llm_graph_result_ptr llama_model::build_graph(
              {
                  llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
              } break;
@@ -387,7 +387,7 @@ index b74dd72c..5fbd0055 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params, gf);
-@@ -13139,6 +13345,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13191,6 +13397,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE:
          case LLM_ARCH_GRANITE_MOE:
          case LLM_ARCH_CHAMELEON:
@@ -396,7 +396,7 @@ index b74dd72c..5fbd0055 100644
              return LLAMA_ROPE_TYPE_NORM;
  
 diff --git a/src/llama-model.h b/src/llama-model.h
-index 0f18dac1..e08d4ae4 100644
+index fd82d106..5865d5e9 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
 @@ -62,6 +62,7 @@ enum llm_type {
@@ -407,7 +407,7 @@ index 0f18dac1..e08d4ae4 100644
      LLM_TYPE_30B,
      LLM_TYPE_32B,
      LLM_TYPE_34B,
-@@ -305,6 +306,8 @@ struct llama_layer {
+@@ -307,6 +308,8 @@ struct llama_layer {
      struct ggml_tensor * ffn_up_scale   = nullptr;
      struct ggml_tensor * ffn_down_scale = nullptr;
  

llama/patches/0007-add-mllama-support.patch → llama/patches/0006-add-mllama-support.patch

 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
 From: jmorganca <jmorganca@gmail.com>
-Date: Tue, 8 Apr 2025 19:27:12 -0700
+Date: Sun, 20 Apr 2025 16:12:36 -0700
 Subject: [PATCH] add mllama support
 
 adds support for the llama 3.2 vision architecture
@@ -28,7 +28,7 @@ adds support for the llama 3.2 vision architecture
  20 files changed, 475 insertions(+), 22 deletions(-)
 
 diff --git a/examples/llava/gemma3-cli.cpp b/examples/llava/gemma3-cli.cpp
-index 91a07e2a..13127c7b 100644
+index 3d566475..654d1358 100644
 --- a/examples/llava/gemma3-cli.cpp
 +++ b/examples/llava/gemma3-cli.cpp
 @@ -106,7 +106,7 @@ struct decode_embd_batch {
@@ -79,10 +79,10 @@ index 03a22cbb..5eb40bcd 100644
              LOG_ERR("%s : failed to eval\n", __func__);
              return false;
 diff --git a/examples/llava/mtmd.cpp b/examples/llava/mtmd.cpp
-index 114c274b..a0e649ad 100644
+index 3fd5bebc..f0cec596 100644
 --- a/examples/llava/mtmd.cpp
 +++ b/examples/llava/mtmd.cpp
-@@ -213,7 +213,7 @@ struct decode_embd_batch {
+@@ -233,7 +233,7 @@ struct decode_embd_batch {
      std::vector<llama_seq_id *> seq_ids;
      std::vector<int8_t>         logits;
      llama_batch batch;
@@ -91,7 +91,7 @@ index 114c274b..a0e649ad 100644
          pos     .resize(n_tokens);
          n_seq_id.resize(n_tokens);
          seq_ids .resize(n_tokens + 1);
-@@ -225,6 +225,7 @@ struct decode_embd_batch {
+@@ -245,6 +245,7 @@ struct decode_embd_batch {
              /*n_tokens       =*/ n_tokens,
              /*tokens         =*/ nullptr,
              /*embd           =*/ embd,
@@ -99,9 +99,9 @@ index 114c274b..a0e649ad 100644
              /*pos            =*/ pos.data(),
              /*n_seq_id       =*/ n_seq_id.data(),
              /*seq_id         =*/ seq_ids.data(),
-@@ -291,7 +292,8 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
+@@ -311,7 +312,8 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
  
-             int32_t n_tokens = chunk.tokens_image->n_tokens();
+             int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
              float * embd = mtmd_get_output_embd(ctx);
 -            decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
 +            int n_embd  = llama_model_n_embd(llama_get_model(lctx));
@@ -158,7 +158,7 @@ index 5657fbf0..f91896e4 100644
      LLAMA_API void llama_free(struct llama_context * ctx);
  
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 0b0fedcd..c1f78618 100644
+index f754bc8f..0568565f 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
 @@ -6,6 +6,7 @@
@@ -174,10 +174,10 @@ index 0b0fedcd..c1f78618 100644
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
      { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
 +    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,       "%s.attention.cross_attention_layers"       },
+     { 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"              },
-@@ -269,6 +271,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -271,6 +273,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
          },
      },
@@ -218,7 +218,7 @@ index 0b0fedcd..c1f78618 100644
      {
          LLM_ARCH_DECI,
          {
-@@ -1692,6 +1728,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -1681,6 +1717,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
      {LLM_TENSOR_BSKCN_TV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -234,7 +234,7 @@ index 0b0fedcd..c1f78618 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 74aa3dd0..f987844d 100644
+index 439aaeab..6a989034 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
 @@ -11,6 +11,7 @@
@@ -250,10 +250,10 @@ index 74aa3dd0..f987844d 100644
      LLM_KV_ATTENTION_SCALE,
      LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
 +    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-     LLM_KV_ROPE_DIMENSION_COUNT,
-     LLM_KV_ROPE_DIMENSION_SECTIONS,
-@@ -343,6 +345,14 @@ enum llm_tensor {
+@@ -347,6 +349,14 @@ enum llm_tensor {
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
      LLM_TENSOR_BSKCN_TV,
@@ -297,10 +297,10 @@ index 01d5ca57..8682b0e6 100644
          batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
      }
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 65135172..afe6f552 100644
+index 32f59819..0343ba8a 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -858,7 +858,7 @@ float * llama_context::get_logits_ith(int32_t i) {
+@@ -862,7 +862,7 @@ float * llama_context::get_logits_ith(int32_t i) {
              throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
          }
  
@@ -309,7 +309,7 @@ index 65135172..afe6f552 100644
      } catch (const std::exception & err) {
          LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
  #ifndef NDEBUG
-@@ -979,6 +979,10 @@ void llama_context::set_warmup(bool value) {
+@@ -983,6 +983,10 @@ void llama_context::set_warmup(bool value) {
      cparams.warmup = value;
  }
  
@@ -320,7 +320,7 @@ index 65135172..afe6f552 100644
  void llama_context::set_adapter_lora(
              llama_adapter_lora * adapter,
              float scale) {
-@@ -1054,7 +1058,7 @@ int llama_context::encode(llama_batch & inp_batch) {
+@@ -1058,7 +1062,7 @@ int llama_context::encode(llama_batch & inp_batch) {
  
      const int64_t n_embd = hparams.n_embd;
  
@@ -329,7 +329,7 @@ index 65135172..afe6f552 100644
  
      const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
  
-@@ -1194,10 +1198,9 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1198,10 +1202,9 @@ int llama_context::decode(llama_batch & inp_batch) {
  
      const llama_batch & batch = batch_allocr.batch;
  
@@ -341,7 +341,7 @@ index 65135172..afe6f552 100644
  
      const int64_t n_tokens_all = batch.n_tokens;
      const int64_t n_embd       = hparams.n_embd;
-@@ -1245,7 +1248,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1249,7 +1252,7 @@ int llama_context::decode(llama_batch & inp_batch) {
  
      const bool logits_all = n_outputs_all == n_tokens_all;
  
@@ -350,7 +350,7 @@ index 65135172..afe6f552 100644
              /* simple_split */ !kv_self->recurrent,
              /* logits_all   */ logits_all);
  
-@@ -1479,12 +1482,11 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1483,12 +1486,11 @@ int llama_context::decode(llama_batch & inp_batch) {
  
  int32_t llama_context::output_reserve(int32_t n_outputs) {
      const auto & hparams = model.hparams;
@@ -364,7 +364,7 @@ index 65135172..afe6f552 100644
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead
-@@ -1554,7 +1556,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
+@@ -1558,7 +1560,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
  void llama_context::output_reorder() {
      auto & out_ids = sbatch.out_ids;
      if (!out_ids.empty()) {
@@ -373,7 +373,7 @@ index 65135172..afe6f552 100644
          const uint32_t n_embd  = model.hparams.n_embd;
  
          GGML_ASSERT((size_t) n_outputs == out_ids.size());
-@@ -2061,7 +2063,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
+@@ -2065,7 +2067,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
      {
          LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
  
@@ -382,7 +382,7 @@ index 65135172..afe6f552 100644
  
          io.write(&logits_size, sizeof(logits_size));
  
-@@ -2244,6 +2246,7 @@ llama_context_params llama_context_default_params() {
+@@ -2248,6 +2250,7 @@ llama_context_params llama_context_default_params() {
          /*.offload_kqv                 =*/ true,
          /*.flash_attn                  =*/ false,
          /*.no_perf                     =*/ true,
@@ -390,7 +390,7 @@ index 65135172..afe6f552 100644
          /*.abort_callback              =*/ nullptr,
          /*.abort_callback_data         =*/ nullptr,
      };
-@@ -2371,6 +2374,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
+@@ -2375,6 +2378,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
      ctx->set_warmup(warmup);
  }
  
@@ -426,7 +426,7 @@ index 30e550f0..85ad91b9 100644
  
      enum llama_pooling_type pooling_type;
 diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
-index cd955d63..83f3c5a8 100644
+index a85e9728..d740c120 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
 @@ -546,6 +546,12 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
@@ -442,7 +442,7 @@ index cd955d63..83f3c5a8 100644
  //
  // llm_graph_context
  //
-@@ -1495,6 +1501,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
+@@ -1506,6 +1512,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
      return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
  }
  
@@ -469,7 +469,7 @@ index cd955d63..83f3c5a8 100644
          llm_graph_input_attn_cross * inp,
          ggml_cgraph * gf,
 diff --git a/src/llama-graph.h b/src/llama-graph.h
-index 5b6618f9..51993998 100644
+index d192dc14..260a2af2 100644
 --- a/src/llama-graph.h
 +++ b/src/llama-graph.h
 @@ -86,6 +86,7 @@ public:
@@ -518,7 +518,7 @@ index 8a667960..6a02de03 100644
 +    return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
 +}
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index c3147cbc..4567a0e9 100644
+index 6e278945..c8a34d52 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
 @@ -2,6 +2,8 @@
@@ -536,9 +536,9 @@ index c3147cbc..4567a0e9 100644
      uint32_t n_rel_attn_bkts = 0;
 +    uint32_t n_vocab = 0;
  
-     // for WavTokenizer
-     struct llama_hparams_posnet   posnet;
-@@ -52,6 +55,7 @@ struct llama_hparams {
+     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+     uint32_t n_embd_head_k_mla = 0;
+@@ -56,6 +59,7 @@ struct llama_hparams {
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
  
      std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr = {};
@@ -546,7 +546,7 @@ index c3147cbc..4567a0e9 100644
  
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
-@@ -154,6 +158,9 @@ struct llama_hparams {
+@@ -158,6 +162,9 @@ struct llama_hparams {
      // Block skip connection
      bool n_bskcn(uint32_t n, uint32_t il) const;
  
@@ -557,7 +557,7 @@ index c3147cbc..4567a0e9 100644
  };
  
 diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
-index dbf5f118..9310f262 100644
+index 7c9d46d8..69f8d35a 100644
 --- a/src/llama-kv-cache.cpp
 +++ b/src/llama-kv-cache.cpp
 @@ -95,8 +95,16 @@ bool llama_kv_cache_unified::init(
@@ -593,7 +593,7 @@ index a012aeae..2e11507d 100644
      bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
          const int kid = gguf_find_key(meta.get(), key.c_str());
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 5fbd0055..d5ad466e 100644
+index aba42819..d051696c 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
 @@ -419,6 +419,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -650,7 +650,7 @@ index 5fbd0055..d5ad466e 100644
          case LLM_ARCH_DECI:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-@@ -1548,7 +1562,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -1550,7 +1564,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
          const int64_t n_embd_head_v = hparams.n_embd_head_v;
          const int64_t n_ff          = hparams.n_ff();
          const int64_t n_embd_gqa    = n_embd_v_gqa;
@@ -659,7 +659,7 @@ index 5fbd0055..d5ad466e 100644
          const int64_t n_token_types = vocab.n_token_types();
          const int64_t n_rot         = hparams.n_rot;
          const int64_t n_expert      = hparams.n_expert;
-@@ -1801,6 +1815,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -1803,6 +1817,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                          }
                      }
                  } break;
@@ -712,7 +712,7 @@ index 5fbd0055..d5ad466e 100644
              case LLM_ARCH_DECI:
                  {
                      tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
-@@ -4665,6 +4725,246 @@ struct llm_build_llama : public llm_graph_context {
+@@ -4683,6 +4743,246 @@ struct llm_build_llama : public llm_graph_context {
      }
  };
  
@@ -900,7 +900,7 @@ index 5fbd0055..d5ad466e 100644
 +
 +                cur = build_attn(inp_attn, gf,
 +                    model.layers[il].wo, model.layers[il].bo,
-+                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
++                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 +
 +                if (il == n_layer - 1) {
 +                    // skip computing output for unused tokens
@@ -959,7 +959,7 @@ index 5fbd0055..d5ad466e 100644
  struct llm_build_deci : public llm_graph_context {
      llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
          const int64_t n_embd_head = hparams.n_embd_head_v;
-@@ -12965,6 +13265,10 @@ llm_graph_result_ptr llama_model::build_graph(
+@@ -13017,6 +13317,10 @@ llm_graph_result_ptr llama_model::build_graph(
              {
                  llm = std::make_unique<llm_build_llama>(*this, params, gf);
              } break;
@@ -970,7 +970,7 @@ index 5fbd0055..d5ad466e 100644
          case LLM_ARCH_DECI:
              {
                  llm = std::make_unique<llm_build_deci>(*this, params, gf);
-@@ -13325,6 +13629,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13377,6 +13681,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          // use what we call a normal RoPE, operating on pairs of consecutive head values
          case LLM_ARCH_LLAMA:
          case LLM_ARCH_LLAMA4:
@@ -979,7 +979,7 @@ index 5fbd0055..d5ad466e 100644
          case LLM_ARCH_BAICHUAN:
          case LLM_ARCH_STARCODER:
 diff --git a/src/llama-model.h b/src/llama-model.h
-index e08d4ae4..21c4617b 100644
+index 5865d5e9..72bab5be 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
 @@ -11,6 +11,7 @@
@@ -998,7 +998,7 @@ index e08d4ae4..21c4617b 100644
      LLM_TYPE_236B,
      LLM_TYPE_314B,
      LLM_TYPE_671B,
-@@ -308,6 +310,16 @@ struct llama_layer {
+@@ -310,6 +312,16 @@ struct llama_layer {
  
      struct ggml_tensor * bskcn_tv = nullptr;