llama: update to commit e1e8e099 (#10513)

llama/llama.cpp/src/llama-sampling.cpp

@@ -232,7 +232,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
     // }
 
     if (k <= 0) {
-        k = cur_p->size;
+        return;
     }
 
     k = std::min(k, (int) cur_p->size);
@@ -298,6 +298,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
         }
         cur_p->sorted = true;
     }
+
     cur_p->size = k;
 }
 

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

@@ -1497,7 +1497,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "llama3"   ||
                     tokenizer_pre == "llama-v3" ||
                     tokenizer_pre == "llama-bpe"||
-                    tokenizer_pre == "falcon3") {
+                    tokenizer_pre == "falcon3"  ||
+                    tokenizer_pre == "pixtral") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;

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

@@ -85,7 +85,7 @@ index e2617b06..242e50a7 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index a7febef7..31750b6f 100644
+index 9fb2134f..04ce764e 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -534,6 +534,7 @@ struct ggml_backend_cuda_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 266d8af4..12886cd3 100644
+index d92392ed..425524d0 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -4759,6 +4759,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -5077,6 +5077,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
      }
  
      free(ctx);
@@ -149,10 +149,10 @@ index 05a2f4e6..392cc18d 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 a0667b7d..bd83adc5 100644
+index 140a775f..e33c4ba0 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -468,6 +468,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -477,6 +477,7 @@ 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;
@@ -161,10 +161,10 @@ index a0667b7d..bd83adc5 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 1de34c96..4600f61e 100644
+index 66b6f2cc..e3e6deae 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 {
+@@ -317,6 +317,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
      ggml_sycl_set_device(ctx->device);
  
      delete ctx;
@@ -172,7 +172,7 @@ index 1de34c96..4600f61e 100644
  }
  catch (sycl::exception const &exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -761,6 +762,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -762,6 +763,7 @@ struct ggml_backend_sycl_split_buffer_context {
  static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
      delete ctx;
@@ -180,7 +180,7 @@ index 1de34c96..4600f61e 100644
  }
  
  static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1095,6 +1097,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1096,6 +1098,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
  
  static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_sycl_host_free(buffer->context);
@@ -189,10 +189,10 @@ index 1de34c96..4600f61e 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 39f3cd34..c569a8a5 100644
+index c0bdb9e1..03d03064 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -8653,6 +8653,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -8660,6 +8660,7 @@ 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 39f3cd34..c569a8a5 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -8796,6 +8797,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -8803,6 +8804,7 @@ 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 48060517..a35b498c 100644
+index 50ded286..a9ee9f03 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) {
@@ -31,7 +31,7 @@ index 48060517..a35b498c 100644
                  pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -1634,7 +1625,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1635,7 +1626,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                  pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                  clean_spaces = false;
              } else {

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 983385f8..32f59819 100644
+index 5a2eef9b..9c1fe93f 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -1236,7 +1236,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1225,7 +1225,7 @@ int llama_context::decode(llama_batch & inp_batch) {
      int64_t n_outputs_all = 0;
  
      // count outputs
@@ -23,7 +23,7 @@ index 983385f8..32f59819 100644
          for (uint32_t i = 0; i < n_tokens_all; ++i) {
              n_outputs_all += batch.logits[i] != 0;
          }
-@@ -1348,7 +1348,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1337,7 +1337,7 @@ int llama_context::decode(llama_batch & inp_batch) {
          //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
          //}
  
@@ -32,7 +32,7 @@ index 983385f8..32f59819 100644
          auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
  
          if (t_embd && res->get_embd_pooled()) {
-@@ -1492,7 +1492,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
+@@ -1481,7 +1481,7 @@ 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 75970615..d57b4bd6 100644
+index ad3e7df1..b3218c78 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -29,6 +29,19 @@
- #include <limits>
+@@ -30,6 +30,19 @@
  #include <array>
+ #include <numeric>
  
 +#if defined(_WIN32)
 +#define WIN32_LEAN_AND_MEAN
@@ -33,7 +33,7 @@ index 75970615..d57b4bd6 100644
  struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
  
  //#define CLIP_DEBUG_FUNCTIONS
-@@ -1430,7 +1443,29 @@ struct clip_model_loader {
+@@ -1971,7 +1984,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 75970615..d57b4bd6 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -1457,7 +1492,11 @@ struct clip_model_loader {
+@@ -1998,7 +2033,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0005-solar-pro.patch

@@ -15,10 +15,10 @@ 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 62e1480b..f754bc8f 100644
+index f2bc8ca7..5ab3f572 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 = {
+@@ -69,6 +69,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_GRANITE,          "granite"          },
      { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
      { LLM_ARCH_CHAMELEON,        "chameleon"        },
@@ -26,7 +26,7 @@ index 62e1480b..f754bc8f 100644
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -140,6 +141,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+@@ -142,6 +143,7 @@ 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"                        },
@@ -34,7 +34,7 @@ index 62e1480b..f754bc8f 100644
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -1482,6 +1484,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -1502,6 +1504,24 @@ 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 62e1480b..f754bc8f 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -1660,6 +1680,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -1680,6 +1700,7 @@ 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}},
@@ -68,10 +68,10 @@ index 62e1480b..f754bc8f 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 98ca00a1..439aaeab 100644
+index 41a023da..525c1b7d 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -72,6 +72,7 @@ enum llm_arch {
+@@ -73,6 +73,7 @@ enum llm_arch {
      LLM_ARCH_GRANITE,
      LLM_ARCH_GRANITE_MOE,
      LLM_ARCH_CHAMELEON,
@@ -79,7 +79,7 @@ index 98ca00a1..439aaeab 100644
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -144,6 +145,7 @@ enum llm_kv {
+@@ -146,6 +147,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
      LLM_KV_ATTENTION_SLIDING_WINDOW,
      LLM_KV_ATTENTION_SCALE,
@@ -87,7 +87,7 @@ index 98ca00a1..439aaeab 100644
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -344,6 +346,7 @@ enum llm_tensor {
+@@ -346,6 +348,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -115,7 +115,7 @@ 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 80fcd65d..6e278945 100644
+index 7ee6a5b7..48dce407 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
 @@ -55,6 +55,8 @@ struct llama_hparams {
@@ -127,7 +127,7 @@ index 80fcd65d..6e278945 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -153,6 +155,9 @@ struct llama_hparams {
+@@ -154,6 +156,9 @@ 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 6b7bfecf..aba42819 100644
+index 822e2bb2..572378c9 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1374,6 +1374,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1386,6 +1386,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -175,7 +175,7 @@ index 6b7bfecf..aba42819 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -3717,6 +3732,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -3741,6 +3756,34 @@ 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 6b7bfecf..aba42819 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);
-@@ -12296,6 +12339,165 @@ struct llm_build_chameleon : public llm_graph_context {
+@@ -12342,6 +12385,165 @@ struct llm_build_chameleon : public llm_graph_context {
      }
  };
  
@@ -376,7 +376,7 @@ index 6b7bfecf..aba42819 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;
-@@ -13045,6 +13247,10 @@ llm_graph_result_ptr llama_model::build_graph(
+@@ -13092,6 +13294,10 @@ llm_graph_result_ptr llama_model::build_graph(
              {
                  llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
              } break;
@@ -387,7 +387,7 @@ index 6b7bfecf..aba42819 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params, gf);
-@@ -13191,6 +13397,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13238,6 +13444,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE:
          case LLM_ARCH_GRANITE_MOE:
          case LLM_ARCH_CHAMELEON:
@@ -396,18 +396,18 @@ index 6b7bfecf..aba42819 100644
              return LLAMA_ROPE_TYPE_NORM;
  
 diff --git a/src/llama-model.h b/src/llama-model.h
-index fd82d106..5865d5e9 100644
+index 95eca002..856e6042 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
-@@ -62,6 +62,7 @@ enum llm_type {
+@@ -64,6 +64,7 @@ enum llm_type {
      LLM_TYPE_15B,
      LLM_TYPE_16B,
      LLM_TYPE_20B,
 +    LLM_TYPE_22B,
+     LLM_TYPE_27B,
      LLM_TYPE_30B,
      LLM_TYPE_32B,
-     LLM_TYPE_34B,
-@@ -307,6 +308,8 @@ struct llama_layer {
+@@ -311,6 +312,8 @@ struct llama_layer {
      struct ggml_tensor * ffn_up_scale   = nullptr;
      struct ggml_tensor * ffn_down_scale = nullptr;
  

llama/patches/0007-add-unpad-operator.patch

@@ -18,10 +18,10 @@ adds the unpad operator to GGML
  10 files changed, 223 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
-index 8fcc16df..d19fc167 100644
+index 1b8603e7..53ef31b2 100644
 --- a/ggml/include/ggml.h
 +++ b/ggml/include/ggml.h
-@@ -488,6 +488,7 @@ extern "C" {
+@@ -489,6 +489,7 @@ extern "C" {
          GGML_OP_UPSCALE, // nearest interpolate
          GGML_OP_PAD,
          GGML_OP_PAD_REFLECT_1D,
@@ -29,7 +29,7 @@ index 8fcc16df..d19fc167 100644
          GGML_OP_ARANGE,
          GGML_OP_TIMESTEP_EMBEDDING,
          GGML_OP_ARGSORT,
-@@ -1757,6 +1758,15 @@ extern "C" {
+@@ -1777,6 +1778,15 @@ extern "C" {
              int                   p0,
              int                   p1);
  
@@ -46,10 +46,10 @@ index 8fcc16df..d19fc167 100644
      // timesteps: [N,]
      // return: [N, dim]
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 50400328..432942bf 100644
+index 64405449..34624cca 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
-@@ -1960,6 +1960,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
+@@ -1964,6 +1964,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
              {
                  ggml_compute_forward_pad_reflect_1d(params, tensor);
              } break;
@@ -60,7 +60,7 @@ index 50400328..432942bf 100644
          case GGML_OP_ARANGE:
              {
                  ggml_compute_forward_arange(params, tensor);
-@@ -2282,6 +2286,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+@@ -2287,6 +2291,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
          case GGML_OP_UPSCALE:
          case GGML_OP_PAD:
          case GGML_OP_PAD_REFLECT_1D:
@@ -69,10 +69,10 @@ index 50400328..432942bf 100644
          case GGML_OP_TIMESTEP_EMBEDDING:
          case GGML_OP_ARGSORT:
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index 6050147b..66b8da68 100644
+index 7413192b..becdae07 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -6531,6 +6531,61 @@ void ggml_compute_forward_pad_reflect_1d(
+@@ -6703,6 +6703,61 @@ void ggml_compute_forward_pad_reflect_1d(
      }
  }
  
@@ -135,10 +135,10 @@ index 6050147b..66b8da68 100644
  
  static void ggml_compute_forward_arange_f32(
 diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
-index 410a3720..3eca1cf8 100644
+index dc081b9e..a7125555 100644
 --- a/ggml/src/ggml-cpu/ops.h
 +++ b/ggml/src/ggml-cpu/ops.h
-@@ -71,6 +71,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
+@@ -72,6 +72,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
  void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
  void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
  void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -147,10 +147,10 @@ index 410a3720..3eca1cf8 100644
  void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
  void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 31750b6f..0fef9522 100644
+index 04ce764e..491acccb 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -2246,6 +2246,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
+@@ -2223,6 +2223,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
          case GGML_OP_PAD:
              ggml_cuda_op_pad(ctx, dst);
              break;
@@ -160,7 +160,7 @@ index 31750b6f..0fef9522 100644
          case GGML_OP_ARANGE:
              ggml_cuda_op_arange(ctx, dst);
              break;
-@@ -3222,6 +3225,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
+@@ -3197,6 +3200,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
          case GGML_OP_UPSCALE:
              return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
          case GGML_OP_PAD:
@@ -233,7 +233,7 @@ index 8fd386b0..e2ededc3 100644
  void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 +void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 12886cd3..b2e95a66 100644
+index 425524d0..112abef6 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
 @@ -341,6 +341,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
@@ -244,7 +244,7 @@ index 12886cd3..b2e95a66 100644
      GGML_METAL_KERNEL_TYPE_ARANGE_F32,
      GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
      GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
-@@ -1020,6 +1021,7 @@ @implementation GGMLMetalClass
+@@ -1277,6 +1278,7 @@ @implementation GGMLMetalClass
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                     upscale_f32,                     true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                         pad_f32,                         true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,              pad_reflect_1d_f32,              true);
@@ -252,7 +252,7 @@ index 12886cd3..b2e95a66 100644
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,          timestep_embedding_f32,          true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                      arange_f32,                      true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,             argsort_f32_i32_asc,             true);
-@@ -1384,6 +1386,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
+@@ -1647,6 +1649,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
          case GGML_OP_POOL_2D:
          case GGML_OP_PAD:
          case GGML_OP_PAD_REFLECT_1D:
@@ -260,7 +260,7 @@ index 12886cd3..b2e95a66 100644
          case GGML_OP_TIMESTEP_EMBEDDING:
          case GGML_OP_ARGSORT:
          case GGML_OP_LEAKY_RELU:
-@@ -3731,6 +3734,36 @@ static void ggml_metal_encode_node(
+@@ -4047,6 +4050,36 @@ static bool ggml_metal_encode_node(
  
                  const int nth = MIN(1024, ne0);
  
@@ -298,7 +298,7 @@ index 12886cd3..b2e95a66 100644
              } break;
          case GGML_OP_ARANGE:
 diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
-index 8d6e99e6..71f0f97f 100644
+index 9f4147e9..6ceb3cef 100644
 --- a/ggml/src/ggml-metal/ggml-metal.metal
 +++ b/ggml/src/ggml-metal/ggml-metal.metal
 @@ -2975,6 +2975,51 @@ kernel void kernel_pad_reflect_1d_f32(
@@ -354,10 +354,10 @@ index 8d6e99e6..71f0f97f 100644
      device        char * dst,
      constant   ggml_metal_kargs_arange & args,
 diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
-index 950772c7..2276b631 100644
+index 7654ae17..3c57aff8 100644
 --- a/ggml/src/ggml.c
 +++ b/ggml/src/ggml.c
-@@ -963,6 +963,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+@@ -923,6 +923,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
      "UPSCALE",
      "PAD",
      "PAD_REFLECT_1D",
@@ -365,16 +365,16 @@ index 950772c7..2276b631 100644
      "ARANGE",
      "TIMESTEP_EMBEDDING",
      "ARGSORT",
-@@ -993,7 +994,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+@@ -953,7 +954,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
      "OPT_STEP_ADAMW",
  };
  
--static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
-+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
++static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
  
  static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
      "none",
-@@ -1057,6 +1058,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+@@ -1018,6 +1019,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
      "upscale(x)",
      "pad(x)",
      "pad_reflect_1d(x)",
@@ -382,16 +382,16 @@ index 950772c7..2276b631 100644
      "arange(start, stop, step)",
      "timestep_embedding(timesteps, dim, max_period)",
      "argsort(x)",
-@@ -1087,7 +1089,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+@@ -1048,7 +1050,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
      "adamw(x)",
  };
  
--static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
-+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
++static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
  
  static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
  
-@@ -4262,6 +4264,25 @@ struct ggml_tensor * ggml_pad_reflect_1d(
+@@ -4270,6 +4272,25 @@ struct ggml_tensor * ggml_pad_reflect_1d(
      return result;
  }
  

llama/patches/0008-fix-deepseek-deseret-regex.patch

@@ -12,7 +12,7 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index a35b498c..032019c9 100644
+index a9ee9f03..1306864e 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -296,7 +296,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {

llama/patches/0009-maintain-ordering-for-rules-for-grammar.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] maintain ordering for rules for grammar
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
-index 90679822..56043678 100644
+index 5b3059c2..656b3eca 100644
 --- a/common/json-schema-to-grammar.cpp
 +++ b/common/json-schema-to-grammar.cpp
-@@ -346,7 +346,7 @@ private:
+@@ -349,7 +349,7 @@ private:
      friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
      std::function<json(const std::string &)> _fetch_json;
      bool _dotall;

llama/patches/0010-ensure-KV-cache-is-fully-defragmented.patch

@@ -22,10 +22,10 @@ multiple batches of processing until everything is complete.
  4 files changed, 51 insertions(+), 106 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 0343ba8a..4b3e6a83 100644
+index cd06ad91..77177c5e 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -594,13 +594,12 @@ llm_graph_result_ptr llama_context::build_kv_self_shift(
+@@ -583,13 +583,12 @@ llm_graph_result_ptr llama_context::build_kv_self_shift(
  
  llm_graph_result_ptr llama_context::build_kv_self_defrag(
          ggml_context * ctx0,
@@ -41,7 +41,7 @@ index 0343ba8a..4b3e6a83 100644
  #if 0
      // CPU defrag
      //
-@@ -672,32 +671,20 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+@@ -661,32 +660,20 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
          ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size());
      }
  #else
@@ -79,7 +79,7 @@ index 0343ba8a..4b3e6a83 100644
  
              ggml_tensor * view_v_src;
              ggml_tensor * view_v_dst;
-@@ -705,34 +692,30 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+@@ -694,34 +681,30 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
              if (cparams.flash_attn) {
                  // NOTE: the V cache is not transposed when using flash attention
                  view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
@@ -122,7 +122,7 @@ index 0343ba8a..4b3e6a83 100644
  #endif
  
      return res;
-@@ -741,8 +724,6 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+@@ -730,8 +713,6 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
  void llama_context::kv_self_update() {
      auto & kv = kv_self;
  
@@ -131,7 +131,7 @@ index 0343ba8a..4b3e6a83 100644
      if (kv->has_shift) {
          if (!kv->get_can_shift()) {
              GGML_ABORT("The current context does not support K-shift");
-@@ -763,8 +744,6 @@ void llama_context::kv_self_update() {
+@@ -752,8 +733,6 @@ void llama_context::kv_self_update() {
              res->set_inputs(nullptr);
  
              graph_compute(gf, false);
@@ -140,7 +140,7 @@ index 0343ba8a..4b3e6a83 100644
          }
  
          {
-@@ -779,49 +758,28 @@ void llama_context::kv_self_update() {
+@@ -768,49 +747,28 @@ void llama_context::kv_self_update() {
      // defragment the KV cache if needed
      if (kv->do_defrag) {
          LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
@@ -202,7 +202,7 @@ index 0343ba8a..4b3e6a83 100644
  }
  
  enum llama_pooling_type llama_context::pooling_type() const {
-@@ -1305,9 +1263,12 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1294,9 +1252,12 @@ int llama_context::decode(llama_batch & inp_batch) {
          // find KV slot
          {
              if (!kv_self->find_slot(ubatch)) {
@@ -219,7 +219,7 @@ index 0343ba8a..4b3e6a83 100644
  
              if (!kv_self->recurrent) {
 diff --git a/src/llama-context.h b/src/llama-context.h
-index baa03276..a59ff8fd 100644
+index a50c4afa..30f84bfd 100644
 --- a/src/llama-context.h
 +++ b/src/llama-context.h
 @@ -5,6 +5,7 @@
@@ -230,7 +230,7 @@ index baa03276..a59ff8fd 100644
  
  #include "ggml-cpp.h"
  
-@@ -180,7 +181,8 @@ private:
+@@ -179,7 +180,8 @@ private:
  
      llm_graph_result_ptr build_kv_self_defrag(
              ggml_context * ctx0,

llama/patches/0012-add-phony-target-ggml-cpu-for-all-cpu-variants.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] add phony target ggml-cpu for all cpu variants
  1 file changed, 2 insertions(+)
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index f00700da..91d6a7d5 100644
+index 43d9fc4f..4c0d3824 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -278,6 +278,7 @@ function(ggml_add_cpu_backend_variant tag_name)
+@@ -279,6 +279,7 @@ function(ggml_add_cpu_backend_variant tag_name)
      endforeach()
  
      ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -19,11 +19,11 @@ index f00700da..91d6a7d5 100644
  endfunction()
  
  ggml_add_backend(CPU)
-@@ -286,6 +287,7 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -287,6 +288,7 @@ if (GGML_CPU_ALL_VARIANTS)
      if (NOT GGML_BACKEND_DL)
          message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
      endif()
 +    add_custom_target(ggml-cpu)
-     ggml_add_cpu_backend_variant(sandybridge    AVX)
-     ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 BMI2 FMA)
-     ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 BMI2 FMA AVX512)
+     ggml_add_cpu_backend_variant(x64)
+     ggml_add_cpu_backend_variant(sse42        SSE42)
+     ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)

llama/patches/0013-remove-amx.patch

 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
 From: jmorganca <jmorganca@gmail.com>
-Date: Tue, 8 Apr 2025 20:33:01 -0700
+Date: Thu, 1 May 2025 15:05:08 -0700
 Subject: [PATCH] remove amx
 
 disable amx as it reduces performance on some systems
@@ -9,16 +9,16 @@ disable amx as it reduces performance on some systems
  1 file changed, 4 deletions(-)
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index 91d6a7d5..d6b393a2 100644
+index 4c0d3824..79c26312 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -293,10 +293,6 @@ if (GGML_CPU_ALL_VARIANTS)
-     ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 BMI2 FMA AVX512)
-     ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-     ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+@@ -296,10 +296,6 @@ if (GGML_CPU_ALL_VARIANTS)
+     ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+     ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+     ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
 -    if (NOT MSVC)
 -        # MSVC doesn't support AMX
--        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+-        ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
 -    endif()
  elseif (GGML_CPU)
      ggml_add_cpu_backend_variant_impl("")

llama/patches/0014-fix-string-arr-kv-loading.patch

@@ -53,7 +53,7 @@ index 381a9c7d..e45b453d 100644
  }
  
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 032019c9..ba37df35 100644
+index 1306864e..d6515ff6 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1459,7 +1459,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {

llama/patches/0015-ollama-debug-tensor.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] ollama debug tensor
  1 file changed, 6 insertions(+)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 432942bf..6d4abe4c 100644
+index 34624cca..59bd3c62 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
 @@ -15,6 +15,8 @@
@@ -20,7 +20,7 @@ index 432942bf..6d4abe4c 100644
  #if defined(_MSC_VER) || defined(__MINGW32__)
  #include <malloc.h> // using malloc.h with MSC/MINGW
  #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-@@ -2854,6 +2856,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+@@ -2859,6 +2861,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
  
          ggml_compute_forward(&params, node);
  

llama/patches/0016-add-model-quantizations.patch

@@ -13,10 +13,10 @@ models not supported in llama.cpp
  4 files changed, 24 insertions(+)
 
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 0568565f..dd01df60 100644
+index eb7b5325..df42d1a5 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
-@@ -73,6 +73,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+@@ -74,6 +74,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
@@ -24,7 +24,7 @@ index 0568565f..dd01df60 100644
      { LLM_ARCH_UNKNOWN,          "(unknown)"        },
  };
  
-@@ -1586,6 +1587,22 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -1606,6 +1607,22 @@ 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" },
          },
      },
@@ -48,10 +48,10 @@ index 0568565f..dd01df60 100644
          LLM_ARCH_UNKNOWN,
          {
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 6a989034..b6227eeb 100644
+index bc8a4f0b..bda9d071 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -75,6 +75,7 @@ enum llm_arch {
+@@ -76,6 +76,7 @@ enum llm_arch {
      LLM_ARCH_CHAMELEON,
      LLM_ARCH_SOLAR,
      LLM_ARCH_WAVTOKENIZER_DEC,
@@ -60,10 +60,10 @@ index 6a989034..b6227eeb 100644
      LLM_ARCH_BAILINGMOE,
      LLM_ARCH_UNKNOWN,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index d051696c..c8374159 100644
+index 9d099f11..ef70486d 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1425,6 +1425,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1437,6 +1437,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                  }
              } break;
@@ -71,7 +71,7 @@ index d051696c..c8374159 100644
          default: throw std::runtime_error("unsupported model architecture");
      }
  
-@@ -13704,6 +13705,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13751,6 +13752,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_CHAMELEON:
          case LLM_ARCH_SOLAR:
          case LLM_ARCH_BAILINGMOE:

llama/patches/0021-add-ollama-vocab-for-grammar-support.patch → llama/patches/0017-add-ollama-vocab-for-grammar-support.patch

@@ -184,10 +184,10 @@ index f8c291de..2a3a62db 100644
                        const char * grammar_root,
                                bool lazy,
 diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
-index d1497985..b1a9dca3 100644
+index c0a5f934..75731053 100644
 --- a/src/llama-sampling.cpp
 +++ b/src/llama-sampling.cpp
-@@ -1465,7 +1465,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
+@@ -1466,7 +1466,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
          trigger_patterns_c.push_back(trigger_pattern.pattern.c_str());
      }
  
@@ -196,7 +196,7 @@ index d1497985..b1a9dca3 100644
                                                   ctx->grammar->lazy, trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                   ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
  
-@@ -1547,7 +1547,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
+@@ -1548,7 +1548,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
              /* .vocab        = */ vocab,
              /* .grammar_str  = */ grammar_str,
              /* .grammar_root = */ grammar_root,

ml/backend/ggml/ggml/include/ggml-cpu.h

@@ -133,6 +133,11 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_bf16_to_fp32(const ggml_bf16_t *, float *, int64_t);
+
 #ifdef __cplusplus
 }
 #endif

ml/backend/ggml/ggml/include/ggml-rpc.h

@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    2
 #define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16