Merge 5f9c6fb2a4 into 8db003a19d

* Support of converting local models added to convert-hf-to-gguf-update.py

* Description fixed

* shutil added to imports

common/common.cpp

@@ -941,11 +941,37 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 
 #ifdef LLAMA_USE_CURL
 
+#define CURL_MAX_RETRY 3
+#define CURL_RETRY_DELAY_SECONDS 2
+
+
 static bool starts_with(const std::string & str, const std::string & prefix) {
     // While we wait for C++20's std::string::starts_with...
     return str.rfind(prefix, 0) == 0;
 }
 
+static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
+    int remaining_attempts = max_attempts;
+
+    while (remaining_attempts > 0) {
+        fprintf(stderr, "%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+
+        CURLcode res = curl_easy_perform(curl);
+        if (res == CURLE_OK) {
+            return true;
+        }
+
+        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
+        fprintf(stderr, "%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
+
+        remaining_attempts--;
+        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+    }
+
+    fprintf(stderr, "%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
+    return false;
+}
+
 static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
 
     // Initialize libcurl
@@ -1049,9 +1075,8 @@ static bool llama_download_file(const std::string & url, const std::string & pat
         curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-        CURLcode res = curl_easy_perform(curl.get());
-        if (res != CURLE_OK) {
-            fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
             return false;
         }
 
@@ -1126,11 +1151,10 @@ static bool llama_download_file(const std::string & url, const std::string & pat
         };
 
         // start the download
-        fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-                llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        auto res = curl_easy_perform(curl.get());
-        if (res != CURLE_OK) {
-            fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
+        fprintf(stderr, "%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
             return false;
         }
 

convert_hf_to_gguf.py

@@ -1837,6 +1837,58 @@ class MiniCPMModel(Model):
 
         return [(self.map_tensor_name(name), data_torch)]
 
+@Model.register("MiniCPM3ForCausalLM")
+class MiniCPM3Model(Model):
+    model_arch = gguf.MODEL_ARCH.MINICPM3
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        rope_dims = hparams["qk_rope_head_dim"]
+
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
+            self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+        rope_scaling = self.find_hparam(['rope_scaling'], True)
+        if rope_scaling is None:
+            return
+
+        long_factors = rope_scaling.get('long_factor', None)
+        short_factors = rope_scaling.get('short_factor', None)
+
+        if long_factors is None or short_factors is None:
+            raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
+
+        if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
+            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
+        
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+
+    def set_vocab(self):
+        self._set_vocab_llama_hf()
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
 
 @Model.register("QWenLMHeadModel")
 class QwenModel(Model):

convert_hf_to_gguf_update.py

@@ -31,6 +31,7 @@ import re
 import requests
 import sys
 import json
+import shutil
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -125,12 +126,27 @@ def download_model(model):
     if tokt == TOKENIZER_TYPE.UGM:
         files.append("spiece.model")
 
-    for file in files:
-        save_path = f"models/tokenizers/{name}/{file}"
-        if os.path.isfile(save_path):
-            logger.info(f"{name}: File {save_path} already exists - skipping")
-            continue
-        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+    if os.path.isdir(repo):
+        # If repo is a path on the file system, copy the directory
+        for file in files:
+            src_path = os.path.join(repo, file)
+            dst_path = f"models/tokenizers/{name}/{file}"
+            if os.path.isfile(dst_path):
+                logger.info(f"{name}: File {dst_path} already exists - skipping")
+                continue
+            if os.path.isfile(src_path):
+                shutil.copy2(src_path, dst_path)
+                logger.info(f"{name}: Copied {src_path} to {dst_path}")
+            else:
+                logger.warning(f"{name}: Source file {src_path} does not exist")
+    else:
+        # If repo is a URL, download the files
+        for file in files:
+            save_path = f"models/tokenizers/{name}/{file}"
+            if os.path.isfile(save_path):
+                logger.info(f"{name}: File {save_path} already exists - skipping")
+                continue
+            download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
 
 
 for model in models:

examples/batched-bench/batched-bench.cpp

@@ -3,32 +3,10 @@
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-// mutates the input string
-static std::vector<int> parse_list(char * p) {
-    std::vector<int> ret;
-
-    char * q = p;
-
-    while (*p) {
-        if (*p == ',') {
-            *p = '\0';
-            ret.push_back(std::atoi(q));
-            q = p + 1;
-        }
-
-        ++p;
-    }
-
-    ret.push_back(std::atoi(q));
-
-    return ret;
-}
-
 static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);

examples/llava/minicpmv-cli.cpp

@@ -18,8 +18,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    LOG_TEE("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
@@ -255,7 +255,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, show_additional_info)) {
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
         return 1;
     }
 

ggml/src/ggml-cuda/mmq.cu

@@ -26,7 +26,11 @@ void ggml_cuda_op_mul_mat_q(
     // nrows_dst == nrows of the matrix that the kernel writes into
     const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
 
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst};
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
+    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:

ggml/src/ggml-cuda/mmq.cuh

@@ -2742,6 +2742,7 @@ struct mmq_args {
     int64_t ne00; int64_t ne01; int64_t stride01;
     int64_t ne10; int64_t ne11; int64_t stride11;
     int64_t ne0;
+    bool use_stream_k;
 };
 
 template<ggml_type type>
@@ -2777,8 +2778,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
     const dim3 block_nums_xy_tiling(nty, ntx, 1);
 
-    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
-    if (!use_stream_k) {
+    if (!args.use_stream_k) {
         if (args.ne01 % mmq_y == 0) {
             constexpr bool need_check = false;
             mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>

gguf-py/gguf/constants.py

@@ -210,6 +210,7 @@ class MODEL_ARCH(IntEnum):
     ORION        = auto()
     INTERNLM2    = auto()
     MINICPM      = auto()
+    MINICPM3     = auto()
     GEMMA        = auto()
     GEMMA2       = auto()
     STARCODER2   = auto()
@@ -364,6 +365,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.ORION:          "orion",
     MODEL_ARCH.INTERNLM2:      "internlm2",
     MODEL_ARCH.MINICPM:        "minicpm",
+    MODEL_ARCH.MINICPM3:       "minicpm3",
     MODEL_ARCH.GEMMA:          "gemma",
     MODEL_ARCH.GEMMA2:         "gemma2",
     MODEL_ARCH.STARCODER2:     "starcoder2",
@@ -867,6 +869,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.MINICPM3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.GEMMA: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

src/llama.cpp

@@ -193,6 +193,7 @@ enum llm_arch {
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
     LLM_ARCH_MINICPM,
+    LLM_ARCH_MINICPM3,
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
     LLM_ARCH_STARCODER2,
@@ -241,6 +242,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ORION,           "orion"        },
     { LLM_ARCH_INTERNLM2,       "internlm2"    },
     { LLM_ARCH_MINICPM,         "minicpm"      },
+    { LLM_ARCH_MINICPM3,        "minicpm3"     },
     { LLM_ARCH_GEMMA,           "gemma"        },
     { LLM_ARCH_GEMMA2,          "gemma2"       },
     { LLM_ARCH_STARCODER2,      "starcoder2"   },
@@ -1034,6 +1036,29 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
         },
     },
+    {
+        LLM_ARCH_MINICPM3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FACTORS_LONG,  "rope_factors_long" },
+            { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q_A_NORM,      "blk.%d.attn_q_a_norm" },
+            { LLM_TENSOR_ATTN_KV_A_NORM,     "blk.%d.attn_kv_a_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_A,           "blk.%d.attn_q_a" },
+            { 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_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+        },
+    },
     {
         LLM_ARCH_GEMMA,
         {
@@ -5385,6 +5410,17 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MINICPM3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
+                ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
+
+                switch (hparams.n_layer) {
+                    case 62: model.type = e_model::MODEL_4B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_GROK:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -6894,6 +6930,54 @@ static bool llm_load_tensors(
                         }
                     }
                 } break;
+            case LLM_ARCH_MINICPM3:
+                {
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+
+                    const int64_t q_lora_rank  = hparams.n_lora_q;
+                    const int64_t kv_lora_rank = hparams.n_lora_kv;
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_q_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank});
+
+                        layer.attn_kv_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank});
+
+                        layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank});
+                        layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k});
+
+                        layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)});
+                        layer.wkv_b     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)});
+                        layer.wo        = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        
+                        layer.rope_long  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                    }
+                } break;
             case LLM_ARCH_GROK:
                 {
                     if (n_expert == 0) {
@@ -12837,6 +12921,215 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_minicpm3() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        //TODO: if the model varies, these parameters need to be read from the model
+        const int64_t n_embd_base = 256;
+        const float scale_embd  = 12.0f;
+        const float scale_depth = 1.4f;
+        const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k));
+
+        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // scale the input embeddings
+        inpL = ggml_scale(ctx0, inpL, scale_embd);
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            struct ggml_tensor * rope_factors = build_rope_factors(il);
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                struct ggml_tensor * q = NULL;
+                // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                cb(q, "q", il);
+
+                q = llm_build_norm(ctx0, q, hparams,
+                        model.layers[il].attn_q_a_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(q, "q", il);
+
+                // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                cb(q, "q", il);
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        0);
+                cb(q_nope, "q_nope", il);
+
+                // and {n_head * n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        ggml_row_size(q->type, n_embd_head_qk_nope));
+                cb(q_pe, "q_pe", il);
+
+                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+                // split into {kv_lora_rank, n_tokens}
+                struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        0);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // and {n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        kv_pe_compresseed->nb[1],
+                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                cb(k_pe, "k_pe", il);
+
+                kv_compressed = ggml_cont(ctx0, kv_compressed); // TODO: the CUDA backend does not support non-contiguous norm
+                kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams,
+                        model.layers[il].attn_kv_a_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+                struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+                cb(kv, "kv", il);
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        0);
+                cb(k_nope, "k_nope", il);
+
+                // and {n_head * n_embd_head_v, n_tokens}
+                struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_cont(ctx0, v_states);
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
+                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+                    0);
+                cb(v_states, "v_states", il);
+
+                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                q_pe = ggml_rope_ext(
+                    ctx0, q_pe, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(q_pe, "q_pe", il);
+
+                // shared RoPE key
+                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                k_pe = ggml_rope_ext(
+                    ctx0, k_pe, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(k_pe, "k_pe", il);
+
+                struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+                cb(q_states, "q_states", il);
+
+                struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+                cb(k_states, "k_states", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // scale_res - scale the hidden states for residual connection
+            const float scale_res = scale_depth/sqrtf(float(n_layer));
+            cur = ggml_scale(ctx0, cur, scale_res);
+            cb(cur, "hidden_scaled", il);
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            // scale the hidden states for residual connection
+            cur = ggml_scale(ctx0, cur, scale_res);
+            cb(cur, "hidden_scaled_ffn", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head scaling
+        const float scale_lmhead = float(n_embd_base)/float(n_embd);
+        cur = ggml_scale(ctx0, cur, scale_lmhead);
+        cb(cur, "lmhead_scaling", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_gemma() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -15377,6 +15670,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_minicpm();
             } break;
+        case LLM_ARCH_MINICPM3:
+            {
+                result = llm.build_minicpm3();
+            } break;
         case LLM_ARCH_GEMMA:
             {
                 result = llm.build_gemma();
@@ -18597,6 +18894,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_CODESHELL:
         case LLM_ARCH_NEMOTRON:
         case LLM_ARCH_EXAONE:
+        case LLM_ARCH_MINICPM3:
             return LLAMA_ROPE_TYPE_NEOX;
 
         // all model arches should be listed explicitly here