Merge branch 'ggerganov:master' into master

Author: zhiyuan1i

Makefile

@@ -34,6 +34,7 @@ BUILD_TARGETS = \
 	llama-save-load-state \
 	llama-server \
 	llama-simple \
+	llama-simple-chat \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -1287,6 +1288,11 @@ llama-simple: examples/simple/simple.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-simple-chat: examples/simple-chat/simple-chat.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-tokenize: examples/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

README.md

@@ -17,7 +17,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669**
+- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
+- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----

ci/run.sh

@@ -155,7 +155,7 @@ struct common_sampler_params {
 
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =     0; // context size
+    int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt

common/common.h

@@ -72,7 +72,8 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
-                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -87,7 +88,7 @@ def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path,
         self.is_safetensors = len(self.part_names) > 0
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
-        self.hparams = Model.load_hparams(self.dir_model)
+        self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
@@ -1541,6 +1542,17 @@ def set_vocab(self):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -1557,17 +1569,6 @@ def set_gguf_parameters(self):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
-        # Apply to granite small models only
-        if self.hparams.get("vocab_size", 32000) == 49152:
-            self.gguf_writer.add_add_bos_token(False)
-
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
         if n_head_kv is not None and n_head != n_head_kv:

convert_hf_to_gguf.py

@@ -12,6 +12,7 @@
 from math import prod
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
+from transformers import AutoConfig
 
 import torch
 
@@ -256,8 +257,8 @@ def parse_args() -> argparse.Namespace:
         help="only print out what will be done, without writing any new files",
     )
     parser.add_argument(
-        "--base", type=Path, required=True,
-        help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required",
+        "--base", type=Path,
+        help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
     )
     parser.add_argument(
         "lora_path", type=Path,
@@ -267,6 +268,12 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
+def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
+    # normally, adapter does not come with base model config, we need to load it from AutoConfig
+    config = AutoConfig.from_pretrained(hf_model_id)
+    return config.to_dict()
+
+
 if __name__ == '__main__':
     args = parse_args()
     logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
@@ -281,7 +288,7 @@ def parse_args() -> argparse.Namespace:
 
     ftype = ftype_map[args.outtype]
 
-    dir_base_model: Path = args.base
+    dir_base_model: Path | None = args.base
     dir_lora: Path = args.lora_path
     lora_config = dir_lora / "adapter_config.json"
     input_model = dir_lora / "adapter_model.safetensors"
@@ -301,9 +308,29 @@ def parse_args() -> argparse.Namespace:
         input_model = os.path.join(dir_lora, "adapter_model.bin")
         lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
 
+    # load LoRA config
+    with open(lora_config, "r") as f:
+        lparams: dict[str, Any] = json.load(f)
+
     # load base model
-    logger.info(f"Loading base model: {dir_base_model.name}")
-    hparams = Model.load_hparams(dir_base_model)
+    if dir_base_model is None:
+        if "base_model_name_or_path" in lparams:
+            model_id = lparams["base_model_name_or_path"]
+            logger.info(f"Loading base model from Hugging Face: {model_id}")
+            try:
+                hparams = load_hparams_from_hf(model_id)
+            except OSError as e:
+                logger.error(f"Failed to load base model config: {e}")
+                logger.error("Please try downloading the base model and add its path to --base")
+                sys.exit(1)
+        else:
+            logger.error("'base_model_name_or_path' is not found in adapter_config.json")
+            logger.error("Base model config is required. Please download the base model and add its path to --base")
+            sys.exit(1)
+    else:
+        logger.info(f"Loading base model: {dir_base_model.name}")
+        hparams = Model.load_hparams(dir_base_model)
+
     with torch.inference_mode():
         try:
             model_class = Model.from_model_architecture(hparams["architectures"][0])
@@ -323,13 +350,15 @@ def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):
                 self.dir_model_card = dir_lora_model
                 self.lora_alpha = float(lora_alpha)
 
+            def set_vocab(self):
+                pass
+
             def set_type(self):
                 self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
                 self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
 
             def set_gguf_parameters(self):
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
-                super().set_gguf_parameters()
 
             def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
                 # Never add extra tensors (e.g. rope_freqs) for LoRA adapters
@@ -350,7 +379,7 @@ def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
                         if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
                             logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
-                            logger.error("Hint: if you are using TRL, make sure not to call setup_chat_format()")
+                            logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
                         sys.exit(1)
 
                     if base_name in tensor_map:
@@ -384,9 +413,6 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
                     yield (dest_name + ".lora_a", lora_a)
                     yield (dest_name + ".lora_b", lora_b)
 
-        with open(lora_config, "r") as f:
-            lparams: dict[str, Any] = json.load(f)
-
         alpha: float = lparams["lora_alpha"]
 
         model_instance = LoraModel(
@@ -399,6 +425,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             dry_run=args.dry_run,
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
+            hparams=hparams,
         )
 
         logger.info("Exporting model...")

convert_lora_to_gguf.py

@@ -49,6 +49,7 @@ else()
     endif()
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(simple-chat)
     add_subdirectory(speculative)
     add_subdirectory(tokenize)
 endif()

examples/CMakeLists.txt

@@ -247,6 +247,7 @@ struct server_slot {
         if (is_processing()) {
             SLT_INF(*this, "stop processing: n_past = %d, truncated = %d\n", n_past, truncated);
 
+            t_last_used = ggml_time_us();
             t_token_generation = (ggml_time_us() - t_start_generation) / 1e3;
             state = SLOT_STATE_IDLE;
             callback_on_release(id);
@@ -725,12 +726,12 @@ struct server_context {
         return nullptr;
     }
 
-    server_slot * get_available_slot(const std::string & prompt) {
+    server_slot * get_available_slot(const server_task & task) {
         server_slot * ret = nullptr;
 
         // find the slot that has at least n% prompt similarity
-        if (ret == nullptr && slot_prompt_similarity != 0.0f && !prompt.empty()) {
-            int max_lcp_len = 0;
+        if (ret == nullptr && slot_prompt_similarity != 0.0f) {
+            int lcs_len = 0;
             float similarity = 0;
 
             for (server_slot & slot : slots) {
@@ -740,25 +741,26 @@ struct server_context {
                 }
 
                 // skip the slot if it does not contains cached tokens
-                if (slot.prompt_tokens.empty()) {
+                if (slot.cache_tokens.empty()) {
                     continue;
                 }
 
-                // length of the Longest Common Prefix between the current slot's prompt and the input prompt
-                int lcp_len = longest_common_prefix(slot.cache_tokens, slot.prompt_tokens);
+                // length of the Longest Common Subsequence between the current slot's prompt and the input prompt
+                int cur_lcs_len = longest_common_subsequence(slot.cache_tokens, task.prompt_tokens);
 
-                // fraction of the common substring length compared to the current slot's prompt length
-                similarity = static_cast<float>(lcp_len) / static_cast<int>(slot.prompt_tokens.size());
+                // fraction of the common subsequence length compared to the current slot's prompt length
+                float cur_similarity = static_cast<float>(cur_lcs_len) / static_cast<int>(slot.cache_tokens.size());
 
                 // select the current slot if the criteria match
-                if (lcp_len > max_lcp_len && similarity > slot_prompt_similarity) {
-                    max_lcp_len = lcp_len;
+                if (cur_lcs_len > lcs_len && cur_similarity > slot_prompt_similarity) {
+                    lcs_len = cur_lcs_len;
+                    similarity = cur_similarity;
                     ret = &slot;
                 }
             }
 
             if (ret != nullptr) {
-                SLT_DBG(*ret, "selected slot by lcp similarity, max_lcp_len = %d, similarity = %f\n", max_lcp_len, similarity);
+                SLT_DBG(*ret, "selected slot by lcs similarity, lcs_len = %d, similarity = %f\n", lcs_len, similarity);
             }
         }
 
@@ -1514,18 +1516,7 @@ struct server_context {
                 {
                     const int id_slot = json_value(task.data, "id_slot", -1);
 
-                    server_slot * slot;
-
-                    if (id_slot != -1) {
-                        slot = get_slot_by_id(id_slot);
-                    } else {
-                        std::string prompt;
-                        if (task.data.contains("prompt") && task.data.at("prompt").is_string()) {
-                            prompt = json_value(task.data, "prompt", std::string());
-                        }
-
-                        slot = get_available_slot(prompt);
-                    }
+                    server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task);
 
                     if (slot == nullptr) {
                         // if no slot is available, we defer this task for processing later
@@ -2714,8 +2705,8 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) {
-        if (ctx_server.params.embedding || ctx_server.params.reranking) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
+        if (ctx_server.params.embedding) {
+            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2820,8 +2811,8 @@ int main(int argc, char ** argv) {
 
     // TODO: maybe merge this function with "handle_completions_generic"
     const auto handle_chat_completions = [&ctx_server, &params, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) {
-        if (ctx_server.params.embedding || ctx_server.params.reranking) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
+        if (ctx_server.params.embedding) {
+            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2946,11 +2937,6 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        // TODO: somehow clean up this checks in the future
-        if (!ctx_server.params.embedding || ctx_server.params.reranking) {
-            res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings` and without `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
-            return;
-        }
         const json body = json::parse(req.body);
         bool is_openai = false;
 
@@ -3002,10 +2988,11 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params.reranking) {
-            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
+        if (!ctx_server.params.reranking || ctx_server.params.embedding) {
+            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
+
         const json body = json::parse(req.body);
 
         // TODO: implement