straightforward /v1/models endpoint

* Update README.md to use PATH for Windows ROCm

* Update README.md

* Update README.md

README.md

@@ -10,7 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- *No hot topics atm. Open to suggestions about what is hot today*
+- Collecting Apple Silicon performance stats: https://github.com/ggerganov/llama.cpp/discussions/4167
 
 ----
 
@@ -410,19 +410,28 @@ Building the program with BLAS support may lead to some performance improvements
   This provides BLAS acceleration on HIP-supported AMD GPUs.
   Make sure to have ROCm installed.
   You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
-  Windows support is coming soon...
 
   - Using `make`:
     ```bash
     make LLAMA_HIPBLAS=1
     ```
-  - Using `CMake`:
+  - Using `CMake` for Linux:
     ```bash
     mkdir build
     cd build
     CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
     cmake --build .
     ```
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS):
+    ```bash
+    set PATH=%HIP_PATH%\bin;%PATH%
+    mkdir build
+    cd build
+    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ..
+    cmake --build .
+    ```
+    Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+
 
   The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
   If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.

common/common.cpp

@@ -12,6 +12,7 @@
 #include <regex>
 #include <sstream>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 #include <cinttypes>
@@ -495,6 +496,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.chatml = true;
         } else if (arg == "--infill") {
             params.infill = true;
+        } else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
+            params.dump_kv_cache = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -835,6 +838,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #endif // GGML_USE_CUBLAS
 #endif
     printf("  --verbose-prompt      print prompt before generation\n");
+    printf("  -dkvc, --dump-kv-cache\n");
+    printf("                        verbose print of the KV cache\n");
     printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
@@ -1386,3 +1391,77 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }
+
+//
+// KV cache utils
+//
+
+void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
+        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        int seq_count = 0;
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] >= 0) { seq_count++; }
+        }
+        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
+void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
+        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    std::unordered_map<llama_seq_id, size_t> seqs;
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] < 0) { continue; }
+            if (seqs.find(cs_curr[j]) == seqs.end()) {
+                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+                seqs[cs_curr[j]] = seqs.size();
+            }
+        }
+        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+    }
+
+    printf("=== Sequence legend: ");
+    for (const auto & it : seqs) {
+        printf("%zu=%d, ", it.second, it.first);
+    }
+    printf("'+'=other sequence ids");
+
+    c_curr = view.cells;
+    cs_curr = view.cells_sequences;
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] >= 0) {
+                const auto & it = seqs.find(cs_curr[j]);
+                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
+            } else {
+                putchar('.');
+            }
+        }
+        putchar(' ');
+    }
+
+    printf("\n=== Done dumping\n");
+}

common/common.h

@@ -122,6 +122,7 @@ struct gpt_params {
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool infill            = false; // use infill mode
+    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
 
     // multimodal models (see examples/llava)
     std::string mmproj = ""; // path to multimodal projector
@@ -218,3 +219,13 @@ std::string get_sortable_timestamp();
 void dump_non_result_info_yaml(
     FILE * stream, const gpt_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
+
+//
+// KV cache utils
+//
+
+// Dump the KV cache view with the number of sequences per cell.
+void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
+
+// Dump the KV cache view showing individual sequences in each cell (long output).
+void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);

examples/finetune/README.md

@@ -21,7 +21,7 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
 ./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
 ```
 
-Finetune output files will be saved every N iterations (config with `--save-every N`).
+**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
 The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
 So in above example after 10 iterations these files will be written:
 - chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf

examples/parallel/parallel.cpp

@@ -1,5 +1,5 @@
 // A basic application simulating a server with multiple clients.
-// The clients submite requests to the server and they are processed in parallel.
+// The clients submit requests to the server and they are processed in parallel.
 
 #include "common.h"
 #include "llama.h"
@@ -113,6 +113,8 @@ int main(int argc, char ** argv) {
     // insert new requests as soon as the previous one is done
     const bool cont_batching = params.cont_batching;
 
+    const bool dump_kv_cache = params.dump_kv_cache;
+
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("parallel", "log"));
     LOG_TEE("Log start\n");
@@ -172,6 +174,8 @@ int main(int argc, char ** argv) {
     int32_t n_total_gen    = 0;
     int32_t n_cache_miss   = 0;
 
+    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients);
+
     const auto t_main_start = ggml_time_us();
 
     LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
@@ -201,6 +205,11 @@ int main(int argc, char ** argv) {
     LOG_TEE("Processing requests ...\n\n");
 
     while (true) {
+        if (dump_kv_cache) {
+            llama_kv_cache_view_update(ctx, &kvc_view);
+            dump_kv_cache_view_seqs(kvc_view, 40);
+        }
+
         llama_batch_clear(batch);
 
         // decode any currently ongoing sequences

examples/server/README.md

@@ -234,6 +234,55 @@ node index.js
 
 -   **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
 
+-   **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
+
+    *Options:*
+
+    See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such are `mirostat` are supported.
+
+    *Examples:*
+
+    You can use either Python `openai` library with appropriate checkpoints:
+
+    ```python
+    import openai
+
+    client = openai.OpenAI(
+        base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+        api_key = "sk-no-key-required"
+    )
+
+    completion = client.chat.completions.create(
+    model="gpt-3.5-turbo",
+    messages=[
+        {"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
+        {"role": "user", "content": "Write a limerick about python exceptions"}
+    ]
+    )
+
+    print(completion.choices[0].message)
+    ```
+    ... or raw HTTP requests:
+
+    ```shell
+    curl http://localhost:8080/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -H "Authorization: Bearer no-key" \
+    -d '{
+    "model": "gpt-3.5-turbo",
+    "messages": [
+    {
+        "role": "system",
+        "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
+    },
+    {
+        "role": "user",
+        "content": "Write a limerick about python exceptions"
+    }
+    ]
+    }'
+    ```
+
 ## More examples
 
 ### Change system prompt on runtime

examples/server/server.cpp

@@ -29,6 +29,8 @@
 #define SERVER_VERBOSE 1
 #endif
 
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+
 using json = nlohmann::json;
 
 struct server_params
@@ -59,6 +61,10 @@ static bool server_verbose = false;
 #define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
 #define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
 
+json oaicompat_completion_params_parse(const json &body);
+std::string format_chatml(std::vector<json> messages);
+
+
 //
 // base64 utils (TODO: move to common in the future)
 //
@@ -378,6 +384,9 @@ struct llama_client_slot
     bool stopped_word = false;
     bool stopped_limit = false;
 
+    bool oaicompat = false;
+    std::string oaicompat_model;
+
     std::string stopping_word;
 
     // sampling
@@ -477,7 +486,7 @@ struct llama_client_slot
         };
     }
 
-    void print_timings() {
+    void print_timings() const {
         LOG_TEE("\n");
         LOG_TEE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, t_prompt_processing, num_prompt_tokens_processed, t_prompt_processing / num_prompt_tokens_processed, 1e3 / t_prompt_processing * num_prompt_tokens_processed);
@@ -609,6 +618,11 @@ struct llama_server_context
 
     std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
     {
+        // TODO: currently, we tokenize using special tokens by default
+        //       this is not always correct (see https://github.com/ggerganov/llama.cpp/pull/4160#issuecomment-1824826216)
+        //       but it's better compared to completely ignoring ChatML and other chat templates
+        const bool TMP_FORCE_SPECIAL = true;
+
         // If `add_bos` is true, we only add BOS, when json_prompt is a string,
         // or the first element of the json_prompt array is a string.
         std::vector<llama_token> prompt_tokens;
@@ -624,12 +638,12 @@ struct llama_server_context
                     std::vector<llama_token> p;
                     if (first)
                     {
-                        p = ::llama_tokenize(ctx, s, add_bos);
+                        p = ::llama_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
                         first = false;
                     }
                     else
                     {
-                        p = ::llama_tokenize(ctx, s, false);
+                        p = ::llama_tokenize(ctx, s, false, TMP_FORCE_SPECIAL);
                     }
                     prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
                 }
@@ -646,7 +660,7 @@ struct llama_server_context
         else
         {
             auto s = json_prompt.template get<std::string>();
-            prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
+            prompt_tokens = ::llama_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
         }
 
         return prompt_tokens;
@@ -677,6 +691,14 @@ struct llama_server_context
         slot_params default_params;
         llama_sampling_params default_sparams;
 
+        if (data.count("__oaicompat") != 0) {
+            slot->oaicompat = true;
+            slot->oaicompat_model = json_value(data, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+        } else {
+            slot->oaicompat = false;
+            slot->oaicompat_model = "";
+        }
+
         slot->params.stream           = json_value(data, "stream",            false);
         slot->params.cache_prompt     = json_value(data, "cache_prompt",      false);
         slot->params.n_predict        = json_value(data, "n_predict",         default_params.n_predict);
@@ -1095,6 +1117,7 @@ struct llama_server_context
         std::lock_guard<std::mutex> lock(mutex_results);
         task_result res;
         res.id = id;
+        res.stop = false;
         res.error = true;
         res.result_json = { { "content", error } };
         queue_results.push_back(res);
@@ -1169,6 +1192,12 @@ struct llama_server_context
             res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output);
         }
 
+        if (slot.oaicompat)
+        {
+            res.result_json["oaicompat_token_ctr"] = slot.n_decoded;
+            res.result_json["model"] = slot.oaicompat_model;
+        }
+
         queue_results.push_back(res);
     }
 
@@ -1216,6 +1245,12 @@ struct llama_server_context
             res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs);
         }
 
+        if (slot.oaicompat)
+        {
+            res.result_json["oaicompat_token_ctr"] = slot.n_decoded;
+            res.result_json["model"] = slot.oaicompat_model;
+        }
+
         queue_results.push_back(res);
     }
 
@@ -1255,7 +1290,8 @@ struct llama_server_context
         std::lock_guard<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
-        task.data = data;
+        task.target_id = 0;
+        task.data = std::move(data);
         task.infill_mode = infill;
         task.embedding_mode = embedding;
         task.type = COMPLETION_TASK;
@@ -2178,6 +2214,233 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
     }
 }
 
+
+static std::string random_string()
+{
+    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
+
+    std::random_device rd;
+    std::mt19937 generator(rd());
+
+    std::string result(32, ' ');
+
+    for (int i = 0; i < 32; ++i) {
+        result[i] = str[generator() % str.size()];
+    }
+
+    return result;
+}
+
+static std::string gen_chatcmplid()
+{
+    std::stringstream chatcmplid;
+    chatcmplid << "chatcmpl-" << random_string();
+    return chatcmplid.str();
+}
+
+std::string format_chatml(std::vector<json> messages)
+{
+    std::ostringstream chatml_msgs;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
+    }
+
+    chatml_msgs << "<|im_start|>assistant" << '\n';
+
+    return chatml_msgs.str();
+}
+
+/* llama.cpp completion api semantics */
+json oaicompat_completion_params_parse(
+    const json &body /* openai api json semantics */)
+{
+    json llama_params;
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["temperature"]       = json_value(body, "temperature", 0.8);
+    llama_params["top_k"]             = json_value(body, "top_k", 40);
+    llama_params["top_p"]             = json_value(body, "top_p", 0.95);
+    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
+    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
+    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
+    llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
+    llama_params["seed"]              = json_value(body, "seed", 0);
+    llama_params["stream"]            = json_value(body, "stream", false);
+    llama_params["mirostat"]          = json_value(body, "mirostat", false);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", 0.0);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", 0.0);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", false);
+    llama_params["typical_p"]         = json_value(body, "typical_p", 0.0);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", 0);
+    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", 0.0);
+
+    if (llama_params.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body["stop"].is_null()) {
+        llama_params["stop"] = json::array({});
+    } else if (body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
+{
+    json result = response.result_json;
+
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res =
+        json{{"choices", choices},
+            {"created", t},
+            {"model",
+                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+            {"usage",
+                json{{"completion_tokens", num_tokens_predicted},
+                    {"prompt_tokens", num_prompt_tokens},
+                    {"total_tokens", num_tokens_predicted + num_prompt_tokens}}},
+            {"id", gen_chatcmplid()}};
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
+    json result = response.result_json;
+
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({response.result_json});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word", false);
+    bool stopped_eos    = json_value(result, "stopped_eos", false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content", std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json{{"choices", choices},
+                    {"created", t},
+                    {"id", gen_chatcmplid()},
+                    {"model", modelname},
+                    {"object", "chat.completion.chunk"}};
+
+    return std::vector<json>({ret});
+}
+
 static json format_partial_response(
     llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector<completion_token_output> &probs
 ) {
@@ -2354,9 +2617,9 @@ int main(int argc, char **argv)
                             task_result result = llama.next_result(task_id);
                             if (!result.error) {
                                 const std::string str =
-                                "data: " +
-                                result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
-                                "\n\n";
+                                    "data: " +
+                                    result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
+                                    "\n\n";
                                 LOG_VERBOSE("data stream", {
                                     { "to_send", str }
                                 });
@@ -2369,9 +2632,9 @@ int main(int argc, char **argv)
                                 }
                             } else {
                                 const std::string str =
-                                "error: " +
-                                result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
-                                "\n\n";
+                                    "error: " +
+                                    result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
+                                    "\n\n";
                                 LOG_VERBOSE("data stream", {
                                     { "to_send", str }
                                 });
@@ -2396,6 +2659,98 @@ int main(int argc, char **argv)
                 }
             });
 
+
+
+    svr.Get("/v1/models", [&params](const httplib::Request&, httplib::Response& res)
+            {
+                std::time_t t = std::time(0);
+
+                json models = {
+                    {"object", "list"},
+                    {"data", {
+                        {
+                            {"id", params.model_alias},
+                            {"object", "model"},
+                            {"created", t},
+                            {"owned_by", "llamacpp"}
+                        },
+                    }}
+                };
+
+                res.set_content(models.dump(), "application/json");
+            });
+
+    // TODO: add mount point without "/v1" prefix -- how?
+    svr.Post("/v1/chat/completions", [&llama](const httplib::Request &req, httplib::Response &res)
+            {
+                json data = oaicompat_completion_params_parse(json::parse(req.body));
+
+                const int task_id = llama.request_completion(data, false, false);
+
+                if (!json_value(data, "stream", false)) {
+                    std::string completion_text;
+                    task_result result = llama.next_result(task_id);
+
+                    if (!result.error && result.stop) {
+                        json oaicompat_result = format_final_response_oaicompat(data, result);
+
+                        res.set_content(oaicompat_result.dump(-1, ' ', false,
+                                            json::error_handler_t::replace),
+                                            "application/json");
+                    } else {
+                        res.status = 500;
+                        res.set_content(result.result_json["content"], "text/plain");
+                        return;
+                    }
+                } else {
+                    const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink &sink) {
+                        while (true) {
+                            task_result llama_result = llama.next_result(task_id);
+                            if (!llama_result.error) {
+                                std::vector<json> result_array = format_partial_response_oaicompat( llama_result);
+
+                                for (auto it = result_array.begin(); it != result_array.end(); ++it)
+                                {
+                                    if (!it->empty()) {
+                                        const std::string str =
+                                            "data: " +
+                                            it->dump(-1, ' ', false, json::error_handler_t::replace) +
+                                            "\n\n";
+                                        LOG_VERBOSE("data stream", {{"to_send", str}});
+                                        if (!sink.write(str.c_str(), str.size())) {
+                                            return false;
+                                        }
+                                    }
+                                }
+                                if (llama_result.stop) {
+                                    break;
+                                }
+                            } else {
+                                const std::string str =
+                                    "error: " +
+                                    llama_result.result_json.dump(-1, ' ', false,
+                                            json::error_handler_t::replace) +
+                                    "\n\n";
+                                LOG_VERBOSE("data stream", {{"to_send", str}});
+                                if (!sink.write(str.c_str(), str.size())) {
+                                    return false;
+                                }
+                                break;
+                            }
+                        }
+                        sink.done();
+                        return true;
+                    };
+
+                    auto on_complete = [task_id, &llama](bool) {
+                        // cancel request
+                        llama.request_cancel(task_id);
+                    };
+
+                    res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
+                }
+            });
+
     svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res)
             {
                 json data = json::parse(req.body);

ggml-cuda.cu

@@ -1,4 +1,5 @@
 #include <algorithm>
+#include <cinttypes>
 #include <cstddef>
 #include <cstdint>
 #include <limits>
@@ -8057,7 +8058,7 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
     if (tensor->op == GGML_OP_MUL_MAT) {
         if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %d, src1->ne[3] = %d - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
+            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = " PRId64 ", src1->ne[3] = " PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
 #endif
             return false;
         }

llama.cpp

@@ -1280,6 +1280,7 @@ struct llama_kv_cache {
     // cannot be freely changed after a slot has been allocated.
     uint32_t head = 0;
     uint32_t size = 0;
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
 
     // computed before each graph build
     uint32_t n = 0;
@@ -1504,6 +1505,7 @@ static bool llama_kv_cache_init(
 
     cache.head = 0;
     cache.size = n_ctx;
+    cache.used = 0;
 
     cache.cells.clear();
     cache.cells.resize(n_ctx);
@@ -1605,6 +1607,8 @@ static bool llama_kv_cache_find_slot(
         }
     }
 
+    cache.used += n_tokens;
+
     return true;
 }
 
@@ -1625,6 +1629,7 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) {
         cache.cells[i].seq_id.clear();
     }
     cache.head = 0;
+    cache.used = 0;
 }
 
 static void llama_kv_cache_seq_rm(
@@ -1647,6 +1652,9 @@ static void llama_kv_cache_seq_rm(
                 continue;
             }
             if (cache.cells[i].seq_id.empty()) {
+                // keep count of the number of used cells
+                if (cache.cells[i].pos >= 0) cache.used--;
+
                 cache.cells[i].pos = -1;
                 if (new_head == cache.size) new_head = i;
             }
@@ -1654,7 +1662,7 @@ static void llama_kv_cache_seq_rm(
     }
 
     // If we freed up a slot, set head to it so searching can start there.
-    if (new_head != cache.size) cache.head = new_head;
+    if (new_head != cache.size && new_head < cache.head) cache.head = new_head;
 }
 
 static void llama_kv_cache_seq_cp(
@@ -1680,6 +1688,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id
 
     for (uint32_t i = 0; i < cache.size; ++i) {
         if (!cache.cells[i].has_seq_id(seq_id)) {
+            if (cache.cells[i].pos >= 0) cache.used--;
             cache.cells[i].pos = -1;
             cache.cells[i].seq_id.clear();
             if (new_head == cache.size) new_head = i;
@@ -1690,7 +1699,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id
     }
 
     // If we freed up a slot, set head to it so searching can start there.
-    if (new_head != cache.size) cache.head = new_head;
+    if (new_head != cache.size && new_head < cache.head) cache.head = new_head;
 }
 
 static void llama_kv_cache_seq_shift(
@@ -1711,6 +1720,7 @@ static void llama_kv_cache_seq_shift(
             cache.cells[i].delta += delta;
 
             if (cache.cells[i].pos < 0) {
+                if (!cache.cells[i].seq_id.empty()) cache.used--;
                 cache.cells[i].pos = -1;
                 cache.cells[i].seq_id.clear();
                 if (new_head == cache.size) new_head = i;
@@ -4814,92 +4824,34 @@ struct llm_build_context {
             // self-attention
             {
                 // compute Q and K and RoPE them
-                struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                cb(tmpq, "tmpq", il);
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
 
-                struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-                cb(tmpk, "tmpk", il);
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
 
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
 
-                // RoPE the first n_rot of q/k, pass the other half, and concat.
-                struct ggml_tensor * qrot = ggml_cont(ctx0, ggml_view_3d(
-                        ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
-                        ggml_element_size(tmpq) * n_embd_head,
-                        ggml_element_size(tmpq) * n_embd_head * n_head,
-                        0
-                        ));
-                cb(qrot, "qrot", il);
-
-                struct ggml_tensor * krot = ggml_cont(ctx0, ggml_view_3d(
-                        ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
-                        ggml_element_size(tmpk) * n_embd_head,
-                        ggml_element_size(tmpk) * n_embd_head * n_head_kv,
-                        0
-                        ));
-                cb(krot, "krot", il);
-
-                // get the second half of tmpq, e.g tmpq[n_rot:, :, :]
-                struct ggml_tensor * qpass = ggml_view_3d(
-                        ctx0, tmpq, (n_embd_head - hparams.n_rot), n_head, n_tokens,
-                        ggml_element_size(tmpq) * n_embd_head,
-                        ggml_element_size(tmpq) * n_embd_head * n_head,
-                        ggml_element_size(tmpq) * hparams.n_rot
-                        );
-                cb(qpass, "qpass", il);
-
-                struct ggml_tensor * kpass = ggml_view_3d(
-                        ctx0, tmpk, (n_embd_head - hparams.n_rot), n_head_kv, n_tokens,
-                        ggml_element_size(tmpk) * (n_embd_head),
-                        ggml_element_size(tmpk) * (n_embd_head) * n_head_kv,
-                        ggml_element_size(tmpk) * hparams.n_rot
-                        );
-                cb(kpass, "kpass", il);
-
-                struct ggml_tensor * qrotated = ggml_rope_custom(
-                    ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
                 );
-                cb(qrotated, "qrotated", il);
-
-                struct ggml_tensor * krotated = ggml_rope_custom(
-                    ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(krotated, "krotated", il);
-
-                // ggml currently only supports concatenation on dim=2
-                // so we need to permute qrot, qpass, concat, then permute back.
-                qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3));
-                cb(qrotated, "qrotated", il);
-
-                krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3));
-                cb(krotated, "krotated", il);
-
-                qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3));
-                cb(qpass, "qpass", il);
-
-                kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3));
-                cb(kpass, "kpass", il);
-
-                struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass);
                 cb(Qcur, "Qcur", il);
 
-                struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3));
-                cb(Q, "Q", il);
-
-                Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3));
+                Kcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
                 cb(Kcur, "Kcur", il);
 
                 llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
 
                 cur = llm_build_kqv(ctx0, hparams, kv_self,
                         model.layers[il].wo, NULL,
-                        Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5527,6 +5479,12 @@ static int llama_decode_internal(
         batch.seq_id = seq_id_arr.data();
     }
 
+    // if we have enough unused cells before the current head ->
+    //   better to start searching from the beginning of the cache, hoping to fill it
+    if (kv_self.head > kv_self.used + 2*n_tokens) {
+        kv_self.head = 0;
+    }
+
     if (!llama_kv_cache_find_slot(kv_self, batch)) {
         return 1;
     }
@@ -5537,7 +5495,7 @@ static int llama_decode_internal(
     //kv_self.n = std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32));   // TODO: this might be better for CUDA?
     kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, llama_kv_cache_cell_max(kv_self)));
 
-    //printf("kv_self.n = %d\n", kv_self.n);
+    //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
 
     ggml_allocr_reset(lctx.alloc);
 
@@ -8847,8 +8805,107 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha
     }
 }
 
+struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq) {
+    struct llama_kv_cache_view result = {
+        /*.n_cells            = */ 0,
+        /*.n_max_seq          = */ n_max_seq,
+        /*.token_count        = */ 0,
+        /*.used_cells         = */ llama_get_kv_cache_used_cells(ctx),
+        /*.max_contiguous     = */ 0,
+        /*.max_contiguous_idx = */ -1,
+        /*.cells              = */ nullptr,
+        /*.cells_sequences    = */ nullptr,
+    };
+    return result;
+}
+
+void llama_kv_cache_view_free(struct llama_kv_cache_view * view) {
+    if (view->cells != nullptr) {
+        free(view->cells);
+        view->cells = nullptr;
+    }
+    if (view->cells_sequences != nullptr) {
+        free(view->cells_sequences);
+        view->cells_sequences = nullptr;
+    }
+}
+
+void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) {
+    if (uint32_t(view->n_cells) < ctx->kv_self.size || view->cells == nullptr) {
+        view->n_cells = int32_t(ctx->kv_self.size);
+        void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells);
+        GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells");
+        view->cells = (struct llama_kv_cache_view_cell *)p;
+        p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_max_seq * view->n_cells);
+        GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences");
+        view->cells_sequences = (llama_seq_id *)p;
+    }
+
+    const std::vector<llama_kv_cell> & kv_cells = ctx->kv_self.cells;
+    llama_kv_cache_view_cell * c_curr = view->cells;
+    llama_seq_id * cs_curr = view->cells_sequences;
+    int32_t used_cells = 0;
+    int32_t token_count = 0;
+    int32_t curr_contig_idx = -1;
+    uint32_t max_contig = 0;
+    int32_t max_contig_idx = -1;
+
+    for (int32_t i = 0; i < int32_t(ctx->kv_self.size); i++, c_curr++, cs_curr += view->n_max_seq) {
+        const size_t curr_size = kv_cells[i].seq_id.size();
+        token_count += curr_size;
+        c_curr->pos = kv_cells[i].pos + kv_cells[i].delta;
+
+        if (curr_size > 0) {
+            if (curr_contig_idx >= 0 && uint32_t(i - curr_contig_idx) > max_contig) {
+                max_contig = i - curr_contig_idx;
+                max_contig_idx = curr_contig_idx;
+            }
+            curr_contig_idx = -1;
+        } else if (curr_contig_idx < 0) {
+            curr_contig_idx = i;
+        }
+
+        int seq_idx = 0;
+        for (const llama_seq_id it : kv_cells[i].seq_id) {
+            if (seq_idx >= view->n_max_seq) {
+                break;
+            }
+            cs_curr[seq_idx] = it;
+            seq_idx++;
+        }
+        if (seq_idx != 0) {
+            used_cells++;
+        }
+        for (; seq_idx < view->n_max_seq; seq_idx++) {
+            cs_curr[seq_idx] = -1;
+        }
+    }
+    if (curr_contig_idx >= 0 && kv_cells.size() - curr_contig_idx > max_contig) {
+        max_contig_idx = curr_contig_idx;
+        max_contig = kv_cells.size() - curr_contig_idx;
+    }
+    view->max_contiguous = max_contig;
+    view->max_contiguous_idx = max_contig_idx;
+    view->token_count = token_count;
+    view->used_cells = used_cells;
+    if (uint32_t(used_cells) != ctx->kv_self.used) {
+        LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n",
+            __func__, ctx->kv_self.used, used_cells);
+    }
+}
+
 int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
-    return ctx->kv_self.head;
+    int result = 0;
+
+    for (uint32_t i = 0; i < ctx->kv_self.size; i++) {
+        result += ctx->kv_self.cells[i].seq_id.size();
+    }
+
+    return result;
+}
+
+int llama_get_kv_cache_used_cells(const struct llama_context * ctx) {
+    return ctx->kv_self.used;
 }
 
 void llama_kv_cache_clear(struct llama_context * ctx) {
@@ -9018,10 +9075,12 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
         const size_t   kv_buf_size = kv_self.buf.size;
         const uint32_t kv_head     = kv_self.head;
         const uint32_t kv_size     = kv_self.size;
+        const uint32_t kv_used     = kv_self.used;
 
         data_ctx->write(&kv_buf_size, sizeof(kv_buf_size));
         data_ctx->write(&kv_head,     sizeof(kv_head));
         data_ctx->write(&kv_size,     sizeof(kv_size));
+        data_ctx->write(&kv_used,     sizeof(kv_used));
 
         if (kv_buf_size) {
             const size_t elt_size = ggml_element_size(kv_self.k);
@@ -9144,10 +9203,12 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         size_t   kv_buf_size;
         uint32_t kv_head;
         uint32_t kv_size;
+        uint32_t kv_used;
 
         memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size);
         memcpy(&kv_head,     inp, sizeof(kv_head));     inp += sizeof(kv_head);
         memcpy(&kv_size,     inp, sizeof(kv_size));     inp += sizeof(kv_size);
+        memcpy(&kv_used,     inp, sizeof(kv_used));     inp += sizeof(kv_used);
 
         if (kv_buf_size) {
             GGML_ASSERT(kv_self.buf.size == kv_buf_size);
@@ -9182,6 +9243,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
         ctx->kv_self.head = kv_head;
         ctx->kv_self.size = kv_size;
+        ctx->kv_self.used = kv_used;
 
         ctx->kv_self.cells.resize(kv_size);
 

llama.h

@@ -361,9 +361,60 @@ extern "C" {
     // KV cache
     //
 
-    // Returns the number of tokens in the KV cache
-    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
-            "avoid using this, it will be removed in the future, instead - count the tokens in user code");
+    // Information associated with an individual cell in the KV cache view.
+    struct llama_kv_cache_view_cell {
+        // The position for this cell. Takes KV cache shifts into account.
+        // May be negative if the cell is not populated.
+        llama_pos pos;
+    };
+
+    // An updateable view of the KV cache.
+    struct llama_kv_cache_view {
+        // Number of KV cache cells. This will be the same as the context size.
+        int32_t n_cells;
+
+        // Maximum number of sequences that can exist in a cell. It's not an error
+        // if there are more sequences in a cell than this value, however they will
+        // not be visible in the view cells_sequences.
+        int32_t n_max_seq;
+
+        // Number of tokens in the cache. For example, if there are two populated
+        // cells, the first with 1 sequence id in it and the second with 2 sequence
+        // ids then you'll have 3 tokens.
+        int32_t token_count;
+
+        // Number of populated cache cells.
+        int32_t used_cells;
+
+        // Maximum contiguous empty slots in the cache.
+        int32_t max_contiguous;
+
+        // Index to the start of the max_contiguous slot range. Can be negative
+        // when cache is full.
+        int32_t max_contiguous_idx;
+
+        // Information for an individual cell.
+        struct llama_kv_cache_view_cell * cells;
+
+        // The sequences for each cell. There will be n_max_seq items per cell.
+        llama_seq_id * cells_sequences;
+    };
+
+    // Create an empty KV cache view. (use only for debugging purposes)
+    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq);
+
+    // Free a KV cache view. (use only for debugging purposes)
+    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
+
+    // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
+    LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
+
+    // Returns the number of tokens in the KV cache (slow, use only for debug)
+    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
+    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+
+    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
+    LLAMA_API int llama_get_kv_cache_used_cells(const struct llama_context * ctx);
 
     // Clear the KV cache
     LLAMA_API void llama_kv_cache_clear(