Merge eacb0d59c2 into c4e57e0aab

WIP: implement locking of model per request

.github/workflows/build-commit.yml

@@ -92,7 +92,7 @@ jobs:
         submodules: true
 
     - name: Install Dependencies
-      run: sudo apt-get update && sudo apt-get install -yq libboost-dev libasio-dev
+      run: sudo apt-get update && sudo apt-get install -yq libboost-dev libasio-dev libboost-thread-dev
 
     - name: Install OpenBlas
       if: ${{ matrix.build == 'avx2-openblas' }}
@@ -207,6 +207,7 @@ jobs:
           $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
           $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
           & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
+
       - name: Build
         id: cmake_build
         env:
@@ -214,7 +215,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. ${{ matrix.defines }}
+          cmake .. ${{ matrix.defines }} -DBoost_LIBRARYDIRS=${{ steps.install-boost.outputs.BOOST_ROOT }}/lib
           cmake --build . --config Release --target turbopilot
 
       # - name: Add libopenblas.dll

.github/workflows/codacy.yml

@@ -0,0 +1,61 @@
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+# This workflow checks out code, performs a Codacy security scan
+# and integrates the results with the
+# GitHub Advanced Security code scanning feature.  For more information on
+# the Codacy security scan action usage and parameters, see
+# https://github.com/codacy/codacy-analysis-cli-action.
+# For more information on Codacy Analysis CLI in general, see
+# https://github.com/codacy/codacy-analysis-cli.
+
+name: Codacy Security Scan
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    # The branches below must be a subset of the branches above
+    branches: [ "main" ]
+  schedule:
+    - cron: '36 12 * * 3'
+
+permissions:
+  contents: read
+
+jobs:
+  codacy-security-scan:
+    permissions:
+      contents: read # for actions/checkout to fetch code
+      security-events: write # for github/codeql-action/upload-sarif to upload SARIF results
+      actions: read # only required for a private repository by github/codeql-action/upload-sarif to get the Action run status
+    name: Codacy Security Scan
+    runs-on: ubuntu-latest
+    steps:
+      # Checkout the repository to the GitHub Actions runner
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      # Execute Codacy Analysis CLI and generate a SARIF output with the security issues identified during the analysis
+      - name: Run Codacy Analysis CLI
+        uses: codacy/codacy-analysis-cli-action@d840f886c4bd4edc059706d09c6a1586111c540b
+        with:
+          # Check https://github.com/codacy/codacy-analysis-cli#project-token to get your project token from your Codacy repository
+          # You can also omit the token and run the tools that support default configurations
+          project-token: ${{ secrets.CODACY_PROJECT_TOKEN }}
+          verbose: true
+          output: results.sarif
+          format: sarif
+          # Adjust severity of non-security issues
+          gh-code-scanning-compat: true
+          # Force 0 exit code to allow SARIF file generation
+          # This will handover control about PR rejection to the GitHub side
+          max-allowed-issues: 2147483647
+
+      # Upload the SARIF file generated in the previous step
+      - name: Upload SARIF results file
+        uses: github/codeql-action/upload-sarif@v2
+        with:
+          sarif_file: results.sarif

CMakeLists.txt

@@ -1,5 +1,15 @@
 cmake_minimum_required (VERSION 3.0)
 project(turbopilot VERSION 0.1.0)
+set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+find_package(Threads REQUIRED)
+
+
+# option(BUILD_SHARED_LIBS "Build using shared libraries" OFF)
+
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS "on")
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
@@ -27,6 +37,9 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES
 endif()
 
 
+
+
+
 add_subdirectory(extern/ggml)
 add_subdirectory(extern/argparse)
 add_subdirectory(extern/spdlog)
@@ -35,8 +48,11 @@ if (GGML_STATIC)
     SET(CMAKE_FIND_LIBRARY_SUFFIXES ".a")
     SET(BUILD_SHARED_LIBS OFF)
     SET(CMAKE_EXE_LINKER_FLAGS "-static")
-endif()
 
+    # if(GGML_OPENBLAS)
+    #     set(BLA_STATIC ON)
+    # endif()
+endif()
 
 
 add_subdirectory(src)

Dockerfile.cuda11

@@ -7,7 +7,7 @@ RUN apt-get update && apt-get install ca-certificates gpg wget
 RUN wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
 RUN echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' | tee /etc/apt/sources.list.d/kitware.list >/dev/null
 
-RUN apt-get update && apt-get install -y build-essential cmake libboost-dev
+RUN apt-get update && apt-get install -y build-essential cmake libboost-dev libboost-thread-dev
 
 
 ADD ./ /turbopilot

Dockerfile.cuda12

@@ -7,7 +7,7 @@ RUN apt-get update && apt-get install ca-certificates gpg wget
 RUN wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
 RUN echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ focal main' | tee /etc/apt/sources.list.d/kitware.list >/dev/null
 
-RUN apt-get update && apt-get install -y build-essential cmake libboost-dev
+RUN apt-get update && apt-get install -y build-essential cmake libboost-dev libboost-thread-dev
 
 ADD ./ /turbopilot
 

Dockerfile.default

@@ -1,6 +1,6 @@
 FROM alpine AS build
 
-RUN apk add --update alpine-sdk boost-dev cmake asio-dev
+RUN apk add --update alpine-sdk boost-dev cmake asio-dev 
 
 ADD ./ /turbopilot/
 

MODELS.md

@@ -6,9 +6,9 @@
 
 | Model Name          | RAM Requirement | Direct Download  | HF Project Link |
 |---------------------|-----------------|-----------------|-----------------|
-| StarCoder   | ~3GiB        |   [:arrow_down:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/blob/main/stablecode-instruct-alpha-3b.ggmlv1.q4_0.bin)           |   [:hugs:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/)           |
-
+| StableCode   | ~3GiB        |   [:arrow_down:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/blob/main/stablecode-instruct-alpha-3b.ggmlv1.q4_0.bin)           |   [:hugs:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/)           |
 
+To run in Turbopilot set model type `-m stablecode`
 
 ## "Coder" family models
 
@@ -23,7 +23,7 @@ This model is primarily trained on Python, Java and Javscript.
 
 | Model Name          | RAM Requirement | Direct Download  | HF Project Link |
 |---------------------|-----------------|-----------------|-----------------|
-| StarCoder   | ~2GiB        |   [:arrow_down:](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/resolve/main/santacoder-q4_0.bin)           |   [:hugs:](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/)           |
+| SantaCoder   | ~2GiB        |   [:arrow_down:](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/resolve/main/santacoder-q4_0.bin)           |   [:hugs:](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/)           |
 
 To run in Turbopilot set model type `-m starcoder`
 
@@ -39,7 +39,7 @@ Even when quantized, WizardCoder is a large model that takes up a significant am
 |---------------------|-----------------|-----------------|-----------------|
 | WizardCoder   | ~12GiB        |   [:arrow_down:](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/resolve/main/WizardCoder-15B-1.0.ggmlv3.q4_0.bin)           |   [:hugs:](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/)           |
 
-To run in Turbopilot set model type `-m starcoder`
+To run in Turbopilot set model type `-m wizardcoder`
 
 
 ### StarCoder (Released 4/5/2023)

include/turbopilot/gptj.hpp

@@ -71,7 +71,7 @@ public:
     }
     virtual ~GPTJModel();
     bool load_model(std::string path);
-    virtual std::stringstream predict(std::string prompt, int max_length, bool include_prompt);
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt);
 
 private:
     gptj_model *model = NULL;

include/turbopilot/gptneox.hpp

@@ -75,7 +75,7 @@ public:
     }
     virtual ~GPTNEOXModel();
     bool load_model(std::string path);
-    virtual std::stringstream predict(std::string prompt, int max_length, bool include_prompt);
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt);
 
 private:
     gpt_neox_model *model = NULL;

include/turbopilot/model.hpp

@@ -7,6 +7,7 @@
 #include <map>
 #include <vector>
 #include <random>
+#include <mutex>
 
 typedef void (*offload_func_t)(struct ggml_tensor * tensor);
 void ggml_nop(struct ggml_tensor * tensor);
@@ -54,11 +55,16 @@ public:
         rng(rng) 
     {}
     virtual bool load_model(std::string model_path) = 0;
-    virtual std::stringstream predict(std::string prompt, int max_length, bool include_prompt) = 0;
+    std::stringstream predict(std::string prompt, int max_length, bool include_prompt);
+    void lock();
+    void unlock();
+
 
 protected:
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt) = 0;
     ModelConfig config;
     std::mt19937 &rng;
+    std::mutex model_lock;
 };
 
 #endif //__TURBOPILOT_MODEL_H

include/turbopilot/server.hpp

@@ -2,6 +2,8 @@
 #define __TURBOPILOT_SERVER_H
 
 
+#include <spdlog/spdlog.h>
+
 #include "turbopilot/model.hpp"
 
 #include "crow_all.h"
@@ -10,6 +12,46 @@ crow::response handle_openai_request(TurbopilotModel *model, const crow::request
 
 crow::response handle_hf_request(TurbopilotModel *model, const crow::request& req);
 
+class TBPLogger : public crow::ILogHandler {
+ public:
+  TBPLogger() {}
+  void log(std::string message, crow::LogLevel crow_level) {
+    // "message" doesn't contain the timestamp and loglevel
+    // prefix the default logger does and it doesn't end
+    // in a newline.
+
+    spdlog::level::level_enum level = spdlog::level::info;
+
+    switch(crow_level){
+        case crow::LogLevel::Critical:
+            level = spdlog::level::critical;
+            break;
+
+        case crow::LogLevel::Error:
+            level = spdlog::level::err;
+            break;
+
+        case crow::LogLevel::Warning:
+            level = spdlog::level::warn;
+            break;
+
+        case crow::LogLevel::Info:
+            level = spdlog::level::info;
+            break;
+
+        case crow::LogLevel::Debug:
+            level = spdlog::level::debug;
+            break;
+
+        default:
+            // if case is not a known value, assume the worst
+            level = spdlog::level::critical;
+    }
+
+    spdlog::log(level, message);
+  }
+};
+
 
 #endif // __TURBOPILOT_SERVER_H
 

include/turbopilot/starcoder.hpp

@@ -68,7 +68,7 @@ public:
     }
     virtual ~StarcoderModel();
     bool load_model(std::string path);
-    virtual std::stringstream predict(std::string prompt, int max_length, bool include_prompt);
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt);
 
 private:
     starcoder_model *model = NULL;

src/CMakeLists.txt

@@ -1,6 +1,7 @@
 set(TURBOPILOT_TARGET turbopilot)
 
-find_package(Boost REQUIRED)
+find_package(Boost COMPONENTS thread system REQUIRED)
+
 include_directories(${Boost_INCLUDE_DIRS})
 
 add_executable(${TURBOPILOT_TARGET} 
@@ -16,6 +17,9 @@ add_executable(${TURBOPILOT_TARGET}
   ../include/turbopilot/starcoder.hpp
   )
 
+#set(THREADS_PREFER_PTHREAD_FLAG TRUE)
+#find_package(Threads REQUIRED)
+
 
 target_include_directories(${TURBOPILOT_TARGET} PRIVATE
   ../include
@@ -23,8 +27,6 @@ target_include_directories(${TURBOPILOT_TARGET} PRIVATE
   ../extern/crow/include
 )
 
+#target_compile_features(${TURBOPILOT_TARGET} PRIVATE cxx_std_11)
 
-target_link_libraries(${TURBOPILOT_TARGET} PRIVATE ggml argparse)
-
-
-#target_link_libraries(${TURBOPILOT_TARGET} PRIVATE spdlog::spdlog_header_only)
+target_link_libraries(${TURBOPILOT_TARGET} PRIVATE ggml argparse)

src/common.cpp

@@ -4,6 +4,22 @@
 #include <cmath>
 #include <random>
 
+
+void TurbopilotModel::lock(){
+    this->model_lock.lock();
+}
+
+void TurbopilotModel::unlock(){
+    this->model_lock.unlock();
+}
+
+std::stringstream TurbopilotModel::predict(std::string prompt, int max_length, bool include_prompt){
+    lock();
+    auto result = predict_impl(prompt, max_length, include_prompt);
+    unlock();
+    return result;
+}
+
 void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
     (void) tensor;
 }
@@ -163,4 +179,6 @@ gpt_vocab::id gpt_sample_top_k_top_p(
     int idx = dist(rng);
 
     return logits_id[idx].second;
-}
+}
+
+

src/gptj.cpp

@@ -556,7 +556,7 @@ bool GPTJModel::load_model(std::string fname) {
     return true;
 }
 
-std::stringstream GPTJModel::predict(std::string prompt, int max_length, bool include_prompt) {
+std::stringstream GPTJModel::predict_impl(std::string prompt, int max_length, bool include_prompt) {
 
     std::stringstream result;
     // tokenize the prompt

src/gptneox.cpp

@@ -91,6 +91,7 @@ bool gpt_neox_eval(
         const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
         //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
 
+        
         // reallocate
         buf_size = buf_size_new;
         buf = realloc(buf, buf_size);
@@ -98,6 +99,8 @@ bool gpt_neox_eval(
             fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
             return false;
         }
+
+        spdlog::debug("{}: reallocating context buffer {} -> now {} bytes of tokens in prompt = {}", __func__, buf_size, buf_size_new);
     }
 
     struct ggml_init_params params = {
@@ -283,6 +286,7 @@ bool gpt_neox_eval(
     //    ggml_graph_print   (&gf);
     //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
     //}
+    
 
     //embd_w.resize(n_vocab*N);
     //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
@@ -293,7 +297,9 @@ bool gpt_neox_eval(
 
     if (mem_per_token == 0) {
         mem_per_token = ggml_used_mem(ctx0)/N;
+        
     }
+    spdlog::debug("used_mem = {}\n", ggml_used_mem(ctx0));
     //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
 
     ggml_free(ctx0);
@@ -612,7 +618,7 @@ bool GPTNEOXModel::load_model(std::string fname) {
     return true;
 }
 
-std::stringstream GPTNEOXModel::predict(std::string prompt, int max_length, bool include_prompt) {
+std::stringstream GPTNEOXModel::predict_impl(std::string prompt, int max_length, bool include_prompt) {
 
     std::stringstream result;
     // tokenize the prompt
@@ -631,6 +637,8 @@ std::stringstream GPTNEOXModel::predict(std::string prompt, int max_length, bool
 
     std::vector<gpt_vocab::id> embd;
 
+
+
     // determine the required inference memory per token:
     size_t mem_per_token = 0;
 
@@ -717,3 +725,4 @@ std::stringstream GPTNEOXModel::predict(std::string prompt, int max_length, bool
 
     return result;
 }
+

src/main.cpp

@@ -60,6 +60,11 @@ int main(int argc, char **argv)
         .default_value(0.1f)
         .scan<'g', float>();
 
+    program.add_argument("-b", "--batch-size")
+        .help("set batch size for model completion")
+        .default_value(512)
+        .scan<'i',int>();
+
 
     program.add_argument("prompt").remaining();
 
@@ -96,6 +101,7 @@ int main(int argc, char **argv)
     config.n_threads = program.get<int>("--threads");
     config.temp = program.get<float>("--temperature");
     config.top_p = program.get<float>("--top-p");
+    config.n_batch = program.get<int>("--batch-size");
 
     if(model_type.compare("codegen") == 0) {
         spdlog::info("Initializing GPT-J type model for '{}' model", model_type);
@@ -121,16 +127,22 @@ int main(int argc, char **argv)
     }
 
     t_load_us = ggml_time_us() - t_start_us;
+    
 
     spdlog::info("Loaded model in {:0.2f}ms", t_load_us/1000.0f);
 
 
     crow::SimpleApp app;
 
+    TBPLogger logger;
+
+    crow::logger::setHandler(&logger);
+
     CROW_ROUTE(app, "/")([](){
         return "Hello world";
     });
 
+
     CROW_ROUTE(app, "/copilot_internal/v2/token")([](){
         //return "Hello world";
 
@@ -168,5 +180,7 @@ int main(int argc, char **argv)
 
     app.port(program.get<int>("--port")).multithreaded().run();
 
+    
+
     free(model);
 }

src/server.cpp

@@ -37,7 +37,6 @@ crow::response handle_hf_request(TurbopilotModel *model, const crow::request& re
         crow::json::wvalue response = {
             {"generated_text", result.str()},
         };
-        
 
         
         crow::response res;

src/starcoder.cpp

@@ -44,13 +44,13 @@ bool starcoder_eval(
 
     if (mem_per_token > 0 && mem_per_token*N > buf_size) {
         const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
-        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+        spdlog::debug("{}: reallocating buffer from {} to {} bytes\n", __func__, buf_size, buf_size_new);
 
         // reallocate
         buf_size = buf_size_new;
         buf = realloc(buf, buf_size);
         if (buf == nullptr) {
-            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            spdlog::error("{}: failed to allocate {} bytes\n", __func__, buf_size);
             return false;
         }
     }
@@ -681,7 +681,7 @@ bool StarcoderModel::load_model(std::string fname) {
 }
 
 
-std::stringstream StarcoderModel::predict(std::string prompt, int max_length, bool include_prompt) {
+std::stringstream StarcoderModel::predict_impl(std::string prompt, int max_length, bool include_prompt) {
 
     std::stringstream result;
     // tokenize the prompt