Merge branch 'smalton/DOR-535-compat-lib' into 'master'

DOR-535: Compatibility lib

Closes DOR-535

See merge request machine-learning/dorado!818

CMakeLists.txt

@@ -159,6 +159,7 @@ add_subdirectory(${DORADO_3RD_PARTY_SOURCE}/date EXCLUDE_FROM_ALL)
 
 enable_testing()
 
+add_subdirectory(dorado/compat)
 add_subdirectory(dorado/utils)
 add_subdirectory(dorado/models)
 add_subdirectory(dorado/basecall)
@@ -368,6 +369,7 @@ if(NOT DORADO_DISABLE_DORADO)
         dorado_lib
         dorado_io_lib
         dorado_models_lib
+        dorado_compat
         minimap2
     )
 

dorado/cli/basecaller.cpp

@@ -31,6 +31,7 @@
 
 #include <htslib/sam.h>
 #include <spdlog/spdlog.h>
+#include <torch/utils.h>
 
 #include <algorithm>
 #include <cstdlib>

dorado/cli/duplex.cpp

@@ -30,6 +30,7 @@
 
 #include <htslib/sam.h>
 #include <spdlog/spdlog.h>
+#include <torch/utils.h>
 
 #include <cstdlib>
 #include <exception>

dorado/compat/CMakeLists.txt

@@ -0,0 +1,11 @@
+add_library(dorado_compat
+    compat_utils.cpp
+    compat_utils.h
+)
+
+target_include_directories(dorado_compat
+    INTERFACE
+    ${CMAKE_CURRENT_SOURCE_DIR}/..
+)
+
+enable_warnings_as_errors(dorado_compat)

dorado/utils/compat_utils.cpp → dorado/compat/compat_utils.cpp

@@ -20,5 +20,16 @@ char* strptime(const char* s, const char* f, tm* tm) {
     return (char*)(s + input.tellg());
 }
 
+int setenv(const char* name, const char* value, int overwrite) {
+    if (!overwrite) {
+        size_t envsize = 0;
+        int errcode = getenv_s(&envsize, NULL, 0, name);
+        if (errcode || envsize) {
+            return errcode;
+        }
+    }
+    return _putenv_s(name, value);
+}
+
 }  // namespace dorado::utils
 #endif  // _WIN32

dorado/utils/compat_utils.h → dorado/compat/compat_utils.h

@@ -9,16 +9,7 @@ namespace dorado::utils {
 char* strptime(const char* s, const char* f, tm* tm);
 
 // A simple wrapper for setenv, since windows doesn't have it.
-inline int setenv(const char* name, const char* value, int overwrite) {
-    if (!overwrite) {
-        size_t envsize = 0;
-        int errcode = getenv_s(&envsize, NULL, 0, name);
-        if (errcode || envsize) {
-            return errcode;
-        }
-    }
-    return _putenv_s(name, value);
-}
+int setenv(const char* name, const char* value, int overwrite);
 
 }  // namespace dorado::utils
 

dorado/data_loader/DataLoader.cpp

@@ -3,7 +3,6 @@
 #include "models/models.h"
 #include "read_pipeline/ReadPipeline.h"
 #include "utils/PostCondition.h"
-#include "utils/compat_utils.h"
 #include "utils/time_utils.h"
 #include "utils/types.h"
 #include "vbz_plugin_user_utils.h"

dorado/main.cpp

@@ -1,6 +1,6 @@
 #include "Version.h"
 #include "cli/cli.h"
-#include "utils/compat_utils.h"
+#include "compat/compat_utils.h"
 
 #include <minimap.h>
 #include <spdlog/cfg/env.h>

dorado/utils/CMakeLists.txt

@@ -8,8 +8,6 @@ add_library(dorado_utils
     barcode_kits.h
     basecaller_utils.cpp
     basecaller_utils.h
-    compat_utils.cpp
-    compat_utils.h
     dev_utils.cpp
     dev_utils.h
     driver_query.cpp
@@ -41,6 +39,7 @@ add_library(dorado_utils
     thread_utils.h
     time_utils.cpp
     time_utils.h
+    torch_utils.cpp
     torch_utils.h
     trim.cpp
     trim.h
@@ -95,6 +94,7 @@ target_link_libraries(dorado_utils
         edlib
         spdlog::spdlog
     PRIVATE
+        dorado_compat
         minimap2
         OpenSSL::SSL
         htslib

dorado/utils/torch_utils.cpp

@@ -0,0 +1,49 @@
+#include "torch_utils.h"
+
+#include "compat/compat_utils.h"
+
+#if DORADO_CUDA_BUILD
+#include <c10/cuda/CUDACachingAllocator.h>
+#endif
+#include <torch/torch.h>
+
+namespace dorado::utils {
+
+void make_torch_deterministic() {
+#if DORADO_CUDA_BUILD
+    setenv("CUBLAS_WORKSPACE_CONFIG", ":4096:8", true);
+    torch::globalContext().setDeterministicCuDNN(true);
+    torch::globalContext().setBenchmarkCuDNN(false);
+#endif
+
+#if TORCH_VERSION_MAJOR > 1 || TORCH_VERSION_MINOR >= 11
+    torch::globalContext().setDeterministicAlgorithms(true, false);
+#else
+    torch::globalContext().setDeterministicAlgorithms(true);
+#endif
+}
+
+void set_torch_allocator_max_split_size() {
+#if DORADO_CUDA_BUILD && TORCH_VERSION_MAJOR >= 2
+
+    // Do not re-use smaller chunks of large buffers
+    // This prevents small allocations from reusing large sections of cached allocated memory
+    // which can lead to OoM errors when the original large allocation is needed again
+    auto max_split_size_mb = 25;
+    std::string settings = "max_split_size_mb:" + std::to_string(max_split_size_mb);
+
+    const char *pytorch_cuda_alloc_conf = std::getenv("PYTORCH_CUDA_ALLOC_CONF");
+    if (pytorch_cuda_alloc_conf != nullptr) {
+        std::string_view str(pytorch_cuda_alloc_conf);
+        if (str.find("max_split_size_mb") != std::string::npos) {
+            // user has set this via env_var - let torch parse and use their value
+            return;
+        }
+        settings += std::string(",") + pytorch_cuda_alloc_conf;
+    }
+
+    c10::cuda::CUDACachingAllocator::setAllocatorSettings(settings);
+#endif
+}
+
+}  // namespace dorado::utils

dorado/utils/torch_utils.h

@@ -1,49 +1,8 @@
 #pragma once
 
-#include "compat_utils.h"
-
-#if DORADO_CUDA_BUILD
-#include <c10/cuda/CUDACachingAllocator.h>
-#endif
-#include <torch/torch.h>
-
 namespace dorado::utils {
 
-inline void make_torch_deterministic() {
-#if DORADO_CUDA_BUILD
-    setenv("CUBLAS_WORKSPACE_CONFIG", ":4096:8", true);
-    torch::globalContext().setDeterministicCuDNN(true);
-    torch::globalContext().setBenchmarkCuDNN(false);
-#endif
-
-#if TORCH_VERSION_MAJOR > 1 || TORCH_VERSION_MINOR >= 11
-    torch::globalContext().setDeterministicAlgorithms(true, false);
-#else
-    torch::globalContext().setDeterministicAlgorithms(true);
-#endif
-}
-
-inline void set_torch_allocator_max_split_size() {
-#if DORADO_CUDA_BUILD && TORCH_VERSION_MAJOR >= 2
-
-    // Do not re-use smaller chunks of large buffers
-    // This prevents small allocations from reusing large sections of cached allocated memory
-    // which can lead to OoM errors when the original large allocation is needed again
-    auto max_split_size_mb = 25;
-    std::string settings = "max_split_size_mb:" + std::to_string(max_split_size_mb);
-
-    const char *pytorch_cuda_alloc_conf = std::getenv("PYTORCH_CUDA_ALLOC_CONF");
-    if (pytorch_cuda_alloc_conf != nullptr) {
-        std::string_view str(pytorch_cuda_alloc_conf);
-        if (str.find("max_split_size_mb") != std::string::npos) {
-            // user has set this via env_var - let torch parse and use their value
-            return;
-        }
-        settings += std::string(",") + pytorch_cuda_alloc_conf;
-    }
-
-    c10::cuda::CUDACachingAllocator::setAllocatorSettings(settings);
-#endif
-}
+void make_torch_deterministic();
+void set_torch_allocator_max_split_size();
 
 }  // namespace dorado::utils

tests/CMakeLists.txt

@@ -78,6 +78,7 @@ foreach(TEST_BIN dorado_tests dorado_smoke_tests)
         dorado_models_lib
         dorado_basecall
         dorado_modbase
+        dorado_compat
         minimap2
         ${ZLIB_LIBRARIES}
         ${POD5_LIBRARIES}

tests/main.cpp

@@ -1,11 +1,11 @@
 #define CATCH_CONFIG_RUNNER
 
-#include "utils/compat_utils.h"
+#include "compat/compat_utils.h"
 #include "utils/torch_utils.h"
 
 #include <catch2/catch.hpp>
 #include <nvtx3/nvtx3.hpp>
-#include <torch/torch.h>
+#include <torch/utils.h>
 
 #include <clocale>