Breaking Change: Remove the deprecated int8 calibrator related (#3759)

Author: lanluo-nvidia

core/conversion/conversion.cpp

@@ -202,13 +202,7 @@ void AddInputs(ConversionCtx* ctx, c10::ArrayRef<const torch::jit::Value*> input
   TORCHTRT_CHECK(
       profile->isValid(),
       "Optimization profile is invalid, please check the input range provided (conversion.AddInputs)");
-
   ctx->cfg->addOptimizationProfile(profile);
-#if NV_TENSORRT_MAJOR > 7 || (NV_TENSORRT_MAJOR == 7 && NV_TENSORRT_MINOR >= 1)
-  if (ctx->enabled_precisions.find(nvinfer1::DataType::kINT8) != ctx->enabled_precisions.end()) {
-    ctx->cfg->setCalibrationProfile(profile);
-  }
-#endif
 }
 
 void MarkOutputs(ConversionCtx* ctx, at::ArrayRef<const torch::jit::Value*> outputs) {

core/conversion/conversionctx/ConversionCtx.cpp

@@ -31,8 +31,7 @@ std::ostream& operator<<(std::ostream& os, const BuilderSettings& s) {
     if (s.device.device_type == nvinfer1::DeviceType::kDLA) {
     os << "\n    DLACore: " << s.device.dla_core;
     }
-    os << "\n    Engine Capability: " << s.capability                                      \
-       << "\n    Calibrator Created: " << (s.calibrator != nullptr);
+    os << "\n    Engine Capability: " << s.capability;
     return os;
 }
 // clang-format on
@@ -64,15 +63,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
         cfg->setFlag(nvinfer1::BuilderFlag::kFP16);
         break;
       case nvinfer1::DataType::kINT8:
-        TORCHTRT_CHECK(
-            builder->platformHasFastInt8(), "Requested inference in INT8 but platform does not support INT8");
-        cfg->setFlag(nvinfer1::BuilderFlag::kINT8);
-        if (!settings.calibrator) {
-          LOG_INFO(
-              "Int8 precision has been enabled but no calibrator provided. This assumes the network has Q/DQ nodes obtained from Quantization aware training. For more details, refer to https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#work-with-qat-networks");
-        } else {
-          cfg->setInt8Calibrator(settings.calibrator);
-        }
+        LOG_DEBUG("INT8 precision has been enabled, we assume the network has Q/DQ nodes obtained from modelopt");
         break;
       case nvinfer1::DataType::kFLOAT:
         break;

core/conversion/conversionctx/ConversionCtx.h

@@ -26,7 +26,6 @@ struct BuilderSettings {
   bool allow_shape_tensors = false;
   ir::Device device;
   nvinfer1::EngineCapability capability = TRT_ENGINE_CAPABILITY_STANDARD;
-  nvinfer1::IInt8Calibrator* calibrator = nullptr;
   uint64_t num_avg_timing_iters = 1;
   uint64_t workspace_size = 0;
   uint64_t dla_sram_size = DLA_SRAM_SIZE;

cpp/BUILD

@@ -7,14 +7,12 @@ cc_library(
     srcs = [
         "src/compile_spec.cpp",
         "src/logging.cpp",
-        "src/ptq.cpp",
         "src/torch_tensorrt.cpp",
         "src/types.cpp",
     ],
     hdrs = [
         "include/torch_tensorrt/logging.h",
         "include/torch_tensorrt/macros.h",
-        "include/torch_tensorrt/ptq.h",
         "include/torch_tensorrt/torch_tensorrt.h",
     ],
     linkstatic = True,

cpp/CMakeLists.txt

@@ -4,15 +4,13 @@ add_library(${lib_name} OBJECT)
 set(CXX_SRCS
     "${CMAKE_CURRENT_SOURCE_DIR}/src/compile_spec.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/src/logging.cpp"
-    "${CMAKE_CURRENT_SOURCE_DIR}/src/ptq.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/src/torch_tensorrt.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/src/types.cpp"
 )
 
 set(HEADER_FILES
     "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/logging.h"
     "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/macros.h"
-    "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/ptq.h"
     "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/torch_tensorrt.h"
 )
 

cpp/bin/torchtrtc/fileio.h

@@ -23,7 +23,6 @@
 #include "torch/torch.h"
 
 #include "torch_tensorrt/logging.h"
-#include "torch_tensorrt/ptq.h"
 #include "torch_tensorrt/torch_tensorrt.h"
 
 namespace torchtrtc {

cpp/bin/torchtrtc/main.cpp

@@ -7,7 +7,6 @@
 #include "torch/script.h"
 
 #include "torch_tensorrt/logging.h"
-#include "torch_tensorrt/ptq.h"
 #include "torch_tensorrt/torch_tensorrt.h"
 
 #include "accuracy.h"
@@ -335,8 +334,6 @@ int main(int argc, char** argv) {
     calibration_cache_file_path = torchtrtc::fileio::resolve_path(args::get(calibration_cache_file));
   }
 
-  auto calibrator = torchtrt::ptq::make_int8_cache_calibrator(calibration_cache_file_path);
-
   compile_settings.require_full_compilation = require_full_compilation;
 
   if (torch_executed_ops || torch_executed_mods) {
@@ -367,13 +364,9 @@ int main(int argc, char** argv) {
         compile_settings.enabled_precisions.insert(torch::kF16);
       } else if (dtype == torchtrt::DataType::kChar) {
         compile_settings.enabled_precisions.insert(torch::kI8);
-        if (calibration_cache_file) {
-          compile_settings.ptq_calibrator = calibrator;
-        } else {
-          torchtrt::logging::log(
-              torchtrt::logging::Level::kINFO,
-              "Int8 precision has been enabled but no calibrator provided. This assumes the network has Q/DQ nodes obtained from Quantization aware training. For more details, refer to https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#work-with-qat-networks");
-        }
+        torchtrt::logging::log(
+            torchtrt::logging::Level::kDEBUG,
+            "Int8 precision has been enabled which assumes the network has Q/DQ nodes obtained");
       } else {
         std::stringstream ss;
         ss << "Invalid precision given for enabled kernel precision, options are [ float | float32 | f32 | fp32 | half | float16 | f16 | fp16 | char | int8 | i8 ], found: ";

cpp/bin/torchtrtc/parser_util.h

@@ -9,7 +9,6 @@
 #include "torch/torch.h"
 
 #include "torch_tensorrt/logging.h"
-#include "torch_tensorrt/ptq.h"
 #include "torch_tensorrt/torch_tensorrt.h"
 
 namespace torchtrtc {

cpp/include/torch_tensorrt/macros.h

@@ -30,9 +30,6 @@
   STR(TORCH_TENSORRT_MAJOR_VERSION) \
   "." STR(TORCH_TENSORRT_MINOR_VERSION) "." STR(TORCH_TENSORRT_PATCH_VERSION)
 
-#define TORCH_TENSORRT_PTQ_DEPRECATION \
-  [[deprecated(                        \
-      "Int8 PTQ Calibrator has been deprecated by TensorRT, please plan on porting to a NVIDIA Model Optimizer Toolkit based workflow. See: https://pytorch.org/TensorRT/tutorials/_rendered_examples/dynamo/vgg16_ptq.html for more details")]]
 // Setup namespace aliases for ease of use
 namespace torch_tensorrt {
 namespace torchscript {}