feat: Enable sparsity support in TRTorch

Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>

Author: peri044

core/conversion/conversionctx/ConversionCtx.cpp

@@ -89,6 +89,10 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
     cfg->clearFlag(nvinfer1::BuilderFlag::kTF32);
   }
 
+  if (settings.sparse_weights) {
+    cfg->setFlag(nvinfer1::BuilderFlag::kSPARSE_WEIGHTS);
+  }
+
   if (settings.refit) {
     cfg->setFlag(nvinfer1::BuilderFlag::kREFIT);
   }

core/conversion/conversionctx/ConversionCtx.h

@@ -25,7 +25,7 @@ struct Device {
 
 struct BuilderSettings {
   std::set<nvinfer1::DataType> enabled_precisions = {nvinfer1::DataType::kFLOAT};
-  std::vector<nvinfer1::DataType> input_dtypes;
+  bool sparse_weights = false;
   bool disable_tf32 = false;
   bool refit = false;
   bool debug = false;

cpp/api/include/trtorch/trtorch.h

@@ -687,6 +687,11 @@ struct TRTORCH_API CompileSpec {
    */
   bool disable_tf32 = false;
 
+  /**
+   * Enable sparsity for weights of conv and FC layers
+   */
+  bool sparse_weights = false;
+
   /**
    * Build a refitable engine
    */

cpp/api/src/compile_spec.cpp

@@ -364,6 +364,7 @@ core::CompileSpec to_internal_compile_spec(CompileSpec external) {
     }
   }
 
+  internal.convert_info.engine_settings.sparse_weights = external.sparse_weights;
   internal.convert_info.engine_settings.disable_tf32 = external.disable_tf32;
   internal.convert_info.engine_settings.refit = external.refit;
   internal.convert_info.engine_settings.debug = external.debug;

cpp/trtorchexec/main.cpp

@@ -57,12 +57,13 @@ int main(int argc, const char* argv[]) {
 
   auto compile_spec = trtorch::CompileSpec(dims);
   compile_spec.workspace_size = 1 << 24;
+  compile_spec.sparse_weights = true;
 
-  std::cout << "Checking operator support" << std::endl;
-  if (!trtorch::CheckMethodOperatorSupport(mod, "forward")) {
-    std::cerr << "Method is not currently supported by TRTorch" << std::endl;
-    return -1;
-  }
+  // std::cout << "Checking operator support" << std::endl;
+  // if (!trtorch::CheckMethodOperatorSupport(mod, "forward")) {
+  //   std::cerr << "Method is not currently supported by TRTorch" << std::endl;
+  //   return -1;
+  // }
 
   std::cout << "Compiling graph to save as TRT engine (/tmp/engine_converted_from_jit.trt)" << std::endl;
   auto engine = trtorch::ConvertGraphToTRTEngine(mod, "forward", compile_spec);

py/trtorch/_compile_spec.py

@@ -199,6 +199,10 @@ def _parse_compile_spec(compile_spec: Dict[str, Any]) -> trtorch._C.CompileSpec:
     if "calibrator" in compile_spec:
         info.ptq_calibrator = compile_spec["calibrator"]
 
+    if "sparse_weights" in compile_spec:
+        assert isinstance(compile_spec["sparse_weights"], bool)
+        info.sparse_weights = compile_spec["sparse_weights"]
+
     if "disable_tf32" in compile_spec:
         assert isinstance(compile_spec["disable_tf32"], bool)
         info.disable_tf32 = compile_spec["disable_tf32"]
@@ -282,8 +286,8 @@ def TensorRTCompileSpec(compile_spec: Dict[str, Any]) -> torch.classes.tensorrt.
                             "dla_core": 0, # (DLA only) Target dla core id to run engine
                             "allow_gpu_fallback": false, # (DLA only) Allow layers unsupported on DLA to run on GPU
                         },
-                        "op_precision": torch.half, # Operating precision set to FP16
-                        # List of datatypes that should be configured for each input. Supported options torch.{float|half|int8|int32|bool}.
+                        "enabled_precisions": {torch.half}, # Operating precision set to FP16
+                        "sparse_weights": Enable sparsity for convolution and fully connected layers.
                         "disable_tf32": False, # Force FP32 layers to use traditional as FP32 format vs the default behavior of rounding the inputs to 10-bit mantissas before multiplying, but accumulates the sum using 23-bit mantissas
                         "refit": False, # enable refit
                         "debug": False, # enable debuggable engine

py/trtorch/_compiler.py

@@ -44,6 +44,8 @@ def compile(module: torch.jit.ScriptModule, compile_spec: Any) -> torch.jit.Scri
                         "allow_gpu_fallback": false, # (DLA only) Allow layers unsupported on DLA to run on GPU
                     },
                     "enabled_precisions": {torch.float, torch.half}, # Enabling FP16 kernels
+                    "disable_tf32": False, # Force FP32 layers to use traditional as FP32 format vs the default behavior of rounding the inputs to 10-bit mantissas before multiplying, but accumulates the sum using 23-bit mantissas
+                    "sparse_weights": Enable sparsity for convolution and fully connected layers.
                     "refit": false, # enable refit
                     "debug": false, # enable debuggable engine
                     "strict_types": false, # kernels should strictly run in operating precision
@@ -113,6 +115,7 @@ def convert_method_to_trt_engine(module: torch.jit.ScriptModule, method_name: st
                     "enabled_precisions": {torch.float, torch.half}, # Enabling FP16 kernels
                     # List of datatypes that should be configured for each input. Supported options torch.{float|half|int8|int32|bool}.
                     "disable_tf32": False, # Force FP32 layers to use traditional as FP32 format vs the default behavior of rounding the inputs to 10-bit mantissas before multiplying, but accumulates the sum using 23-bit mantissas
+                    "sparse_weights": Enable sparsity for convolution and fully connected layers.
                     "refit": false, # enable refit
                     "debug": false, # enable debuggable engine
                     "strict_types": false, # kernels should strictly run in operating precision

py/trtorch/csrc/register_tensorrt_classes.cpp

@@ -48,6 +48,7 @@ void RegisterTRTCompileSpec() {
           .def("_set_ptq_calibrator", &trtorch::pyapi::CompileSpec::setPTQCalibratorViaHandle)
           .def("__str__", &trtorch::pyapi::CompileSpec::stringify);
 
+  ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, sparse_weights);
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, disable_tf32);
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, refit);
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, debug);

py/trtorch/csrc/tensorrt_classes.cpp

@@ -183,6 +183,7 @@ core::CompileSpec CompileSpec::toInternalCompileSpec() {
   }
 
   info.convert_info.engine_settings.calibrator = ptq_calibrator;
+  info.convert_info.engine_settings.sparse_weights = sparse_weights;
   info.convert_info.engine_settings.disable_tf32 = disable_tf32;
   info.convert_info.engine_settings.refit = refit;
   info.convert_info.engine_settings.debug = debug;
@@ -222,6 +223,7 @@ std::string CompileSpec::stringify() {
   }
   ss << "    ]" << std::endl;
   ss << "    \"TF32 Disabled\": " << disable_tf32 << std::endl;
+  ss << "    \"Sparsity\": " << sparse_weights << std::endl;
   ss << "    \"Refit\": " << refit << std::endl;
   ss << "    \"Debug\": " << debug << std::endl;
   ss << "    \"Strict Types\": " << strict_types << std::endl;

py/trtorch/csrc/tensorrt_classes.h

@@ -139,6 +139,7 @@ struct CompileSpec : torch::CustomClassHolder {
   }
 
   ADD_FIELD_GET_SET(disable_tf32, bool);
+  ADD_FIELD_GET_SET(sparse_weights, bool);
   ADD_FIELD_GET_SET(refit, bool);
   ADD_FIELD_GET_SET(debug, bool);
   ADD_FIELD_GET_SET(strict_types, bool);
@@ -155,6 +156,7 @@ struct CompileSpec : torch::CustomClassHolder {
   std::vector<Input> inputs;
   nvinfer1::IInt8Calibrator* ptq_calibrator = nullptr;
   std::set<DataType> enabled_precisions = {DataType::kFloat};
+  bool sparse_weights = false;
   bool disable_tf32 = false;
   bool refit = false;
   bool debug = false;