Refit bug fix (#3097)

Co-authored-by: Dheeraj Peri <[email protected]>

Author: cehongwang

examples/dynamo/mutable_torchtrt_module_example.py

@@ -34,7 +34,7 @@
     "make_refitable": True,
 }
 
-model = models.resnet18(pretrained=False).eval().to("cuda")
+model = models.resnet18(pretrained=True).eval().to("cuda")
 mutable_module = torch_trt.MutableTorchTensorRTModule(model, **settings)
 # You can use the mutable module just like the original pytorch module. The compilation happens while you first call the mutable module.
 mutable_module(*inputs)
@@ -45,7 +45,7 @@
 
 # %%
 # Making changes to mutable module can trigger refit or re-compilation. For example, loading a different state_dict and setting new weight values will trigger refit, and adding a module to the model will trigger re-compilation.
-model2 = models.resnet18(pretrained=True).eval().to("cuda")
+model2 = models.resnet18(pretrained=False).eval().to("cuda")
 mutable_module.load_state_dict(model2.state_dict())
 
 

examples/dynamo/refit_engine_example.py

@@ -39,7 +39,7 @@
 # Compile the module for the first time and save it.
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-model = models.resnet18(pretrained=False).eval().to("cuda")
+model = models.resnet18(pretrained=True).eval().to("cuda")
 exp_program = torch.export.export(model, tuple(inputs))
 enabled_precisions = {torch.float}
 debug = False
@@ -68,7 +68,7 @@
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 # Create and compile the updated model
-model2 = models.resnet18(pretrained=True).eval().to("cuda")
+model2 = models.resnet18(pretrained=False).eval().to("cuda")
 exp_program2 = torch.export.export(model2, tuple(inputs))
 
 

py/torch_tensorrt/dynamo/_compiler.py

@@ -56,9 +56,9 @@ def compile(
     disable_tf32: bool = _defaults.DISABLE_TF32,
     assume_dynamic_shape_support: bool = _defaults.ASSUME_DYNAMIC_SHAPE_SUPPORT,
     sparse_weights: bool = _defaults.SPARSE_WEIGHTS,
-    enabled_precisions: (
-        Set[torch.dtype | dtype] | Tuple[torch.dtype | dtype]
-    ) = _defaults.ENABLED_PRECISIONS,
+    enabled_precisions: Union[
+        Set[Union[torch.dtype, dtype]], Tuple[Union[torch.dtype, dtype]]
+    ] = _defaults.ENABLED_PRECISIONS,
     engine_capability: EngineCapability = _defaults.ENGINE_CAPABILITY,
     make_refitable: bool = _defaults.MAKE_REFITABLE,
     debug: bool = _defaults.DEBUG,

py/torch_tensorrt/dynamo/_refit.py

@@ -34,7 +34,7 @@
     TorchTensorRTModule,
 )
 from torch_tensorrt.dynamo.utils import (
-    check_output,
+    check_module_output,
     get_torch_inputs,
     set_log_level,
     to_torch_device,
@@ -115,19 +115,8 @@ def construct_refit_mapping_from_weight_name_map(
     for engine_weight_name, (sd_weight_name, np_weight_type) in weight_name_map.items():
         trt_dtype = dtype.try_from(np_weight_type).to(trt.DataType)
         torch_dtype = dtype.try_from(np_weight_type).to(torch.dtype)
-        if engine_weight_name.split(" ")[-1] in ["SCALE", "SHIFT"]:
-            # Batch Norm Layer
-            params = {}
-            for w in sd_weight_name:
-                params[w.split(".")[-1]] = state_dict[w]
-            scale = params["weight"] / torch.sqrt(params["running_var"] + 1e-7)
-            shift = params["bias"] - params["running_mean"] * scale
-            # Set scale to scale or shift to shift
-            engine_weight_map[engine_weight_name] = eval(
-                engine_weight_name.split(" ")[-1].lower()
-            )
 
-        elif sd_weight_name not in state_dict:
+        if sd_weight_name not in state_dict:
             # If weights is not in sd, we can leave it unchanged
             continue
         else:
@@ -157,16 +146,25 @@ def _refit_single_trt_engine_with_gm(
     """
 
     refitted = set()
-
+    torch_device = list(new_gm.state_dict().values())[0].device.type
     refitter = trt.Refitter(old_engine, TRT_LOGGER)
     weight_list = refitter.get_all_weights()
 
     if weight_name_map:
         # Get the refitting mapping
-        trt_wt_location = trt.TensorLocation.DEVICE
+        trt_wt_location = (
+            trt.TensorLocation.DEVICE
+            if torch_device == "cuda"
+            else trt.TensorLocation.HOST
+        )
         mapping = construct_refit_mapping_from_weight_name_map(
             weight_name_map, new_gm.state_dict()
         )
+
+        # Debug Use
+        # correct = construct_refit_mapping(new_gm, input_list, settings)
+        # comparison = {k: (np.allclose(correct[k][0], mapping[k][0].cpu().numpy(), 1e-2, 1e-2), correct[k][0], mapping[k][0]) for k in mapping if k in correct}
+
         for layer_name in weight_list:
             if layer_name not in mapping:
                 logger.warning(f"{layer_name} is not found in weight mapping.")
@@ -235,7 +233,7 @@ def refit_module_weights(
         compiled_module = copy.deepcopy(compiled_module)
     elif inline_module:
         raise AssertionError(
-            "Exported program does not support modifying in place. Please set inplace to false and use the returned graph module."
+            "Exported program does not support modifying in place. Please set in_place to false and use the returned graph module."
         )
 
     # Get the settings and check the setting to be uniform
@@ -283,6 +281,7 @@ def refit_module_weights(
             arg_inputs = [arg_inputs]
         torch_inputs = get_torch_inputs(arg_inputs, device)
 
+    torch_kwarg_inputs: Any = {}
     if kwarg_inputs:
         torch_kwarg_inputs = get_torch_inputs(kwarg_inputs, device)
     runtime = trt.Runtime(TRT_LOGGER)
@@ -436,6 +435,7 @@ def refit_module_weights(
                 settings=settings,
                 weight_name_map=weight_name_map,
             )
+
         except AssertionError as e:
             # If fast_refit is used and failed, we fall back to regular refit
             logger.warning(e)
@@ -463,14 +463,27 @@ def refit_module_weights(
             setattr(compiled_module, f"{name}_engine", refitted_engine)
 
     if verify_output and arg_inputs is not None:
-        if check_output(
+        if check_module_output(
             new_module=new_gm,
             refitted_module=compiled_module,
             arg_inputs=torch_inputs,
             kwarg_inputs=torch_kwarg_inputs,
         ):
             logger.info("Refitting Succeed!")
         else:
+            if weight_name_map:
+                logger.warning(
+                    "Refitting with weight_name_map yielded incorrect result! The outputs do not match."
+                )
+                return refit_module_weights(
+                    compiled_module,
+                    new_weight_module,
+                    arg_inputs,
+                    kwarg_inputs,
+                    verify_output,
+                    use_weight_map_cache=False,
+                    in_place=in_place,
+                )
             logger.error("Refitting Failed! The outputs do not match.")
     else:
         logger.info("Refitting Completed! Output verification skipped.")

py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py

@@ -1,9 +1,21 @@
+import gc
 import io
 import logging
 import os
 import warnings
 from datetime import datetime
-from typing import Any, Callable, Dict, List, NamedTuple, Optional, Sequence, Set, Tuple
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    List,
+    NamedTuple,
+    Optional,
+    Sequence,
+    Set,
+    Tuple,
+    Union,
+)
 
 import numpy as np
 import torch
@@ -26,7 +38,7 @@
     get_node_name,
     get_trt_tensor,
 )
-from torch_tensorrt.dynamo.utils import DYNAMIC_DIM
+from torch_tensorrt.dynamo.utils import DYNAMIC_DIM, to_torch_device
 from torch_tensorrt.fx.observer import Observer
 from torch_tensorrt.logging import TRT_LOGGER
 
@@ -327,6 +339,39 @@ def _construct_trt_network_def(self) -> None:
             f"TRT INetwork construction elapsed time: {datetime.now() - run_module_start_time}"
         )
 
+    @staticmethod
+    def find_weight(
+        weight_name: str, np_map: dict[str, Any], state_dict: dict[str, Any]
+    ) -> str:
+        """
+        We need to build map from engine weight name to state_dict weight name.
+        The purpose of this function is to find the corresponding weight name in module state_dict.
+
+        weight_name: the target weight name we want to search for
+        np_map: the map from weight name to np values in INetworkDefinition
+        state_dict: state of the graph module
+        """
+        network_weight = np_map[weight_name]
+        network_weight = torch.from_numpy(np_map[weight_name]).cuda()
+        for sd_w_name, sd_weight in state_dict.items():
+            if TRTInterpreter.check_weight_equal(sd_weight, network_weight):
+                del state_dict[sd_w_name]
+                return sd_w_name
+        return ""
+
+    @staticmethod
+    def check_weight_equal(
+        sd_weight: torch.tensor, network_weight: Union[torch.Tensor, np.ndarray]
+    ) -> Any:
+        if not isinstance(network_weight, torch.Tensor):
+            network_weight = torch.from_numpy(network_weight).cuda()
+        try:
+            return sd_weight.shape == network_weight.shape and torch.all(
+                torch.abs(sd_weight - network_weight) < 0.01
+            )
+        except Exception:
+            return torch.all(sd_weight == network_weight)
+
     def _save_weight_mapping(self) -> None:
         """
         Construct the weight name mapping from engine weight name to state_dict weight name.
@@ -336,23 +381,6 @@ def _save_weight_mapping(self) -> None:
         2. Value mapping that, for each weight in INetworkDefinition search for identical weight in state_dict
         """
 
-        def find_weight(
-            weight_name: str, np_map: dict[str, Any], sd: dict[str, Any]
-        ) -> str:
-            network_weight = np_map[weight_name]
-            for sd_w_name, sd_weight in sd.items():
-                if check_weight_equal(sd_weight, network_weight):
-                    return sd_w_name
-            return ""
-
-        def check_weight_equal(
-            sd_weight: torch.tensor, network_weight: np.ndarray
-        ) -> Any:
-            sd_weight = sd_weight.reshape(-1).cpu().numpy()
-            return sd_weight.size == network_weight.size and np.allclose(
-                sd_weight, network_weight, 1e-1, 1e-1
-            )
-
         MODULE_MAP = {
             "SCALE": (
                 trt.IScaleLayer,
@@ -398,8 +426,19 @@ def check_weight_equal(
             )
         }
         """
+        _LOGGER.info("Building weight name mapping...")
         # Stage 1: Name mapping
         sd = self.module.state_dict()
+        torch_device = to_torch_device(self.compilation_settings.device)
+        gm_is_on_cuda = list(sd.values())[0].device.type == "cuda"
+        if not gm_is_on_cuda:
+            # If the model original position is on CPU, move it GPU
+            sd = {
+                k: v.reshape(-1).to(torch_device)
+                for k, v in self.module.state_dict().items()
+            }
+        else:
+            sd = {k: v.reshape(-1) for k, v in self.module.state_dict().items()}
         weight_name_map: dict[str, Any] = {}
         np_map = {}
         net = self.ctx.net
@@ -448,10 +487,10 @@ def check_weight_equal(
             if "SCALE" in engine_weight_name:
                 # There is no direct connection in batch_norm layer. So skip it
                 pass
-            elif sd_weight_name not in sd or not check_weight_equal(
+            elif sd_weight_name not in sd or not TRTInterpreter.check_weight_equal(
                 sd[sd_weight_name], np_map[engine_weight_name]
             ):
-                weight_name_map[engine_weight_name] = find_weight(
+                weight_name_map[engine_weight_name] = TRTInterpreter.find_weight(
                     engine_weight_name, np_map, sd
                 )
 
@@ -462,6 +501,10 @@ def check_weight_equal(
 
         self.weight_name_map = weight_name_map
 
+        del np_map, sd
+        gc.collect()
+        torch.cuda.empty_cache()
+
     def run(
         self,
         strict_type_constraints: bool = False,

py/torch_tensorrt/dynamo/conversion/_conversion.py

@@ -130,14 +130,13 @@ def convert_module(
     from torch_tensorrt.dynamo._refit import _refit_single_trt_engine_with_gm
     from torch_tensorrt.logging import TRT_LOGGER
 
-    runtime = trt.Runtime(TRT_LOGGER)
-    refit_test_engine = runtime.deserialize_cuda_engine(
-        interpreter_result.serialized_engine
-    )
     weight_name_map: Any = None
     # Do the test refit with cached map if make_refitable is enabled
     if settings.make_refitable:
-        weight_name_map = interpreter_result.weight_name_map
+        runtime = trt.Runtime(TRT_LOGGER)
+        refit_test_engine = runtime.deserialize_cuda_engine(
+            interpreter_result.serialized_engine
+        )
         try:
             _refit_single_trt_engine_with_gm(
                 new_gm=module,
@@ -146,9 +145,13 @@ def convert_module(
                 settings=settings,
                 weight_name_map=interpreter_result.weight_name_map,
             )
+            weight_name_map = interpreter_result.weight_name_map
         except AssertionError:
             logger.warning("Fast refit test failed. Removing the weight map caching.")
 
+        del refit_test_engine
+        torch.cuda.empty_cache()
+
     rt_cls = PythonTorchTensorRTModule
 
     if ENABLED_FEATURES.torch_tensorrt_runtime and not settings.use_python_runtime: