[FX] Changes done internally at Facebook (#1603)

Co-authored-by: Wei Wei <[email protected]>

Author: frank-wei

.circleci/config.yml

@@ -263,7 +263,7 @@ commands:
     parameters:
       torch-build:
         type: string
-        default: "2.0.0.dev20230103+cu117"
+        default: "2.0.0.dev20230120+cu117"
       torch-build-index:
         type: string
         default: "https://download.pytorch.org/whl/nightly/cu117"
@@ -992,7 +992,7 @@ parameters:
   # Nightly platform config
   torch-build:
     type: string
-    default: "2.0.0.dev20230103+cu117"
+    default: "2.0.0.dev20230120+cu117"
   torch-build-index:
     type: string
     default: "https://download.pytorch.org/whl/nightly/cu117"

py/torch_tensorrt/fx/README.md

@@ -5,7 +5,7 @@ FX2TRT is merged as FX module in Torch-TensorRT
 
 * Method 1. Follow the instrucions for Torch-TensorRT
 * Method 2. To install FX path only (Python path) and avoid the C++ build for torchscript path
-```
+`
     $ conda create --name python_env python=3.8
     $ conda activate python_env
     # Recommend to install PyTorch 1.12 and later
@@ -18,4 +18,4 @@ FX2TRT is merged as FX module in Torch-TensorRT
     $ pyton -c "import torch_tensorrt.fx"
     # Test an example by
     $ python py/torch_tensorrt/fx/example/lower_example.py
-```
+`

py/torch_tensorrt/fx/converters/acc_ops_converters.py

@@ -2802,7 +2802,7 @@ def acc_ops_linear(
 
     if isinstance(kwargs["weight"], torch.Tensor):
         weight = get_trt_tensor(network, kwargs["weight"].t(), f"{name}_weight")
-        if target is not acc_ops.linear:
+        if target not in (acc_ops.linear, torch.ops.aten.linear):
             weight_op = trt.MatrixOperation.TRANSPOSE
         else:
             weight_op = trt.MatrixOperation.NONE

py/torch_tensorrt/fx/converters/aten_ops_converters.py

@@ -187,27 +187,20 @@ def aten_ops_fmod(
     return acc_ops_converters.acc_ops_fmod(network, target, None, kwargs_new, name)
 
 
-@tensorrt_converter(torch.ops.aten.mm.default)
-@tensorrt_converter(torch.ops.aten.addmm.default)
+@tensorrt_converter(torch.ops.aten.linear)
 def aten_ops_linear(
     network: TRTNetwork,
     target: Target,
     args: Tuple[Argument, ...],
     kwargs: Dict[str, Argument],
     name: str,
 ) -> Union[TRTTensor, Sequence[TRTTensor]]:
-    if target == torch.ops.aten.addmm.default:
-        kwargs_new = {
-            "bias": args[0],
-            "input": args[1],
-            "weight": args[2],
-        }
-    elif target == torch.ops.aten.mm.default:
-        kwargs_new = {
-            "bias": None,
-            "input": args[0],
-            "weight": args[1],
-        }
+    kwargs_new = {
+        "input": args[0],
+        "weight": args[1],
+        "bias": args[2],
+    }
+
     return acc_ops_converters.acc_ops_linear(network, target, None, kwargs_new, name)
 
 
@@ -320,3 +313,35 @@ def aten_ops_reshape(
         "acc_out_ty": acc_utils.build_raw_tensor_meta(shape=args[1]),
     }
     return acc_ops_converters.acc_ops_reshape(network, target, None, kwargs_new, name)
+
+
+@tensorrt_converter(torch.ops.aten.cat.default)
+def aten_ops_cat(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> Union[TRTTensor, Sequence[TRTTensor]]:
+    kwargs_new = {
+        "tensors": args[0],
+        "dim": args[1],
+    }
+    return acc_ops_converters.acc_ops_cat(network, target, None, kwargs_new, name)
+
+
+@tensorrt_converter(torch.ops.aten.expand.default)
+def aten_ops_expand(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> Union[TRTTensor, Sequence[TRTTensor]]:
+    kwargs_new = {
+        "input": args[0],
+        "sizes": args[1],
+    }
+    return acc_ops_converters.acc_ops_expand_tensor(
+        network, target, None, kwargs_new, name
+    )

py/torch_tensorrt/fx/fx2trt.py

@@ -341,7 +341,12 @@ def call_method(self, target, args, kwargs):
 
     def output(self, target, args, kwargs):
         assert len(args) == 1
-        outputs = args[0] if isinstance(args[0], tuple) else (args[0],)
+        if isinstance(args[0], tuple):
+            outputs = args[0]
+        elif isinstance(args[0], list):
+            outputs = tuple(args[0])
+        else:
+            outputs = (args[0],)
 
         if not all(isinstance(output, trt.tensorrt.ITensor) for output in outputs):
             raise RuntimeError("TensorRT requires all outputs to be Tensor!")

py/torch_tensorrt/fx/lower.py

@@ -1,7 +1,5 @@
 import dataclasses as dc
 import logging
-import dataclasses as dc
-import logging
 from typing import Any, Callable, Optional, Sequence
 
 # @manual=//deeplearning/trt/python:py_tensorrt
@@ -180,8 +178,9 @@ def lower_pass(
         interp_res: TRTInterpreterResult = interpreter(mod, input, module_name)
         if lower_setting.use_experimental_rt:
             import io
-            from torch_tensorrt._TRTModuleNext import TRTModuleNext
+
             from torch_tensorrt._Device import Device
+            from torch_tensorrt._TRTModuleNext import TRTModuleNext
 
             with io.BytesIO() as engine_bytes:
                 engine_bytes.write(interp_res.engine.serialize())

py/torch_tensorrt/fx/lower_setting.py

@@ -66,8 +66,8 @@ class LowerSetting(LowerSettingBasic):
     cuda_graph_batch_size (int): Cuda graph batch size, default to be -1.
     preset_lowerer (str): when specified, use a preset logic to build the
     instance of Lowerer.
-    opt_profile_replica (int): the number of opt profile set for TensorRT engine, this field is
-    only used by explicit batch dim with dynamic shape mode.
+    only used by explicit batch dim with dynamic shape mode. In general, we use 2 GPU setting with
+    2 stream on each. Set total number to 8 as a safe default value.
     dynamic_batch: enable the dynamic shape in TRT with dim=-1 for the 1st dimension.
     tactic_sources: tactic sources for TensorRT kernel selection. Default to None,
     meaning all possible tactic sources.
@@ -81,17 +81,21 @@ class LowerSetting(LowerSettingBasic):
     explicit_precision: bool = False
     max_workspace_size: int = 1 << 30
     strict_type_constraints: bool = False
-    customized_fuse_pass: PassManager = PassManager.build_from_passlist([])
-    lower_basic_fuse_pass: PassManager = PassManager.build_from_passlist(
-        [fuse_permute_matmul, fuse_permute_linear]
+    customized_fuse_pass: PassManager = dc.field(
+        default_factory=lambda: PassManager.build_from_passlist([])
+    )
+    lower_basic_fuse_pass: PassManager = dc.field(
+        default_factory=lambda: PassManager.build_from_passlist(
+            [fuse_permute_matmul, fuse_permute_linear]
+        )
     )
     verbose_log: bool = False
     algo_selector = None
     timing_cache_prefix: str = ""
     save_timing_cache: bool = False
     cuda_graph_batch_size: int = -1
     preset_lowerer: str = ""
-    opt_profile_replica: int = 1
+    opt_profile_replica: int = 8
     dynamic_batch: bool = True
     tactic_sources: Optional[int] = None
     correctness_atol: float = 0.1

py/torch_tensorrt/fx/passes/lower_basic_pass.py

@@ -608,3 +608,25 @@ def _get_shape(node: fx.Node) -> Optional[torch.Size]:
         # shape info not available
         return None
     return node.meta["tensor_meta"].shape
+
+
+@log_before_after
+@validate_inference(atol=1e-3, rtol=1e-2)
+def fix_clamp_numerical_limits_to_fp16(
+    mod: torch.fx.GraphModule, input: Input
+) -> torch.fx.GraphModule:
+    MIN_FP16 = -65504.0
+    MAX_FP16 = 65504.0
+    for node in mod.graph.nodes:
+        if node.op == "call_function" and "clamp" in str(node.target):
+            input_kwargs = node.kwargs
+            if input_kwargs["min"] < MIN_FP16 and input_kwargs["max"] > MAX_FP16:
+                new_kwargs = {
+                    "input": input_kwargs["input"],
+                    "min": MIN_FP16,
+                    "max": MAX_FP16,
+                }
+                node.kwargs = new_kwargs
+
+    mod.recompile()
+    return mod