empty_memory_format evaluator #2745

apbose · 2024-04-12T00:22:16Z

In this PR I am facing issue in the test case of

    (
        "empty_four_dimension_memformat",
        [1, 2, 2, 1],
        torch.float32,
        "cuda",
        torch.channels_last,
    ),
    (
        "empty_five_dimension_memformat",
        [1, 2, 2, 2, 1],
        torch.float32,
        "cuda",
        torch.channels_last_3d,
    ),

In the TRTInterpreter.run(), the stride of the empty_tensor
torch.ops.aten.empty.memory_format([1,2,2,1], dtype = torch.int32, memory_format = torch.channels_last)
is (4,1,2,2) (the desired torch output. This is an evaluator so could check the output in this line -
However when it comes to line the stride is (4,2,1,1). This is the output from the TRT engine from the the above INetwork.

Would anyone know what is going wrong?

fixes: #2738

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:12:59.638980+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:14:55.849732+00:00
@@ -87,10 +87,11 @@
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.randn(*args[0])

+
def randperm_validator(randperm_node: Node) -> bool:
    dtype = randperm_node.kwargs.get("dtype", None)
    layout = randperm_node.kwargs.get("layout", None)
    input = randperm_node.args[0]
    if not isinstance(input, int):
@@ -116,10 +117,11 @@
    args: Tuple[Argument, ...],
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.permutation(args[0])
+

def empty_validator(empty_node: Node) -> bool:
    layout = empty_node.kwargs.get("layout", None)
    pin_memory = empty_node.kwargs.get("pin_memory", None)
    memory_format = empty_node.kwargs.get("memory_format", None)
@@ -162,8 +164,5 @@
    elif memory_format == torch.channels_last_3d:
        # shape of args[0] must be 5
        empty_tensor = empty_tensor.to(memory_format=torch.channels_last_3d)

    return empty_tensor
-
-
-

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:13:13.183345+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:15:09.256602+00:00
@@ -87,10 +87,11 @@
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.randn(*args[0])

+
def randperm_validator(randperm_node: Node) -> bool:
    dtype = randperm_node.kwargs.get("dtype", None)
    layout = randperm_node.kwargs.get("layout", None)
    input = randperm_node.args[0]
    if not isinstance(input, int):
@@ -116,10 +117,11 @@
    args: Tuple[Argument, ...],
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.permutation(args[0])
+

def empty_validator(empty_node: Node) -> bool:
    layout = empty_node.kwargs.get("layout", None)
    pin_memory = empty_node.kwargs.get("pin_memory", None)
    memory_format = empty_node.kwargs.get("memory_format", None)
@@ -162,8 +164,5 @@
    elif memory_format == torch.channels_last_3d:
        # shape of args[0] must be 5
        empty_tensor = empty_tensor.to(memory_format=torch.channels_last_3d)

    return empty_tensor
-
-
-

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:13:23.345164+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:15:15.812085+00:00
@@ -87,10 +87,11 @@
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.randn(*args[0])

+
def randperm_validator(randperm_node: Node) -> bool:
    dtype = randperm_node.kwargs.get("dtype", None)
    layout = randperm_node.kwargs.get("layout", None)
    input = randperm_node.args[0]
    if not isinstance(input, int):
@@ -116,10 +117,11 @@
    args: Tuple[Argument, ...],
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.permutation(args[0])
+

def empty_validator(empty_node: Node) -> bool:
    layout = empty_node.kwargs.get("layout", None)
    pin_memory = empty_node.kwargs.get("pin_memory", None)
    memory_format = empty_node.kwargs.get("memory_format", None)
@@ -162,8 +164,5 @@
    elif memory_format == torch.channels_last_3d:
        # shape of args[0] must be 5
        empty_tensor = empty_tensor.to(memory_format=torch.channels_last_3d)

    return empty_tensor
-
-
-

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:13:45.282455+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-05-24 00:15:36.375546+00:00
@@ -87,10 +87,11 @@
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.randn(*args[0])

+
def randperm_validator(randperm_node: Node) -> bool:
    dtype = randperm_node.kwargs.get("dtype", None)
    layout = randperm_node.kwargs.get("layout", None)
    input = randperm_node.args[0]
    if not isinstance(input, int):
@@ -116,10 +117,11 @@
    args: Tuple[Argument, ...],
    kwargs: Dict[str, Argument],
    name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
    return np.random.permutation(args[0])
+

def empty_validator(empty_node: Node) -> bool:
    layout = empty_node.kwargs.get("layout", None)
    pin_memory = empty_node.kwargs.get("pin_memory", None)
    memory_format = empty_node.kwargs.get("memory_format", None)
@@ -162,8 +164,5 @@
    elif memory_format == torch.channels_last_3d:
        # shape of args[0] must be 5
        empty_tensor = empty_tensor.to(memory_format=torch.channels_last_3d)

    return empty_tensor
-
-
-

zewenli98 · 2024-05-29T23:26:52Z

py/torch_tensorrt/dynamo/conversion/ops_evaluators.py

+    pin_memory = empty_node.kwargs.get("pin_memory", None)
+    memory_format = empty_node.kwargs.get("memory_format", None)


It seems pin_memory and memory_format are not used in the validator.

zewenli98 · 2024-05-29T23:28:55Z

tests/py/dynamo/conversion/test_empty_aten.py

+
+class TestRandConverter(DispatchTestCase):
+    @parameterized.expand(
+        [(empty_op[0], empty_op[1], empty_op[2], empty_op[3]) for empty_op in empty_ops]


Since each test has 5 arguments, do we need empty_op[4] here?

zewenli98 · 2024-05-29T23:34:20Z

tests/py/dynamo/conversion/test_empty_aten.py

+
+            def forward(self, x):
+                shape_or_input[0] = x.shape[0]
+                return torch.empty(shape_or_input)


For consistence, I think it's better to use torch.ops.aten.empty.memory_format with necessary args.

zewenli98 · 2024-05-29T23:35:04Z

tests/py/dynamo/conversion/test_empty_aten.py

+    @parameterized.expand(
+        [(empty_op[0], empty_op[1], empty_op[2], empty_op[3]) for empty_op in empty_ops]
+    )
+    def test_empty(self, name, shape_or_input, data_type, device):


It seems data_type and device are not tested.

zewenli98 · 2024-05-29T23:35:56Z

tests/py/dynamo/conversion/test_empty_aten.py

+import torch
+import torch.nn as nn
+import torch_tensorrt
+from harness import DispatchTestCase


I think it should be from .harness import DispatchTestCase

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_empty_aten.py	2024-06-05 01:02:35.600566+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_empty_aten.py	2024-06-05 01:04:39.444401+00:00
@@ -87,11 +87,14 @@
                super().__init__()

            def forward(self, x):
                shape_or_input[0] = x.shape[0]
                return torch.ops.aten.empty.memory_format(
-                    shape_or_input, dtype=data_type, memory_format=memory_format, device=device
+                    shape_or_input,
+                    dtype=data_type,
+                    memory_format=memory_format,
+                    device=device,
                )

        empty_model = TestModule()

        inputs = [torch.randint(1, 3, shape_or_input, dtype=torch.int32)]

zewenli98

Just a minor comment. The other LGTM

zewenli98 · 2024-06-10T20:37:18Z

tests/py/dynamo/conversion/test_empty_aten.py

+            def __init__(self):
+                super().__init__()


The init seems not necessary.

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-06-13 17:55:17.730470+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-06-13 17:57:14.508152+00:00
@@ -130,12 +130,14 @@
    if layout is not None:
        _LOGGER.debug(f"Currently we don't support specifying layout, got {layout}.")
        return False
    memory_format = empty_node.kwargs.get("memory_format", None)
    if memory_format is not None:
-        _LOGGER.debug(f"Currently we don't support specifying memory_format, got {memory_format}.")
-        return False    
+        _LOGGER.debug(
+            f"Currently we don't support specifying memory_format, got {memory_format}."
+        )
+        return False
    return True


@dynamo_tensorrt_converter(
    torch.ops.aten.empty.memory_format, capability_validator=empty_validator

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-06-14 21:26:46.844227+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-06-14 21:29:08.372195+00:00
@@ -130,12 +130,14 @@
    if layout is not None:
        _LOGGER.debug(f"Currently we don't support specifying layout, got {layout}.")
        return False
    memory_format = empty_node.kwargs.get("memory_format", None)
    if memory_format is not None:
-        _LOGGER.debug(f"Currently we don't support specifying memory_format, got {memory_format}.")
-        return False    
+        _LOGGER.debug(
+            f"Currently we don't support specifying memory_format, got {memory_format}."
+        )
+        return False
    return True


@dynamo_tensorrt_converter(
    torch.ops.aten.empty.memory_format, capability_validator=empty_validator

github-actions

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-06-14 21:37:02.453645+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/ops_evaluators.py	2024-06-14 21:38:54.121244+00:00
@@ -177,12 +177,14 @@
    if layout is not None:
        _LOGGER.debug(f"Currently we don't support specifying layout, got {layout}.")
        return False
    memory_format = empty_node.kwargs.get("memory_format", None)
    if memory_format is not None:
-        _LOGGER.debug(f"Currently we don't support specifying memory_format, got {memory_format}.")
-        return False    
+        _LOGGER.debug(
+            f"Currently we don't support specifying memory_format, got {memory_format}."
+        )
+        return False
    return True


@dynamo_tensorrt_converter(
    torch.ops.aten.empty.memory_format, capability_validator=empty_validator

Review comments- adding cases for stride, correcting validator and changing call to torch.ops.aten.empty.memory_format Review comments- adding cases for stride, correcting validator and changing call to torch.ops.aten.empty.memory_format Removing stride and device since - 1. converting to torch Tensor would lead to faketensor 2. Also get_trt_tensor in the forward pass while creating the engine does not retain the memory_format Removing init from empty test

apbose self-assigned this Apr 12, 2024

facebook-github-bot added the cla signed label Apr 12, 2024

apbose marked this pull request as draft April 12, 2024 00:22

github-actions bot added component: tests Issues re: Tests component: conversion Issues re: Conversion stage component: api [Python] Issues re: Python API component: dynamo Issues relating to the `torch.compile` or `torch._dynamo.export` paths labels Apr 12, 2024

github-actions bot requested a review from gs-olive April 12, 2024 00:23

apbose mentioned this pull request Apr 12, 2024

empty_permute decomposition #2698

Merged

apbose force-pushed the empty_memory_format_evaluator branch 2 times, most recently from ed97b61 to ad5b467 Compare May 24, 2024 00:12

apbose marked this pull request as ready for review May 24, 2024 00:13

apbose requested review from chohk88 and zewenli98 May 24, 2024 00:13

github-actions bot requested changes May 24, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch from ad5b467 to fd330cf Compare May 24, 2024 01:50

zewenli98 requested changes May 29, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch from f97b696 to 52fc6a9 Compare June 5, 2024 01:02

github-actions bot requested changes Jun 5, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch 2 times, most recently from 1854b76 to a135bab Compare June 5, 2024 01:14

zewenli98 approved these changes Jun 10, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch from 5bc1798 to efe6745 Compare June 13, 2024 17:55

github-actions bot requested changes Jun 13, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch from efe6745 to 0180ae2 Compare June 14, 2024 21:26

github-actions bot requested changes Jun 14, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch from 0180ae2 to 1719a13 Compare June 14, 2024 21:36

github-actions bot requested changes Jun 14, 2024

View reviewed changes

apbose force-pushed the empty_memory_format_evaluator branch from 1719a13 to 659195e Compare June 14, 2024 21:55

apbose merged commit 0387757 into main Jun 14, 2024

		pin_memory = empty_node.kwargs.get("pin_memory", None)
		memory_format = empty_node.kwargs.get("memory_format", None)

empty_memory_format evaluator #2745

empty_memory_format evaluator #2745

Uh oh!

Conversation

apbose commented Apr 12, 2024 • edited by narendasan Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

zewenli98 May 29, 2024

Choose a reason for hiding this comment

Uh oh!

zewenli98 May 29, 2024

Choose a reason for hiding this comment

Uh oh!

zewenli98 May 29, 2024

Choose a reason for hiding this comment

Uh oh!

zewenli98 May 29, 2024

Choose a reason for hiding this comment

Uh oh!

zewenli98 May 29, 2024

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

zewenli98 left a comment

Choose a reason for hiding this comment

Uh oh!

zewenli98 Jun 10, 2024

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

apbose commented Apr 12, 2024 •

edited by narendasan

Loading