Upgrade torch nightly version to 2.8.0.dev20250408+cpu by dayo09 · Pull Request #30 · Samsung/TICO

dayo09 · 2025-04-09T07:45:33Z

Let's upgrade torch nightly version to the latest.

TICO-DCO-1.0-Signed-off-by: Dayoung Lee dayoung.lee@samsung.com

Related to: #9

dayo09 · 2025-04-09T07:59:15Z

.github/workflows/check-pr.yaml

    strategy:
      matrix:
-        torch-version: ["2.5", "2.6"]
+        torch-version: ["2.5", "2.6", "2.7", "nightly"]


NOTE: pytorch 2.7.0 release date is planed to 4/25, due to https://dev-discuss.pytorch.org/t/pytorch-release-2-7-0-call-for-features/2780.

test/modules/net/interpolate.py

tico/passes/lower_to_resize_nearest_neighbor.py

mhs4670go

jinevening · 2025-04-10T01:27:47Z

It seems that this PR includes some chanages about ResizeNearestNeighbor (not sure this is the only change), but they are hidden behind the commit message of "Upgrade torch nightly version".

I suggest to make an issue whenever we upgrade torch version, and list up related issues there. That will be helpful to track our decisions for supporting new torch versions.

@dayo09 Could you make an issue for this update?

dayo09 · 2025-04-10T01:31:55Z

@jinevening Sorry that I forgot to link the issue #9.

dayo09 · 2025-04-10T01:32:44Z

tico/utils/validate_args_kwargs.py


    input: torch.fx.Node
-    dim: int
+    dim: int = 0


Some operators have void dim field in torch 2.8.dev

Does it occur only in SliceArgs? If not, how about to split the PR to set the default value for dim arg?

dayo09 · 2025-04-10T01:33:20Z

tico/utils/convert.py

            SegmentIndexSelectConst(),
            LegalizeCausalMaskValue(enabled=config.get("legalize_causal_mask_value")),
-            ConvertIndexToResizeNearestNeighbor(),
+            LowerToResizeNearestNeighbor(),


Now that Upsample is added to the Lowering to Resize Nearest Neighbor target list, I intergrate them and rename the pass.

dayo09 · 2025-04-10T01:34:03Z

tico/utils/convert.py

    # This pass should be run before 'RestoreLinear' and after 'decompose_quantize_op'.
    # TODO run pass regardless of the orders.
-    with SuppressWarning(UserWarning, ".*quantize_per_tensor"):
+    with SuppressWarning(UserWarning, ".*quantize_per_tensor"), SuppressWarning(


To suppress the warning from 2.8. It seems oneDNN kinda thirdparty feature is added. It's very unnecessary warning for us.

dayo09 · 2025-04-10T01:34:21Z

tico/passes/lower_to_resize_nearest_neighbor.py

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


This part is almost the same to the original code.

dayo09 · 2025-04-10T01:34:39Z

tico/passes/lower_to_resize_nearest_neighbor.py

+
+        return resize_nearest_neighbor
+
+    def convert_upsample_nearest2d_to_resize_nearest_neighbor(


This part is newly added.

dayo09 · 2025-04-10T01:39:06Z

test/modules/net/interpolate.py

+@tag.skip_if(
+    not HAS_TORCH_OVER_28_DEV,
+    reason="The case isn't supported yet. It will be supported from torch 2.8.0.dev",
+)
+class InterpolateOnePointFive(torch.nn.Module):


Q. Why it's not supported in lower version?

A.

CASE(torch < 2.8.0.dev)

it's converted to index. -> It's interpreted to "Multiple indices" case. -> it's not supported in tico yet.

CASE(torch >= 2.8.0.dev)

it's converted to upsample nearest 2d. -> it converts directly to RNN(ResizeNearsestNeighbor) in tico. -> Can support 1.5 times.

Can I ask what is "Multiple indices" case?

@jinevening IndexTensorVisitor don't support multiple indices.

# TODO Support multiple indices if len(indices) - indices.count(None) > 1: # type: ignore[arg-type] raise NotYetSupportedError( "Multiple indices is not supported yet in aten.index.Tensor" )

This is current implementaion. I left it as TODO.
I didn't look into what values were. 🤔

tico/passes/lower_to_resize_nearest_neighbor.py

jinevening · 2025-04-10T02:28:58Z

test/modules/net/interpolate.py

+from test.utils import tag


 class InterpolateDouble(torch.nn.Module):


(not in this PR) BTW, it seems that this file can be moved to op, not net.

jinevening · 2025-04-10T02:31:40Z

Ah, I missed #9 (added a back link to the main comment). I left some comments. PTAL

dayo09 · 2025-04-10T02:43:41Z

#37 must be merged to resolve the lint error.

miusic · 2025-04-10T02:48:05Z

tico/utils/validate_args_kwargs.py

 class ResizeNearestNeighborArgs:
    """
-    # Only consider `torch.nn.functional.interpolate(x, scale_factor=2.0, mode='nearest')` case.
+    # Maps from `torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode='nearest')` case.


Suggested change

# Maps from `torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode='nearest')` case.

# Mapped from `torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode='nearest')` case.

(nit)
May be..?

Let's upgrade torch nightly version to the latest. TICO-DCO-1.0-Signed-off-by: Dayoung Lee <dayoung.lee@samsung.com>

Co-authored-by: seongwoo chae <mhs4670go@naver.com>

miusic · 2025-04-10T04:24:28Z

tico/passes/lower_to_resize_nearest_neighbor.py

+        [BEFORE PASS]
+        input - aten.index - output


Suggested change

[BEFORE PASS]

input - aten.index - output

[BEFORE PASS]

case 1.

input - aten.index - output

case 2.

input - aten.upsample_nearest2d.vec - output

It would be nice to add a case for aten.upsample_nearest2d.vec :D

Let me apply in the next PR :-D

miusic

LGTM with some comments 👍

mhs4670go · 2025-04-10T05:31:15Z

test/pt2_to_circle_test/test_pt2_to_circle.py

    model: torch.nn.Module, example_inputs: tuple, pt2_model_path: str
 ):
    # Create .pt2 model
-    with torch.no_grad(), SuppressWarning(UserWarning, ".*quantize_per_tensor"):


Not related with this PR but seems that quantize_per_tensor warning has been removed.

mhs4670go

jinevening

LGTM

dayo09 commented Apr 9, 2025

View reviewed changes

dayo09 force-pushed the 0409-upgrade branch 2 times, most recently from bd7971e to 8a05322 Compare April 9, 2025 08:26

dayo09 marked this pull request as ready for review April 9, 2025 08:26

dayo09 added the ready for review label Apr 9, 2025

mhs4670go reviewed Apr 9, 2025

View reviewed changes

test/modules/net/interpolate.py Outdated Show resolved Hide resolved

tico/passes/lower_to_resize_nearest_neighbor.py Outdated Show resolved Hide resolved

dayo09 force-pushed the 0409-upgrade branch from c377ad1 to 5c7a77f Compare April 9, 2025 10:08

mhs4670go previously approved these changes Apr 9, 2025

View reviewed changes

dayo09 requested a review from a team April 10, 2025 01:09

dayo09 commented Apr 10, 2025

View reviewed changes

dayo09 requested a review from jinevening April 10, 2025 01:39

jinevening reviewed Apr 10, 2025

View reviewed changes

tico/passes/lower_to_resize_nearest_neighbor.py Show resolved Hide resolved

jinevening reviewed Apr 10, 2025

View reviewed changes

dayo09 dismissed mhs4670go’s stale review via f2c94ee April 10, 2025 02:41

miusic reviewed Apr 10, 2025

View reviewed changes

dayo09 and others added 3 commits April 10, 2025 13:08

Upgrade torch nightly version to 2.8.0.dev20250408+cpu

81e1c5b

Let's upgrade torch nightly version to the latest. TICO-DCO-1.0-Signed-off-by: Dayoung Lee <dayoung.lee@samsung.com>

Update tico/passes/lower_to_resize_nearest_neighbor.py

f94f9f0

Co-authored-by: seongwoo chae <mhs4670go@naver.com>

Apply feedback

797dcc0

dayo09 force-pushed the 0409-upgrade branch from f2c94ee to 797dcc0 Compare April 10, 2025 04:08

miusic reviewed Apr 10, 2025

View reviewed changes

miusic approved these changes Apr 10, 2025

View reviewed changes

dayo09 requested review from jinevening and mhs4670go April 10, 2025 04:44

mhs4670go reviewed Apr 10, 2025

View reviewed changes

mhs4670go approved these changes Apr 10, 2025

View reviewed changes

jinevening approved these changes Apr 10, 2025

View reviewed changes

jinevening merged commit f0aeefc into Samsung:main Apr 10, 2025
6 checks passed


		return resize_nearest_neighbor

		def convert_upsample_nearest2d_to_resize_nearest_neighbor(

		from test.utils import tag


		class InterpolateDouble(torch.nn.Module):

	# Maps from `torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode='nearest')` case.
	# Mapped from `torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode='nearest')` case.

Conversation

dayo09 commented Apr 9, 2025 • edited by jinevening Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

mhs4670go left a comment

Choose a reason for hiding this comment

Uh oh!

jinevening commented Apr 10, 2025

Uh oh!

dayo09 commented Apr 10, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jinevening commented Apr 10, 2025

Uh oh!

dayo09 commented Apr 10, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

miusic left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mhs4670go left a comment

Choose a reason for hiding this comment

Uh oh!

jinevening left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

dayo09 commented Apr 9, 2025 •

edited by jinevening

Loading