Replace mypy with ty for static type checking by svij-sc · Pull Request #585 · Snapchat/GiGL

svij-sc · 2026-04-13T22:50:33Z

ty (from Astral, makers of ruff/uv) replaces mypy as the project's static type checker.

The ty fixes are introduced here: #586

Benchmarked on the code base (631 source files):

Metric	mypy	ty	Speedup
Cold cache	184.6s	0.79s	~234x
Warm cache	13.9s	0.83s	~17x

Changes:

Add ty to lint deps, remove mypy + mypy-extensions (keep mypy-protobuf for protoc stub generation)
Migrate mypy.ini config to [tool.ty.*] sections in pyproject.toml
Update Makefile type_check target: uv run ty check
Downgrade 12 rule categories to "warn" for initial migration; a follow-up PR will fix the 226 warnings and promote them back to errors

Scope of work done

Where is the documentation for this feature?: N/A

Did you add automated tests or write a test plan?

Updated Changelog.md? NO

Ready for code review?: YES

ty (from Astral, makers of ruff/uv) replaces mypy as the project's static type checker. Benchmark on this codebase (631 source files): | Metric | mypy | ty | Speedup | |-------------|----------|-------|---------| | Cold cache | 184.6s | 0.79s | ~234x | | Warm cache | 13.9s | 0.83s | ~17x | Changes: - Add ty to lint deps, remove mypy + mypy-extensions (keep mypy-protobuf for protoc stub generation) - Migrate mypy.ini config to [tool.ty.*] sections in pyproject.toml - Update Makefile type_check target: `uv run ty check` - Downgrade 12 rule categories to "warn" for initial migration; a follow-up PR will fix the 226 warnings and promote them back to errors - Update docs (CLAUDE.md, README.md, CHANGELOG.md, .claude/formatting.md) - Update .gitignore/.dockerignore cache directory references - Replace "mypy" mentions in source comments with generic "type checker" - Delete mypy.ini Co-Authored-By: shubhamvij <25601958+shubhamvij@users.noreply.github.com>

- Pin ty~=0.0.29 for reproducible builds - Add comment clarifying mypy-protobuf is for stub generation only - Exclude notebooks (*.ipynb) from type checking - Add comment about torch lacking inline types/stubs - Promote all ty rules back to default error severity (remove warn downgrades — fixes come in a follow-up PR) - Add [[tool.ty.overrides]] to suppress errors in generated *_pb2.py files that cannot be fixed Co-Authored-By: shubhamvij <25601958+shubhamvij@users.noreply.github.com>

svij-sc · 2026-04-15T01:53:48Z

+[tool.ty.src]
+exclude = ["*_pb2.py", "*_pb2.pyi", "**/*.ipynb"]
+
+[tool.ty.analysis]


Migrated from mypy.ini - removed any that could be fixed - thank you robots!

kmontemayor2-sc · 2026-04-15T02:54:59Z


    # Need to do this "backwards" so the parent class can be defined first.
-    # Otherwise, mypy complains that:
+    # Otherwise, the type checker complains that:


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

I threw bot at it, looks like we still might need these.

kmontemayor2-sc · 2026-04-15T02:55:02Z


        if is_positive:
-            # This assert is added to pass mypy type check, in practice we will not see this fail
+            # This assert is added to pass the type checker, in practice we will not see this fail


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

kmontemayor2-sc · 2026-04-15T02:55:09Z

                self._positive_label_edge_index = None
        else:
-            # This assert is added to pass mypy type check, in practice we will not see this fail
+            # This assert is added to pass the type checker, in practice we will not see this fail


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

kmontemayor2-sc · 2026-04-15T02:55:13Z

        else None,
        # This should be `aiplatform.gapic.Scheduling.Strategy[inferencer_resource_config.scheduling_strategy]`
-        # But mypy complains otherwise...
+        # But the type checker complains otherwise...


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

kmontemayor2-sc · 2026-04-15T02:55:19Z



-# Below type ignores are due to mypy star expansion issues: https://github.com/python/mypy/issues/6799
+# Below type ignores are due to star expansion issues with the type checker: https://github.com/python/mypy/issues/6799


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

fixed - nice

Wait did we keep the link as is?

kmontemayor2-sc · 2026-04-15T02:55:34Z

 from gigl.src.data_preprocessor.lib.types import InstanceDictPTransform

-# Type hints for abstract dataclasses are currently not supported. https://github.com/python/mypy/issues/5374
+# Type hints for abstract dataclasses may have limited support in type checkers. https://github.com/python/mypy/issues/5374


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

Also update the TODO from python/mypy?

this was fixed with ty - updated

ditto on link.

kmontemayor2-sc · 2026-04-15T02:55:42Z

+    # We cannot at the moment as we type-ignore GLT
+    # And adding it as a type here will break the type checker.
    # PartitionBook


Shall we check that ty still complains?

(I am going to leave identical comments on all of the changes like this - that way the robots can pick them up and check automatically)

kmontemayor2-sc · 2026-04-15T02:56:40Z

+python-version = "3.11"
+
+[tool.ty.src]
+exclude = ["*_pb2.py", "*_pb2.pyi", "**/*.ipynb"]


should we add a TODO to include our notebooks here eventually?

Tbh, I think its fine if we don't
Notebooks need not have verbose static type checking

Most act as documentation

kmontemayor2-sc · 2026-04-15T02:59:04Z

+  "kfp.**",
+  "tensorflow.**",
+  "tensorflow_metadata.**",
+  "torch.**",


hmmmm, we weren't type-ignoring torch before.

IIRC I saw some issues with torch when I looked into this earlier, the ty people suggested to enable some flag - did you try this? astral-sh/ty#2244

Since we're so reliant on torch I'd rather not type-ignore it if possible.

Unforunately the flag is not the problem

From the robots analysis, there are 200+ issues that were not being caught before; I think we should reserve for a future change?
Seems might just be using the apis in an incorrect way according to types - event hough it seems to be working.

Group 1: Tensor.getitem with NodeType/EdgeType/CondensedNodeType keys (~64 errors)

These happen because dict[NodeType, Tensor] access returns Tensor, and then a second indexing
like result[tensor_index] hits Tensor's strict getitem signature. The custom NewType keys
(NodeType = str, CondensedNodeType = int, EdgeType = NamedTuple) don't satisfy Tensor's index
types.

Fix: These are dict-on-dict access patterns — the real issue is that ty is resolving the chained
subscript incorrectly or the intermediate type needs explicit annotation. Likely fixable via
local variable annotations or targeted # type: ignore comments.

Group 2: FloatTensor/LongTensor vs Tensor return type mismatches (~25 errors)

Functions annotated to return FloatTensor/LongTensor but torch.gather(), torch.stack(),
torch.cat(), torch.mean(), torch.tensor() all return generic Tensor.

Fix: Change return type annotations from FloatTensor/LongTensor to Tensor. These deprecated
aliases are not what torch operations actually return.

Group 3: init argument type mismatches (~20 errors)

Parameters using torch.dtype, torch.device, etc. where ty resolves the stub types more strictly.

Fix: Explicit casts or type annotations at call sites.

Group 4: HeteroData/dict access returning object/Unknown (~15 errors)

When iterating over graph data structures, values come back as object rather than typed tensors.

Fix: Add explicit type annotations to loop variables.

Group 5: List invariance issues (~8 errors)

list[Tensor | None] assigned to list[Tensor], or Optional[Tensor] appended to list[FloatTensor].

Fix: Widen the list type annotations.

Group 6: Misc (operators, spawn, call-non-callable, etc.) (~10 errors)

Fix: Case-by-case — spawn import fix, operator type annotations, etc.

I think I'd prefer to keep the torch typing around (if my reading is correct that this infact is equivalent to type-ignoring torch) and just add a bunch of # type-ignore and TODOs to fix.

IDK how feasible that'd be to setup tho.

kmontemayor2-sc · 2026-04-15T02:59:32Z

+# https://docs.astral.sh/ty/reference/configuration/#replace-imports-with-any
+replace-imports-with-any = [
+  "apache_beam.**",
+  "google.cloud.**",


likewise, is there a reason we're blanket ignoring all gcp imports?

svij-sc

I updated the downstream PR where I addressed the ty fixes and to it added some new fixes identified by @kmontemayor2-sc below.

Ref: #586

svij-sc · 2026-04-15T06:35:06Z

+  "kfp.**",
+  "tensorflow.**",
+  "tensorflow_metadata.**",
+  "torch.**",


Unforunately the flag is not the problem

From the robots analysis, there are 200+ issues that were not being caught before; I think we should reserve for a future change?
Seems might just be using the apis in an incorrect way according to types - event hough it seems to be working.

Group 1: Tensor.getitem with NodeType/EdgeType/CondensedNodeType keys (~64 errors)

These happen because dict[NodeType, Tensor] access returns Tensor, and then a second indexing
like result[tensor_index] hits Tensor's strict getitem signature. The custom NewType keys
(NodeType = str, CondensedNodeType = int, EdgeType = NamedTuple) don't satisfy Tensor's index
types.

Fix: These are dict-on-dict access patterns — the real issue is that ty is resolving the chained
subscript incorrectly or the intermediate type needs explicit annotation. Likely fixable via
local variable annotations or targeted # type: ignore comments.

Group 2: FloatTensor/LongTensor vs Tensor return type mismatches (~25 errors)

Functions annotated to return FloatTensor/LongTensor but torch.gather(), torch.stack(),
torch.cat(), torch.mean(), torch.tensor() all return generic Tensor.

Fix: Change return type annotations from FloatTensor/LongTensor to Tensor. These deprecated
aliases are not what torch operations actually return.

Group 3: init argument type mismatches (~20 errors)

Parameters using torch.dtype, torch.device, etc. where ty resolves the stub types more strictly.

Fix: Explicit casts or type annotations at call sites.

Group 4: HeteroData/dict access returning object/Unknown (~15 errors)

When iterating over graph data structures, values come back as object rather than typed tensors.

Fix: Add explicit type annotations to loop variables.

Group 5: List invariance issues (~8 errors)

list[Tensor | None] assigned to list[Tensor], or Optional[Tensor] appended to list[FloatTensor].

Fix: Widen the list type annotations.

Group 6: Misc (operators, spawn, call-non-callable, etc.) (~10 errors)

Fix: Case-by-case — spawn import fix, operator type annotations, etc.

svij-sc · 2026-04-17T23:11:31Z


    # Need to do this "backwards" so the parent class can be defined first.
-    # Otherwise, mypy complains that:
+    # Otherwise, the type checker complains that:


I threw bot at it, looks like we still might need these.

svij-sc · 2026-04-17T23:13:58Z

 from gigl.src.data_preprocessor.lib.types import InstanceDictPTransform

-# Type hints for abstract dataclasses are currently not supported. https://github.com/python/mypy/issues/5374
+# Type hints for abstract dataclasses may have limited support in type checkers. https://github.com/python/mypy/issues/5374


this was fixed with ty - updated

svij-sc · 2026-04-17T23:14:28Z



-# Below type ignores are due to mypy star expansion issues: https://github.com/python/mypy/issues/6799
+# Below type ignores are due to star expansion issues with the type checker: https://github.com/python/mypy/issues/6799


fixed - nice

svij-sc · 2026-04-17T23:15:43Z

        else None,
        # This should be `aiplatform.gapic.Scheduling.Strategy[inferencer_resource_config.scheduling_strategy]`
-        # But mypy complains otherwise...
+        # But the type checker complains otherwise...


svij-sc · 2026-04-17T23:16:06Z

                self._positive_label_edge_index = None
        else:
-            # This assert is added to pass mypy type check, in practice we will not see this fail
+            # This assert is added to pass the type checker, in practice we will not see this fail


svij-sc · 2026-04-17T23:16:24Z


        if is_positive:
-            # This assert is added to pass mypy type check, in practice we will not see this fail
+            # This assert is added to pass the type checker, in practice we will not see this fail


svij-sc · 2026-04-17T23:20:02Z

+python-version = "3.11"
+
+[tool.ty.src]
+exclude = ["*_pb2.py", "*_pb2.pyi", "**/*.ipynb"]


Tbh, I think its fine if we don't
Notebooks need not have verbose static type checking

Most act as documentation

svij-sc · 2026-04-17T23:26:15Z

+# https://docs.astral.sh/ty/reference/configuration/#replace-imports-with-any
+replace-imports-with-any = [
+  "apache_beam.**",
+  "google.cloud.**",


svij-sc requested review from kmontemayor2-sc, mkolodner-sc, nshah-sc, xgao4-sc, yliu2-sc and zfan3-sc as code owners April 13, 2026 22:50

svij-sc and others added 3 commits April 15, 2026 00:59

Merge remote-tracking branch 'origin/main' into svij/use-ty

60dd366

revert changelog

1136c2d

svij-sc mentioned this pull request Apr 15, 2026

Implement ty fixes #586

Open

svij-sc commented Apr 15, 2026

View reviewed changes

svij-sc added 4 commits April 15, 2026 02:14

update

ada8d42

update pyproject.toml

0e40212

update pyproject

789f1c6

format

00c05e9

kmontemayor2-sc reviewed Apr 15, 2026

View reviewed changes

svij-sc commented Apr 17, 2026

View reviewed changes

svij-sc added 2 commits April 17, 2026 23:52

address pr comments

3fd253a

merge main

db6e523



		# Below type ignores are due to mypy star expansion issues: https://github.com/python/mypy/issues/6799
		# Below type ignores are due to star expansion issues with the type checker: https://github.com/python/mypy/issues/6799

Conversation

svij-sc commented Apr 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

svij-sc Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

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!

svij-sc left a comment

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!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

svij-sc commented Apr 13, 2026 •

edited

Loading

svij-sc Apr 15, 2026 •

edited

Loading