[Relax][PyTroch] Add randn.default and randn_like.default support by guan404ming · Pull Request #18815 · apache/tvm

guan404ming · 2026-02-23T18:47:34Z

Why

PyTorch models using torch.randn() or torch.randn_like() fail to convert via from_exported_program (part of #18476).

How

Add _randn and _randn_like handlers that emit constant tensors sampled
Add tests verifying conversion produces correct output shape and dtype

gemini-code-assist · 2026-02-23T18:47:50Z

Summary of Changes

Hello @guan404ming, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses a critical issue where PyTorch models utilizing torch.randn() or torch.randn_like() failed to convert properly using the from_exported_program utility. By introducing specific handlers for these operations, the PR ensures that models incorporating random number generation can now be successfully translated, improving the compatibility and robustness of the conversion process.

Highlights

PyTorch randn and randn_like support: Implemented handlers for torch.randn() and torch.randn_like() operations to enable their conversion via from_exported_program.
Test coverage: Added dedicated test cases to verify the correct conversion and output shape/dtype for randn and randn_like operations.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

python/tvm/relax/frontend/torch/exported_program_translator.py
- Added _randn method to handle torch.randn operations, emitting constant tensors sampled using NumPy.
- Added _randn_like method to handle torch.randn_like operations, determining shape and dtype from the input tensor.
- Registered randn.default and randn_like.default with their respective new handlers in the create_convert_map.
tests/python/relax/test_frontend_from_exported_program.py
- Added test_randn to verify the conversion of a model using torch.randn and assert its output shape and dtype.
- Added test_randn_like to verify the conversion of a model using torch.randn_like and assert its output shape and dtype.

Activity

No human activity has been recorded on this pull request yet.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

Signed-off-by: Guan-Ming Chiu <guanmingchiu@gmail.com>

gemini-code-assist

Code Review

This pull request adds support for torch.randn and torch.randn_like in the PyTorch frontend by emitting constant tensors. This is a solid approach for handling these operations during model import. The accompanying tests correctly verify the output shape and data type for static shapes.

My main feedback concerns the handling of dynamic shapes. The current implementations for _randn and _randn_like will fail if they encounter symbolic dimensions, as they rely on numpy which requires concrete shapes. I've suggested adding explicit checks to raise a NotImplementedError in such cases to prevent crashes and make the limitation clear.

gemini-code-assist · 2026-02-23T18:51:55Z

python/tvm/relax/frontend/torch/exported_program_translator.py

+    def _randn(self, node: fx.Node) -> relax.Var:
+        import numpy as np
+
+        args = self.retrieve_args(node)
+        size = args[0] if isinstance(args[0], (list, tuple)) else (args[0],)
+        dtype = self._convert_data_type(
+            node.kwargs.get("dtype", torch.get_default_dtype()), self.env
+        )
+        data = np.random.randn(*size).astype(dtype)
+        return self.block_builder.emit(relax.const(data, dtype))


This implementation of _randn will fail if size contains symbolic dimensions (i.e., tir.Var from torch.export.Dim), as np.random.randn requires concrete integer shapes. To prevent a crash with dynamic shapes, it's better to add an explicit check and raise a NotImplementedError.

Suggested change

def _randn(self, node: fx.Node) -> relax.Var:

import numpy as np

args = self.retrieve_args(node)

size = args[0] if isinstance(args[0], (list, tuple)) else (args[0],)

dtype = self._convert_data_type(

node.kwargs.get("dtype", torch.get_default_dtype()), self.env

)

data = np.random.randn(*size).astype(dtype)

return self.block_builder.emit(relax.const(data, dtype))

def _randn(self, node: fx.Node) -> relax.Var:

import numpy as np

args = self.retrieve_args(node)

size = args[0] if isinstance(args[0], (list, tuple)) else (args[0],)

if any(isinstance(s, tir.Var) for s in size):

raise NotImplementedError("torch.randn with dynamic shapes is not supported yet.")

dtype = self._convert_data_type(

node.kwargs.get("dtype", torch.get_default_dtype()), self.env

)

data = np.random.randn(*size).astype(dtype)

return self.block_builder.emit(relax.const(data, dtype))

int(s) will already raise a TypeError with a clear message. Adding explicit checks here would be inconsistent with the rest of the codebase.

gemini-code-assist · 2026-02-23T18:51:56Z

python/tvm/relax/frontend/torch/exported_program_translator.py

+    def _randn_like(self, node: fx.Node) -> relax.Var:
+        import numpy as np
+
+        x = self.env[node.args[0]]
+        x_sinfo = x.struct_info
+        shape = [int(s) for s in x_sinfo.shape]
+        dtype = self._convert_data_type(node.kwargs.get("dtype", None) or x_sinfo.dtype, self.env)
+        data = np.random.randn(*shape).astype(dtype)
+        return self.block_builder.emit(relax.const(data, dtype))
+
    def _zeros(self, node: fx.Node) -> relax.Var:


Similar to _randn, this implementation for _randn_like does not support dynamic shapes. The line shape = [int(s) for s in x_sinfo.shape] will raise a TypeError if any shape dimension s is a symbolic tir.Var. Please add a check to handle this case gracefully.

Suggested change

def _randn_like(self, node: fx.Node) -> relax.Var:

import numpy as np

x = self.env[node.args[0]]

x_sinfo = x.struct_info

shape = [int(s) for s in x_sinfo.shape]

dtype = self._convert_data_type(node.kwargs.get("dtype", None) or x_sinfo.dtype, self.env)

data = np.random.randn(*shape).astype(dtype)

return self.block_builder.emit(relax.const(data, dtype))

def _zeros(self, node: fx.Node) -> relax.Var:

def _randn_like(self, node: fx.Node) -> relax.Var:

import numpy as np

x = self.env[node.args[0]]

x_sinfo = x.struct_info

if any(isinstance(s, tir.Var) for s in x_sinfo.shape):

raise NotImplementedError("torch.randn_like with dynamic shapes is not supported yet.")

shape = [int(s) for s in x_sinfo.shape]

dtype = self._convert_data_type(

node.kwargs.get("dtype", None) or x_sinfo.dtype, self.env

)

data = np.random.randn(*shape).astype(dtype)

return self.block_builder.emit(relax.const(data, dtype))

guan404ming · 2026-02-24T04:16:39Z

cc @tlopex @mshr-h

guan404ming · 2026-02-24T16:03:21Z

Thanks!

[Relax][PyTroch] Add randn.default and randn_like.default support

bdf8da8

Signed-off-by: Guan-Ming Chiu <guanmingchiu@gmail.com>

guan404ming force-pushed the feat/pytorch-frontend-randn-support branch from a2d36fc to bdf8da8 Compare February 23, 2026 18:51

gemini-code-assist bot reviewed Feb 23, 2026

View reviewed changes

guan404ming marked this pull request as ready for review February 24, 2026 04:15

mshr-h approved these changes Feb 24, 2026

View reviewed changes

mshr-h merged commit 6d46fa7 into apache:main Feb 24, 2026
13 checks passed

guan404ming deleted the feat/pytorch-frontend-randn-support branch February 24, 2026 16:03

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Relax][PyTroch] Add randn.default and randn_like.default support#18815

[Relax][PyTroch] Add randn.default and randn_like.default support#18815
mshr-h merged 1 commit intoapache:mainfrom
guan404ming:feat/pytorch-frontend-randn-support

guan404ming commented Feb 23, 2026

Uh oh!

gemini-code-assist bot commented Feb 23, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Feb 23, 2026

Uh oh!

guan404ming Feb 23, 2026

Uh oh!

gemini-code-assist bot Feb 23, 2026

Uh oh!

guan404ming Feb 23, 2026

Uh oh!

guan404ming commented Feb 24, 2026

Uh oh!

Uh oh!

guan404ming commented Feb 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

guan404ming commented Feb 23, 2026

Uh oh!

gemini-code-assist bot commented Feb 23, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Feb 23, 2026

Choose a reason for hiding this comment

Uh oh!

guan404ming Feb 23, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Feb 23, 2026

Choose a reason for hiding this comment

Uh oh!

guan404ming Feb 23, 2026

Choose a reason for hiding this comment

Uh oh!

guan404ming commented Feb 24, 2026

Uh oh!

Uh oh!

guan404ming commented Feb 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants