Skip to content

update autotuner input tensor random range #2116

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yzh119 merged 3 commits into from
Nov 22, 2025
Merged

update autotuner input tensor random range #2116

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9d7dd0e
uddapte autotuner input random tensor
jiahanc Nov 19, 2025
cad6e85
rever unrelated change
jiahanc Nov 19, 2025
3e9bc87
per code rabbit
jiahanc Nov 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions flashinfer/autotuner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,8 +61,8 @@ def __post_init__(self):
# Set default tensor_initializers if not provided
if self.tensor_initializers is None:
self.tensor_initializers = [
lambda shapes, dtype, device: torch.randn(shapes, device=device).to(
Copy link
Collaborator

@yzh119 yzh119 Nov 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

randn is gaussian distribution, which is different from your desciption:

input tensor random range from [0,1) to [-5,5) for larger range

where [0, 1) is a uniform distribution.

I have no idea about the real data distribution tbh and changing it to [-5, 5) seems fine. Just a heads up to make sure it's not a typo.

Copy link
Collaborator Author

@jiahanc jiahanc Nov 20, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

thanks for pointing out. Is there a reason why randn here but rand in the other place?
Reason to change to [-5,5) is @rosenrodt did some work on MXFP4 tuning experiment and found this range can get better autotuner config than [0,1)

Copy link

@rosenrodt rosenrodt Nov 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I speculated [-5, 5) is than [0, 1) because the latter could truncate to 0s, thus affecting the power profile during autotune and less representative of the power profile of the actual workload.

Copy link
Collaborator

@yzh119 yzh119 Nov 22, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes I think if [-5, 5) is better let's use it, data distribution affects kernel execution time.

dtype
lambda shapes, dtype, device: (
torch.randn(shapes, device=device).to(dtype) * 10 - 5
)
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Nov 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

This change has a couple of issues:

  1. Inconsistent random distribution: This uses torch.randn, which generates values from a normal distribution. With the * 10 - 5 transformation, this will produce values from a normal distribution with a mean of -5 and a standard deviation of 10. This is inconsistent with the change in _prepare_input_tensors which uses torch.rand to generate values in a uniform [-5, 5) range, and also doesn't seem to match the PR description's goal of a [-5, 5) range. For consistency, torch.rand should be used here as well.

  2. Potential precision loss: The type conversion .to(dtype) is applied before the multiplication and subtraction. This can cause a loss of precision, especially with lower-precision dtypes like float16. The arithmetic operations should be performed in float32 (the default for torch.rand/torch.randn) before casting to the target dtype.

The suggested change below addresses both points for consistency and correctness.

                lambda shapes, dtype, device: (torch.rand(shapes, device=device) * 10 - 5).to(dtype)

for _ in range(len(self.input_idx))
]
Expand Down Expand Up @@ -761,8 +761,8 @@ def _create_tensor_like(
def _prepare_input_tensors(
self, profile: OptimizationProfile, inputs: List[torch.Tensor]
) -> List[torch.Tensor]:
default_initializer = lambda shapes, dtype, device: torch.rand(
Copy link
Collaborator

@yzh119 yzh119 Nov 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this is indeed uniform distribution [0, 1)

shapes, device=device
default_initializer = lambda shapes, dtype, device: (
torch.rand(shapes, device=device) * 10 - 5
).to(dtype)
tensors = []
for i, p in enumerate(profile.shapes):
Expand Down