[Operator] Develop forward function for cumprod operator.

[RubiaCx]

PR Category

Operator cumprod

Type of Change

New Feature

Description

Develop forward function for cumprod operator.

• Add support for cumulative product along any dimension
• Handle edge cases (empty tensors, single elements) Implement scan-and-propagate pattern for 1-Dim and N-Dims cases

Issue

#319

Test & Performance

Run test cases for:

• Basic functionality (1D/2D/3D tensors)
• Different data types
• Edge cases (empty/single element)
• Large tensors
• Negative dimensions

[RubiaCx]

Benchmarked on NV A100 GPUs comparing Torch and Triton for cumprod. Tested across sizes, data types, edge cases, and special values like NaNs.
Triton matches Torch in accuracy with differences between 1e−8 and 1e−15:

[iclementine]

Please add unit tests and benchmark for cumprod according to the contribution guide.

[RubiaCx]

Please add unit tests and benchmark for cumprod according to the contribution guide.

Hi, I've implemented the Cumprod operator with accuracy tests and performance benchmarks.
However, I encountered an error with the torch backend device: AttributeError: 'NoneType' object has no attribute 'matmul'