Skip to content

Add 23 new operators for 2026 Spring Qiyuan AI Competition (T1-1-1)#81

Open
YZYY95K wants to merge 1 commit into
InfiniTensor:masterfrom
YZYY95K:2026-spring-YZYY95K-T1-1-1
Open

Add 23 new operators for 2026 Spring Qiyuan AI Competition (T1-1-1)#81
YZYY95K wants to merge 1 commit into
InfiniTensor:masterfrom
YZYY95K:2026-spring-YZYY95K-T1-1-1

Conversation

@YZYY95K
Copy link
Copy Markdown

@YZYY95K YZYY95K commented May 20, 2026

Summary

This PR adds 23 new operators to ntops as part of the 2026 Spring Qiyuan AI Competition.

New Operators

Activation Functions:

  • leaky_relu, hardswish, hardtanh, elu, selu, mish, softplus, softsign, tanhshrink, relu6, prelu

Reduction Operators:

  • log_softmax (with FP16/FP32 support)

Backward Operators:

  • gelu_backward, sigmoid_backward, relu_backward, tanh_backward

Utility Operators:

  • where, cast, log, sqrt, reciprocal, ceil, floor

Key Features

  • All operators support both FP16 and FP32 with proper type casting for numerical stability
  • FP16 operators use float32 intermediate computation to avoid precision issues
  • 232 tests passing across all operators

Testing

All operators have been tested with multiple shapes and dtypes:

  • Shapes: (128,), (16, 128), (2, 16, 128), (2, 4, 16, 128)
  • Dtypes: float32 (rtol=0.001), float16 (rtol=0.01)

Team: YZYY95K

…erators

- Activation: leaky_relu, hardswish, hardtanh, elu, selu, mish, softplus, softsign, tanhshrink, relu6, prelu
- Reduction: log_softmax (with FP16 support)
- Backward: gelu_backward, sigmoid_backward, relu_backward, tanh_backward
- Utility: where, cast, log, sqrt, reciprocal, ceil, floor
- All operators support FP16/FP32 with proper type casting
- 232 tests passing
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant