tilus.target

tilus.target

Tilus automatically detects the GPU at runtime and selects the most capable compilation target for the installed hardware. You can query or override the target using the functions below.

tilus.target.get_current_target()

Get the current compilation target.

tilus.target.set_current_target(target)

Set the current compilation target.

Predefined Targets

Each target represents a GPU architecture with its compute capability and feature suffix. Tilus uses these to determine which instructions and optimizations are available.

Target

SM

GPU Examples

nvgpu_sm70

7.0

V100

nvgpu_sm75

7.5

T4, RTX 2080

nvgpu_sm80

8.0

A100

nvgpu_sm86

8.6

RTX 3090, A40

nvgpu_sm89

8.9

L4, L40, RTX 4090

nvgpu_sm90 / sm90a

9.0

H100, H200

nvgpu_sm100 / sm100f / sm100a

10.0

B200, GB200

nvgpu_sm103 / sm103f / sm103a

10.3

GB300

nvgpu_sm120 / sm120f / sm120a

12.0

RTX 5080, RTX 5090

Feature Suffixes

Starting with SM 9.0, NVIDIA targets can carry a feature suffix that controls which architecture-conditional features are enabled. Understanding the suffixes is important when writing kernels that use advanced hardware features (e.g., Tensor Memory, tcgen05 MMA, TMA).

No suffix: The base architecture (e.g., sm_100). Only instructions guaranteed on every chip with that compute capability are available. Use this when you only need baseline capabilities and want maximum hardware compatibility.

`a` (architecture-specific): The full-featured variant for that exact architecture (e.g., sm_100a). Enables all architecture-conditional features, such as Tensor Memory allocation modes, tcgen05 Tensor Core operations, and special TMA behaviors. Use this when you want every available hardware feature and are targeting a specific GPU.

`f` (family-portable): The family-profile variant (e.g., sm_100f). Enables features that are portable across all implementations within the same SM family. A feature supported on sm_100f is guaranteed on any chip in the sm_100 family that advertises f-level support. Use this when you want advanced features with portability across family variants.

The compatibility relationship is:

  • sm_100a supports everything in sm_100f and sm_100.

  • sm_100f supports everything in sm_100, but not sm_100a-only features.

  • sm_100 supports only baseline features.

Note

When Tilus auto-detects the target, it picks the most capable variant available: a first, then f, then the base. For example, on a B200 GPU (SM 10.0), Tilus selects nvgpu_sm100a.

Usage

import tilus
from tilus.target import set_current_target, nvgpu_sm100a

# Query the auto-detected target
target = tilus.get_current_target()
print(target)  # e.g., nvgpu/sm100a

# Override the target (e.g., for cross-compilation)
set_current_target(nvgpu_sm100a)
Contents