vllm.kernels.helion.ops ¶
Helion kernel implementation.
Importing this package does NOT register any kernels. Runtime code imports the specific op module it needs, e.g.::
from vllm.kernels.helion.ops import scaled_mm # noqa: F401
which triggers that op's @register_kernel as an import side effect.
Tools that need the full registry (e.g. scripts/autotune_helion_kernels.py) call import_all_ops() to force every op module to register.
Modules:
-
dynamic_per_token_scaled_fp8_quant– -
fused_qk_norm_rope– -
per_token_group_fp8_quant– -
rms_norm_dynamic_per_token_quant– -
rms_norm_per_block_quant– -
silu_and_mul_per_block_quant– -
silu_mul_fp8–
Functions:
-
import_all_kernels–Import every kernel submodule so all
@register_kerneldecorators run.
import_all_kernels() ¶
Import every kernel submodule so all @register_kernel decorators run.
Returns: