submission 839293
vinu0163 · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 31 lines, June 9 Researcher Reciprocity License v1.0.
2026-06-27-15-18-45-eigh_b200-worker-3-951b2066593c.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-eigh-839293?include=source"interfacepython
Compatibility
measured onNVIDIA B200
declared hardwareNVIDIA B200
architecturessm_100
dtypesfp32
Benchmark evidence
1 measurement across 1 GPU, fastest first.
Operation / workload
Hardware
Latency
Rank
Observed
Reported · How evidence levels are derived →
Source and license
sourceavailable
revision digestsha256:3fd8e54ac0d46db635922864be725f78b77612e2743e84ce0953ac8bdec4f4bb
license declaredunknown
license concludedunknown
authorsvinu0163
imported2026-08-26
Kernel source
2026-06-27-15-18-45-eigh_b200-worker-3-951b2066593c.py31 lines
import torch
from task import input_t, output_t
def _identity(batch: int, n: int, data: torch.Tensor) -> torch.Tensor:
return torch.eye(n, device=data.device, dtype=torch.float32).expand(batch, n, n)
def _diagonal_eigh(data: torch.Tensor) -> output_t | None:
batch, n, _ = data.shape
if batch > 2:
return None
if n > 1 and bool(data[..., 0, 1].abs().amax() != 0):
return None
diag = torch.diagonal(data, dim1=-2, dim2=-1)
offdiag_l1 = data.abs().sum(dim=(-2, -1)) - diag.abs().sum(dim=-1)
if bool(offdiag_l1.amax() != 0):
return None
values, order = torch.sort(diag, dim=-1)
vectors = _identity(batch, n, data).gather(2, order.unsqueeze(1).expand(batch, n, n))
return vectors.contiguous(), values.contiguous()
def custom_kernel(data: input_t) -> output_t:
diagonal = _diagonal_eigh(data)
if diagonal is not None:
return diagonal
values, vectors = torch.linalg.eigh(data)
return vectors, values
scrolls · 31 lines total
Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0
Best evidence level for this revision: reported
JSON