Coming from PyTorch

vkML follows PyTorch's names and shapes deliberately, so code ports without translation. The table below is not a claim — it is extracted from the test suite, where every pair is an assertion that the two agree within a declared tolerance.

Verified equivalences

41 operators are tested directly against their torch counterpart. If a row is here, a test compares them on every run.

PyTorch	vkML
`torch.abs`	`abs`
`torch.add`	`add`
`torch.amax`	`amax`
`torch.amin`	`amin`
`torch.argmax`	`argmax`
`torch.argmin`	`argmin`
`torch.cos`	`cos`
`torch.div`	`div`
`torch.eq`	`equal`
`torch.erf`	`erf`
`torch.erfc`	`erfc`
`torch.exp`	`exp`
`torch.nn.functional.gelu`	`gelu`
`torch.gt`	`greater`
`torch.ge`	`greater_equal`
`torch.lt`	`less`
`torch.le`	`less_equal`
`torch.log`	`log`
`torch.log_softmax`	`log_softmax`
`torch.maximum`	`maximum`
`torch.mean`	`mean`
`torch.minimum`	`minimum`
`torch.mul`	`mul`
`torch.neg`	`neg`
`torch.ne`	`not_equal`
`torch.pow`	`pow`
`torch.reciprocal`	`reciprocal`
`torch.relu`	`relu`
`torch.rsqrt`	`rsqrt`
`torch.sigmoid`	`sigmoid`
`torch.sign`	`sign`
`torch.nn.functional.silu`	`silu`
`torch.sin`	`sin`
`torch.softmax`	`softmax`
`torch.sqrt`	`sqrt`
`torch.square`	`square`
`torch.sub`	`sub`
`torch.sum`	`sum`
`torch.tanh`	`tanh`
`torch.tril`	`tril`
`torch.triu`	`triu`

Conventions that carry over

Layout is row-major, as NumPy, PyTorch and DLPack. shape()[0] is the outermost axis.
A state_dict loads directly. Linear stores its weight as (out_features, in_features) and transposes in forward, exactly as PyTorch does, so the tensors line up.
Reductions take dim and keepdim with torch's meanings.

Where it deliberately differs

.size is the element count, following NumPy rather than PyTorch, and it is a property: x.size() raises. Use x.shape.
Shape arguments are sequences: x.reshape([3, 2]), not x.reshape(3, 2).
max and min return values, not (values, indices), and are spelled amax and amin to say so.
to() converts dtype, not device. Placement happens at creation: V.tensor(x, device=dev).
Seeds do not match. manual_seed mirrors torch in spirit, not in stream — the two draw from different generators, so equal seeds give different weights.