Coming from PyTorch
SonicBoom is not PyTorch, but many of its names and conventions carry over. Type a PyTorch name below to see what the Guile binding offers for it:
- supported — there is a function with the same meaning; the note says how it differs, if it does.
- recipe — there is no single function, but the operation is a few lines of existing ones; the recipe runs as written.
- missing — nothing in the binding does this yet.
- out of scope — the feature belongs to a part of PyTorch that SonicBoom does not aim to cover; the note says why.
The lookup matches full names and common spellings (nn.Linear, F.relu), and a name typed in the address bar, as in pytorch.html?q=torch.add, runs the same query.
Names with a SonicBoom counterpart
The table below lists only the names that are supported or have a recipe, grouped by the PyTorch module they come from. ### torch
| Name | Status | SonicBoom |
|---|---|---|
torch.Tensor |
supported | <tensor>, tensor?, tensor-dtype, tensor-shape, tensor-requires-grad?, tensor-grad, set-tensor-grad!, ctensor?, ctensor-dtype, ctensor-shape |
torch.abs |
recipe | (ctensor->list (where (lt x 0.0) (neg x) x)) |
torch.add |
supported | add |
torch.amax |
supported | amax |
torch.bmm |
supported | bmm |
torch.broadcast_shapes |
supported | broadcast-shapes |
torch.clamp |
recipe | (ctensor->list (where (lt x -0.5) -0.5 (where (gt x 0.5) 0.5 x))) |
torch.concat |
supported | concat |
torch.div |
supported | div |
torch.eq |
supported | equal-to |
torch.exp |
supported | texp |
torch.flatten |
recipe | (ctensor->list (reshape x (list (ctensor-numel x)))) |
torch.frombuffer |
supported | bytevector->tensor, bytevector->ctensor, tensor-from-bytes |
torch.gather |
supported | gather |
torch.gt |
supported | gt |
torch.index_select |
supported | index-select |
torch.load |
supported | load-params! |
torch.log |
supported | tlog |
torch.logical_and |
supported | logical-and |
torch.logical_not |
supported | logical-not |
torch.logical_or |
supported | logical-or |
torch.logical_xor |
recipe | (ctensor->list (logical-not (equal-to a b))) |
torch.lt |
supported | lt |
torch.matmul |
supported | matmul |
torch.maximum |
recipe | (ctensor->list (where (gt a b) a b)) |
torch.mean |
supported | mean |
torch.minimum |
recipe | (ctensor->list (where (lt a b) a b)) |
torch.mul |
supported | mul |
torch.narrow |
recipe | (ctensor->list (slice x dim start (+ start length))) |
torch.ne |
recipe | (ctensor->list (logical-not (equal-to a b))) |
torch.neg |
supported | neg |
torch.ones |
supported | ones |
torch.permute |
supported | permute |
torch.randn |
supported | randn |
torch.reciprocal |
recipe | (ctensor->list (div 1.0 x)) |
torch.reshape |
supported | reshape |
torch.save |
supported | save-params! |
torch.sigmoid |
recipe | (ctensor->list (div 1.0 (add 1.0 (texp (neg x))))) |
torch.sqrt |
supported | tsqrt |
torch.square |
recipe | (ctensor->list (mul x x)) |
torch.sub |
supported | sub |
torch.sum |
supported | sum |
torch.tanh |
supported | ttanh |
torch.tensor |
supported | make-tensor, list->tensor, make-ctensor, list->ctensor, scalar->ctensor, tensor-from-list |
torch.transpose |
supported | transpose |
torch.where |
supported | where |
torch.zeros |
supported | zeros |
torch.Tensor
| Name | Status | SonicBoom |
|---|---|---|
torch.Tensor.backward |
supported | autograd-backward |
torch.Tensor.clone |
supported | tensor-copy, ctensor-copy |
torch.Tensor.expand |
supported | expand |
torch.Tensor.grad |
supported | parameter-grad, autograd-grad |
torch.Tensor.index_add_ |
supported | index-add |
torch.Tensor.is_cuda |
supported | param-on-device? |
torch.Tensor.item |
supported | ctensor-item |
torch.Tensor.nbytes |
supported | ctensor-nbytes |
torch.Tensor.numel |
supported | tensor-numel, ctensor-numel |
torch.Tensor.scatter_add_ |
supported | scatter-add |
torch.Tensor.tolist |
supported | tensor->list, ctensor->list, ctensor->nested, tensor->list |
torch.autograd
| Name | Status | SonicBoom |
|---|---|---|
torch.autograd.backward |
supported | backward! |
torch.autograd.grad |
supported | grad |
torch.func
| Name | Status | SonicBoom |
|---|---|---|
torch.func.grad_and_value |
supported | value-and-grad |
torch.nn
| Name | Status | SonicBoom |
|---|---|---|
torch.nn.Linear |
supported | linear |
torch.nn.Module |
supported | nn-model?, nn-model-structure, nn-model-params, nn-model-input-dims |
torch.nn.Sequential |
supported | model?, model-layers, model-parameters, sequential |
torch.nn.Module
| Name | Status | SonicBoom |
|---|---|---|
torch.nn.Module.forward |
supported | model-forward-native, model-forward |
torch.nn.Module.get_parameter |
supported | layer-parameter-ref, model-parameter-ref |
torch.nn.functional
| Name | Status | SonicBoom |
|---|---|---|
torch.nn.functional.cross_entropy |
supported | fused-cross-entropy, cross-entropy |
torch.nn.functional.embedding |
supported | embedding |
torch.nn.functional.gelu |
supported | gelu |
torch.nn.functional.hardtanh |
recipe | (ctensor->list (where (lt x -1.0) -1.0 (where (gt x 1.0) 1.0 x))) |
torch.nn.functional.layer_norm |
supported | layer-norm |
torch.nn.functional.leaky_relu |
recipe | (ctensor->list (where (lt x 0.0) (mul 0.01 x) x)) |
torch.nn.functional.linear |
supported | linear, autograd-linear |
torch.nn.functional.log_softmax |
supported | log-softmax |
torch.nn.functional.mse_loss |
supported | autograd-mse |
torch.nn.functional.relu |
supported | relu, relu, autograd-relu |
torch.nn.functional.relu6 |
recipe | (ctensor->list (where (lt x 0.0) 0.0 (where (gt x 6.0) 6.0 x))) |
torch.nn.functional.rms_norm |
supported | fused-rms-norm, rms-norm |
torch.nn.functional.silu |
recipe | (ctensor->list (mul x (div 1.0 (add 1.0 (texp (neg x)))))) |
torch.nn.functional.softmax |
supported | fused-softmax, softmax |
torch.nn.parameter
| Name | Status | SonicBoom |
|---|---|---|
torch.nn.parameter.Parameter |
supported | make-param, parameter, leaf?, leaf-value, set-leaf-value!, leaf-grad, set-leaf-grad! |
torch.nn.utils
| Name | Status | SonicBoom |
|---|---|---|
torch.nn.utils.clip_grad_norm_ |
supported | fused-clip-grad-norm, clip-grad-norm!, clip-grad-norm! |
torch.optim
| Name | Status | SonicBoom |
|---|---|---|
torch.optim.Adam |
supported | adam-init, adam-update |
torch.optim.AdamW |
supported | adamw-step! |
torch.optim.SGD |
supported | sgd-update |
torch.optim.AdamW
| Name | Status | SonicBoom |
|---|---|---|
torch.optim.AdamW.step |
supported | param-step! |
torch.optim.Optimizer
| Name | Status | SonicBoom |
|---|---|---|
torch.optim.Optimizer.zero_grad |
supported | param-zero-grad!, zero-grad!, tape-zero-grad! |
torch.optim.SGD
| Name | Status | SonicBoom |
|---|---|---|
torch.optim.SGD.step |
supported | autograd-sgd-step! |