vollo_torch

• vollo_torch.fx
  • prepare_shape()
  • save()
  • load()
  • is_vollo_tracing
• vollo_torch.nn
  • PaddedConv1d
  • LSTM
  • RecurrentStateLSTM
  • LSTMCell
    • LSTMCell.forward()
    • LSTMCell.reset()
  • Ones
  • Zeros
  • Scan
    • Scan.forward()
  • RMSNorm
    • RMSNorm.extra_repr()

class vollo_torch.Fp8Weights

Weights used inside this context will be quantized to an 8-bit format, which may be useful for getting a model to fit on the board.

Note that this only applies to constant weight matrix multiplications. To ensure predictable behaviour, all MatMuls and Linears that would use dynamic weights are rejected inside this context.

class vollo_torch.Fp32Activations

Activations/computations inside this context will be in 32-bit precision. See https://vollo.myrtle.ai/latest/supported-models.html for which operations support this.

class vollo_torch.CorePartition(cores: Sequence[int], override=False)

Computations inside this context will be limited to the cores specified in the context’s constructor.

If override=False (default), the CorePartition must specify a subset of the enclosing partition’s cores; a ValueError is raised at context entry if this is violated.

If override=True, the partition overrides the enclosing partition.

Warning

This class is experimental and is subject to change in future versions.