For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
CausalConv1DVarlenUpdate
struct CausalConv1DVarlenUpdate[activation: StringSlice[StaticConstantOrigin]]
Varlen causal conv1d update for autoregressive decoding.
Performs incremental convolution update for token-by-token generation. Updates the conv_state in-place with new input values.
Tensor Shapes: - output: (batch, dim, seqlen) - Output tensor - x: (batch, dim, seqlen) - Input tensor - weight: (dim, width) - Convolution weights - bias: (dim,) - Per-channel bias - conv_state: (batch, dim, state_len) - Conv state (in/out) - cache_seqlens: (batch,) - Current sequence lengths (optional) - conv_state_indices: (batch,) - Indices into conv_state (optional)
Parameters
- activation (
StringSlice[StaticConstantOrigin]): Activation function - "none" or "silu".
Implemented traits
AnyType,
ImplicitlyDestructible
Methods
execute
static execute[dtype: DType, target: StringSlice[StaticConstantOrigin]](output: ManagedTensorSlice[Output, static_spec=output.static_spec], conv_state: ManagedTensorSlice[Output, static_spec=conv_state.static_spec], x: ManagedTensorSlice[Input, static_spec=x.static_spec], weight: ManagedTensorSlice[Input, static_spec=weight.static_spec], bias: ManagedTensorSlice[Input, static_spec=bias.static_spec], cache_seqlens: ManagedTensorSlice[Input, static_spec=cache_seqlens.static_spec], conv_state_indices: ManagedTensorSlice[Input, static_spec=conv_state_indices.static_spec], ctx: DeviceContext)
shape
static shape[dtype: DType](x: ManagedTensorSlice[Input, static_spec=x.static_spec], weight: ManagedTensorSlice[Input, static_spec=weight.static_spec], bias: ManagedTensorSlice[Input, static_spec=bias.static_spec], cache_seqlens: ManagedTensorSlice[Input, static_spec=cache_seqlens.static_spec], conv_state_indices: ManagedTensorSlice[Input, static_spec=conv_state_indices.static_spec]) -> IndexList[3]
Returns:
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!