For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo function
causal_conv1d_update_gpu_functional
def causal_conv1d_update_gpu_functional[x_dtype: DType, conv_state_dtype: DType, weight_dtype: DType, output_dtype: DType, bias_dtype: DType, kNThreads: Int, x_LT: TensorLayout, conv_state_LT: TensorLayout, weight_LT: TensorLayout, output_LT: TensorLayout, bias_LT: TensorLayout, x_engine: TensorEngine, conv_state_engine: TensorEngine, weight_engine: TensorEngine, output_engine: TensorEngine, bias_engine: TensorEngine, conv_state_in_LT: TensorLayout, conv_state_in_engine: TensorEngine](batch: Int32, dim: Int32, seqlen: Int32, width: Int32, state_len: Int32, x: TileTensor[x_dtype, x_LT, ImmUntrackedOrigin, Engine=x_engine], conv_state: TileTensor[conv_state_dtype, conv_state_LT, MutUntrackedOrigin, Engine=conv_state_engine], conv_state_in: TileTensor[conv_state_dtype, conv_state_in_LT, ImmUntrackedOrigin, Engine=conv_state_in_engine], weight: TileTensor[weight_dtype, weight_LT, ImmUntrackedOrigin, Engine=weight_engine], output: TileTensor[output_dtype, output_LT, MutUntrackedOrigin, Engine=output_engine], bias: TileTensor[bias_dtype, bias_LT, ImmUntrackedOrigin, Engine=bias_engine], silu_activation: Int8)
Reads the previous state separately while writing the complete next state.
Uses the original scalar channel-per-thread update and arithmetic. Identical input/output state views may alias: convolution finishes before state writes, and the ascending shift only reads the current or a later state position. State input loads are non-invariant because the output may alias them. Distinct state layouts are supported when the allocations do not overlap. Partially overlapping state views or aliases with different layouts are unsupported.