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
MatmulSwiGLUBias
struct MatmulSwiGLUBias
Fused GEMM+SwiGLU+bias on SM100 for BF16 inputs.
Like mo.matmul_swiglu but adds a 1D bias vector before the activation.
The bias must be sigma-permuted to match the weight layout: even element
bias[2h] is added to the gate column and odd element bias[2h+1] to
the up column before silu(gate) * up is computed.
Implemented traits
Methods
execute
static def execute[target: StringSlice[ImmStaticOrigin]](output: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output.static_spec], a: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=a.static_spec], b: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=b.static_spec], bias: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=bias.static_spec], ctx: DeviceContext)
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!