Struct GpuOutputHead 

pub struct GpuOutputHead { /* private fields */ }

The output projection, resident on the GPU.

None from Self::try_build is the ordinary case, not a failure: no GPU, a weight the kernel does not handle, or a device that cannot bind it. The caller keeps using the CPU path.

Implementations

impl GpuOutputHead

pub fn try_build(model: &LoadedModel, executor: &Executor) -> Option<Self>

Uploads the model’s output projection if — and only if — every condition for it to be a win holds.

Returns None (never an error) when offload is simply not available: there is no GPU, the tensor is not Q8_0 (the 1.7B ships Q6_K, which this kernel does not decode), the tensor is missing, or the adapter cannot bind it. Every one of those is a normal machine configuration, not a fault, and the CPU path is always correct.

pub fn forward( &self, executor: &Executor, x: &[f32], m: usize, ) -> Option<Vec<f32>>

logits = x @ w^T for m rows of hidden states.

Errors

Any GPU failure, which the caller should treat as “use the CPU path this time” rather than as fatal.