matmul

Operand QDQ and explicit arithmetic schedules for linear-attention matmuls.

Classes

LinearAttentionMatmulSites

The eight prefill matmuls, with independent QDQ on their actual operands.

class LinearAttentionMatmulSites

Bases: ModuleDict

The eight prefill matmuls, with independent QDQ on their actual operands.

Inputs to each handle are four-dimensional, with reduction on the last axis. The state-read LHS uses the parent’s gdn_w_quantizer or kda_w_quantizer; no second W quantizer is registered here. New handles start disabled.

__init__()

Create disabled operand handles for the eight prefill sites.

property is_enabled

Whether any of the additional operand handles is enabled.

matmul(name, lhs, rhs, policy, *, w_quantizer=None)

Multiply LHS by transposed RHS using the named numerical site.

validate()

Reject formats whose scale grouping has not been qualified at these sites.