Class ov::pass::BroadcastMatMulFusion#

class BroadcastMatMulFusion : public ov::pass::MatcherPass#

Removes a redundant Broadcast that expands one MatMul input’s batch dimensions.

Matches the Data -> Broadcast -> MatMul pattern, with the Broadcast on either MatMul input and Data being an arbitrary input (not necessarily a Constant). MatMul broadcasts the batch (leading) dimensions of its operands implicitly, so an explicit Broadcast that only expands those dimensions is redundant. The Broadcast is detached and Data is connected to the MatMul directly; the now-dangling Broadcast and its target-shape subgraph are left unreferenced and removed by later clean-up.

A common source is the rotary-embedding branch, where a constant (e.g. inv_freq) is expanded to the batch dimension taken from another input’s shape. Detaching the Broadcast also lets that shape-providing input become removable when it has no other consumers (for example attention_mask during PagedAttention conversion).

Before: 

Data          Other
  │             │
┌───┴─────┐ │ │Broadcast│ │ └───┬─────┘ │ │ │ └──────┬──────┘ ┌──┴───┐ │MatMul│ └──────┘

After: 

Data          Other
  │             │
  └──────┬──────┘
      ┌──┴───┐
      │MatMul│
      └──────┘
The Broadcast is removed only when it does not change the MatMul result:

  • the matrix (last two) dimensions are left intact by the Broadcast;

  • for every expanded batch dimension, the other MatMul operand carries the same dimension, proven equal by static value or by shape symbol; an unlabeled dynamic dimension is never assumed compatible, since that could hide a runtime batch mismatch the Broadcast would have rejected.

The Broadcast node itself is not removed and may keep other consumers; only the matched MatMul input is rewired to Data directly.