Match RMSNorm pattern starting from final mul(weight, normalized).
(cls, head: Node, **context)
| 67 | |
| 68 | @classmethod |
| 69 | def maybe_create(cls, head: Node, **context) -> Optional["RMSNormMatch"]: |
| 70 | """Match RMSNorm pattern starting from final mul(weight, normalized).""" |
| 71 | # Head must be mul |
| 72 | if not match_target(head, torch.ops.aten.mul.Tensor): |
| 73 | return None |
| 74 | |
| 75 | if len(head.args) < 2: |
| 76 | return None |
| 77 | |
| 78 | # Try both orderings: mul(weight, normalized) or mul(normalized, weight) |
| 79 | for weight_idx, norm_idx in [(0, 1), (1, 0)]: |
| 80 | weight_node = head.args[weight_idx] |
| 81 | norm_node = head.args[norm_idx] |
| 82 | |
| 83 | if not isinstance(norm_node, Node): |
| 84 | continue |
| 85 | |
| 86 | # Match entire chain with single walk_back: |
| 87 | # [_to_copy] -> mul(input, rsqrt) -> rsqrt -> add -> mean -> pow -> [_to_copy] |
| 88 | # The mul follows arg_index=1 to get rsqrt (not input) |
| 89 | result = walk_back( |
| 90 | norm_node, |
| 91 | [ |
| 92 | OpStep( |
| 93 | op=torch.ops.aten._to_copy.default, |
| 94 | optional=True, |
| 95 | kwargs={ |
| 96 | "dtype", |
| 97 | "layout", |
| 98 | "device", |
| 99 | "pin_memory", |
| 100 | "non_blocking", |
| 101 | "memory_format", |
| 102 | }, |
| 103 | ), |
| 104 | OpStep(op=torch.ops.aten.mul.Tensor, nargs=2, arg_index=1), |
| 105 | OpStep(op=torch.ops.aten.rsqrt.default), |
| 106 | OpStep(op=torch.ops.aten.add.Tensor, nargs=2), |
| 107 | OpStep(op=torch.ops.aten.mean.dim, nargs=(2, 3), kwargs={"dtype"}), |
| 108 | OpStep(op=torch.ops.aten.pow.Tensor_Scalar, nargs=2), |
| 109 | OpStep( |
| 110 | op=torch.ops.aten._to_copy.default, |
| 111 | optional=True, |
| 112 | require_single_user=False, # _to_copy output used by both pow and mul |
| 113 | kwargs={ |
| 114 | "dtype", |
| 115 | "layout", |
| 116 | "device", |
| 117 | "pin_memory", |
| 118 | "non_blocking", |
| 119 | "memory_format", |
| 120 | }, |
| 121 | ), |
| 122 | ], |
| 123 | ) |
| 124 | if result is None: |
| 125 | continue |
| 126 |
no test coverage detected