|
| 1 | +# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 | +# All rights reserved. |
| 3 | +# |
| 4 | +# This source code is licensed under the BSD-style license found in the |
| 5 | +# LICENSE file in the root directory of this source tree. |
| 6 | + |
| 7 | +from typing import Optional |
| 8 | + |
| 9 | +import torch |
| 10 | + |
| 11 | +from executorch.backends.vulkan.patterns.pattern_registry import ( |
| 12 | + PatternMatch, |
| 13 | + register_pattern_detector, |
| 14 | + register_pattern_replacement, |
| 15 | +) |
| 16 | + |
| 17 | +from executorch.exir import ExportedProgram |
| 18 | +from executorch.exir.dialects._ops import ops as exir_ops |
| 19 | + |
| 20 | + |
| 21 | +_CAST_OPS = { |
| 22 | + exir_ops.edge.aten._to_copy.default, |
| 23 | + exir_ops.edge.aten.to.dtype, |
| 24 | +} |
| 25 | + |
| 26 | + |
| 27 | +def _skip_casts(node: torch.fx.Node) -> torch.fx.Node: |
| 28 | + """Unwrap chains of dtype-cast nodes to find the underlying value.""" |
| 29 | + while node.target in _CAST_OPS: |
| 30 | + arg0 = node.args[0] if node.args else None |
| 31 | + if not isinstance(arg0, torch.fx.Node): |
| 32 | + break |
| 33 | + node = arg0 |
| 34 | + # pyre-ignore[7]: node is always a Node; Pyre cannot narrow through loops |
| 35 | + return node |
| 36 | + |
| 37 | + |
| 38 | +class RmsNormMatch(PatternMatch): |
| 39 | + """ |
| 40 | + Detects the decomposed RMSNorm pattern, including variants where dtype |
| 41 | + casts (to_copy) are inserted around the computation. |
| 42 | +
|
| 43 | + The canonical pattern emitted by the Llama RMSNorm implementation is: |
| 44 | +
|
| 45 | + x_orig (any dtype) |
| 46 | + -> to_copy(fp32) -> x_f32 |
| 47 | + -> mul(x_f32, x_f32) -> mean(dim=-1, keepdim=True) |
| 48 | + -> add(eps) -> rsqrt -> rstd_f32 |
| 49 | + -> mul(x_f32, rstd_f32) -> norm_f32 |
| 50 | + -> to_copy(orig dtype) -> norm_cast |
| 51 | + weight -> to_copy(orig dtype) -> weight_cast |
| 52 | + -> mul(norm_cast, weight_cast) ← anchor node |
| 53 | +
|
| 54 | + We look through to_copy nodes when comparing tensor identities so that |
| 55 | + the match also handles fp32-only models where no casts are present. |
| 56 | +
|
| 57 | + The anchor node is the final mul (scale by weight). |
| 58 | + """ |
| 59 | + |
| 60 | + def __init__(self, final_mul_node: torch.fx.Node) -> None: # noqa: C901 |
| 61 | + self.anchor_node = final_mul_node |
| 62 | + self.match_found = False |
| 63 | + self.all_nodes = [self.anchor_node] |
| 64 | + |
| 65 | + # final_mul: mul(normalized_cast, weight_cast) |
| 66 | + # Unwrap casts to reach the underlying norm_mul and weight. |
| 67 | + norm_mul_node, self.weight_node = self._identify_norm_mul_and_weight( |
| 68 | + final_mul_node |
| 69 | + ) |
| 70 | + if norm_mul_node is None: |
| 71 | + return |
| 72 | + |
| 73 | + self.all_nodes.append(norm_mul_node) |
| 74 | + |
| 75 | + # norm_mul: mul(x_f32, rstd_f32) |
| 76 | + rsqrt_node, x_for_norm = self._identify_rsqrt_and_input(norm_mul_node) |
| 77 | + if rsqrt_node is None: |
| 78 | + return |
| 79 | + |
| 80 | + self.all_nodes.append(rsqrt_node) |
| 81 | + |
| 82 | + # rsqrt -> add(mean_sq, eps) -> mean(x_sq, dim=-1, keepdim=True) |
| 83 | + add_node = self._get_single_arg_node( |
| 84 | + rsqrt_node, exir_ops.edge.aten.rsqrt.default |
| 85 | + ) |
| 86 | + if add_node is None or add_node.target != exir_ops.edge.aten.add.Tensor: |
| 87 | + return |
| 88 | + |
| 89 | + self.all_nodes.append(add_node) |
| 90 | + |
| 91 | + self.eps_node = None |
| 92 | + mean_node = None |
| 93 | + for arg in add_node.args[:2]: |
| 94 | + if ( |
| 95 | + isinstance(arg, torch.fx.Node) |
| 96 | + and arg.target == exir_ops.edge.aten.mean.dim |
| 97 | + ): |
| 98 | + mean_node = arg |
| 99 | + else: |
| 100 | + self.eps_node = arg |
| 101 | + |
| 102 | + if mean_node is None or self.eps_node is None: |
| 103 | + return |
| 104 | + |
| 105 | + self.all_nodes.append(mean_node) |
| 106 | + |
| 107 | + # Verify mean has keepdim=True and dim=[-1] |
| 108 | + if len(mean_node.args) < 3: |
| 109 | + return |
| 110 | + mean_dims = mean_node.args[1] |
| 111 | + if mean_dims != [-1]: |
| 112 | + return |
| 113 | + if not mean_node.args[2]: |
| 114 | + return |
| 115 | + |
| 116 | + # mean's input should be x_sq = mul(x, x) or pow(x, 2) |
| 117 | + sq_node = mean_node.args[0] |
| 118 | + if not isinstance(sq_node, torch.fx.Node): |
| 119 | + return |
| 120 | + |
| 121 | + self.all_nodes.append(sq_node) |
| 122 | + |
| 123 | + # Use the fp32 x (x_for_norm) as the canonical fp32 input. |
| 124 | + # Both mul(x,x) and the norm mul should share the same fp32 source. |
| 125 | + x_f32 = ( |
| 126 | + _skip_casts(x_for_norm) |
| 127 | + if isinstance(x_for_norm, torch.fx.Node) |
| 128 | + else x_for_norm |
| 129 | + ) |
| 130 | + |
| 131 | + if sq_node.target == exir_ops.edge.aten.mul.Tensor: |
| 132 | + if sq_node.args[0] != sq_node.args[1]: |
| 133 | + return |
| 134 | + sq_input = sq_node.args[0] |
| 135 | + if not isinstance(sq_input, torch.fx.Node): |
| 136 | + return |
| 137 | + if _skip_casts(sq_input) != x_f32 and sq_input != x_for_norm: |
| 138 | + return |
| 139 | + elif sq_node.target == exir_ops.edge.aten.pow.Tensor_Scalar: |
| 140 | + sq_input = sq_node.args[0] |
| 141 | + if not isinstance(sq_input, torch.fx.Node): |
| 142 | + return |
| 143 | + if _skip_casts(sq_input) != x_f32 and sq_input != x_for_norm: |
| 144 | + return |
| 145 | + if sq_node.args[1] != 2 and sq_node.args[1] != 2.0: |
| 146 | + return |
| 147 | + else: |
| 148 | + return |
| 149 | + |
| 150 | + # The canonical input node to expose to the fused op is the original |
| 151 | + # tensor before any fp32 upcast (i.e. the input to the first to_copy). |
| 152 | + # If there's no cast, x_for_norm is already the original input. |
| 153 | + self.input_node = ( |
| 154 | + _skip_casts(x_for_norm) |
| 155 | + if isinstance(x_for_norm, torch.fx.Node) |
| 156 | + else x_for_norm |
| 157 | + ) |
| 158 | + # Also collect the intermediate cast nodes so they can be cleaned up |
| 159 | + cast_node = x_for_norm |
| 160 | + while ( |
| 161 | + isinstance(cast_node, torch.fx.Node) |
| 162 | + and cast_node.target in _CAST_OPS |
| 163 | + and cast_node not in self.all_nodes |
| 164 | + ): |
| 165 | + self.all_nodes.append(cast_node) |
| 166 | + cast_node = cast_node.args[0] if cast_node.args else cast_node |
| 167 | + |
| 168 | + self.match_found = True |
| 169 | + |
| 170 | + def _identify_norm_mul_and_weight(self, final_mul_node): |
| 171 | + """From mul(norm_cast, weight_cast), unwrap casts and find the |
| 172 | + underlying norm-mul node and the weight source node.""" |
| 173 | + if len(final_mul_node.args) < 2: |
| 174 | + return None, None |
| 175 | + |
| 176 | + a, b = final_mul_node.args[0], final_mul_node.args[1] |
| 177 | + |
| 178 | + for norm_candidate_raw, weight_candidate_raw in [(a, b), (b, a)]: |
| 179 | + if not isinstance(norm_candidate_raw, torch.fx.Node): |
| 180 | + continue |
| 181 | + norm_candidate = _skip_casts(norm_candidate_raw) |
| 182 | + if ( |
| 183 | + isinstance(norm_candidate, torch.fx.Node) |
| 184 | + and norm_candidate.target == exir_ops.edge.aten.mul.Tensor |
| 185 | + and self._has_rsqrt_ancestor(norm_candidate) |
| 186 | + ): |
| 187 | + return norm_candidate, weight_candidate_raw |
| 188 | + |
| 189 | + return None, None |
| 190 | + |
| 191 | + def _has_rsqrt_ancestor(self, mul_node): |
| 192 | + """Check if one of mul_node's args is an rsqrt node (possibly through casts).""" |
| 193 | + for arg in mul_node.args[:2]: |
| 194 | + if not isinstance(arg, torch.fx.Node): |
| 195 | + continue |
| 196 | + if _skip_casts(arg).target == exir_ops.edge.aten.rsqrt.default: |
| 197 | + return True |
| 198 | + return False |
| 199 | + |
| 200 | + def _identify_rsqrt_and_input(self, norm_mul_node): |
| 201 | + """From mul(x, rstd), find the rsqrt node and the input x. |
| 202 | + The rsqrt may be wrapped in a cast node.""" |
| 203 | + if len(norm_mul_node.args) < 2: |
| 204 | + return None, None |
| 205 | + |
| 206 | + a, b = norm_mul_node.args[0], norm_mul_node.args[1] |
| 207 | + |
| 208 | + for rsqrt_candidate_raw, input_candidate in [(a, b), (b, a)]: |
| 209 | + if not isinstance(rsqrt_candidate_raw, torch.fx.Node): |
| 210 | + continue |
| 211 | + rsqrt_candidate = _skip_casts(rsqrt_candidate_raw) |
| 212 | + if ( |
| 213 | + isinstance(rsqrt_candidate, torch.fx.Node) |
| 214 | + and rsqrt_candidate.target == exir_ops.edge.aten.rsqrt.default |
| 215 | + ): |
| 216 | + return rsqrt_candidate, input_candidate |
| 217 | + |
| 218 | + return None, None |
| 219 | + |
| 220 | + def _get_single_arg_node(self, node, expected_target): |
| 221 | + """Get the single input arg of a unary op node.""" |
| 222 | + if node.target != expected_target: |
| 223 | + return None |
| 224 | + if len(node.args) < 1 or not isinstance(node.args[0], torch.fx.Node): |
| 225 | + return None |
| 226 | + return node.args[0] |
| 227 | + |
| 228 | + |
| 229 | +@register_pattern_detector("rms_norm") |
| 230 | +def find_rms_norm_patterns( |
| 231 | + node: torch.fx.Node, |
| 232 | +) -> Optional[RmsNormMatch]: |
| 233 | + if node.target != exir_ops.edge.aten.mul.Tensor: |
| 234 | + return None |
| 235 | + |
| 236 | + matched_pattern = RmsNormMatch(node) |
| 237 | + if matched_pattern.match_found: |
| 238 | + return matched_pattern |
| 239 | + |
| 240 | + return None |
| 241 | + |
| 242 | + |
| 243 | +## |
| 244 | +## Pattern Replacement |
| 245 | +## |
| 246 | + |
| 247 | + |
| 248 | +def _extract_eps_value(eps_node) -> float: |
| 249 | + if isinstance(eps_node, (int, float)): |
| 250 | + return float(eps_node) |
| 251 | + if isinstance(eps_node, torch.fx.Node) and "val" in eps_node.meta: |
| 252 | + val = eps_node.meta["val"] |
| 253 | + if isinstance(val, torch.Tensor): |
| 254 | + return float(val.item()) |
| 255 | + if isinstance(val, (int, float)): |
| 256 | + return float(val) |
| 257 | + raise ValueError(f"Cannot extract epsilon value from {eps_node}") |
| 258 | + |
| 259 | + |
| 260 | +@register_pattern_replacement("rms_norm") |
| 261 | +def replace_rms_norm_with_fused_op( |
| 262 | + ep: ExportedProgram, |
| 263 | + graph_module: torch.fx.GraphModule, |
| 264 | + match: RmsNormMatch, |
| 265 | +): |
| 266 | + eps_val = _extract_eps_value(match.eps_node) |
| 267 | + |
| 268 | + with graph_module.graph.inserting_before(match.anchor_node): |
| 269 | + rms_norm_node = graph_module.graph.create_node( |
| 270 | + "call_function", |
| 271 | + exir_ops.edge.et_vk.rms_norm.default, |
| 272 | + args=( |
| 273 | + match.input_node, |
| 274 | + match.weight_node, |
| 275 | + eps_val, |
| 276 | + ), |
| 277 | + ) |
| 278 | + |
| 279 | + rms_norm_node.meta["val"] = match.anchor_node.meta["val"] |
| 280 | + match.anchor_node.replace_all_uses_with(rms_norm_node) |
0 commit comments