File size: 28,986 Bytes
066effd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 |
# ------------------------------------------------------------------------
# RF-DETR
# Copyright (c) 2025 Roboflow. All Rights Reserved.
# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
# ------------------------------------------------------------------------
# Modified from LW-DETR (https://github.com/Atten4Vis/LW-DETR)
# Copyright (c) 2024 Baidu. All Rights Reserved.
# ------------------------------------------------------------------------
"""
OnnxOptimizer
"""
import os
from collections import OrderedDict
from copy import deepcopy
import numpy as np
import onnx
import torch
from onnx import shape_inference
import onnx_graphsurgeon as gs
from polygraphy.backend.onnx.loader import fold_constants
from onnx_graphsurgeon.logger.logger import G_LOGGER
from .symbolic import CustomOpSymbolicRegistry
class OnnxOptimizer():
def __init__(
self,
input,
severity=G_LOGGER.INFO
):
if isinstance(input, str):
onnx_graph = self.load_onnx(input)
else:
onnx_graph = input
self.graph = gs.import_onnx(onnx_graph)
self.severity = severity
self.set_severity(severity)
def set_severity(self, severity):
G_LOGGER.severity = severity
def load_onnx(self, onnx_path:str):
"""Load onnx from file
"""
assert os.path.isfile(onnx_path), f"not found onnx file: {onnx_path}"
onnx_graph = onnx.load(onnx_path)
G_LOGGER.info(f"load onnx file: {onnx_path}")
return onnx_graph
def save_onnx(self, onnx_path:str):
onnx_graph = gs.export_onnx(self.graph)
G_LOGGER.info(f"save onnx file: {onnx_path}")
onnx.save(onnx_graph, onnx_path)
def info(self, prefix=''):
G_LOGGER.verbose(f"{prefix} .. {len(self.graph.nodes)} nodes, {len(self.graph.tensors().keys())} tensors, {len(self.graph.inputs)} inputs, {len(self.graph.outputs)} outputs")
def cleanup(self, return_onnx=False):
self.graph.cleanup().toposort()
if return_onnx:
return gs.export_onnx(self.graph)
def select_outputs(self, keep, names=None):
self.graph.outputs = [self.graph.outputs[o] for o in keep]
if names:
for i, name in enumerate(names):
self.graph.outputs[i].name = name
def find_node_input(self, node, name:str=None, value=None) -> int:
for i, inp in enumerate(node.inputs):
if isinstance(name, str) and inp.name == name:
index = i
elif inp == value:
index = i
assert index >= 0, f"not found {name}({value}) in node.inputs"
return index
def find_node_output(self, node, name:str=None, value=None) -> int:
for i, inp in enumerate(node.outputs):
if isinstance(name, str) and inp.name == name:
index = i
elif inp == value:
index = i
assert index >= 0, f"not found {name}({value}) in node.outputs"
return index
def common_opt(self, return_onnx=False):
for fn in CustomOpSymbolicRegistry._OPTIMIZER:
fn(self)
self.cleanup()
onnx_graph = fold_constants(gs.export_onnx(self.graph), allow_onnxruntime_shape_inference=False)
if onnx_graph.ByteSize() > 2147483648:
raise TypeError("ERROR: model size exceeds supported 2GB limit")
else:
onnx_graph = shape_inference.infer_shapes(onnx_graph)
self.graph = gs.import_onnx(onnx_graph)
self.cleanup()
if return_onnx:
return onnx_graph
def resize_fix(self):
'''
This function loops through the graph looking for Resize nodes that uses scales for resize (has 3 inputs).
It substitutes found Resize with Resize that takes the size of the output tensor instead of scales.
It adds Shape->Slice->Concat
Shape->Slice----^ subgraph to the graph to extract the shape of the output tensor.
This fix is required for the dynamic shape support.
'''
mResizeNodes = 0
for node in self.graph.nodes:
if node.op == "Resize" and len(node.inputs) == 3:
name = node.name + "/"
add_node = node.o().o().i(1)
div_node = node.i()
shape_hw_out = gs.Variable(name=name + "shape_hw_out", dtype=np.int64, shape=[4])
shape_hw = gs.Node(op="Shape", name=name+"shape_hw", inputs=[add_node.outputs[0]], outputs=[shape_hw_out])
const_zero = gs.Constant(name=name + "const_zero", values=np.array([0], dtype=np.int64))
const_two = gs.Constant(name=name + "const_two", values=np.array([2], dtype=np.int64))
const_four = gs.Constant(name=name + "const_four", values=np.array([4], dtype=np.int64))
slice_hw_out = gs.Variable(name=name + "slice_hw_out", dtype=np.int64, shape=[2])
slice_hw = gs.Node(op="Slice", name=name+"slice_hw", inputs=[shape_hw_out, const_two, const_four, const_zero], outputs=[slice_hw_out])
shape_bc_out = gs.Variable(name=name + "shape_bc_out", dtype=np.int64, shape=[2])
shape_bc = gs.Node(op="Shape", name=name+"shape_bc", inputs=[div_node.outputs[0]], outputs=[shape_bc_out])
slice_bc_out = gs.Variable(name=name + "slice_bc_out", dtype=np.int64, shape=[2])
slice_bc = gs.Node(op="Slice", name=name+"slice_bc", inputs=[shape_bc_out, const_zero, const_two, const_zero], outputs=[slice_bc_out])
concat_bchw_out = gs.Variable(name=name + "concat_bchw_out", dtype=np.int64, shape=[4])
concat_bchw = gs.Node(op="Concat", name=name+"concat_bchw", attrs={"axis": 0}, inputs=[slice_bc_out, slice_hw_out], outputs=[concat_bchw_out])
none_var = gs.Variable.empty()
resize_bchw = gs.Node(op="Resize", name=name+"resize_bchw", attrs=node.attrs, inputs=[node.inputs[0], none_var, none_var, concat_bchw_out], outputs=[node.outputs[0]])
self.graph.nodes.extend([shape_hw, slice_hw, shape_bc, slice_bc, concat_bchw, resize_bchw])
node.inputs = []
node.outputs = []
mResizeNodes += 1
self.cleanup()
return mResizeNodes
def adjustAddNode(self):
nAdjustAddNode = 0
for node in self.graph.nodes:
# Change the bias const to the second input to allow Gemm+BiasAdd fusion in TRT.
if node.op in ["Add"] and isinstance(node.inputs[0], gs.ir.tensor.Constant):
tensor = node.inputs[1]
bias = node.inputs[0]
node.inputs = [tensor, bias]
nAdjustAddNode += 1
self.cleanup()
return nAdjustAddNode
def decompose_instancenorms(self):
nRemoveInstanceNorm = 0
for node in self.graph.nodes:
if node.op == "InstanceNormalization":
name = node.name + "/"
input_tensor = node.inputs[0]
output_tensor = node.outputs[0]
mean_out = gs.Variable(name=name + "mean_out")
mean_node = gs.Node(op="ReduceMean", name=name + "mean_node", attrs={"axes": [-1]}, inputs=[input_tensor], outputs=[mean_out])
sub_out = gs.Variable(name=name + "sub_out")
sub_node = gs.Node(op="Sub", name=name + "sub_node", attrs={}, inputs=[input_tensor, mean_out], outputs=[sub_out])
pow_out = gs.Variable(name=name + "pow_out")
pow_const = gs.Constant(name=name + "pow_const", values=np.array([2.0], dtype=np.float32))
pow_node = gs.Node(op="Pow", name=name + "pow_node", attrs={}, inputs=[sub_out, pow_const], outputs=[pow_out])
mean2_out = gs.Variable(name=name + "mean2_out")
mean2_node = gs.Node(op="ReduceMean", name=name + "mean2_node", attrs={"axes": [-1]}, inputs=[pow_out], outputs=[mean2_out])
epsilon_out = gs.Variable(name=name + "epsilon_out")
epsilon_const = gs.Constant(name=name + "epsilon_const", values=np.array([node.attrs["epsilon"]], dtype=np.float32))
epsilon_node = gs.Node(op="Add", name=name + "epsilon_node", attrs={}, inputs=[mean2_out, epsilon_const], outputs=[epsilon_out])
sqrt_out = gs.Variable(name=name + "sqrt_out")
sqrt_node = gs.Node(op="Sqrt", name=name + "sqrt_node", attrs={}, inputs=[epsilon_out], outputs=[sqrt_out])
div_out = gs.Variable(name=name + "div_out")
div_node = gs.Node(op="Div", name=name + "div_node", attrs={}, inputs=[sub_out, sqrt_out], outputs=[div_out])
constantScale = gs.Constant("InstanceNormScaleV-" + str(nRemoveInstanceNorm), np.ascontiguousarray(node.inputs[1].inputs[0].attrs["value"].values.reshape(1, 32, 1)))
constantBias = gs.Constant("InstanceBiasV-" + str(nRemoveInstanceNorm), np.ascontiguousarray(node.inputs[2].inputs[0].attrs["value"].values.reshape(1, 32, 1)))
mul_out = gs.Variable(name=name + "mul_out")
mul_node = gs.Node(op="Mul", name=name + "mul_node", attrs={}, inputs=[div_out, constantScale], outputs=[mul_out])
add_node = gs.Node(op="Add", name=name + "add_node", attrs={}, inputs=[mul_out, constantBias], outputs=[output_tensor])
self.graph.nodes.extend([mean_node, sub_node, pow_node, mean2_node, epsilon_node, sqrt_node, div_node, mul_node, add_node])
node.inputs = []
node.outputs = []
nRemoveInstanceNorm += 1
self.cleanup()
return nRemoveInstanceNorm
def insert_groupnorm_plugin(self):
nGroupNormPlugin = 0
for node in self.graph.nodes:
if node.op == "Reshape" and node.outputs != [] and \
node.o().op == "ReduceMean" and node.o(1).op == "Sub" and node.o().o() == node.o(1) and \
node.o().o().o().o().o().o().o().o().o().o().o().op == "Mul" and \
node.o().o().o().o().o().o().o().o().o().o().o().o().op == "Add" and \
len(node.o().o().o().o().o().o().o().o().inputs[1].values.shape) == 3:
# "node.outputs != []" is added for VAE
inputTensor = node.inputs[0]
gammaNode = node.o().o().o().o().o().o().o().o().o().o().o()
index = [type(i) == gs.ir.tensor.Constant for i in gammaNode.inputs].index(True)
gamma = np.array(deepcopy(gammaNode.inputs[index].values.tolist()), dtype=np.float32)
constantGamma = gs.Constant("groupNormGamma-" + str(nGroupNormPlugin), np.ascontiguousarray(gamma.reshape(-1))) # MUST use np.ascontiguousarray, or TRT will regard the shape of this Constant as (0) !!!
betaNode = gammaNode.o()
index = [type(i) == gs.ir.tensor.Constant for i in betaNode.inputs].index(True)
beta = np.array(deepcopy(betaNode.inputs[index].values.tolist()), dtype=np.float32)
constantBeta = gs.Constant("groupNormBeta-" + str(nGroupNormPlugin), np.ascontiguousarray(beta.reshape(-1)))
epsilon = node.o().o().o().o().o().inputs[1].values.tolist()[0]
if betaNode.o().op == "Sigmoid": # need Swish
bSwish = True
lastNode = betaNode.o().o() # Mul node of Swish
else:
bSwish = False
lastNode = betaNode # Cast node after Group Norm
if lastNode.o().op == "Cast":
lastNode = lastNode.o()
inputList = [inputTensor, constantGamma, constantBeta]
groupNormV = gs.Variable("GroupNormV-" + str(nGroupNormPlugin), np.dtype(np.float16), inputTensor.shape)
groupNormN = gs.Node("GroupNorm", "GroupNormN-" + str(nGroupNormPlugin), inputs=inputList, outputs=[groupNormV], attrs=OrderedDict([('epsilon', epsilon), ('bSwish', int(bSwish))]))
self.graph.nodes.append(groupNormN)
for subNode in self.graph.nodes:
if lastNode.outputs[0] in subNode.inputs:
index = subNode.inputs.index(lastNode.outputs[0])
subNode.inputs[index] = groupNormV
node.inputs = []
lastNode.outputs = []
nGroupNormPlugin += 1
self.cleanup()
return nGroupNormPlugin
def insert_layernorm_plugin(self):
nLayerNormPlugin = 0
for node in self.graph.nodes:
if node.op == 'ReduceMean' and \
node.o().op == 'Sub' and node.o().inputs[0] == node.inputs[0] and \
node.o().o(0).op =='Pow' and node.o().o(1).op =='Div' and \
node.o().o(0).o().op == 'ReduceMean' and \
node.o().o(0).o().o().op == 'Add' and \
node.o().o(0).o().o().o().op == 'Sqrt' and \
node.o().o(0).o().o().o().o().op == 'Div' and node.o().o(0).o().o().o().o() == node.o().o(1) and \
node.o().o(0).o().o().o().o().o().op == 'Mul' and \
node.o().o(0).o().o().o().o().o().o().op == 'Add' and \
len(node.o().o(0).o().o().o().o().o().inputs[1].values.shape) == 1:
if node.i().op == "Add":
inputTensor = node.inputs[0] # CLIP
else:
inputTensor = node.i().inputs[0] # UNet and VAE
gammaNode = node.o().o().o().o().o().o().o()
index = [type(i) == gs.ir.tensor.Constant for i in gammaNode.inputs].index(True)
gamma = np.array(deepcopy(gammaNode.inputs[index].values.tolist()), dtype=np.float32)
constantGamma = gs.Constant("LayerNormGamma-" + str(nLayerNormPlugin), np.ascontiguousarray(gamma.reshape(-1))) # MUST use np.ascontiguousarray, or TRT will regard the shape of this Constant as (0) !!!
betaNode = gammaNode.o()
index = [type(i) == gs.ir.tensor.Constant for i in betaNode.inputs].index(True)
beta = np.array(deepcopy(betaNode.inputs[index].values.tolist()), dtype=np.float32)
constantBeta = gs.Constant("LayerNormBeta-" + str(nLayerNormPlugin), np.ascontiguousarray(beta.reshape(-1)))
inputList = [inputTensor, constantGamma, constantBeta]
layerNormV = gs.Variable("LayerNormV-" + str(nLayerNormPlugin), np.dtype(np.float32), inputTensor.shape)
layerNormN = gs.Node("LayerNorm", "LayerNormN-" + str(nLayerNormPlugin), inputs=inputList, attrs=OrderedDict([('epsilon', 1.e-5)]), outputs=[layerNormV])
self.graph.nodes.append(layerNormN)
nLayerNormPlugin += 1
if betaNode.outputs[0] in self.graph.outputs:
index = self.graph.outputs.index(betaNode.outputs[0])
self.graph.outputs[index] = layerNormV
else:
if betaNode.o().op == "Cast":
lastNode = betaNode.o()
else:
lastNode = betaNode
for subNode in self.graph.nodes:
if lastNode.outputs[0] in subNode.inputs:
index = subNode.inputs.index(lastNode.outputs[0])
subNode.inputs[index] = layerNormV
lastNode.outputs = []
self.cleanup()
return nLayerNormPlugin
def fuse_kv(self, node_k, node_v, fused_kv_idx, heads, num_dynamic=0):
# Get weights of K
weights_k = node_k.inputs[1].values
# Get weights of V
weights_v = node_v.inputs[1].values
# Input number of channels to K and V
C = weights_k.shape[0]
# Number of heads
H = heads
# Dimension per head
D = weights_k.shape[1] // H
# Concat and interleave weights such that the output of fused KV GEMM has [b, s_kv, h, 2, d] shape
weights_kv = np.dstack([weights_k.reshape(C, H, D), weights_v.reshape(C, H, D)]).reshape(C, 2 * H * D)
# K and V have the same input
input_tensor = node_k.inputs[0]
# K and V must have the same output which we feed into fmha plugin
output_tensor_k = node_k.outputs[0]
# Create tensor
constant_weights_kv = gs.Constant("Weights_KV_{}".format(fused_kv_idx), np.ascontiguousarray(weights_kv))
# Create fused KV node
fused_kv_node = gs.Node(op="MatMul", name="MatMul_KV_{}".format(fused_kv_idx), inputs=[input_tensor, constant_weights_kv], outputs=[output_tensor_k])
self.graph.nodes.append(fused_kv_node)
# Connect the output of fused node to the inputs of the nodes after K and V
node_v.o(num_dynamic).inputs[0] = output_tensor_k
node_k.o(num_dynamic).inputs[0] = output_tensor_k
for i in range(0,num_dynamic):
node_v.o().inputs.clear()
node_k.o().inputs.clear()
# Clear inputs and outputs of K and V to ge these nodes cleared
node_k.outputs.clear()
node_v.outputs.clear()
node_k.inputs.clear()
node_v.inputs.clear()
self.cleanup()
return fused_kv_node
def insert_fmhca(self, node_q, node_kv, final_tranpose, mhca_idx, heads, num_dynamic=0):
# Get inputs and outputs for the fMHCA plugin
# We take an output of reshape that follows the Q GEMM
output_q = node_q.o(num_dynamic).o().inputs[0]
output_kv = node_kv.o().inputs[0]
output_final_tranpose = final_tranpose.outputs[0]
# Clear the inputs of the nodes that follow the Q and KV GEMM
# to delete these subgraphs (it will be substituted by fMHCA plugin)
node_kv.outputs[0].outputs[0].inputs.clear()
node_kv.outputs[0].outputs[0].inputs.clear()
node_q.o(num_dynamic).o().inputs.clear()
for i in range(0,num_dynamic):
node_q.o(i).o().o(1).inputs.clear()
weights_kv = node_kv.inputs[1].values
dims_per_head = weights_kv.shape[1] // (heads * 2)
# Reshape dims
shape = gs.Constant("Shape_KV_{}".format(mhca_idx), np.ascontiguousarray(np.array([0, 0, heads, 2, dims_per_head], dtype=np.int64)))
# Reshape output tensor
output_reshape = gs.Variable("ReshapeKV_{}".format(mhca_idx), np.dtype(np.float16), None)
# Create fMHA plugin
reshape = gs.Node(op="Reshape", name="Reshape_{}".format(mhca_idx), inputs=[output_kv, shape], outputs=[output_reshape])
# Insert node
self.graph.nodes.append(reshape)
# Create fMHCA plugin
fmhca = gs.Node(op="fMHCA", name="fMHCA_{}".format(mhca_idx), inputs=[output_q, output_reshape], outputs=[output_final_tranpose])
# Insert node
self.graph.nodes.append(fmhca)
# Connect input of fMHCA to output of Q GEMM
node_q.o(num_dynamic).outputs[0] = output_q
if num_dynamic > 0:
reshape2_input1_out = gs.Variable("Reshape2_fmhca{}_out".format(mhca_idx), np.dtype(np.int64), None)
reshape2_input1_shape = gs.Node("Shape", "Reshape2_fmhca{}_shape".format(mhca_idx), inputs=[node_q.inputs[0]], outputs=[reshape2_input1_out])
self.graph.nodes.append(reshape2_input1_shape)
final_tranpose.o().inputs[1] = reshape2_input1_out
# Clear outputs of transpose to get this subgraph cleared
final_tranpose.outputs.clear()
self.cleanup()
def fuse_qkv(self, node_q, node_k, node_v, fused_qkv_idx, heads, num_dynamic=0):
# Get weights of Q
weights_q = node_q.inputs[1].values
# Get weights of K
weights_k = node_k.inputs[1].values
# Get weights of V
weights_v = node_v.inputs[1].values
# Input number of channels to Q, K and V
C = weights_k.shape[0]
# Number of heads
H = heads
# Hidden dimension per head
D = weights_k.shape[1] // H
# Concat and interleave weights such that the output of fused QKV GEMM has [b, s, h, 3, d] shape
weights_qkv = np.dstack([weights_q.reshape(C, H, D), weights_k.reshape(C, H, D), weights_v.reshape(C, H, D)]).reshape(C, 3 * H * D)
input_tensor = node_k.inputs[0] # K and V have the same input
# Q, K and V must have the same output which we feed into fmha plugin
output_tensor_k = node_k.outputs[0]
# Concat and interleave weights such that the output of fused QKV GEMM has [b, s, h, 3, d] shape
constant_weights_qkv = gs.Constant("Weights_QKV_{}".format(fused_qkv_idx), np.ascontiguousarray(weights_qkv))
# Created a fused node
fused_qkv_node = gs.Node(op="MatMul", name="MatMul_QKV_{}".format(fused_qkv_idx), inputs=[input_tensor, constant_weights_qkv], outputs=[output_tensor_k])
self.graph.nodes.append(fused_qkv_node)
# Connect the output of the fused node to the inputs of the nodes after Q, K and V
node_q.o(num_dynamic).inputs[0] = output_tensor_k
node_k.o(num_dynamic).inputs[0] = output_tensor_k
node_v.o(num_dynamic).inputs[0] = output_tensor_k
for i in range(0,num_dynamic):
node_q.o().inputs.clear()
node_k.o().inputs.clear()
node_v.o().inputs.clear()
# Clear inputs and outputs of Q, K and V to ge these nodes cleared
node_q.outputs.clear()
node_k.outputs.clear()
node_v.outputs.clear()
node_q.inputs.clear()
node_k.inputs.clear()
node_v.inputs.clear()
self.cleanup()
return fused_qkv_node
def insert_fmha(self, node_qkv, final_tranpose, mha_idx, heads, num_dynamic=0):
# Get inputs and outputs for the fMHA plugin
output_qkv = node_qkv.o().inputs[0]
output_final_tranpose = final_tranpose.outputs[0]
# Clear the inputs of the nodes that follow the QKV GEMM
# to delete these subgraphs (it will be substituted by fMHA plugin)
node_qkv.outputs[0].outputs[2].inputs.clear()
node_qkv.outputs[0].outputs[1].inputs.clear()
node_qkv.outputs[0].outputs[0].inputs.clear()
weights_qkv = node_qkv.inputs[1].values
dims_per_head = weights_qkv.shape[1] // (heads * 3)
# Reshape dims
shape = gs.Constant("Shape_QKV_{}".format(mha_idx), np.ascontiguousarray(np.array([0, 0, heads, 3, dims_per_head], dtype=np.int64)))
# Reshape output tensor
output_shape = gs.Variable("ReshapeQKV_{}".format(mha_idx), np.dtype(np.float16), None)
# Create fMHA plugin
reshape = gs.Node(op="Reshape", name="Reshape_{}".format(mha_idx), inputs=[output_qkv, shape], outputs=[output_shape])
# Insert node
self.graph.nodes.append(reshape)
# Create fMHA plugin
fmha = gs.Node(op="fMHA_V2", name="fMHA_{}".format(mha_idx), inputs=[output_shape], outputs=[output_final_tranpose])
# Insert node
self.graph.nodes.append(fmha)
if num_dynamic > 0:
reshape2_input1_out = gs.Variable("Reshape2_{}_out".format(mha_idx), np.dtype(np.int64), None)
reshape2_input1_shape = gs.Node("Shape", "Reshape2_{}_shape".format(mha_idx), inputs=[node_qkv.inputs[0]], outputs=[reshape2_input1_out])
self.graph.nodes.append(reshape2_input1_shape)
final_tranpose.o().inputs[1] = reshape2_input1_out
# Clear outputs of transpose to get this subgraph cleared
final_tranpose.outputs.clear()
self.cleanup()
def mha_mhca_detected(self, node, mha):
# Go from V GEMM down to the S*V MatMul and all way up to K GEMM
# If we are looking for MHCA inputs of two matmuls (K and V) must be equal.
# If we are looking for MHA inputs (K and V) must be not equal.
if node.op == "MatMul" and len(node.outputs) == 1 and \
((mha and len(node.inputs[0].inputs) > 0 and node.i().op == "Add") or \
(not mha and len(node.inputs[0].inputs) == 0)):
if node.o().op == 'Shape':
if node.o(1).op == 'Shape':
num_dynamic_kv = 3 if node.o(2).op == 'Shape' else 2
else:
num_dynamic_kv = 1
# For Cross-Attention, if batch axis is dynamic (in QKV), assume H*W (in Q) is dynamic as well
num_dynamic_q = num_dynamic_kv if mha else num_dynamic_kv + 1
else:
num_dynamic_kv = 0
num_dynamic_q = 0
o = node.o(num_dynamic_kv)
if o.op == "Reshape" and \
o.o().op == "Transpose" and \
o.o().o().op == "Reshape" and \
o.o().o().o().op == "MatMul" and \
o.o().o().o().i(0).op == "Softmax" and \
o.o().o().o().i(1).op == "Reshape" and \
o.o().o().o().i(0).i().op == "Mul" and \
o.o().o().o().i(0).i().i().op == "MatMul" and \
o.o().o().o().i(0).i().i().i(0).op == "Reshape" and \
o.o().o().o().i(0).i().i().i(1).op == "Transpose" and \
o.o().o().o().i(0).i().i().i(1).i().op == "Reshape" and \
o.o().o().o().i(0).i().i().i(1).i().i().op == "Transpose" and \
o.o().o().o().i(0).i().i().i(1).i().i().i().op == "Reshape" and \
o.o().o().o().i(0).i().i().i(1).i().i().i().i().op == "MatMul" and \
node.name != o.o().o().o().i(0).i().i().i(1).i().i().i().i().name:
# "len(node.outputs) == 1" to make sure we are not in the already fused node
node_q = o.o().o().o().i(0).i().i().i(0).i().i().i()
node_k = o.o().o().o().i(0).i().i().i(1).i().i().i().i()
node_v = node
final_tranpose = o.o().o().o().o(num_dynamic_q).o()
# Sanity check to make sure that the graph looks like expected
if node_q.op == "MatMul" and final_tranpose.op == "Transpose":
return True, num_dynamic_q, num_dynamic_kv, node_q, node_k, node_v, final_tranpose
return False, 0, 0, None, None, None, None
def fuse_kv_insert_fmhca(self, heads, mhca_index, sm):
nodes = self.graph.nodes
# Iterate over graph and search for MHCA pattern
for idx, _ in enumerate(nodes):
# fMHCA can't be at the 2 last layers of the network. It is a guard from OOB
if idx + 1 > len(nodes) or idx + 2 > len(nodes):
continue
# Get anchor nodes for fusion and fMHCA plugin insertion if the MHCA is detected
detected, num_dynamic_q, num_dynamic_kv, node_q, node_k, node_v, final_tranpose = \
self.mha_mhca_detected(nodes[idx], mha=False)
if detected:
assert num_dynamic_q == 0 or num_dynamic_q == num_dynamic_kv + 1
# Skip the FMHCA plugin for SM75 except for when the dim per head is 40.
if sm == 75 and node_q.inputs[1].shape[1] // heads == 160:
continue
# Fuse K and V GEMMS
node_kv = self.fuse_kv(node_k, node_v, mhca_index, heads, num_dynamic_kv)
# Insert fMHCA plugin
self.insert_fmhca(node_q, node_kv, final_tranpose, mhca_index, heads, num_dynamic_q)
return True
return False
def fuse_qkv_insert_fmha(self, heads, mha_index):
nodes = self.graph.nodes
# Iterate over graph and search for MHA pattern
for idx, _ in enumerate(nodes):
# fMHA can't be at the 2 last layers of the network. It is a guard from OOB
if idx + 1 > len(nodes) or idx + 2 > len(nodes):
continue
# Get anchor nodes for fusion and fMHA plugin insertion if the MHA is detected
detected, num_dynamic_q, num_dynamic_kv, node_q, node_k, node_v, final_tranpose = \
self.mha_mhca_detected(nodes[idx], mha=True)
if detected:
assert num_dynamic_q == num_dynamic_kv
# Fuse Q, K and V GEMMS
node_qkv = self.fuse_qkv(node_q, node_k, node_v, mha_index, heads, num_dynamic_kv)
# Insert fMHA plugin
self.insert_fmha(node_qkv, final_tranpose, mha_index, heads, num_dynamic_kv)
return True
return False
def insert_fmhca_plugin(self, num_heads, sm):
mhca_index = 0
while self.fuse_kv_insert_fmhca(num_heads, mhca_index, sm):
mhca_index += 1
return mhca_index
def insert_fmha_plugin(self, num_heads):
mha_index = 0
while self.fuse_qkv_insert_fmha(num_heads, mha_index):
mha_index += 1
return mha_index
|