-
Notifications
You must be signed in to change notification settings - Fork 0
/
modeling_qwen.py
79 lines (64 loc) · 2.57 KB
/
modeling_qwen.py
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
from typing import TYPE_CHECKING
import numpy as np
import openvino as ov
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
if TYPE_CHECKING:
from transformers.generation.streamers import BaseStreamer
try:
from einops import rearrange
except ImportError:
rearrange = None
from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List, Any, Generator
qwen_compiled_models = {}
DEVICE = "GPU"
def create_ov_mlp_model(np_hidden_states, np_w1_weight, np_w2_weight, np_c_proj_weight):
import openvino.runtime as ovrt
hidden_states = ovrt.opset1.parameter(
np_hidden_states.shape, np_hidden_states.dtype
)
w1_weight = ovrt.opset1.parameter(np_w1_weight.shape, np_w1_weight.dtype)
w2_weight = ovrt.opset1.parameter(np_w2_weight.shape, np_w2_weight.dtype)
c_proj_weight = ovrt.opset1.parameter(
np_c_proj_weight.shape, np_c_proj_weight.dtype
)
w1 = ovrt.opset1.matmul(hidden_states, w1_weight, False, True)
w2 = ovrt.opset1.matmul(hidden_states, w2_weight, False, True)
silu = ovrt.opset4.swish(w2)
mul = ovrt.opset1.multiply(w1, silu)
c_proj = ovrt.opset1.matmul(mul, c_proj_weight, False, True)
out = ovrt.opset1.result(c_proj)
model = ovrt.Model(
[out], [hidden_states, w1_weight, w2_weight, c_proj_weight], "QWenMLP"
)
return model
def qwenmlp_forward(self, hidden_states):
TensorClass = hidden_states.__class__
np_hidden_states = hidden_states.numpy()
np_w1 = self.w1.weight.numpy()
np_w2 = self.w2.weight.numpy()
np_c_proj = self.c_proj.weight.numpy()
model_spec = f"{np_hidden_states.shape}_mlp"
if model_spec not in qwen_compiled_models:
print(f"compiling {model_spec}")
model = create_ov_mlp_model(np_hidden_states, np_w1, np_w2, np_c_proj)
qwen_compiled_models[model_spec] = ov.compile_model(model, DEVICE)
output = qwen_compiled_models[model_spec].infer(
[np_hidden_states, np_w1, np_w2, np_c_proj]
)[0]
return TensorClass(output)
def get_modeling_qwen():
import inspect
import transformers
# get modeling_qwen module and hacking QWenAttention by QWenAttentionNPU
QWenLMHeadModel = transformers.dynamic_module_utils.get_class_from_dynamic_module(
"modeling_qwen.QWenLMHeadModel", "Qwen/Qwen-1_8B-Chat"
)
modeling_qwen = inspect.getmodule(QWenLMHeadModel)
return modeling_qwen
def hack_mlp():
modeling_qwen = get_modeling_qwen()
qwenmlp_forward.hacking_target = modeling_qwen.QWenMLP.forward
modeling_qwen.QWenMLP.forward = qwenmlp_forward