昇腾910B部署Qwen2-7B-Instruct进行流式输出【pytorch框架】NPU推理_qwen2 910b

前情提要

torch_npu框架

官方未适配

mindsport框架

官方未适配

mindnlp框架

官方适配了，但是速度非常非常慢，10秒一个字

下载模型

国外

Hugging Face

国内

modelscope

环境设置

pip install transformers==4.39.2
pip3 install torch==2.1.0
pip3 install torch-npu==2.1.0.post4
pip3 install accelerate==0.24.1
pip3 install transformers-stream-generator==0.0.5
1
2
3
4
5

代码适配（非流式）

Main

import torch
import torch_npu
import os
import platform
torch_device = "npu:1" # 0~7
torch.npu.set_device(torch.device(torch_device))
torch.npu.set_compile_mode(jit_compile=False)
option = {}
option["NPU_FUZZY_COMPILE_BLACKLIST"] = "Tril"
torch.npu.set_option(option)
from transformers import AutoModelForCausalLM, AutoTokenizer
# device = "cuda" # the device to load the model onto
DEFAULT_CKPT_PATH = '/root/.cache/modelscope/hub/qwen/Qwen2-7B-Instruct'
model = AutoModelForCausalLM.from_pretrained(
    DEFAULT_CKPT_PATH,
    torch_dtype=torch.float16,
    device_map=torch_device
).npu().eval()
tokenizer = AutoTokenizer.from_pretrained(DEFAULT_CKPT_PATH)
while True:
    prompt = input("user:")
    if prompt == "exit":
        break
    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": prompt}
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(torch_device)

    generated_ids = model.generate(
        model_inputs.input_ids,
        max_new_tokens=512
    )
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    print("Qwen2-7B-Instruct:",response)
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

Branch

找到自己虚拟环境

which python
1

我的是/root/anaconda3/envs/sakura/bin/python
找到/lib/python3.9/site-packages/transformers/generation/utils.py示例：

/root/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/generation/utils.py
1

找到第2708行，注释掉2708行~2712行
在2709行添加

next_token_scores = outputs.logits[:, -1, :]
1

示例：

出错就是在这里，如果进行了pre-process distribution，就会报错

/root/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/generation/logits_process.py:455: UserWarning: AutoNonVariableTypeMode is deprecated and will be removed in 1.10 release. For kernel implementations please use AutoDispatchBelowADInplaceOrView instead, If you are looking for a user facing API to enable running your inference-only workload, please use c10::InferenceMode. Using AutoDispatchBelowADInplaceOrView in user code is under risk of producing silent wrong result in some edge cases. See Note [AutoDispatchBelowAutograd] for more details. (Triggered internally at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:74.)
  sorted_indices_to_remove[..., -self.min_tokens_to_keep :] = 0
Traceback (most recent call last):
  File "/root/Qwen_test.py", line 63, in <module>
    generated_ids = model.generate(
  File "/root/anaconda3/envs/sakura/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/root/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/generation/utils.py", line 1576, in generate
    result = self._sample(
  File "/root/anaconda3/envs/sakura/lib/python3.9/site-packages/transformers/generation/utils.py", line 2736, in _sample
    next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
RuntimeError: Sync:build/CMakeFiles/torch_npu.dir/compiler_depend.ts:158 NPU error, error code is 507018
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
E39999: Inner Error!
E39999: 2024-07-02-14:14:50.735.070  An exception occurred during AICPU execution, stream_id:23, task_id:2750, errcode:21008, msg:inner error[FUNC:ProcessAicpuErrorInfo][FILE:device_error_proc.cc][LINE:730]
        TraceBack (most recent call last):
        rtStreamSynchronizeWithTimeout execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
        synchronize stream failed, runtime result = 507018[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]


DEVICE[1] PID[864803]:
EXCEPTION TASK:
  Exception info:TGID=864803, model id=65535, stream id=23, stream phase=SCHEDULE, task id=2750, task type=aicpu kernel, recently received task id=2750, recently send task id=2749, task phase=RUN
  Message info[0]:aicpu=0,slot_id=0,report_mailbox_flag=0x5a5a5a5a,state=0x5210
    Other info[0]:time=2024-07-02-14:14:50.091.974, function=proc_aicpu_task_done, line=970, error code=0x2a
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EZ9999: Inner Error!
EZ9999: 2024-07-02-14:14:50.743.702  Kernel task happen error, retCode=0x2a, [aicpu exception].[FUNC:PreCheckTaskErr][FILE:task_info.cc][LINE:1776]
        TraceBack (most recent call last):
        Aicpu kernel execute failed, device_id=1, stream_id=23, task_id=2750, errorCode=2a.[FUNC:PrintAicpuErrorInfo][FILE:task_info.cc][LINE:1579]
        Aicpu kernel execute failed, device_id=1, stream_id=23, task_id=2750, fault op_name=[FUNC:GetError][FILE:stream.cc][LINE:1512]
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
        wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.745.695  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.747.300  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.814.377  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.816.023  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.817.628  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.819.236  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.820.843  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
[W compiler_depend.ts:368] Warning: NPU warning, error code is 507018[Error]:
[Error]: The aicpu execution is abnormal.
        Rectify the fault based on the error information in the ascend log.
EH9999: Inner Error!
        rtDeviceSynchronize execute failed, reason=[aicpu exception][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999: 2024-07-02-14:14:50.822.422  wait for compute device to finish failed, runtime result = 507018.[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:161]
        TraceBack (most recent call last):
 (function npuSynchronizeDevice)
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

结果展示

最后运行Main文件

代码适配（流式）

code

import torch
import torch_npu
import os
import platform
torch_device = "npu:1" # 0~7
torch.npu.set_device(torch.device(torch_device))
torch.npu.set_compile_mode(jit_compile=False)
option = {}
option["NPU_FUZZY_COMPILE_BLACKLIST"] = "Tril"
torch.npu.set_option(option)
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, TextStreamer
from threading import Thread
# device = "cuda" # the device to load the model onto
DEFAULT_CKPT_PATH = '/root/.cache/modelscope/hub/qwen/Qwen2-7B-Instruct'

model = AutoModelForCausalLM.from_pretrained(
    DEFAULT_CKPT_PATH,
    torch_dtype=torch.float16,
    device_map=torch_device
).npu().eval()
tokenizer = AutoTokenizer.from_pretrained(DEFAULT_CKPT_PATH)

def generate_with_npu_device(**generation_kwargs):
    torch.npu.set_device(torch.device(torch_device))
    model.generate(**generation_kwargs)

streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
turn_count = 0

while True:
    query = input("user:")
    if query.strip() == "exit":
        break
    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": query}
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(torch_device)

    generation_kwargs = dict(model_inputs, streamer=streamer, max_new_tokens=512)
    thread = Thread(target=generate_with_npu_device, kwargs=generation_kwargs)
    thread.start()
    generated_text = ""
    position = 0
    # 流式输出
    print('Qwen2-7B-Instruct:', end='', flush=True)
    for new_text in streamer:
        generated_text += new_text
        print(generated_text[position:], end='', flush=True)
        position = len(generated_text)
    print('\n')
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