【StableDiffusion】秋叶包，安装 tensorRT 各种坑，包含请求超时、Exporting to ONNX failed、Failed to parse ONNX model

总览

1.无法安装
2.报错解决
3.显存小的就别玩 tensorRT 了

一、无法安装

1.使用魔法

我知道你要说，魔法也不行巴拉巴拉
我也很疑惑，最开始使用本地网络和魔法都不行
后来发现是概率问题
使用魔法，然后多试几次（我试了大概7、8次）然后下载成功了

2.疑难杂症

什么请求超时、服务器无法回应、回应错误之类的巴拉巴拉
直接使用魔法多试几次就好了

二、报错

1.报错情况1

ERROR:root:Exporting to ONNX failed. Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)

完整报错内容：

Disabling attention optimization
Exporting boleromixPony_v14 to TensorRT using - Batch Size: 1-1-1
Height: 768-768-768
Width: 768-768-768
Token Count: 75-75-75
Disabling attention optimization
ERROR:root:Exporting to ONNX failed. Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)
Building TensorRT engine... This can take a while, please check the progress in the terminal.
Building TensorRT engine for S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\models\Unet-onnx\boleromixPony_v14.onnx: S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\models\Unet-trt\boleromixPony_v14_ad5d6010_cc86_sample=2x4x96x96-timesteps=2-encoder_hidden_states=2x77x2048-y=2x2816.trt
Could not open file S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\models\Unet-onnx\boleromixPony_v14.onnx
Could not open file S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\models\Unet-onnx\boleromixPony_v14.onnx
[E] ModelImporter.cpp:773: Failed to parse ONNX model from file: S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\models\Unet-onnx\boleromixPony_v14.onnx
[!] Failed to parse ONNX model. Does the model file exist and contain a valid ONNX model?
Traceback (most recent call last):
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\gradio\routes.py", line 488, in run_predict
    output = await app.get_blocks().process_api(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\gradio\blocks.py", line 1431, in process_api
    result = await self.call_function(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\gradio\blocks.py", line 1103, in call_function
    prediction = await anyio.to_thread.run_sync(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\anyio\to_thread.py", line 31, in run_sync
    return await get_asynclib().run_sync_in_worker_thread(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\anyio\_backends\_asyncio.py", line 937, in run_sync_in_worker_thread
    return await future
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\anyio\_backends\_asyncio.py", line 867, in run
    result = context.run(func, *args)
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\gradio\utils.py", line 707, in wrapper
    response = f(*args, **kwargs)
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\extensions\Stable-Diffusion-WebUI-TensorRT\ui_trt.py", line 126, in export_unet_to_trt
    ret = export_trt(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\extensions\Stable-Diffusion-WebUI-TensorRT\exporter.py", line 231, in export_trt
    ret = engine.build(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\extensions\Stable-Diffusion-WebUI-TensorRT\utilities.py", line 227, in build
    network = network_from_onnx_path(
  File "<string>", line 3, in network_from_onnx_path
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\polygraphy\backend\base\loader.py", line 40, in __call__
    return self.call_impl(*args, **kwargs)
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\polygraphy\util\util.py", line 710, in wrapped
    return func(*args, **kwargs)
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\polygraphy\backend\trt\loader.py", line 247, in call_impl
    trt_util.check_onnx_parser_errors(parser, success)
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\polygraphy\backend\trt\util.py", line 88, in check_onnx_parser_errors
    G_LOGGER.critical(
  File "S:\app_AI\stableDiffusion-webui-aki\sd-webui-aki-v4.8\python\lib\site-packages\polygraphy\logger\logger.py", line 605, in critical
    raise ExceptionType(message) from None
polygraphy.exception.exception.PolygraphyException: Failed to parse ONNX model. Does the model file exist and contain a valid ONNX model?

2.解决办法

去这个目录：

X:\XXX\sd-webui-aki-v4.8\extensions\Stable-Diffusion-WebUI-TensorRT

替换这两个文件：

ui_trt.py
exporter.py

替换的代码如下：
ui_trt.py 替换代码（直接复制就好了，别忘了删掉 CSDN 水印）：

import os
import gc
import json
import logging
from collections import defaultdict

import torch
from safetensors.torch import save_file
import gradio as gr

from modules.shared import cmd_opts
from modules.ui_components import FormRow
from modules import sd_hijack, sd_models, shared
from modules.ui_common import refresh_symbol
from modules.ui_components import ToolButton

from model_helper import UNetModel
from exporter import export_onnx, export_trt, export_lora
from model_manager import modelmanager, cc_major, TRT_MODEL_DIR
from datastructures import SDVersion, ProfilePrests, ProfileSettings


profile_presets = ProfilePrests()

logging.basicConfig(level=logging.INFO)


def get_context_dim():
    if shared.sd_model.is_sd1:
        return 768
    elif shared.sd_model.is_sd2:
        return 1024
    elif shared.sd_model.is_sdxl:
        return 2048


def is_fp32():
    use_fp32 = False
    if cc_major < 7:
        use_fp32 = True
        print("FP16 has been disabled because your GPU does not support it.")
    return use_fp32


def export_unet_to_trt(
    batch_min,
    batch_opt,
    batch_max,
    height_min,
    height_opt,
    height_max,
    width_min,
    width_opt,
    width_max,
    token_count_min,
    token_count_opt,
    token_count_max,
    force_export,
    static_shapes,
    preset,
):
    sd_hijack.model_hijack.apply_optimizations("None")

    is_xl = shared.sd_model.is_sdxl
    model_name = shared.sd_model.sd_checkpoint_info.model_name

    profile_settings = ProfileSettings(
        batch_min,
        batch_opt,
        batch_max,
        height_min,
        height_opt,
        height_max,
        width_min,
        width_opt,
        width_max,
        token_count_min,
        token_count_opt,
        token_count_max,
    )
    if preset == "Default":
        profile_settings = profile_presets.get_default(is_xl=is_xl)
    use_fp32 = is_fp32()

    print(f"Exporting {
     
     model_name} to TensorRT using - {
     
     profile_settings}")
    profile_settings.token_to_dim(static_shapes)

    model_hash = shared.sd_model.sd_checkpoint_info.hash
    model_name = shared.sd_model.sd_checkpoint_info.model_name

    onnx_filename, onnx_path = modelmanager.get_onnx_path(model_name)
    timing_cache = modelmanager.get_timing_cache()

    diable_optimizations = is_xl
    embedding_dim = get_context_dim()

    modelobj = UNetModel(
        shared.sd_model.model.diffusion_model,
        embedding_dim,
        text_minlen=profile_settings.t_min,
        is_xl=is_xl,
    )
    modelobj.apply_torch_model()

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    profile = modelobj.get_input_profile(profile_settings)
    modelobj.unet.to(device)
    for key, value in profile.items():
        if isinstance(value, torch.Tensor):
            profile[key] = value.to(device)
        elif isinstance(value, tuple):
            profile[key] = tuple(v.to(device) for v in value)

    export_onnx(
        onnx_path,
        modelobj,
        profile_settings,
        diable_optimizations=diable_optimizations,
    )
    gc.collect()
    torch.cuda.empty_cache()

    trt_engine_filename, trt_path = modelmanager.get_trt_path(
        model_name, model_hash, profile, static_shapes
    )

    if not os.path.exists(trt_path) or force_export:
        print(
            "Building TensorRT engine... This can take a while, please check the progress in the terminal."
        )
        gr.Info(
            "Building TensorRT engine... This can take a while, please check the progress in the terminal."
        )
        ret = export_trt(
            trt_path,
            onnx_path,
            timing_cache,
            profile=profile,
            use_fp16=not use_fp32,
        )
        if ret:
            return "## Export Failed due to unknown reason. See shell for more information. \n"

        print("TensorRT engines has been saved to disk.")
        modelmanager.add_entry(
            model_name,
            model_hash,
            profile,
            static_shapes,
            fp32=use_fp32,
            inpaint=True if modelobj.in_channels == 6 else False,
            refit=True,
            vram=0,
            unet_hidden_dim=modelobj.in_channels,
            lora=False,
        )
    else:
        print(
            "TensorRT engine found. Skipping build. You can enable Force Export in the Advanced Settings to force a rebuild if needed."
        )

    gc.collect()
    torch.cuda.empty_cache()

    return "## Exported Successfully \n"


def export_lora_to_trt(lora_name, force_export):
    is_xl = shared.sd_model.is_sdxl

    available_lora_models = get_lora_checkpoints()
    lora_name = lora_name.split(" ")[0]
    lora_model