# Why doesn't the official tutorial work? **URL:** \ **Category:** Questions\ **Created:** [March 25, 2025, 3:01am UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209 "2025-03-25T03:01:51Z")\ **Posts on this page:** 8\ **Page:** 1
**Author:** ![Jianhua-Cui](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/jianhua-cui/32/5764_2.png) [@Jianhua-Cui](https://discuss.tvm.apache.org/u/Jianhua-Cui)\ **Post date:** [March 25, 2025, 3:01am UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/1 "2025-03-25T03:01:51Z")
I installed tvm from the source code according to the official website and passed the Python-based installation verification steps. Subsequently, I tried running this chapter from the tutorial: [End-to-End Optimize Model — tvm 0.20.dev0 documentation](https://tvm.apache.org/docs/how_to/tutorials/e2e_opt_model.html). To quickly verify it, I replaced the resnet18 imported from torchvision with the following simple Torch model: class SimpleCNN(nn.Module): def **init** (self): super(SimpleCNN, self). **init** () self.conv1 = nn.Conv2d(1, 16, kernel\_size=3, padding=1) self.relu = nn.ReLU() ``` def forward(self, x): x = self.relu(self.conv1(x)) return x ``` And I modified TOTAL\_TRIALS to 10. However, at this step: ex = tvm.compile(mod, target=“cuda”), I encountered an error: TVMError: Traceback (most recent call last): 0: operator() at /home/cuijianhua/llm\_workspace/mlc\_workspace/learning\_tvm/tvm/src/tir/analysis/verify\_memory.cc:203 Did you forget to bind? Variable `p_conv1_bias` is directly accessed by host memory (it is not contained in a thread environment or in the function arguments. Variable `T_reshape` is directly accessed by host memory (it is not contained in a thread environment or in the function arguments. File “/home/cuijianhua/learning\_tvm/tvm/src/tir/analysis/verify\_memory.cc”, line 203 RuntimeError: Memory verification failed with the following errors: As a beginner, I’m very confused and don’t know where to start troubleshooting this kind of issue. Theoretically, I only modified a small, seemingly unimportant part of the tutorial, so why isn’t it working? Are there any relevant documents I can refer to? I couldn’t find effective help from Google or ChatGPT. ---
**Author:** ![AQSingh](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/aqsingh/32/4666_2.png) [@AQSingh](https://discuss.tvm.apache.org/u/AQSingh)\ **Post date:** [March 25, 2025, 4:03am UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/2 "2025-03-25T04:03:44Z")
Can you provide a full self-contained example that one can use to reproduce? The lack of formatting here makes it unclear. ---
**Author:** ![Hzfengsy](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/hzfengsy/32/2961_2.png) [@Hzfengsy](https://discuss.tvm.apache.org/u/Hzfengsy)\ **Post date:** [March 25, 2025, 7:47am UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/3 "2025-03-25T07:47:53Z")
> [@Jianhua-Cui](#): > > And I modified TOTAL\_TRIALS to 10 10 is not large enough 🙂 ---
**Author:** ![Jianhua-Cui](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/jianhua-cui/32/5764_2.png) [@Jianhua-Cui](https://discuss.tvm.apache.org/u/Jianhua-Cui)\ **Post date:** [March 25, 2025, 11:43pm UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/4 "2025-03-25T23:43:39Z")
Thank you very much for your response. I’ll immediately adjust TOTAL\_TRIALS to 8000 to match the tutorial and give it a try. By the way, I’d like to ask about this kind of issue (i.e., whether “TOTAL\_TRAILS” affects binding in kernel). I’m not sure where to find similar information. I searched for both “get\_pipeline” and “static\_shape\_tuning” on [Apache TVM Documentation — tvm 0.20.dev0 documentation](https://tvm.apache.org/docs/) but couldn’t find relevant details. Have I possibly overlooked some documentation? ---
**Author:** ![Jianhua-Cui](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/jianhua-cui/32/5764_2.png) [@Jianhua-Cui](https://discuss.tvm.apache.org/u/Jianhua-Cui)\ **Post date:** [March 25, 2025, 11:43pm UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/5 "2025-03-25T23:43:41Z")
Sorry for the poor format, I’ve reorganized my code and pasted it here: ```python import tvm import torch import torch.nn as nn import tvm from tvm import relax from tvm.relax.frontend.onnx import from_onnx from tvm.relax.frontend.torch import from_exported_program import numpy as np import os import shutil class SimpleCNN(nn.Module): def __init__ (self): super(SimpleCNN, self). __init__ () self.conv1 = nn.Conv2d(1, 16, kernel_size=3, padding=1) self.relu = nn.ReLU() def forward(self, x): x = self.relu(self.conv1(x)) return x from torch.export import export torch_model = SimpleCNN() torch_model.eval() example_args = (torch.randn(1, 1, 28, 28), ) with torch.no_grad(): exported_program = export(torch_model, example_args) mod = from_exported_program(exported_program, keep_params_as_input=True) mod, params = relax.frontend.detach_params(mod) mod.show() TOTAL_TRIALS = 10 # Change to 20000 for better performance if needed target = tvm.target.Target("nvidia/geforce-gtx-1080-ti") # Change to your target device work_dir = "tuning_logs" mod = relax.get_pipeline("static_shape_tuning", target=target, total_trials=TOTAL_TRIALS)(mod) # Only show the main function mod["main"].show() ex = tvm.compile(mod, target="cuda") ``` Thank you for the reminder! ---
**Author:** ![Hzfengsy](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/hzfengsy/32/2961_2.png) [@Hzfengsy](https://discuss.tvm.apache.org/u/Hzfengsy)\ **Post date:** [March 26, 2025, 5:09am UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/6 "2025-03-26T05:09:11Z")
During pipeline tuning, schedules (which determine thread binding) are only generated as part of the tuning process. If `TOTAL_TRIALS` is insufficient for all kernels to discover valid schedules, the remaining kernels will be left unscheduled (without thread binding), causing such issue. In this case, the value 20 for `TOTAL_TRIALS` is too small, likely allowing only the first kernel to be properly tuned while leaving subsequent kernels unscheduled. This is the root cause of the problem. ---
**Author:** ![Jianhua-Cui](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/jianhua-cui/32/5764_2.png) [@Jianhua-Cui](https://discuss.tvm.apache.org/u/Jianhua-Cui)\ **Post date:** [March 27, 2025, 5:31pm UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/7 "2025-03-27T17:31:43Z")
Thanks for the explanation, it works well after adjusting TOTAL\_TRIALS to 8000. ---
**Author:** ![erdayang](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/erdayang/32/5740_2.png) [@erdayang](https://discuss.tvm.apache.org/u/erdayang)\ **Post date:** [June 9, 2025, 3:17pm UTC](https://discuss.tvm.apache.org/t/why-doesnt-the-official-tutorial-work/18209/8 "2025-06-09T15:17:12Z")
@Hzfengsy hello，do you have an example of auto tuning for dynamic shape model? The example I found is using static\_shape\_tuning。