# Questions

**URL:** https://discuss.tvm.apache.org/c/questions/7.md

[Latest](https://discuss.tvm.apache.org/latest.md) · [Categories](https://discuss.tvm.apache.org/categories.md)

---

## [About the Questions category](https://discuss.tvm.apache.org/t/about-the-questions-category/20)

<div class="topic-metadata">

**Author:** [@tqchen](https://discuss.tvm.apache.org/u/tqchen)\
**Replies:** 1\
**Last updated:** [July 16, 2018, 4:23am UTC](https://discuss.tvm.apache.org/t/about-the-questions-category/20 "2018-07-16T04:23:29Z")

</div>

Any questions related to the tvm stack

---

## [Seeking Community Input: Embedded Device Support in Relax](https://discuss.tvm.apache.org/t/seeking-community-input-embedded-device-support-in-relax/18664)

<div class="topic-metadata">

**Author:** [@a\_j](https://discuss.tvm.apache.org/u/a_j)\
**Replies:** 2\
**Last updated:** [October 29, 2026, 5:27pm UTC](https://discuss.tvm.apache.org/t/seeking-community-input-embedded-device-support-in-relax/18664 "2026-10-29T17:27:18Z")

</div>

Hi TVM community, We’ve been successfully using Relay-based TVM for ML compilation across our embedded DSP and MCU devices (including Cortex-M) for several years. The Relay flow has been excellent for our use cases, par…

---

## [\[Question\] Can self-contained BYOC be achieved via micro-kernel Injection?](https://discuss.tvm.apache.org/t/question-can-self-contained-byoc-be-achieved-via-micro-kernel-injection/18947)

<div class="topic-metadata">

**Author:** [@garfield0xff-cochl](https://discuss.tvm.apache.org/u/garfield0xff-cochl)\
**Replies:** 0\
**Last updated:** [March 9, 2026, 3:10pm UTC](https://discuss.tvm.apache.org/t/question-can-self-contained-byoc-be-achieved-via-micro-kernel-injection/18947 "2026-03-09T15:10:17Z")

</div>

I am currently exploring a “runtime-less” or “minimal-dependency” deployment strategy using TVM. While the standard BYOC (Bring Your Own Codegen) pipeline is powerful, it often requires linking against external librarie…

---

## [Build Apache TVM runtime for Adreno GPU on Android ARM64v8](https://discuss.tvm.apache.org/t/build-apache-tvm-runtime-for-adreno-gpu-on-android-arm64v8/17602)

<div class="topic-metadata">

**Author:** [@varunnaw](https://discuss.tvm.apache.org/u/varunnaw)\
**Replies:** 0\
**Last updated:** [August 14, 2024, 4:46pm UTC](https://discuss.tvm.apache.org/t/build-apache-tvm-runtime-for-adreno-gpu-on-android-arm64v8/17602 "2024-08-14T16:46:05Z")

</div>

I have two host platforms: a linux amd64 docker container and a macos m1 mac. My target platform is an android device with a qualcomm soc: adreno gpu and hexagon dsp. I would like to just work with the adreno gpu for no…

---

## [C++ relay frontend](https://discuss.tvm.apache.org/t/c-relay-frontend/18917)

<div class="topic-metadata">

**Author:** [@binzhang1008](https://discuss.tvm.apache.org/u/binzhang1008)\
**Replies:** 0\
**Last updated:** [February 2, 2026, 2:38pm UTC](https://discuss.tvm.apache.org/t/c-relay-frontend/18917 "2026-02-02T14:38:08Z")

</div>

Currently, it seems that in TVM, the ONNX input is first received through the Python interface relay.frontend.from\_onnx, and then the model is compiled using relay.build. However, my environment is purely C++, and I have…

---

## [import tvm error, when initialized module 'tvm\_ffi'(most likely due to a circular import)](https://discuss.tvm.apache.org/t/import-tvm-error-when-initialized-module-tvm-ffi-most-likely-due-to-a-circular-import/18892)

<div class="topic-metadata">

**Author:** [@shujuliu818](https://discuss.tvm.apache.org/u/shujuliu818)\
**Replies:** 0\
**Last updated:** [January 15, 2026, 6:44pm UTC](https://discuss.tvm.apache.org/t/import-tvm-error-when-initialized-module-tvm-ffi-most-likely-due-to-a-circular-import/18892 "2026-01-15T18:44:29Z")

</div>

hello: when i use python “import tvm” it’s some error: File “”, line 1, in import tvm; print(tvm.base.\_LIB) ^^^^^^^^^^ File “/home/ll/tvm/python/tvm/init.py”, line 24, in from tvm\_ffi import register\_object, register\_…

---

## [tvm 0.20.0 compile onnx, infer time slow.](https://discuss.tvm.apache.org/t/tvm-0-20-0-compile-onnx-infer-time-slow/18874)

<div class="topic-metadata">

**Author:** [@zhiguangli](https://discuss.tvm.apache.org/u/zhiguangli)\
**Replies:** 0\
**Last updated:** [January 6, 2026, 3:08am UTC](https://discuss.tvm.apache.org/t/tvm-0-20-0-compile-onnx-infer-time-slow/18874 "2026-01-06T03:08:35Z")

</div>

Subject: Why is inference so slow when compiling ONNX with TVM and using automatic optimization? Background: I am using TVM version 0.20.0 to compile an ONNX model into a .so library and optimize it with relax.transform…

---

## [\[C++ RPC + AOT\] Error "Can not pass in local device" when deploying AOT-compiled model via C++ RPC client](https://discuss.tvm.apache.org/t/c-rpc-aot-error-can-not-pass-in-local-device-when-deploying-aot-compiled-model-via-c-rpc-client/18845)

<div class="topic-metadata">

**Author:** [@markwester](https://discuss.tvm.apache.org/u/markwester)\
**Replies:** 1\
**Last updated:** [December 22, 2025, 2:40pm UTC](https://discuss.tvm.apache.org/t/c-rpc-aot-error-can-not-pass-in-local-device-when-deploying-aot-compiled-model-via-c-rpc-client/18845 "2025-12-22T14:40:58Z")

</div>

Hi TVM community, I’m trying to deploy an AOT-compiled graph executor model using the TVM C++ RPC API, but I keep encountering a runtime error related to device handling. Unfortunately, documentation and examples for th…

---

## [torch.nn.Dropout is not converted into Relax](https://discuss.tvm.apache.org/t/torch-nn-dropout-is-not-converted-into-relax/18862)

<div class="topic-metadata">

**Author:** [@AugustineYang](https://discuss.tvm.apache.org/u/AugustineYang)\
**Replies:** 0\
**Last updated:** [December 22, 2025, 1:43pm UTC](https://discuss.tvm.apache.org/t/torch-nn-dropout-is-not-converted-into-relax/18862 "2025-12-22T13:43:51Z")

</div>

Hi. I encountered an issue when trying to convert a PyTorch FCN model into Relax IRModule following the tutorial Import from existing models: torch.nn.dropout seems to be discarded. In Relay IR, torch.nn.Dropout is conv…

---

## [How to pass mutable python objects to C++ ffi functions](https://discuss.tvm.apache.org/t/how-to-pass-mutable-python-objects-to-c-ffi-functions/18848)

<div class="topic-metadata">

**Author:** [@mamodev](https://discuss.tvm.apache.org/u/mamodev)\
**Replies:** 3\
**Last updated:** [December 20, 2025, 12:21pm UTC](https://discuss.tvm.apache.org/t/how-to-pass-mutable-python-objects-to-c-ffi-functions/18848 "2025-12-20T12:21:43Z")

</div>

Hi, i’m building a tvm ffi extension to integrate it with fastflow hpc parallel engine. I cannot understand how to pass Mutable objects from python to c++ Maps and Arrays are immutable by default… It would be nice to …

---

## [Did you forget to bind?](https://discuss.tvm.apache.org/t/did-you-forget-to-bind/17553)

<div class="topic-metadata">

**Author:** [@pedro](https://discuss.tvm.apache.org/u/pedro)\
**Replies:** 0\
**Last updated:** [August 5, 2024, 12:56am UTC](https://discuss.tvm.apache.org/t/did-you-forget-to-bind/17553 "2024-08-05T00:56:53Z")

</div>

I am trying to load one of my “whisper” models this way model = model.jit(spec=mod\_spec, device="cuda", out\_format="torch", debug=True) To this I get Traceback (most recent call last): File "/algo/users/ppp/GoodsDet…

---

## [bash: cd: /home/tvm/apps/uma: No such file or directory](https://discuss.tvm.apache.org/t/bash-cd-home-tvm-apps-uma-no-such-file-or-directory/18846)

<div class="topic-metadata">

**Author:** [@wangyangyang](https://discuss.tvm.apache.org/u/wangyangyang)\
**Replies:** 0\
**Last updated:** [December 15, 2025, 9:33am UTC](https://discuss.tvm.apache.org/t/bash-cd-home-tvm-apps-uma-no-such-file-or-directory/18846 "2025-12-15T09:33:46Z")

</div>

Could someone please explain what’s going on? Additionally, are there any tutorials for adapting hardware accelerators to TVM using UMA? I’m a beginner, thank you! Urgent!

---

## [Alternative for the bundle\_deploy](https://discuss.tvm.apache.org/t/alternative-for-the-bundle-deploy/18832)

<div class="topic-metadata">

**Author:** [@Xunflash](https://discuss.tvm.apache.org/u/Xunflash)\
**Replies:** 0\
**Last updated:** [December 11, 2025, 5:04am UTC](https://discuss.tvm.apache.org/t/alternative-for-the-bundle-deploy/18832 "2025-12-11T05:04:40Z")

</div>

After version 0.20.0, what is the recommended alternative for the bundle\_deploy functionality that was removed? I am referring to the code that was previously located here:

---

## [How to tile buffers to fit into on-chip memory?](https://discuss.tvm.apache.org/t/how-to-tile-buffers-to-fit-into-on-chip-memory/15119)

<div class="topic-metadata">

**Author:** [@Necrotos](https://discuss.tvm.apache.org/u/Necrotos)\
**Replies:** 4\
**Last updated:** [December 11, 2025, 5:03am UTC](https://discuss.tvm.apache.org/t/how-to-tile-buffers-to-fit-into-on-chip-memory/15119 "2025-12-11T05:03:43Z")

</div>

Hello, I am trying to understand how to use TVM for conv2d on a custom accelerator. For this I have created a toy example of a conv2d operation. So far I understand how to tile the computation, reorder loops and how to …

---

## [How to bring an ASIC Accelerator to TVM](https://discuss.tvm.apache.org/t/how-to-bring-an-asic-accelerator-to-tvm/18808)

<div class="topic-metadata">

**Author:** [@sexjun](https://discuss.tvm.apache.org/u/sexjun)\
**Replies:** 2\
**Last updated:** [December 9, 2025, 1:38am UTC](https://discuss.tvm.apache.org/t/how-to-bring-an-asic-accelerator-to-tvm/18808 "2025-12-09T01:38:33Z")

</div>

At present, the AI ​​compiler we wrote ourselves uses the front end of relax, and then we have re-implemented the optimization, memory allocation, scheduling and code generation ourselves. This is because we did not hav…

---

## [Does tvm Bring Tensorflow support to tvm.relax.frontend from v0.21.0](https://discuss.tvm.apache.org/t/does-tvm-bring-tensorflow-support-to-tvm-relax-frontend-from-v0-21-0/18804)

<div class="topic-metadata">

**Author:** [@MankalaSrujan](https://discuss.tvm.apache.org/u/MankalaSrujan)\
**Replies:** 0\
**Last updated:** [December 1, 2025, 2:16am UTC](https://discuss.tvm.apache.org/t/does-tvm-bring-tensorflow-support-to-tvm-relax-frontend-from-v0-21-0/18804 "2025-12-01T02:16:12Z")

</div>

As of TVM version v0.22.0 we don’t have support in tvm.relax.frontend to convert any tensorflow model to tvm relax ir module. Does TVM has any plans in future to add support for tensorflow model in tvm.relax.frontend?

---

## [Error with tune\_tir](https://discuss.tvm.apache.org/t/error-with-tune-tir/18801)

<div class="topic-metadata">

**Author:** [@a333222](https://discuss.tvm.apache.org/u/a333222)\
**Replies:** 0\
**Last updated:** [December 1, 2025, 2:15am UTC](https://discuss.tvm.apache.org/t/error-with-tune-tir/18801 "2025-12-01T02:15:46Z")

</div>

mod, params = relax.frontend.detach\_params(mod) mod = relax.get\_pipeline(“zero”)(mod) db = ms.tir\_integration.tune\_tir( mod=mod, target=target, work\_dir=“./tuning\_spr\_logs”, max\_trials\_global=8000, # SPR은 코어 많아서…

---

## [New releases on PyPI?](https://discuss.tvm.apache.org/t/new-releases-on-pypi/18448)

<div class="topic-metadata">

**Author:** [@mattpitkin](https://discuss.tvm.apache.org/u/mattpitkin)\
**Replies:** 3\
**Last updated:** [November 30, 2025, 12:04pm UTC](https://discuss.tvm.apache.org/t/new-releases-on-pypi/18448 "2025-11-30T12:04:34Z")

</div>

Hi, I wanted to ask whether new releases of apache-tvm will ever by published on PyPI? The last stable release was v0.12.0 back in June 2023, with dev releases up to 0.14dev273 in November 2023. The version in the GitHub…

---

## [MetaScheduleTuneTIR does not work well](https://discuss.tvm.apache.org/t/metascheduletunetir-does-not-work-well/18726)

<div class="topic-metadata">

**Author:** [@alen](https://discuss.tvm.apache.org/u/alen)\
**Replies:** 3\
**Last updated:** [November 29, 2025, 1:46pm UTC](https://discuss.tvm.apache.org/t/metascheduletunetir-does-not-work-well/18726 "2025-11-29T13:46:55Z")

</div>

code version: 0.21.0(latest release) torch version: 2.4.0+cu121 import tvm from tvm import IRModule, relax from tvm.relax.frontend import nn class RelaxModel(nn.Module): def \_\_init\_\_(self): super(RelaxMode…

---

## [TVM building issue on AArch64 architecture](https://discuss.tvm.apache.org/t/tvm-building-issue-on-aarch64-architecture/18779)

<div class="topic-metadata">

**Author:** [@andy](https://discuss.tvm.apache.org/u/andy)\
**Replies:** 2\
**Last updated:** [November 28, 2025, 6:47pm UTC](https://discuss.tvm.apache.org/t/tvm-building-issue-on-aarch64-architecture/18779 "2025-11-28T18:47:31Z")

</div>

Ubuntu22, AArch64. gcc version is 12.3.0 Could anyone help me? The error info when building: This warning is for project developers. Use -Wno-dev to suppress it. – Found CUDA\_TOOLKIT\_ROOT\_DIR=/usr/local/cuda-12.2 – F…

---

## [How to keep data in local buffer between matrix and vector ops (avoid extra global memory copies)?](https://discuss.tvm.apache.org/t/how-to-keep-data-in-local-buffer-between-matrix-and-vector-ops-avoid-extra-global-memory-copies/18797)

<div class="topic-metadata">

**Author:** [@Haleski](https://discuss.tvm.apache.org/u/Haleski)\
**Replies:** 0\
**Last updated:** [November 28, 2025, 1:10pm UTC](https://discuss.tvm.apache.org/t/how-to-keep-data-in-local-buffer-between-matrix-and-vector-ops-avoid-extra-global-memory-copies/18797 "2025-11-28T13:10:03Z")

</div>

Hi everyone, I’m working on a custom accelerator backend for TVM and ran into a scheduling / memory issue that I’m not sure how to model correctly. Hardware background On our hardware we have: A matrix compute unit A…

---

## [Quantization in Post Relay TVM](https://discuss.tvm.apache.org/t/quantization-in-post-relay-tvm/18770)

<div class="topic-metadata">

**Author:** [@pedrot123](https://discuss.tvm.apache.org/u/pedrot123)\
**Replies:** 0\
**Last updated:** [November 17, 2025, 8:34pm UTC](https://discuss.tvm.apache.org/t/quantization-in-post-relay-tvm/18770 "2025-11-17T20:34:47Z")

</div>

Hello, I was wondering if there are options for quantization in post Relay TVM. To be able to build support and research for custom accelerators, integer support through quantization is needed. In the old VTA flow, relay…

---

## [Advice on Learning TVM for Embedded Development](https://discuss.tvm.apache.org/t/advice-on-learning-tvm-for-embedded-development/18705)

<div class="topic-metadata">

**Author:** [@nielskoe](https://discuss.tvm.apache.org/u/nielskoe)\
**Replies:** 3\
**Last updated:** [November 7, 2025, 12:36am UTC](https://discuss.tvm.apache.org/t/advice-on-learning-tvm-for-embedded-development/18705 "2025-11-07T00:36:32Z")

</div>

Hi all, sorry to reach out to you with such a general question. I having been reading and watching things about TVM and I am very interested! Both conceptually as I think the underlying tech is really amazing, and for u…

---

## [Do we have plan to introduce step attribute to ForNode?](https://discuss.tvm.apache.org/t/do-we-have-plan-to-introduce-step-attribute-to-fornode/18685)

<div class="topic-metadata">

**Author:** [@LeiWang1999](https://discuss.tvm.apache.org/u/LeiWang1999)\
**Replies:** 3\
**Last updated:** [November 5, 2025, 6:58am UTC](https://discuss.tvm.apache.org/t/do-we-have-plan-to-introduce-step-attribute-to-fornode/18685 "2025-11-05T06:58:49Z")

</div>

In many CUDA kernels, the conventional pattern for thread iteration looks like this: for (int i = thread\_idx; i \< numel; i += num\_threads) out\[i\] = 0; However, in TileLang we currently have to write: for i in T.se…

---

## [CodeGen of Nimble](https://discuss.tvm.apache.org/t/codegen-of-nimble/9489)

<div class="topic-metadata">

**Author:** [@fengyuICT](https://discuss.tvm.apache.org/u/fengyuICT)\
**Replies:** 8\
**Last updated:** [November 4, 2025, 7:48am UTC](https://discuss.tvm.apache.org/t/codegen-of-nimble/9489 "2025-11-04T07:48:17Z")

</div>

I read the paper “NIMBLE: EFFICIENTLY COMPILING DYNAMIC NEURAL NETWORKS FOR MODEL INFERENCE”, and I am confused about section 3.5. Specifically, the technique “the residues modulo of the tiling factor” is not understood.…

---

## [Where is nimble code in tvm mainstream](https://discuss.tvm.apache.org/t/where-is-nimble-code-in-tvm-mainstream/18729)

<div class="topic-metadata">

**Author:** [@Stone1](https://discuss.tvm.apache.org/u/Stone1)\
**Replies:** 0\
**Last updated:** [November 4, 2025, 2:20am UTC](https://discuss.tvm.apache.org/t/where-is-nimble-code-in-tvm-mainstream/18729 "2025-11-04T02:20:35Z")

</div>

I tried a dynamic shape model compilation with tvm. However, it just complited in a few second, which confuse me. I had always expected it could have cost more time for tunning. After read the article “Nimble”， I w…

---

## [Tutorial Code does not work](https://discuss.tvm.apache.org/t/tutorial-code-does-not-work/18689)

<div class="topic-metadata">

**Author:** [@garfield0xff](https://discuss.tvm.apache.org/u/garfield0xff)\
**Replies:** 3\
**Last updated:** [November 3, 2025, 9:35am UTC](https://discuss.tvm.apache.org/t/tutorial-code-does-not-work/18689 "2025-11-03T09:35:58Z")

</div>

Hello, I’m exploring TVM to optimize and port AI models on embedded devices. While running the tutorial code in TVM version 0.22.dev0, I encountered an issue that I’d like some guidance on. When converting the ResNet m…

---

## [Export Executable](https://discuss.tvm.apache.org/t/export-executable/18710)

<div class="topic-metadata">

**Author:** [@confusedRoll](https://discuss.tvm.apache.org/u/confusedRoll)\
**Replies:** 2\
**Last updated:** [October 29, 2025, 4:28pm UTC](https://discuss.tvm.apache.org/t/export-executable/18710 "2025-10-29T16:28:45Z")

</div>

Is there a way to export relax or other imported pytorch/ONNX code into an executable file? I seem to be unable to find any documentation on the website for this. Thanks in advance

---

## [Meta\_schedule tuning for GPU does not work](https://discuss.tvm.apache.org/t/meta-schedule-tuning-for-gpu-does-not-work/18682)

<div class="topic-metadata">

**Author:** [@andy](https://discuss.tvm.apache.org/u/andy)\
**Replies:** 2\
**Last updated:** [October 29, 2025, 3:36pm UTC](https://discuss.tvm.apache.org/t/meta-schedule-tuning-for-gpu-does-not-work/18682 "2025-10-29T15:36:53Z")

</div>

I just build and install the latest TVM according to the instruction, Install from Source — tvm 0.22.dev0 documentation. (I did the \[echo “set(USE\_CUDA ON)” \>\> config.cmake\]) It works well for CPU kernel; but when I …

---

## [Has anyone attempted to compile the int8 model of yolov5s](https://discuss.tvm.apache.org/t/has-anyone-attempted-to-compile-the-int8-model-of-yolov5s/18715)

<div class="topic-metadata">

**Author:** [@zejialin](https://discuss.tvm.apache.org/u/zejialin)\
**Replies:** 0\
**Last updated:** [October 29, 2025, 3:36pm UTC](https://discuss.tvm.apache.org/t/has-anyone-attempted-to-compile-the-int8-model-of-yolov5s/18715 "2025-10-29T15:36:31Z")

</div>

I quantified the yolov5s model into the int8 tflite format and then compiled it via tvmc: The command I use: tvmc compile --target=ethos-u,cmsis-nn,c \\ --target-ethos-u-accelerator\_config=ethos-u55-256 \\ …

[Next page](https://discuss.tvm.apache.org/c/questions/7.md?page=1)
