Mlc thor support Thor and Orin

[kbenkhaled]

Follow-up to #1315: Ensure build env vars are set correctly

Summary

This PR is a follow-up to #1315 and updates the environment-variable setup to ensure the build completes successfully.

Testing

• Thor: Build completes successfully.

• Orin: Not yet tested.

[johnnynunez]

LGTM (GH200)

[Copilot]

Pull Request Overview

This PR updates the environment variable setup for MLC and TVM packages to ensure build completion on Thor and Orin platforms. It introduces a new Thor-specific configuration and refactors the TVM build system to use external TVM dependencies.

• Refactors TVM package to support both pip installation and source builds with proper environment setup
• Updates MLC to use external TVM dependencies instead of bundled TVM for Thor platform
• Adds Thor-specific MLC configuration with SM110 CUDA architecture support

Reviewed Changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
packages/ml/tvm/config.py	Introduces TVM package configuration function with commit/version support
packages/ml/tvm/install.sh	New script for attempting pip install before falling back to source build
packages/ml/tvm/build.sh	New source build script with CUDA architecture configuration
packages/ml/tvm/Dockerfile	Refactored to use new install/build scripts and proper environment setup
packages/ml/tvm/test.py	Added success confirmation message
packages/llm/mlc/config.py	Added Thor-specific configuration and TVM dependency
packages/llm/mlc/build.sh	Updated to use external TVM and simplified build process
packages/llm/mlc/test.sh	Fixed module invocation and benchmark path
packages/llm/mlc/benchmark.py	Improved API compatibility detection
packages/llm/mlc/Dockerfile	Updated environment variables for external TVM
packages/llm/mlc/patches/empty.diff	Empty patch file for Thor configuration
packages/cuda/cuda-python/utils.py	Added compatibility for cuda-python 13.x layout
packages/cuda/cuda-python/test_runtime.py	Updated version detection and import handling
packages/cuda/cuda-python/test_driver.py	Fixed context management and improved compatibility

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

[nitpick] Remove the trailing blank lines at the end of the file to improve code cleanliness.

Suggested change

[Copilot]

[nitpick] This file contains only empty lines. Consider using a truly empty file or adding a comment explaining why this empty patch is needed for the Thor configuration.

Suggested change

	# This empty patch file is required for Thor configuration. No changes are needed.

[OriNachum]

@johnnynunez did you try running on Orin? If so, what configurations?

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[kbenkhaled]

@OriNachum I ran the build on Orin with this command
LSB_RELEASE=24.04 ./jetson-containers build mlc:thor
build fails in the stage 27/28 in the flashinfer stage
I will be checking it out right now will also update the version so we won't have to add :thor

[johnnynunez]

@OriNachum I ran the build on Orin with this command LSB_RELEASE=24.04 ./jetson-containers build mlc:thor build fails in the stage 27/28 in the flashinfer stage I will be checking it out right now will also update the version so we won't have to add :thor

Flashinfer is not compatible > 0.2.9 with cuda 12.6.
On last release they starting to support lower cuda versions again
flashinfer-ai/flashinfer@d6c3e33

[johnnynunez]

@kbenkhaled i just upload new flashinfer version

[johnnynunez]

@kbenkhaled i added recently cuda 13 support in tvm. It is working with cutlass 4.2. Also i fixed flash attention builds for cuda 13. This PR should work now @OriNachum

[kbenkhaled]

@johnnynunez Thank you, I opened this PR because it was not working on orin, but I think it should work now as per my previous fix from #1315

[johnnynunez]

@johnnynunez Thank you, I opened this PR because it was not working on orin, but I think it should work now as per my previous fix from #1315

Yes, i will try it. It can be cleaned with new flash infer 3.1