Getting Started using Nix

Getting Started using Nix#

Install Nix (for the first time)#

If you don’t have nix installed yet, fetch the download link from this page.

nix is a package manager that allows us to configure the system beyond pythonic libraries.

Install environment using Nix#

  1. Clone the MASE repository:

git clone git@github.com:DeepWok/mase.git

  1. Create your own branch to work on:

cd mase
git checkout -b your_branch_name

  1. Activate the nix shell:

nix-shell

Tested Systems#

  • darwin aarch64 mac

  • linux x86_64 cuda-enabled

  • windows x86_64 cuda-enabled

Troubleshooting#

  1. Clang problem on darwin aarch64 systems

    There is some legacy issue with porting clang on nix, or generally nix shells and Python packages with C++ Extensions on macOS, you can find the issue detailed here.

    MacOS users should follow the standard procedure to install xcode , which provides you with Apple's clang/clang++ compilation tools . In our setup.py, several installation would use this local system clang, you should be able to verify this by typing

    [nix-shell:~/Projects/mase]$ which clang
# The following is the expected output
# /usr/bin/clang
[nix-shell:~/Projects/mase]$ clang --version
# The following is the expected output
# Apple clang version 15.0.0 (clang-1500.3.9.4)
# Target: arm64-apple-darwin23.5.0
# Thread model: posix
# InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin

    Using other clang variants, especially the llvm-backed nix clang will cause installation or running issues with cocotb and verilator because of the confusion in std library paths.

  2. g++ or glibc problem on WSL (Ubuntu-24.04 system)

    While using verilator in the nix shell, some of the generated files were compiled by g++ (or gcc), you might encounter issues about g++ can not find or glibc-xxx cannot find, etc.

    This is because the nix shell does not include any g++ or glibc in the environment. When compiling files, it automatically calls the related libraries in your local environment.

    When facing this kind of problem, you need to configure the local environment to match nix shell requirement. For instance, install or update the related packages to the verilator required version would work:

    (.venv) cx922@DESKTOP-UAFT8QR:~/mase$ ldd --version 
ldd (Ubuntu GLIBC 2.39-0ubuntu8.2) 2.39
...

  3. verilator installation and its cocotb integration

    We expect users to self-install verilator on their local system. The nix-shell will not install verilator for you. You can find the installation guide here.

    One common problem we found with using verilator backed cocotb is that the cocotb flow lacks resolution when your python is running in an advanced virtual environment. The cocotb flow makes mistakes in identifying certain paths. One error is something in the following form:

    - Verilator: Built from 0.000 MB sources in 0 modules, into 0.000 MB in 0 C++ files needing 0.000 MB
- Verilator: Walltime 0.003 s (elab=0.000, cvt=0.000, bld=0.000); cpu 0.000 s on 8 threads
INFO: Running command make -C /Users/name/Projects/mase/src/mase_components/activations/test/build/fixed_gelu/test_0 -f Vtop.mk in directory /Users/name/Projects/mase/src/mase_components/activations/test/build/fixed_gelu/test_0
make: *** No rule to make target `/usr/local/lib/python3.11/dist-packages/cocotb/share/lib/verilator/verilator.cpp', needed by `verilator.o'.  Stop.

    This error message is basically saying that the cocotb flow is looking for the verilator.cpp file in the wrong path. The verilator.cpp file is actually located in the cocotb package, which is installed in the dite-packages directory of your python environment. The cocotb flow is looking for the verilator.cpp file in the native directory, which is incorrect. One hack for this problem is to change the check in the cocotb Makefile.

    Edit the file /Users/yz10513/anaconda3/envs/mase/lib/python3.11/site-packages/cocotb/share/makefiles/Makefile.inc and change the following line:

    # Our comments: this is exactly the problem, sometimes we would like to use 	
# this ensures we use the same python as the one cocotb was installed into
ifeq ($(IS_VENV),True)
		# In a virtual environment, the Python binary may be a symlink, so it should not use realpath
		PYTHON_BIN ?= $(shell cocotb-config --python-bin)
else
		# disable the use of realpath!	
		# realpath to convert windows paths to unix paths, like cygpath -u
		#PYTHON_BIN ?= $(realpath $(shell cocotb-config --python-bin))
		PYTHON_BIN ?= $(shell cocotb-config --python-bin)
endif

  4. GPU-enabled torch install It is possible that the torch package is not installed correctly with GPU support in the nix-shell. You should have noticed that we have created a virtual environment in the nix-shell and installed the torch package in it. However, the torch package may not be installed with GPU support becuase of operating system compatibility issues. In these cases, you may choose to install the torch package yourself using pip

    # check your python
# it should point you to the python3 in the nix-shell
# ../mase/.venv/bin/python3
which python3
python3 -m pip install torch torchvision torchaudio

  5. CUDA_HOME problem with deepspeed on cuda enabled systems

    When installing deepspeed on cuda enabled systems, you might encounter the following error:

    	× python setup.py egg_info did not run successfully.
	│ exit code: 1
	╰─> [9 lines of output]
			Traceback (most recent call last):
				File "<string>", line 2, in <module>
				File "<pip-setuptools-caller>", line 34, in <module>
				File "/tmp/pip-install-lmczkuc5/deepspeed_2de5ecce4b1e495ea5546f4a526749f4/setup.py", line 101, in <module>
					cuda_major_ver, cuda_minor_ver = installed_cuda_version()
																					^^^^^^^^^^^^^^^^^^^^^^^^
				File "/tmp/pip-install-lmczkuc5/deepspeed_2de5ecce4b1e495ea5546f4a526749f4/op_builder/builder.py", line 50, in installed_cuda_version
					raise MissingCUDAException("CUDA_HOME does not exist, unable to compile CUDA op(s)")
			op_builder.builder.MissingCUDAException: CUDA_HOME does not exist, unable to compile CUDA op(s)
			[end of output]
	note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed
× Encountered error while generating package metadata.
╰─> See above for output.

    This normally means that the cuda toolkit is missing or not installed properly, which you can chek by running the following command:

    nvcc --version
which nvcc

    If there is an error, this is an indication to reinstall the cuda toolkit. You might need to run sudo apt-get install cuda-toolkit on Ubuntu systems. You may, in fact, need to install normal build tools for deepspeedd and pycuda too, these can be gcc, g++, make, cmake, etc. The tensorrt installation may also trigger an independent install if you are on wsl.

Contents