I needed to build the zero variant of the HotSpot JVM for the Windows platform recently. libffi is one of the prerequisites for the zero variant. It provides “a portable, high level programming interface to various calling conventions.” I decided to build libffi/libffi at v3.4.8 since it looks like the latest version. I used a Windows x64 machine for this entire process. Visual C++ and MSYS need to be installed to do this. Launch MSYS2 and get the sources from GitHub:
Install automake and libtool using these commands:
pacman -S automake
pacman -S libtool
The Visual C++ compiler needs to be available in the path as well. Run cl without any parameters to ensure the compiler is available. It most likely won’t be by default. If it isn’t, add it to the path as follows:
Note that the name of the Visual C++ linker is link.exe, which clashes with the built in “link” command. Prepending the C++ compiler path means that the built-in link command will not be available. Appending the C++ compiler path means that the linker cannot be invoked without specifying its full path.
Generating the configure file
With the MSYS prerequisites installed, run the autogen.sh script:
user@machine /d/repos/libffi
$ ./autogen.sh
This creates a configure script in the root of the repository. Run it using bash:
Running configure takes about a minute and a half on my 24-core (32 logical processor) machine with 128GB RAM.
Building the Source Code
Simply run make in the root of the repo. The generated LIB and DLL files should be in the x86_64-w64-mingw32/.libs/ subdirectory of the repo root. There will also be ffi.h and ffitarget.h include files in the x86_64-w64-mingw32/include/ subdirectory of the repo root. These 4 files are typically what will be required by other projects with a libffi dependency (like OpenJDK).
$ ls -1 x86_64-w64-mingw32/.libs/
libffi.la
libffi.lai
libffi_convenience.la
libffi_convenience.lib
libffi-8.dll*
libffi-8.exp
libffi-8.lib
$ ls -1 x86_64-w64-mingw32/include/
ffi.h
ffitarget.h
Makefile
My Motivation for Building libffi
I was trying to configure an OpenJDK build (at commit c3de94cee12471) using this command line:
...
checking if hsdis should be bundled... no
checking for --enable-libffi-bundling... disabled, default
checking for LIBFFI... checking for ffi.h... no
configure: error: Could not find libffi!
configure exiting with result code 1
...
checking for --enable-libffi-bundling... disabled, default
checking if libffi works... no
configure: error: Found libffi but could not link and compile with it.
configure exiting with result code 1
This was my hint that I probably need to build libffi first. libffi/README.md at v3.4.8 · libffi/libffi explains that the configure script can be generated by running autogen.sh. I first need to fix the line endings. This copilot prompt “convert all existing files in a repo from windows to unix line endings” gets me the solution:
# Tells Git to convert CRLF to LF on commit
# but not the other way around on checkout.
git config core.autocrlf input
# resets the working directory and re-checks
# out the files using the current core.autocrlf setting
git reset --hard
Now autogen.sh can be executed. I didn’t read the instructions all the way through to see what prerequisites are required. Even so, which ones can I get away without?
user@machine /cygdrive/d/repos/libffi
$ ./autogen.sh
autoreconf-2.71: export WARNINGS=
autoreconf-2.71: Entering directory '.'
autoreconf-2.71: configure.ac: not using Gettext
autoreconf-2.71: running: aclocal -I m4
Can't exec "aclocal": No such file or directory at /usr/share/autoconf2.7/Autom4te/FileUtils.pm line 274.
autoreconf-2.71: error: aclocal failed with exit status: 1
configure: loading site script /etc/config.site
checking build system type... x86_64-w64-mingw32
checking host system type... x86_64-w64-mingw32
checking target system type... x86_64-w64-mingw32
continue configure in default builddir "./x86_64-w64-mingw32"
....exec /bin/sh ../configure "--srcdir=.." "--enable-builddir=x86_64-w64-mingw32" "mingw32"
configure: loading site script /etc/config.site
checking build system type... x86_64-w64-mingw32
checking host system type... x86_64-w64-mingw32
checking target system type... x86_64-w64-mingw32
checking for gsed... sed
checking for a BSD-compatible install... /usr/bin/install -c
checking whether sleep supports fractional seconds... yes
checking filesystem timestamp resolution... 0.01
checking whether build environment is sane... yes
checking for a race-free mkdir -p... /usr/bin/mkdir -p
checking for gawk... gawk
checking whether make sets $(MAKE)... yes
checking whether make supports nested variables... yes
checking xargs -n works... yes
checking for gcc... /d/repos/libffi/msvcc.sh -m64
checking whether the C compiler works... no
configure: error: in '/d/repos/libffi/x86_64-w64-mingw32':
configure: error: C compiler cannot create executables
See 'config.log' for more details
In Cygwin, that command failed with “configure: error: cannot run /bin/sh ./config.sub“. What could be going wrong in the configure script? M365 Copilot prompt: “change build system type in msys2” refers to gcc – Configuration x86_64-pc-msys not supported – Stack Overflow but those flags seem unnecessary given my platform. I tried removing some of the compiler setting flags to no avail:
$ time bash configure CPPFLAGS="-DFFI_BUILDING_DLL" CPP="cl -nologo -EP" CXXCPP="cl -nologo -EP" --disable-docs --prefix=/d/temp/libffi
configure: loading site script /etc/config.site
checking build system type... x86_64-w64-mingw32
checking host system type... x86_64-w64-mingw32
checking target system type... x86_64-w64-mingw32
continue configure in default builddir "./x86_64-w64-mingw32"
....exec /bin/sh ../configure "--srcdir=.." "--enable-builddir=x86_64-w64-mingw32" "mingw32"
configure: loading site script /etc/config.site
checking build system type... x86_64-w64-mingw32
checking host system type... x86_64-w64-mingw32
checking target system type... x86_64-w64-mingw32
checking for gsed... sed
checking for a BSD-compatible install... /usr/bin/install -c
checking whether sleep supports fractional seconds... yes
checking filesystem timestamp resolution... 0.01
checking whether build environment is sane... yes
checking for a race-free mkdir -p... /usr/bin/mkdir -p
checking for gawk... gawk
checking whether make sets $(MAKE)... yes
checking whether make supports nested variables... yes
checking xargs -n works... yes
checking for gcc... no
checking for cc... no
checking for cl.exe... no
checking for clang... no
configure: error: in '/d/repos/libffi/x86_64-w64-mingw32':
configure: error: no acceptable C compiler found in $PATH
See 'config.log' for more details
The config.log file is in the x86_64-w64-mingw32 folder in the repo root. What I should have verified is that I could run cl.exe in MSYS before trying any of this stuff. That was the primary reason for launch ucrt64.exe from a developer command prompt. Unfortunately, that didn’t work for whatever reason.
user@machine UCRT64 /d/repos/libffi
$ cl
-bash: cl: command not found
user@machine UCRT64 /d/repos/libffi
$ echo $PATH
/ucrt64/bin:/usr/local/bin:/usr/bin:/bin:/c/Windows/System32:/c/Windows:/c/Windows/System32/Wbem:/c/Windows/System32/WindowsPowerShell/v1.0/:/usr/bin/site_perl:/usr/bin/vendor_perl:/usr/bin/core_perl
I tried manually fixing the path as follows but this didn’t work (cl.exe could still not be found):
The dir command can show the short name equivalents of a file name, e.g. dir /x C:\Program Files.
dir /x C:\
...
05/24/2025 11:42 AM <DIR> PROGRA~1 Program Files
04/09/2025 01:31 AM <DIR> PROGRA~2 Program Files (x86)
...
dir /x "C:\Program Files"
11/30/2023 04:40 PM <DIR> MIB055~1 Microsoft Visual Studio
Sure enough, I could now find cl.exe and the configure script worked!
$ where cl.exe
C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Tools\MSVC\14.43.34808\bin\Hostx64\x64\cl.exe
$ bash configure \
CC="/d/repos/libffi/msvcc.sh -m64" \
CXX="/d/repos/libffi/msvcc.sh -m64" \
CPPFLAGS="-DFFI_BUILDING_DLL" \
LD=link CPP="cl -nologo -EP" CXXCPP="cl -nologo -EP" \
--disable-docs \
--prefix=/d/temp/libffi
configure: loading site script /etc/config.site
checking build system type... x86_64-w64-mingw32
checking host system type... x86_64-w64-mingw32
checking target system type... x86_64-w64-mingw32
continue configure in default builddir "./x86_64-w64-mingw32"
....exec /bin/sh ../configure "--srcdir=.." "--enable-builddir=x86_64-w64-mingw32" "mingw32"
configure: loading site script /etc/config.site
checking build system type... x86_64-w64-mingw32
checking host system type... x86_64-w64-mingw32
checking target system type... x86_64-w64-mingw32
checking for gsed... sed
checking for a BSD-compatible install... /usr/bin/install -c
checking whether sleep supports fractional seconds... yes
checking filesystem timestamp resolution... 0.01
checking whether build environment is sane... yes
checking for a race-free mkdir -p... /usr/bin/mkdir -p
checking for gawk... gawk
checking whether make sets $(MAKE)... yes
checking whether make supports nested variables... yes
checking xargs -n works... yes
checking for gcc... /d/repos/libffi/msvcc.sh -m64
checking whether the C compiler works... yes
checking for C compiler default output file name... conftest.exe
checking for suffix of executables... .exe
checking whether we are cross compiling... no
checking for suffix of object files... obj
checking whether the compiler supports GNU C... no
checking whether /d/repos/libffi/msvcc.sh -m64 accepts -g... yes
checking for /d/repos/libffi/msvcc.sh -m64 option to enable C11 features... unsupported
checking for /d/repos/libffi/msvcc.sh -m64 option to enable C99 features... unsupported
checking for /d/repos/libffi/msvcc.sh -m64 option to enable C89 features... unsupported
checking whether /d/repos/libffi/msvcc.sh -m64 understands -c and -o together... yes
checking whether make supports the include directive... yes (GNU style)
checking dependency style of /d/repos/libffi/msvcc.sh -m64... none
checking whether the compiler supports GNU C++... no
checking whether /d/repos/libffi/msvcc.sh -m64 accepts -g... yes
checking for /d/repos/libffi/msvcc.sh -m64 option to enable C++11 features... unsupported
checking for /d/repos/libffi/msvcc.sh -m64 option to enable C++98 features... unsupported
checking dependency style of /d/repos/libffi/msvcc.sh -m64... none
checking dependency style of /d/repos/libffi/msvcc.sh -m64... none
checking for grep that handles long lines and -e... /usr/bin/grep
checking for egrep... /usr/bin/grep -E
checking how to print strings... printf
checking for a sed that does not truncate output... /usr/bin/sed
checking for fgrep... /usr/bin/grep -F
checking for non-GNU ld... link
checking if the linker (link) is GNU ld... no
checking for BSD- or MS-compatible name lister (nm)... /usr/bin/nm -B
checking the name lister (/usr/bin/nm -B) interface... BSD nm
checking whether ln -s works... no, using cp -pR
checking the maximum length of command line arguments... 8192
checking how to convert x86_64-w64-mingw32 file names to x86_64-w64-mingw32 format... func_convert_file_msys_to_w32
checking how to convert x86_64-w64-mingw32 file names to toolchain format... func_convert_file_msys_to_w32
checking for link option to reload object files... -r
checking for file... file
checking for objdump... objdump
checking how to recognize dependent libraries... file_magic ^x86 archive import|^x86 DLL
checking for dlltool... dlltool
checking how to associate runtime and link libraries... func_cygming_dll_for_implib
checking for ranlib... ranlib
checking for ar... ar
checking for archiver @FILE support... @
checking for strip... strip
checking command to parse /usr/bin/nm -B output from /d/repos/libffi/msvcc.sh -m64 object... ok
checking for sysroot... no
checking for a working dd... /usr/bin/dd
checking how to truncate binary pipes... /usr/bin/dd bs=4096 count=1
checking for mt... no
checking if : is a manifest tool... no
checking for stdio.h... yes
checking for stdlib.h... yes
checking for string.h... yes
checking for inttypes.h... yes
checking for stdint.h... yes
checking for strings.h... no
checking for sys/stat.h... yes
checking for sys/types.h... yes
checking for unistd.h... no
checking for dlfcn.h... no
checking for objdir... .libs
checking for /d/repos/libffi/msvcc.sh -m64 option to produce PIC... -DDLL_EXPORT -DPIC
checking if /d/repos/libffi/msvcc.sh -m64 PIC flag -DDLL_EXPORT -DPIC works... yes
checking if /d/repos/libffi/msvcc.sh -m64 static flag works... yes
checking if /d/repos/libffi/msvcc.sh -m64 supports -c -o file.obj... yes
checking if /d/repos/libffi/msvcc.sh -m64 supports -c -o file.obj... (cached) yes
checking whether the /d/repos/libffi/msvcc.sh -m64 linker (link) supports shared libraries... yes
checking dynamic linker characteristics... Win32 ld.exe
checking how to hardcode library paths into programs... immediate
checking whether stripping libraries is possible... yes
checking if libtool supports shared libraries... yes
checking whether to build shared libraries... yes
checking whether to build static libraries... yes
checking how to run the C++ preprocessor... cl -nologo -EP
checking whether the /d/repos/libffi/msvcc.sh -m64 linker (link) supports shared libraries... no
checking for /d/repos/libffi/msvcc.sh -m64 option to produce PIC... -DDLL_EXPORT -DPIC
checking if /d/repos/libffi/msvcc.sh -m64 PIC flag -DDLL_EXPORT -DPIC works... yes
checking if /d/repos/libffi/msvcc.sh -m64 static flag works... yes
checking if /d/repos/libffi/msvcc.sh -m64 supports -c -o file.obj... yes
checking if /d/repos/libffi/msvcc.sh -m64 supports -c -o file.obj... (cached) yes
checking whether the /d/repos/libffi/msvcc.sh -m64 linker (link) supports shared libraries... no
checking dynamic linker characteristics... Win32 ld.exe
checking how to hardcode library paths into programs... immediate
checking for readelf... readelf
checking size of size_t... 8
checking for C compiler vendor... microsoft
checking whether C compiler accepts -O2... yes
checking CFLAGS for most reasonable warnings...
checking whether to enable maintainer-specific portions of Makefiles... no
checking for sys/memfd.h... no
checking for memfd_create... no
checking for egrep... (cached) /usr/bin/grep -E
checking for memcpy... no
checking for alloca.h... no
checking size of double... 8
checking size of long double... 8
checking whether byte ordering is bigendian... no
checking assembler .cfi pseudo-op support... no
checking assembler supports pc related relocs... yes
checking whether compiler supports pointer authentication... no
checking for _ prefix in compiled symbols... no
configure: versioning on shared library symbols is no
checking that generated files are newer than configure... done
configure: creating ./config.status
config.status: creating include/Makefile
config.status: creating include/ffi.h
config.status: creating Makefile
config.status: creating testsuite/Makefile
config.status: creating man/Makefile
config.status: creating doc/Makefile
config.status: creating libffi.pc
config.status: creating fficonfig.h
config.status: executing buildir commands
config.status: create top_srcdir/Makefile guessed from local Makefile
config.status: build in x86_64-w64-mingw32 (HOST=)
config.status: executing depfiles commands
config.status: executing libtool commands
config.status: executing include commands
config.status: executing src commands
I could now run make as instructed by the readme. Here is the tail of the resulting output:
I manually copied these files to set up the libffi repo for building OpenJDK (the expected LIB filename does not have the -8 suffix by default). I’m guessing make install or something like that is the proper way to do this but I had what I needed so this was good enough for me.
I tested this file to ensure I could compile it. The linking step failed (cl.exe without the -c option).
$ cl -c -I include ffi_test.c
Microsoft (R) C/C++ Optimizing Compiler Version 19.43.34810 for x64
Copyright (C) Microsoft Corporation. All rights reserved.
ffi_test.c
$ cl -I include ffi_test.c
Microsoft (R) C/C++ Optimizing Compiler Version 19.43.34810 for x64
Copyright (C) Microsoft Corporation. All rights reserved.
ffi_test.c
Microsoft (R) Incremental Linker Version 14.43.34810.0
Copyright (C) Microsoft Corporation. All rights reserved.
/out:ffi_test.exe
ffi_test.obj
ffi_test.obj : error LNK2019: unresolved external symbol __imp_ffi_call referenced in function main
ffi_test.exe : fatal error LNK1120: 1 unresolved externals
I tried manually running link.exe but this failed because the wrong link.exe is called.
$ where link.exe
C:\software\msys64\usr\bin\link.exe
C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Tools\MSVC\14.43.34808\bin\Hostx64\x64\link.exe
Prepending the compiler path to $PATH resolved this.
$ cl -I include ffi_test.c -link -libpath:lib libffi.lib
Microsoft (R) C/C++ Optimizing Compiler Version 19.43.34810 for x64
Copyright (C) Microsoft Corporation. All rights reserved.
ffi_test.c
Microsoft (R) Incremental Linker Version 14.43.34810.0
Copyright (C) Microsoft Corporation. All rights reserved.
/out:ffi_test.exe
-libpath:lib
libffi.lib
ffi_test.obj
$ ./ffi_test.exe
At this point, things were in good enough shape to build OpenJDK. However, I could not successfully run bash configure ... in Cygwin (to build OpenJDK) now. Perhaps it’s because I had been mucking around with the Cygwin setup. I tried removing automake and libtool but that didn’t fix the problem.
The “Analysis Board” command looks like the place to enter it. Click on the “Edit Position” button.
En Croissant Analysis Board Window
A FEN section appears with a textbox for your FEN. You can click on the EMPTY button to clear the board.
En Croissant Edit Position Window
I pasted my FEN into the FEN position text box then tabbed out of it and it just reverts to the empty board. Very unintuitive behavior – I accidentally discover that I need to press ENTER in that text box to accept my new position. An error appears about this being an invalid position. It would have been nice to have this validation happen instead of deleting my entry on TAB.
Invalid board
This is the corrected FEN:
KR1r4/n7/k7/8/8/8/8/8 w - - 0 1
This leaves me confident that I get the computer to help me analyze this position.
I select “C:\software\stockfish-windows-x86-64-avx2\stockfish\stockfish-windows-x86-64-avx2.exe” as per the wiki.
Add EngineStockfish 17.1 Engine Added
Now that I’ve set the engine to play, why aren’t the arrows labeled? I actually switched to Arena (see below) before returning with a renewed determination to get this to work. On the “Play from Here” screen, notice that the Engine button now displays Stockfish 17.1. Perhaps it’s the UI that’s confusing, looks like a bunch of controls just dumped into a panel. Switch both controls to Engine and adjust the time as desired. I’m using a 3s limit to keep things moving. Finally, click on “Start Game” and watch the engine battle itself.
Here are a few other FEN positions (from various Facebook posts) to play with:
KR1r4/n7/k/8/8/8/8/8 w - - 0 0
kr6/1r1N4/2Q5/8/8/8/8/K7 w - - 0 0
8/7p/1k4p1/4K3/6P1/8/8/8 w - - 0 0
k7/b7/2K5/4B3/8/8/8/Q7 w - - 0 0
Arena (3.5.1)
I had actually installed Arena before En Croissant but didn’t like that it wasn’t open source. I initially struggled getting the engine to play the game in En Croissant and went back to Arena.
Arena 3.5.1 Setup Wizard
I was pleasantly surprised that I could just load the FEN by going to Position > Load FEN from clipboard then have the computer play by going to Game > Move Now.
However, I didn’t like how long it took to make a move. After some exploration, I found that I could go to Levels > Adjust and change from Tournament mode to Time per move mode. It was ironic that I was using Arena without Stockfish for my analysis but still wanted to document my exploration of En Croissant. Perhaps it was this experience with the time setting in Arena that allowed me to adjust the time settings for the engine in En Croissant and actually see the Stockfish engine in action!
Improvements?
I wonder if I could add this Load FEN from clipboard option to En Croissant since it is open source. Such a command should automatically open the Analysis Board window. As I’m wrapping up this post, I just noticed the error below when I switched back to the En Croissant window.
Minified React Error
Making that UI change could be a good entry for me into the React world.
Cellular networks are a ubiquitous part of modern daily life. The history of cellular technology is definitely worth knowing, even if only at the high level presented in this video on the Evolution of Mobile Standards (1G, 2G, 3G, 4G, and 5G).
Evolution of Mobile Standards [1G, 2G, 3G, 4G, 5G]
2G introduced digital modulation and came in variants like TDMA, CDMA, and GSM. 3G uses spread spectrum in the radio portion of the network whereas 4G uses Orthogonal Frequency Division Mutiplexing (OFDM). 4G also separates the user and control planes whereas they were on the same hardware in 3G (and therefore couldn’t scale independently). The following videos from Sunny Classroom are brief but helpful explanations of these communications concepts.
FHSS – Frequency Hopping Spread Spectrum
DSSS – Direct Sequence Spread Spectrum
OFDM – Orthogonal Frequency Division Multiplexing
5G offers lower end-to-end latency and higher uplink and downlink throughput than 4G because it has more bands (low, mid, and high) vs just low and high with 4G. It is also a programmable network, which lets developers access network stats via APIs.
Mobility
The discussion in the class proceeded to mobility, introducing the concept of cellular handoff, which can broadly be classified into mobile assisted and mobile controlled handover. See Handoff in Wireless Mobile Networks for more details. Another classification of types of handover is based on when the UE disconnects from one cell: hard handoff vs soft handoff. Soft handoff ensures that calls are not dropped. The professor was drawing hexagonal cells when illustrating these and I realized I had no idea why they are hexagonal. Here’s why:
antennas in a coverage area are in a hexagonal pattern… because it requires fewer cells to represent a hexagon compared to triangle or square – meaning network carriers can cover a wider area with less base stations. – The Fundamentals of Cellular System Design
An interesting aspect of the LTE packet core is that a 5G base station can be attached to it. Contrast this mode, known as 5G non-standalone to 5G standalone mode, where a 5G radio is connected to a 5G packet core. See this post on Non-standalone and Standalone: two standards-based paths to 5G for a detailed review of these modes. One advantage of the 5G packet core is that it allows for cloud-based implementations. The Boost Mobile Network, for example, is 100% implemented in the cloud. Is AWS set to flex cloud on telecom? has a discussion of such a transition (to the cloud) in the telecom space.
The antennas used on the base stations can be of multiple types, e.g. SISO and MIMO. MIMO Antennas Explained: An In-Depth Guide provides more details on the differences between these designs. A key benefit of MIMO is that it eliminates performance degradation caused by multipath wave propagation.
We also dug into the LTE and 5G network evolution, from network deployment, to network growth, then finally coverage and capacity optimization. Deployment may involve dual-radio in UEs and EPC capabilities to support interoperability with earlier generations like 2G/3G. 5G is currently in the deployment phase since 5G SA has not yet been fully rolled out. Network growth may involve cell splitting in the RAN for capacity as well as expansion of the core network. Coverage and capacity optimization may involve spectrum aggregation, advanced network topologies, and advanced antenna techniques.
The continued growth in application and device diversity, RAN complexity, and QoS variance is making networks more complex and thus harder to optimize under the current network management paradigm. Self-organizing networks (SON) were designed to address this problem. Here is an overview of SON.
3GPP SON Series: An Introduction to Self-Organizing Networks (SON)
SON is used to set many required configuration parameters when introducing a new eNB or gNB to a network e.g. IP addresses from DHCP, transmit power, beam width, supported connections, connecting to neighboring base stations via the X2 (4G) or XN (5G) connection, etc. SON can also be used for driving energy savings by shutting down carriers when less capacity is required e.g. in the middle of the night (without dropping emergency calls). Another application is coverage and capacity optimization, which involves adjusting transmission power and continuously adjusting antenna tilt to increase capacity (thus decreasing coverage) or to increase coverage (thus decreasing capacity). Mobility handover optimization is also required to avoid too early/too late handover or a ping pong between base stations. The SON architecture can be centralized, distributed, or hybrid.
5G
Finally, we took a look at 5G technology, which has much lower latencies, much higher throughput, and high capacity. Some of the key technologies I learn about include Fixed Wireless Access, in which the UE does not move, and mmWave. More info on the latter is available at What is mmWave and how does it relate to 5G? 5G also supports modified air interfaces (modified OFDM), massive MIMO, device-to-device communication, separated user and control planes, and network virtualization.
An important capability that 5G introduced is positioning, which has many potential use cases e.g. industrial, automotive, and AR/VR. See 5G positioning: What you need to know for more details. In the industrial setting, for example, 5g all in one boxes are deployed in the 5G private networks. They have a base station and a packet core in a single piece of hardware, e.g. RAK All-in-One 5G box (the first one in the search results).
The 5G core network architecture is significantly different from the LTE packet core (eNB, SGW, PGW, MME, HSS, and PCRF). It moved to a service based architecture where microservices expose functionality via APIs. This makes the 5G network programmable and extensible. This 5G System Overview covers the overall 5G architecture. These are a few of the 5G components:
Network Slice Selection Function (NSSF), which allocates slices to users
Authentication Server Function (AUSF)
Policy Control Function (PCF)
Unified Data Management (UDM), which is functionally similar to 3G and 4G’s HSS
Application Function (AF), which can let applications retrieve data like latencies
Summary
This is the final post in the Introduction to Networks series of posts. It has been an extremely enlightening course. I have appreciated how much more extensive it was than I expected from an introductory course as I try to stay on top of the fast moving tech space.
The 8353009: Improve documentation for Windows AArch64 builds PR has a comment stating that “the BuildTools distribution of Visual Studio do not include aarch64-hosted compilers, so to be able to run native builds without the Prism emulation, you need to install the full Visual Studio, including the IDE.” This post describes how I determined this to be false.
Running bash configure --with-boot-jdk=<PATH> failed with the error that it could not find a C compiler:
...
checking for cacerts file... default
checking for cacerts source... default
checking for --enable-unlimited-crypto... enabled, default
checking for jni library path... default
configure: Using default toolchain microsoft (Microsoft Visual Studio)
configure: Found Visual Studio installation at /cygdrive/c/progra~2/micros~2/2022/BuildTools using well-known name
configure: Found Microsoft Visual Studio 2022
configure: Trying to extract Visual Studio environment variables for aarch64
configure: using /cygdrive/c/progra~2/micros~2/2022/BuildTools/vc/auxiliary/build/vcvarsarm64.bat
configure: Setting extracted environment variables for aarch64
checking that Visual Studio variables have been correctly extracted... ok
checking for cl... [not found]
configure: error: Could not find a C compiler.
configure exiting with result code 1
The TOOLCHAIN_FIND_COMPILER macro calls the UTIL_LOOKUP_TOOLCHAIN_PROGS macro to find the C compiler. I verified that the last argument is “cl” with an AC_MSG_NOTICE. At this point, I compared the TOOLCHAIN_PATH in config.log with that on a different ARM64 machine with a full VS install. Sure enough, it didn’t contain the bin/hostarm64/arm64 path with the buildtools setup, even though the path exists on disk. TOOLCHAIN_PATH is coming from VS_PATH in toolchain_microsoft.m4. Here is the build\windows-aarch64-server-slowdebug\configure-support\vs-env-aarch64\set-vs-env.sh file.
Notice that VS_PATH only has what VS_ENV_CMD added to the PATH! This was a clue that I need to take another step back – I realized that I couldn’t even run cl.exe in the developer command prompt! Then again, the command line for the terminal is:
Changing the host architecture to arm64 did not help. I launched the VS installer and noticed that the “Desktop development with C++” workload was not installed so I must have been missing additional components.
Visual Studio Build Tools 2022 LTSC 17.12 Workloads
I didn’t want to install the whole workload though, just the necessary individual components. I noticed the C++ Build Tools core features component wasn’t installed so I selected it. The Windows Universal C Runtime component is automatically selected as well:
Visual Studio Build Tools 2022 LTSC 17.12 Individual Components
Once the installation completed, I could run cl.exe in the developer command prompt!
**********************************************************************
** Visual Studio 2022 Developer Command Prompt v17.12.7
** Copyright (c) 2022 Microsoft Corporation
**********************************************************************
C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools>cl
Microsoft (R) C/C++ Optimizing Compiler Version 19.42.34441 for ARM64
Copyright (C) Microsoft Corporation. All rights reserved.
usage: cl [ option... ] filename... [ /link linkoption... ]
C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools>
The VS installer log in %TEMP% contained these components:
Copilot informed me that the caret was the way to split a command across multiple lines in the Windows Command Prompt. This was the final command I used to complete the 8353009: Improve documentation for Windows AArch64 builds PR.
While learning about large language models, the issue of vanishing gradients came up. What is a gradient? I attempted to describe it as the difference between where you are and where you want to be (the target), which is given by the loss function. This led to the question of what exactly a loss function is. The video below from IBM Technology explains loss functions: a loss function is an evaluation metric (how well is the model performing) and/or a guide that directs the model’s learning process.
What is a Loss Function? Understanding How AI Models Learn
The primary reason for calculating the loss function is to guide the model’s learning process. It provides a numeric value that indicates how far off the model’s predictions are from the actual results. By analyzing the loss, the model’s parameters can be adjusted (optimization) since the loss function is a feedback mechanism to the model, telling it how well it is performing and where it needs to improve. – What is a Loss Function? Understanding How AI Models Learn
A smaller value of the loss function indicates that the performance of the model has improved.
The gradient of the loss function is useful because it enables algorithms to determine which adjustments (e.g. to weights) will result in a smaller loss. The next video on Gradient descent, how neural networks learn is a helpful introduction to how loss functions are used to guide learning.
Gradient descent, how neural networks learn | DL2
Backpropagation is the algorithm used to compute the gradient. This video from 3Blue1Brown is a helpful explanation of what backpropagation is:
Backpropagation, intuitively | DL3
Two important phenomena in gradient descent are the problems of vanishing and exploding gradients. The Vanishing & Exploding Gradient explained | A problem resulting from backpropagationvideo describes these problems as follows: vanishing gradients mean that updated weights earlier in the network barely change (stuck) which means that the rest of the network cannot really minimize the loss function (i.e. learn). Exploding gradients mean that the earlier weights now increase so much that the optimal value of the loss function will never be achieved because weights become too big too quickly.
Vanishing & Exploding Gradient explained | A problem resulting from backpropagation
The focus of part 3 of this series is on the different types of wired networks. A key aspet of many networks is that QoS is just one concern, another key issue being how to meet guarantees for delivery of voice services. How is voice delivered? Integrated Services Digital Network (ISDN) is an international standard for voice, video, and data transmission over digital telephone or copper lines. It has two service levels. The first is Basic Rate Interface (BRI), which supports 2 bearer channels at 64kbps each and 1 D channel at 16 kbps. The second is the Primary Rate Interface (PRI), which supports 23 bearer channels (in the US) at 64 kbps each and 1 D channel at 16 kbps. The signaling/data (D) channel runs the ISDN signaling protocol based on Q.931. This video is a good high level introduction of ISDN. T1 and ISDN are used in access networks, together with technologies like IP and MPLS.
ISDN – Integrated Services Digital Network
Optical Networks
In newer generations of networks, the core is fiber (instead of copper) because it can deliver terabits per second. Installation and management of fiber networks is also much easier than copper networks. Fiber optic signals are analog – (in the infrared range).
What is the ELECTROMAGNETIC SPECTRUM
Light sources used for fiber optic communication include light-emitting diodes (LEDs), laser diodes, vertical cavity surface emitting lasers (VCSELs), Fabry-Perot lasers, and distributed feedback lasers.
How LED Works – Unravel the Mysteries of How LEDs Work!
The packet transport network is another key piece to understand. Customers send traffic to metro access, aggregation, and core portions of the network where voice and data are converged. In the packet core, wavelengths are being added and dropped by add-drop multiplexers. There are several types of ADMs with links to explanations about them from various vendors:
Open ROADM, which works to address the fact that optical systems have been proprietary (e.g. because SD FEC algorithms) on transponders are not interoperable and there are proprietary control loops between transponders and other optical components).
The next video gave me a better understanding of customer concerns with ROADMs and FOADMs.
Tutorial: To ROADM or Not to ROADM: When does a FOADM make sense in your optical network?
Other major types of network components include amplifiers, regenerators, and equalization nodes. Transponders map client side signals to wavelengths for high speed transport. They can be contained in a field-replaceable unit (FRU). Common types of pluggable optics include SFP+ (Small Form-factor Pluggable), CFP4, and QSFP28. Amplification is an analog process of boosting signal strength and is done in the optical domain (no conversion to electrical). Any impairments in the signal are boosted as well. A single pump laser is used for this. Regeneration can reshape and retime the optical signal but requires conversion to the electrical domain then back to the optical domain, making it more expensive to implement.
How a Fiber Laser Works
Major types of amplifiers in optical networks include EDFA (Erbium Doped Fiber Amplifer), Raman amplifier, and Hybrid Raman-EDFA amplifier. These are great explanations of these amplifiers:
Working Principle of Erbium Doped Fiber Amplifier (EDFA)
The EDFA – how it was developed.
Wavelength Selective Switch (WSS) was first implemented using MEMS but did’t work well because the hinges would fail. Liquid Crystal on Silicon (LCoS) is now commonly used to implement WSS since it has no moving parts. It can also support Flexgrid.
What is LCoS Based Wavelength Selective Switch – FO4SALE.COM
Optical patch panels are another component in fiber networks. They are used to join optical fibers where a connect/disconnect capability is required.
Handling Failure
There are 2 types of protection in networks:
Network protection: ensures that customer SLAs are met by preventing failures. Optical protection examples include mesh restoration (GMPLS, SDN), SNCP (OTN), UPSR & BLSR for SONET, and 1+1 or 1:1 circuits (active vs inactive backup circuit). Packet protection examples include MPLS fast reroute, LAG, G.8031, G.8032.
Equipment protection: focuses on protecting individual nodes.
I couldn’t emphasize this enough: this is such a broad field with so many technologies! What an introduction to networking!
The previous post introduced different types of networks and some of their architectural details. In this post, we look at the biggest problem network engineers work on: congestion. How are networks designed to address it? The professor starts tackling this area with a discussion of Quality of Service (QoS). Quality is defined in terms of the underlying requirements e.g. throughput, delay, jitter, packet loss, service availability, and per-flow sequence preservation. Services can be best effort, or other classes like gold service. Cisco’s Quality of Service (QoS) document discusses four levels of metal policy (platinum, gold, silver, and bronze), for example.
Class of Service (CoS) is a traffic classification that enables different actions to be taken on individual classes of traffic. Contrast this to type of service (ToS), which is a specific field in the IPv4 header (used to implement CoS). Juniper Networks post on Understanding Class of Service (CoS) Profiles equates QoS and CoS, but the professor explains that QoS is a more abstract concept than CoS.
QoS is a set of actions that the network takes to deliver the right delay, throughput, etc. QoS timeframes affect the way congestion is handled. For example, scheduling and dropping techniques and per-hop queuing are useful for the low millisecond time regime common in web traffic. Congestion over hundreds of milliseconds typically affects TCP (e.g. round trip times, closed-loop feedback) and this is addressed via methods like active queue management (AQM) and congestion control techniques like random early detection (RED). Congestion that occurs in the tens of seconds to minutes range is addressed by capacity planning.
How is QoS achieved in the data and control planes? By queuing, scheduling, policing, and dropping. The roles of the data and control planes are quite extensive as per the router diagram used to describe them. This is without getting into the details of the management plane e.g. the element management systems (per node) and the network management systems they communicate with. Control plane QoS mechanisms handle admission control and resource reservation and are typically implemented in software. Resource Reservation Protocol (RSVP) is the protocol mostly used in practice for control plane QoS. There are many explanations on RSVP, e.g. this Introduction to RSVP and this RSVP Overview. The primary QoS architectures are integrated services (Intserv) and differentiated services (Diffserv). Intserv uses RSVP and although it doesn’t scale, it is useful when guaranteed service is required.
We start a deep dive into the QoS techniques with queuing. There are different types of queues: first come first served (FCFS/FIFO), priority queues, and weighted queues. Packet schedulers can have a mix of these approches, e.g. 1 priority queue and N other weighted queues. Performance modeling can be done on queues. For voice traffic, the distribution of the arrival rate of traffic is a Poisson distribution. Therefore, the delay of packets and the length of the queue can be accurately modeled/predicted! See M/M/1 queues as a starting point (M/M/1 is Kendall notation and is more fully described in the next video).
Queuing Theory Tutorial – Queues/Lines, Characteristics, Kendall Notation, M/M/1 Queues
Data Plane QoS Mechanisms
These data plane QoS mechanisms are applied at each network node: classification, marking, policing and shaping, prioritization, minimum rate assurance. Below are more details about each.
Classification
This is the process of identifying flows of packets and grouping individual traffic flows into aggregated streams such that actions can be applied to those flow streams. Up to this point, I have had a vague idea of what a flow is but not a proper definition. The instructor defines a flow as a 5-tuple of source & destination IP addresses and TCP/UDP ports and a transport protocol. What is a Network Traffic Flow? discusses various ways of defining a flow, and this is just one of many. Classification needs to avoid fragmentation because the 5-tuple information is only in the first packet. There are 4 ways of classifying traffic:
Simple classification – the use of fields designed for QoS classification in IP headers e.g. the type of service (TOS) byte in IPv4. There are complications with using the DTRM bits of the TOS (e.g. minimizing delay and maximizing throughput could conflict).
Implicit classification – done without inspecting packet header or content, e.g. by examining layer 1 or 2 identifiers.
Complex classification – using fields not designed for QoS classification or layer 2 criteria like MAC addresses.
This is simply setting the fields assigned for QoS classification in IP packet headers (DSCP field) or MPLS packet headers (EXP field).
Source marking is applied at the source of the packets
Ingress marking: used when source cannot mark correctly or cannot be trusted to do so.
Rate Enforcement
This is done to avoid congestion. Policing is a mechanism to ensure that a traffic stream does not exceed a defined maximum rate. It stands in constrast to shaping, which is typically accomplished by queuing (delays traffic, never drops it). One type of policer is the token bucket policer. It never delays traffic and cannot reorder or reprioritize traffic. See Cisco’s Policing and Shaping Overview and QoS Policing documents for details. This is one of the rate limiting algorithms discussed in the video below (I found this video’s explanation more intuitive).
Five Rate Limiting Algorithms ~ Key Concepts in System Design
The next stage is prioritization of the traffic. 4 possible approaches: with prioritiy queues, e.g. where VoIP traffic always has highest priority, other queues can be starved by the scheduler. Weighted round robbin will take more packets from the high priority queues but still cycle through the other queues, taking fewer packets from them. Weighted bandwidth scheduling considers the packet sizes instead of just packet counts per queue (e.g. just taking 1 packet from a low priority queue can have negative impact if the packet is huge). Deficit round robbin is the one used in practice. It keeps track of the history of the number of packets services, and not just instantaneous values. I found the next video to expand on these brief explanations of scheduling algorithms.
How Do Schedulers in Routers Work? Understanding RR, WRR, WFQ, and DRR Through Simple Examples
One of the points that came up in discussion was that the schedulers use Run-to-completion scheduling, which means that a packet must be fully processed before starting on another packet. Routers have an interface FIFO (Tx buffer) on the physical link. When it fills up, this signals to the scheduler that there may be congestion downstream, thereby allowing for back pressure flow control. There is also multi-level strict policy queuing which allows for multiple priority queues instead of just 1 (e.g. voice & video) but not as common today.
Routers also drop packets to prevent unacceptable delays caused by buffering too many packets. There are different dropping strategies, e.g. tail dropping (dropping from the back of the queue), weighted tail dropping (>1 queue limit via heuristics), and head dropping (rare).
Active queue management (AQM) is a congestion avoidance technique. It works by detecting congestion before queues overflow. These are some techniques for AQM:
These QoS mechanisms operate in the context of an overriding architecture, integrated services (Intserv) or differentiated services (Diffserv). IntServ can be used in the financial industry or medical health facilities, for example. These are delay sensitive applications where unbounded scaling is not a real requirement. IntServ explicitly manages bandwidth resources on a per flow basis. DiffServ was developed to support (as the name suggests) differentiated treatment of packets in large scale environments. It does this using a 6-bit differentiated services code point (DSCP) in the IPv4 ToS header or the IPv6 traffic class octet. Classification and conditioning happen at the edge of the DiffServ domain. Actions are performed on behavior aggregates (contrast this to the per flow actions of IntServ). The next technology we learn about is Multiprotocol Label Switching, defined as follows on Wikipedia:
MPLS is similar to IntServ in that it lets you define an end-to-end path through the network for traffic but without reserving resources. It is a hop by hop forwarding mechanism, which stands in contrast to IP which works by making next hop routing decisions without regard to the end-to-end path taken by the packets. MPLS can be deployed on any layer 2 technology (multiprotocol). Benefits of MPLS include fast rerouting in case of failures and providing QoS support. One of the settings in which MPLS is used is in SD-WAN. This article provides a helpful contrast: What is the difference between SD-WAN and MPLS? These are the main applications of MPLS:
Traffic Engineering: allows network administrator to make the path deterministic (normal hop-by-hop routing is not). In other words, a frame forwarding policy can be used instead of relying on dynamic routing protocols.
QoS: the MPLS EXP bits are used for marking traffic per the labels.
This is quite the array of topics, especially for an introduction to networks course. I have a greater appreciation of how broad this space is.
I’m taking an online introductory course on networks. I have been surprised by how much ground this course is covering. I didn’t expect to cover wireless (mobile) networks, for example. I looked for videos on some of the topics to learn more, e.g. 4g network architecture – YouTube. Networking is turning out to be much cooler and more interesting than I thought possible. This post is a compilation of all the key topics introduced in the course (in the general order they were introduced, but not particularly organized into a coherent story).
My main takeaway from this first video is that 4G networks are entirely packet switched (basic, but new to me).
4G LTE Network Architecture Simplified
The next video on how messages are transmitted to the cell phone tower is insightful as well. I appreciated the high-level discussion of antennas.
How WiFi and Cell Phones Work | Wireless Communication Explained
The concept of control plane and data plane came up as well. One advantage of this separation as per the overview below are independent evolution and development of each (e.g. control software can be upgraded without changing the hardware).
M2.1: Overview of Control and Data Plane Separation
5G Service Based Architecture | Telecoms Bytes – Mpirical
Then of course there are the fundamental concepts of throughput, delay, and packet loss error. Jim Kurose’s book (and video below) covers these topics but it’s been a while since I read that book.
The professor also clarified the difference between bandwidth and throughput. The next video briefly touches on this distinction:
The course has also introduced me to the concept of spectral efficiency as part of understanding the difference between bandwidth and throughput. There is no shortage of concepts to learn about, from the different types of lines like T1 and T3 to bit robbing to the existence of network interface devices. The video below is good intro to T1.
DS1 (T1) Fundamentals
There was also a discussion about cable networks, with an onslaught of concepts like Hybrid fiber-coaxial. This Cable 101 video is a helpful resource.
The HFC Cable Systems Introduction video below starts out with a comparison of coax and fiber then explains the flow of signals from the core network to the home.
HFC Cable Systems Introduction
I still need to learn more information about the Cable modem termination system (CMTS) and the next resource is perfect. It mentions CMTS vendors like Arris, Cisco, and Motorola, which inspires me to look up the Cisco CMTS.
Cable Modem Termination System Tutorial (CMTS)
I have never researched how most of these systems work so I am greatly appreciating this introduction to networks course! Here’s a video on how cable modems work, including their interactions with the CMTS.
How Cable Modems Work
The communication between the CMTS and the CMs is done via DOCSIS. Here is the reference I found with insight into DOCSIS.
DOCSIS® 3.1 – An Overview
Something I picked up is that CableLabs does a lot of the research for these systems. Other concepts to know include wavelength-division multiplexing (WDM), which was used in the traditional coax networks. The following explanation is an example of WDM in fiber.
Next, we get into the 7-layer OSI model. The example given for the physical layer is SONET technology. Another foray into T1 technology reveals the fact that bipolar transmission is used for T1 since it is more power efficient.
Multiplexing is the next interesting topic introduced. I have included some videos below on the different types of multiplexing employed in communications.
FDM involves modulating message signals over carrier frequencies then using bandpass filters to extract the individual signals.
The course also addresses transmission fundamentals like the difference between bit rate and baud rate, the Shannon–Hartley theorem, the Nyquist–Shannon sampling theorem, modulation, modems, and codecs. I have compiled a few videos covering these topics below.
Here is an explanation of the Shannon–Hartley theorem:
Channel Capacity by Shannon-Hartley | Basics, Proof & Maximum Bandwidth Condition
We then start getting into network addressing. One of the important concepts here is how the exhaustion of IPv4 addresses is handled: private IP addresses, DHCP, subnetting, and IPv6. One particularly interesting point was the difference between IPv4 and IPv6 headers:
In a discussion of the impact of TCP on throughput, the professor called out TCP global synchronization as an issue that networks need to avoid. Here’s one video about it.
Avoiding packet reordering is another important aspect of TCP. The contrast with UDP is especially interesting when other protocols like Google’s QUIC are designed. The RTP protocol (a relative of UDP, informally speaking) is used for VoIP. This is a good description of RTP:
I mentioned in my last post (Learning about Large Language Models) that I recently started going through Sebastian Raschka’s Build a Large Language Model from Scratch book. It had been a while since I ran python code on my laptop so I needed to do some cleanup to restore my environment. I cloned the repo and started executing the first cell in Chapter 2:
from importlib.metadata import version
print("torch version:", version("torch"))
print("tiktoken version:", version("tiktoken"))
I got this error: PackageNotFoundError: No package metadata was found for torch".
{
"name": "PackageNotFoundError",
"message": "No package metadata was found for torch",
"stack": "---------------------------------------------------------------------------
StopIteration Traceback (most recent call last)
File /opt/homebrew/Cellar/python@3.11/3.11.1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/importlib/metadata/__init__.py:563, in Distribution.from_name(cls, name)
562 try:
--> 563 return next(cls.discover(name=name))
564 except StopIteration:
StopIteration:
During handling of the above exception, another exception occurred:
PackageNotFoundError Traceback (most recent call last)
Cell In[2], line 7
3 print(sys.version)
5 from importlib.metadata import version
----> 7 print(\"torch version:\", version(\"torch\"))
8 print(\"tiktoken version:\", version(\"tiktoken\"))
File /opt/homebrew/Cellar/python@3.11/3.11.1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/importlib/metadata/__init__.py:1008, in version(distribution_name)
1001 def version(distribution_name):
1002 \"\"\"Get the version string for the named package.
1003
1004 :param distribution_name: The name of the distribution package to query.
1005 :return: The version string for the package as defined in the package's
1006 \"Version\" metadata key.
1007 \"\"\"
-> 1008 return distribution(distribution_name).version
File /opt/homebrew/Cellar/python@3.11/3.11.1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/importlib/metadata/__init__.py:981, in distribution(distribution_name)
975 def distribution(distribution_name):
976 \"\"\"Get the ``Distribution`` instance for the named package.
977
978 :param distribution_name: The name of the distribution package as a string.
979 :return: A ``Distribution`` instance (or subclass thereof).
980 \"\"\"
--> 981 return Distribution.from_name(distribution_name)
File /opt/homebrew/Cellar/python@3.11/3.11.1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/importlib/metadata/__init__.py:565, in Distribution.from_name(cls, name)
563 return next(cls.discover(name=name))
564 except StopIteration:
--> 565 raise PackageNotFoundError(name)
PackageNotFoundError: No package metadata was found for torch"
}
How is this showing version 3.10.9 but VS Code is using 3.11 from brew (per the file paths in the error messages? This is how to print the version from a python script as per Printing Python version in output – Stack Overflow
Since I have brew installed, this might be what I need: the comment that “If you brew-install Python, but pip is still not in your path, you might need to re-link, like this brew unlink python && brew link python” at python – How do I install pip on macOS or OS X? – Stack Overflow.
saint@MacBookPro LLMs-from-scratch % brew update
==> Downloading https://ghcr.io/v2/homebrew/portable-ruby/portable-ruby/blobs/sha256:d9faa506c014dedc0b034a68103ba75c9a58242f4d6c67b6ca0f649c39602bcf
######################################################################################################################################################################################################################################### 100.0%
==> Pouring portable-ruby-3.3.7.arm64_big_sur.bottle.tar.gz
==> Homebrew collects anonymous analytics.
Read the analytics documentation (and how to opt-out) here:
https://docs.brew.sh/Analytics
No analytics have been recorded yet (nor will be during this `brew` run).
==> homebrew/core is old and unneeded, untapping to save space...
Untapping homebrew/core...
Untapped 3 commands and 7398 formulae (7,130 files, 1GB).
==> homebrew/cask is old and unneeded, untapping to save space...
Untapping homebrew/cask...
Untapped 7333 casks (4,415 files, 487.2MB).
==> Downloading https://formulae.brew.sh/api/formula_tap_migrations.jws.json
Updated 4 taps (microsoft/git, homebrew/cask-versions, homebrew/core and homebrew/cask).
==> New Formulae
...<hundreds of lines omitted but included python entries below>
python-argcomplete
python-freethreading
python-gdbm@3.12
python-gdbm@3.13
python-matplotlib
python-packaging
python-setuptools
python-tk@3.12
python-tk@3.13
python@3.12
python@3.13
pyupgrade
...
==> Deleted Installed Formulae
icu4c ✘
==> Deleted Installed Casks
git-credential-manager-core ✘ microsoft-openjdk11 ✘
Error: Unexpected method 'appcast' called on Cask adoptopenjdk16.
Follow the instructions here:
https://github.com/Homebrew/homebrew-cask#reporting-bugs
==> Downloading https://formulae.brew.sh/api/cask_tap_migrations.jws.json
==> Outdated Formulae
aom fb303 fribidi gnuplot jasper libheif libtool lua openblas pstoedit sqlite xorgproto
arpack fbthrift gcc graphicsmagick jbig2dec libidn libunistring lz4 openexr pyqt@5 suite-sparse xz
autoconf fig2dev gd harfbuzz jpeg-turbo libidn2 libvmaf maven openjdk python@3.10 sundials zstd
boost fizz gdbm hdf5 jpeg-xl liblqr libx11 mpdecimal openjpeg python@3.11 tcl-tk
brotli flac gettext highway libaec libomp libxau mpfr openssl@1.1 python@3.9 texinfo
ca-certificates fltk ghostscript hwloc libavif libpng libxcb mpg123 openssl@3 qscintilla2 wangle
cairo fmt giflib icu4c@76 libcerf libraw libxdmcp netpbm opus qt@5 watchman
cmake folly git-gui imagemagick libde265 libsndfile libxext ninja pango readline webp
double-conversion fontconfig glib imath libevent libsodium libxrender octave pcre2 shared-mime-info wget
edencommon freetype gmp isl libffi libtiff little-cms2 open-mpi pixman snappy x265
==> Outdated Casks
git-credential-manager microsoft-openjdk microsoft-openjdk@11
You have 113 outdated formulae and 3 outdated casks installed.
You can upgrade them with brew upgrade
or list them with brew outdated.
Error: Unexpected method 'appcast' called on Cask adoptopenjdk16.
Follow the instructions here:
https://github.com/Homebrew/homebrew-cask#reporting-bugs
==> Migrating cask git-credential-manager-core to git-credential-manager
Error: inreplace failed
/opt/homebrew/Caskroom/git-credential-manager/.metadata/2.1.2/20230703191748.675/Casks/git-credential-manager.rb:
expected replacement of /\A\s*cask\s+"git\-credential\-manager\-core"/ with "cask \"git-credential-manager\""
python3 --version is still 3.10.9 after this. I tried running pip but zsh sayd command not found. Unfortunately, linking or unlinking of either python or python3 fails with the errors below (despite ls -l `which python3` showing the same path as before).
saint@MacBookPro LLMs-from-scratch % brew unlink python3
Error: Unexpected method 'appcast' called on Cask adoptopenjdk16.
Follow the instructions here:
https://github.com/Homebrew/homebrew-cask#reporting-bugs
Error: No such keg: /opt/homebrew/Cellar/python3
I decided to install python3 again.
saint@MacBookPro LLMs-from-scratch % brew install python3
==> Downloading https://formulae.brew.sh/api/formula.jws.json
==> Downloading https://formulae.brew.sh/api/cask.jws.json
==> Downloading https://ghcr.io/v2/homebrew/core/python/3.13/manifests/3.13.1
######################################################################################################################################################################################################################################### 100.0%
==> Fetching dependencies for python@3.13: mpdecimal, ca-certificates, openssl@3, readline, sqlite and xz
==> Downloading https://ghcr.io/v2/homebrew/core/mpdecimal/manifests/4.0.0-1
######################################################################################################################################################################################################################################### 100.0%
==> Fetching mpdecimal
==> Downloading https://ghcr.io/v2/homebrew/core/mpdecimal/blobs/sha256:0f5f269bed0e6be2de3edfc4b52867e656f993e5bcff40717f26ee94dd0d2211
######################################################################################################################################################################################################################################### 100.0%
==> Downloading https://ghcr.io/v2/homebrew/core/ca-certificates/manifests/2024-12-31
######################################################################################################################################################################################################################################### 100.0%
==> Fetching ca-certificates
...
<lots of omitted lines>
...
==> Fetching harfbuzz
==> Downloading https://ghcr.io/v2/homebrew/core/harfbuzz/blobs/sha256:2f892566c02b3c8c61aed6f7867b4405e5c814df8500ef4bc4ca91a9e40205a9
######################################################################################################################################################################################################################################### 100.0%
==> Fetching openjdk
==> Downloading https://ghcr.io/v2/homebrew/core/openjdk/blobs/sha256:1285eadf2b5998cda49e4470ee3875e855b0be199765401ad77dc38aea573f49
######################################################################################################################################################################################################################################### 100.0%
Error: can't modify frozen String: "The bottle needs the Xcode Command Line Tools to be installed at /Library/Developer/CommandLineTools.\nDevelopment tools provided by Xcode.app are not sufficient.\n\nYou can install the Xcode Command Line Tools, if desired, with:\n xcode-select --install\n"
This was the new state of affairs is that command (failed):
saint@MacBookPro LLMs-from-scratch % python3 --version
Python 3.13.1
saint@MacBookPro LLMs-from-scratch % which python3
/opt/homebrew/bin/python3
saint@MacBookPro LLMs-from-scratch % ls -l `which python3`
lrwxr-xr-x 1 saint admin 40 Feb 4 17:02 /opt/homebrew/bin/python3 -> ../Cellar/python@3.13/3.13.1/bin/python3
saint@MacBookPro LLMs-from-scratch % which pip
pip not found
Ah, all that agonizing and look at this – did I need to be using pip3 all this time?
saint@MacBookPro LLMs-from-scratch % which pip3
/opt/homebrew/bin/pip3
saint@MacBookPro LLMs-from-scratch % ls -l `which pip3`
lrwxr-xr-x 1 saint admin 37 Feb 4 17:02 /opt/homebrew/bin/pip3 -> ../Cellar/python@3.13/3.13.1/bin/pip3
saint@MacBookPro LLMs-from-scratch %
Interestingly, I still can’t install pytorch using pip3?
saint@MacBookPro LLMs-from-scratch % pip3 install pytorch
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
error: externally-managed-environment
× This environment is externally managed
╰─> To install Python packages system-wide, try brew install
xyz, where xyz is the package you are trying to
install.
If you wish to install a Python library that isn't in Homebrew,
use a virtual environment:
python3 -m venv path/to/venv
source path/to/venv/bin/activate
python3 -m pip install xyz
If you wish to install a Python application that isn't in Homebrew,
it may be easiest to use 'pipx install xyz', which will manage a
virtual environment for you. You can install pipx with
brew install pipx
You may restore the old behavior of pip by passing
the '--break-system-packages' flag to pip, or by adding
'break-system-packages = true' to your pip.conf file. The latter
will permanently disable this error.
If you disable this error, we STRONGLY recommend that you additionally
pass the '--user' flag to pip, or set 'user = true' in your pip.conf
file. Failure to do this can result in a broken Homebrew installation.
Read more about this behavior here: <https://peps.python.org/pep-0668/>
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.
hint: See PEP 668 for the detailed specification.
saint@MacBookPro LLMs-from-scratch % xcode-select --install
xcode-select: note: install requested for command line developer tools
I installed the command line developer tools when prompted below.
Trying to run the first cell in VS Code with the updated setup now gave this error:
{
"name": "",
"message": "",
"stack": "Running cells with 'Python 3.13.1' requires the ipykernel package.
Run the following command to install 'ipykernel' into the Python environment.
Command: '/opt/homebrew/bin/python3 -m pip install ipykernel -U --user --force-reinstall'"
}
saint@MacBookPro LLMs-from-scratch % /opt/homebrew/bin/python3 -m pip install ipykernel -U --user --force-reinstall
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
error: externally-managed-environment
× This environment is externally managed
╰─> To install Python packages system-wide, try brew install
xyz, where xyz is the package you are trying to
install.
If you wish to install a Python library that isn't in Homebrew,
use a virtual environment:
python3 -m venv path/to/venv
source path/to/venv/bin/activate
python3 -m pip install xyz
If you wish to install a Python application that isn't in Homebrew,
it may be easiest to use 'pipx install xyz', which will manage a
virtual environment for you. You can install pipx with
brew install pipx
You may restore the old behavior of pip by passing
the '--break-system-packages' flag to pip, or by adding
'break-system-packages = true' to your pip.conf file. The latter
will permanently disable this error.
If you disable this error, we STRONGLY recommend that you additionally
pass the '--user' flag to pip, or set 'user = true' in your pip.conf
file. Failure to do this can result in a broken Homebrew installation.
Read more about this behavior here: <https://peps.python.org/pep-0668/>
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.
hint: See PEP 668 for the detailed specification.
saint@MacBookPro LLMs-from-scratch % brew install ipykernel
==> Downloading https://formulae.brew.sh/api/formula.jws.json
==> Downloading https://formulae.brew.sh/api/cask.jws.json
Warning: No available formula with the name "ipykernel".
==> Searching for similarly named formulae and casks...
Error: No formulae or casks found for ipykernel.
I override the warning using the --break-system-packages flag and the VS Code notebook now runs.
saint@MacBookPro LLMs-from-scratch % python3 -m pip install ipykernel -U --user --force-reinstall --break-system-package
Collecting ipykernel
Downloading ipykernel-6.29.5-py3-none-any.whl.metadata (6.3 kB)
Collecting appnope (from ipykernel)
Downloading appnope-0.1.4-py2.py3-none-any.whl.metadata (908 bytes)
Collecting comm>=0.1.1 (from ipykernel)
Downloading comm-0.2.2-py3-none-any.whl.metadata (3.7 kB)
Collecting debugpy>=1.6.5 (from ipykernel)
Downloading debugpy-1.8.12-cp313-cp313-macosx_14_0_universal2.whl.metadata (1.3 kB)
Collecting ipython>=7.23.1 (from ipykernel)
Downloading ipython-8.32.0-py3-none-any.whl.metadata (5.0 kB)
Collecting jupyter-client>=6.1.12 (from ipykernel)
Downloading jupyter_client-8.6.3-py3-none-any.whl.metadata (8.3 kB)
Collecting jupyter-core!=5.0.*,>=4.12 (from ipykernel)
Downloading jupyter_core-5.7.2-py3-none-any.whl.metadata (3.4 kB)
Collecting matplotlib-inline>=0.1 (from ipykernel)
Downloading matplotlib_inline-0.1.7-py3-none-any.whl.metadata (3.9 kB)
Collecting nest-asyncio (from ipykernel)
Downloading nest_asyncio-1.6.0-py3-none-any.whl.metadata (2.8 kB)
Collecting packaging (from ipykernel)
Downloading packaging-24.2-py3-none-any.whl.metadata (3.2 kB)
Collecting psutil (from ipykernel)
Downloading psutil-6.1.1-cp36-abi3-macosx_11_0_arm64.whl.metadata (22 kB)
Collecting pyzmq>=24 (from ipykernel)
Downloading pyzmq-26.2.1-cp313-cp313-macosx_10_15_universal2.whl.metadata (6.2 kB)
Collecting tornado>=6.1 (from ipykernel)
Downloading tornado-6.4.2-cp38-abi3-macosx_10_9_universal2.whl.metadata (2.5 kB)
Collecting traitlets>=5.4.0 (from ipykernel)
Downloading traitlets-5.14.3-py3-none-any.whl.metadata (10 kB)
Collecting decorator (from ipython>=7.23.1->ipykernel)
Downloading decorator-5.1.1-py3-none-any.whl.metadata (4.0 kB)
Collecting jedi>=0.16 (from ipython>=7.23.1->ipykernel)
Downloading jedi-0.19.2-py2.py3-none-any.whl.metadata (22 kB)
Collecting pexpect>4.3 (from ipython>=7.23.1->ipykernel)
Downloading pexpect-4.9.0-py2.py3-none-any.whl.metadata (2.5 kB)
Collecting prompt_toolkit<3.1.0,>=3.0.41 (from ipython>=7.23.1->ipykernel)
Downloading prompt_toolkit-3.0.50-py3-none-any.whl.metadata (6.6 kB)
Collecting pygments>=2.4.0 (from ipython>=7.23.1->ipykernel)
Downloading pygments-2.19.1-py3-none-any.whl.metadata (2.5 kB)
Collecting stack_data (from ipython>=7.23.1->ipykernel)
Downloading stack_data-0.6.3-py3-none-any.whl.metadata (18 kB)
Collecting python-dateutil>=2.8.2 (from jupyter-client>=6.1.12->ipykernel)
Downloading python_dateutil-2.9.0.post0-py2.py3-none-any.whl.metadata (8.4 kB)
Collecting platformdirs>=2.5 (from jupyter-core!=5.0.*,>=4.12->ipykernel)
Downloading platformdirs-4.3.6-py3-none-any.whl.metadata (11 kB)
Collecting parso<0.9.0,>=0.8.4 (from jedi>=0.16->ipython>=7.23.1->ipykernel)
Downloading parso-0.8.4-py2.py3-none-any.whl.metadata (7.7 kB)
Collecting ptyprocess>=0.5 (from pexpect>4.3->ipython>=7.23.1->ipykernel)
Downloading ptyprocess-0.7.0-py2.py3-none-any.whl.metadata (1.3 kB)
Collecting wcwidth (from prompt_toolkit<3.1.0,>=3.0.41->ipython>=7.23.1->ipykernel)
Downloading wcwidth-0.2.13-py2.py3-none-any.whl.metadata (14 kB)
Collecting six>=1.5 (from python-dateutil>=2.8.2->jupyter-client>=6.1.12->ipykernel)
Downloading six-1.17.0-py2.py3-none-any.whl.metadata (1.7 kB)
Collecting executing>=1.2.0 (from stack_data->ipython>=7.23.1->ipykernel)
Downloading executing-2.2.0-py2.py3-none-any.whl.metadata (8.9 kB)
Collecting asttokens>=2.1.0 (from stack_data->ipython>=7.23.1->ipykernel)
Downloading asttokens-3.0.0-py3-none-any.whl.metadata (4.7 kB)
Collecting pure-eval (from stack_data->ipython>=7.23.1->ipykernel)
Downloading pure_eval-0.2.3-py3-none-any.whl.metadata (6.3 kB)
Downloading ipykernel-6.29.5-py3-none-any.whl (117 kB)
Downloading comm-0.2.2-py3-none-any.whl (7.2 kB)
Downloading debugpy-1.8.12-cp313-cp313-macosx_14_0_universal2.whl (2.5 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2.5/2.5 MB 36.8 MB/s eta 0:00:00
Downloading ipython-8.32.0-py3-none-any.whl (825 kB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 825.5/825.5 kB 54.9 MB/s eta 0:00:00
Downloading jupyter_client-8.6.3-py3-none-any.whl (106 kB)
Downloading jupyter_core-5.7.2-py3-none-any.whl (28 kB)
Downloading matplotlib_inline-0.1.7-py3-none-any.whl (9.9 kB)
Downloading pyzmq-26.2.1-cp313-cp313-macosx_10_15_universal2.whl (1.3 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.3/1.3 MB 62.7 MB/s eta 0:00:00
Downloading tornado-6.4.2-cp38-abi3-macosx_10_9_universal2.whl (436 kB)
Downloading traitlets-5.14.3-py3-none-any.whl (85 kB)
Downloading appnope-0.1.4-py2.py3-none-any.whl (4.3 kB)
Downloading nest_asyncio-1.6.0-py3-none-any.whl (5.2 kB)
Downloading packaging-24.2-py3-none-any.whl (65 kB)
Downloading psutil-6.1.1-cp36-abi3-macosx_11_0_arm64.whl (248 kB)
Downloading jedi-0.19.2-py2.py3-none-any.whl (1.6 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.6/1.6 MB 79.1 MB/s eta 0:00:00
Downloading pexpect-4.9.0-py2.py3-none-any.whl (63 kB)
Downloading platformdirs-4.3.6-py3-none-any.whl (18 kB)
Downloading prompt_toolkit-3.0.50-py3-none-any.whl (387 kB)
Downloading pygments-2.19.1-py3-none-any.whl (1.2 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.2/1.2 MB 58.1 MB/s eta 0:00:00
Downloading python_dateutil-2.9.0.post0-py2.py3-none-any.whl (229 kB)
Downloading decorator-5.1.1-py3-none-any.whl (9.1 kB)
Downloading stack_data-0.6.3-py3-none-any.whl (24 kB)
Downloading asttokens-3.0.0-py3-none-any.whl (26 kB)
Downloading executing-2.2.0-py2.py3-none-any.whl (26 kB)
Downloading parso-0.8.4-py2.py3-none-any.whl (103 kB)
Downloading ptyprocess-0.7.0-py2.py3-none-any.whl (13 kB)
Downloading six-1.17.0-py2.py3-none-any.whl (11 kB)
Downloading pure_eval-0.2.3-py3-none-any.whl (11 kB)
Downloading wcwidth-0.2.13-py2.py3-none-any.whl (34 kB)
Installing collected packages: wcwidth, pure-eval, ptyprocess, traitlets, tornado, six, pyzmq, pygments, psutil, prompt_toolkit, platformdirs, pexpect, parso, packaging, nest-asyncio, executing, decorator, debugpy, asttokens, appnope, stack_data, python-dateutil, matplotlib-inline, jupyter-core, jedi, comm, jupyter-client, ipython, ipykernel
WARNING: The script pygmentize is installed in '/Users/saint/Library/Python/3.13/bin' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
WARNING: The script debugpy is installed in '/Users/saint/Library/Python/3.13/bin' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
WARNING: The scripts jupyter, jupyter-migrate and jupyter-troubleshoot are installed in '/Users/saint/Library/Python/3.13/bin' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
WARNING: The scripts jupyter-kernel, jupyter-kernelspec and jupyter-run are installed in '/Users/saint/Library/Python/3.13/bin' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
WARNING: The scripts ipython and ipython3 are installed in '/Users/saint/Library/Python/3.13/bin' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
Successfully installed appnope-0.1.4 asttokens-3.0.0 comm-0.2.2 debugpy-1.8.12 decorator-5.1.1 executing-2.2.0 ipykernel-6.29.5 ipython-8.32.0 jedi-0.19.2 jupyter-client-8.6.3 jupyter-core-5.7.2 matplotlib-inline-0.1.7 nest-asyncio-1.6.0 packaging-24.2 parso-0.8.4 pexpect-4.9.0 platformdirs-4.3.6 prompt_toolkit-3.0.50 psutil-6.1.1 ptyprocess-0.7.0 pure-eval-0.2.3 pygments-2.19.1 python-dateutil-2.9.0.post0 pyzmq-26.2.1 six-1.17.0 stack_data-0.6.3 tornado-6.4.2 traitlets-5.14.3 wcwidth-0.2.13
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
saint@MacBookPro LLMs-from-scratch %
First cell now fails because the torch package cannot be found:
{
"name": "PackageNotFoundError",
"message": "No package metadata was found for torch",
"stack": "---------------------------------------------------------------------------
StopIteration Traceback (most recent call last)
File /opt/homebrew/Cellar/python@3.13/3.13.1/Frameworks/Python.framework/Versions/3.13/lib/python3.13/importlib/metadata/__init__.py:407, in Distribution.from_name(cls, name)
406 try:
--> 407 return next(iter(cls.discover(name=name)))
408 except StopIteration:
StopIteration:
During handling of the above exception, another exception occurred:
PackageNotFoundError Traceback (most recent call last)
Cell In[1], line 3
1 from importlib.metadata import version
----> 3 print(\"torch version:\", version(\"torch\"))
4 print(\"tiktoken version:\", version(\"tiktoken\"))
File /opt/homebrew/Cellar/python@3.13/3.13.1/Frameworks/Python.framework/Versions/3.13/lib/python3.13/importlib/metadata/__init__.py:987, in version(distribution_name)
980 def version(distribution_name: str) -> str:
981 \"\"\"Get the version string for the named package.
982
983 :param distribution_name: The name of the distribution package to query.
984 :return: The version string for the package as defined in the package's
985 \"Version\" metadata key.
986 \"\"\"
--> 987 return distribution(distribution_name).version
File /opt/homebrew/Cellar/python@3.13/3.13.1/Frameworks/Python.framework/Versions/3.13/lib/python3.13/importlib/metadata/__init__.py:960, in distribution(distribution_name)
954 def distribution(distribution_name: str) -> Distribution:
955 \"\"\"Get the ``Distribution`` instance for the named package.
956
957 :param distribution_name: The name of the distribution package as a string.
958 :return: A ``Distribution`` instance (or subclass thereof).
959 \"\"\"
--> 960 return Distribution.from_name(distribution_name)
File /opt/homebrew/Cellar/python@3.13/3.13.1/Frameworks/Python.framework/Versions/3.13/lib/python3.13/importlib/metadata/__init__.py:409, in Distribution.from_name(cls, name)
407 return next(iter(cls.discover(name=name)))
408 except StopIteration:
--> 409 raise PackageNotFoundError(name)
PackageNotFoundError: No package metadata was found for torch"
}
These are the commands I tried to install pytorch before finding the correct one: pip3 install torch --break-system-packages.
saint@MacBookPro LLMs-from-scratch % brew install torch
==> Downloading https://formulae.brew.sh/api/formula.jws.json
==> Downloading https://formulae.brew.sh/api/cask.jws.json
Warning: No available formula with the name "torch". Did you mean tor, pytorch or orc?
==> Searching for similarly named formulae and casks...
==> Formulae
pytorch ✔ torchvision tor orc
To install pytorch ✔, run:
brew install pytorch ✔
saint@MacBookPro LLMs-from-scratch % brew install pytorch
Warning: pytorch 2.5.1_4 is already installed and up-to-date.
To reinstall 2.5.1_4, run:
brew reinstall pytorch
saint@MacBookPro LLMs-from-scratch % pip3 install torch
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
error: externally-managed-environment
× This environment is externally managed
╰─> To install Python packages system-wide, try brew install
xyz, where xyz is the package you are trying to
install.
If you wish to install a Python library that isn't in Homebrew,
use a virtual environment:
python3 -m venv path/to/venv
source path/to/venv/bin/activate
python3 -m pip install xyz
If you wish to install a Python application that isn't in Homebrew,
it may be easiest to use 'pipx install xyz', which will manage a
virtual environment for you. You can install pipx with
brew install pipx
You may restore the old behavior of pip by passing
the '--break-system-packages' flag to pip, or by adding
'break-system-packages = true' to your pip.conf file. The latter
will permanently disable this error.
If you disable this error, we STRONGLY recommend that you additionally
pass the '--user' flag to pip, or set 'user = true' in your pip.conf
file. Failure to do this can result in a broken Homebrew installation.
Read more about this behavior here: <https://peps.python.org/pep-0668/>
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.
hint: See PEP 668 for the detailed specification.
saint@MacBookPro LLMs-from-scratch %
saint@MacBookPro LLMs-from-scratch % pip3 install torch --break-system-packages
Collecting torch
Downloading torch-2.6.0-cp313-none-macosx_11_0_arm64.whl.metadata (28 kB)
Collecting filelock (from torch)
Downloading filelock-3.17.0-py3-none-any.whl.metadata (2.9 kB)
Collecting typing-extensions>=4.10.0 (from torch)
Downloading typing_extensions-4.12.2-py3-none-any.whl.metadata (3.0 kB)
Collecting networkx (from torch)
Downloading networkx-3.4.2-py3-none-any.whl.metadata (6.3 kB)
Collecting jinja2 (from torch)
Downloading jinja2-3.1.5-py3-none-any.whl.metadata (2.6 kB)
Collecting fsspec (from torch)
Downloading fsspec-2025.2.0-py3-none-any.whl.metadata (11 kB)
Collecting setuptools (from torch)
Downloading setuptools-75.8.0-py3-none-any.whl.metadata (6.7 kB)
Collecting sympy==1.13.1 (from torch)
Downloading sympy-1.13.1-py3-none-any.whl.metadata (12 kB)
Collecting mpmath<1.4,>=1.1.0 (from sympy==1.13.1->torch)
Downloading mpmath-1.3.0-py3-none-any.whl.metadata (8.6 kB)
Collecting MarkupSafe>=2.0 (from jinja2->torch)
Downloading MarkupSafe-3.0.2-cp313-cp313-macosx_11_0_arm64.whl.metadata (4.0 kB)
Downloading torch-2.6.0-cp313-none-macosx_11_0_arm64.whl (66.5 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 66.5/66.5 MB 74.3 MB/s eta 0:00:00
Downloading sympy-1.13.1-py3-none-any.whl (6.2 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 6.2/6.2 MB 75.0 MB/s eta 0:00:00
Downloading typing_extensions-4.12.2-py3-none-any.whl (37 kB)
Downloading filelock-3.17.0-py3-none-any.whl (16 kB)
Downloading fsspec-2025.2.0-py3-none-any.whl (184 kB)
Downloading jinja2-3.1.5-py3-none-any.whl (134 kB)
Downloading networkx-3.4.2-py3-none-any.whl (1.7 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.7/1.7 MB 63.7 MB/s eta 0:00:00
Downloading setuptools-75.8.0-py3-none-any.whl (1.2 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.2/1.2 MB 37.0 MB/s eta 0:00:00
Downloading MarkupSafe-3.0.2-cp313-cp313-macosx_11_0_arm64.whl (12 kB)
Downloading mpmath-1.3.0-py3-none-any.whl (536 kB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 536.2/536.2 kB 27.9 MB/s eta 0:00:00
Installing collected packages: mpmath, typing-extensions, sympy, setuptools, networkx, MarkupSafe, fsspec, filelock, jinja2, torch
Successfully installed MarkupSafe-3.0.2 filelock-3.17.0 fsspec-2025.2.0 jinja2-3.1.5 mpmath-1.3.0 networkx-3.4.2 setuptools-75.8.0 sympy-1.13.1 torch-2.6.0 typing-extensions-4.12.2
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
saint@MacBookPro LLMs-from-scratch %
The pytorch import finally works! The next error is also a PackageNotFoundError: "No package metadata was found for tiktoken" which I addressed with the same installation steps:
saint@MacBookPro LLMs-from-scratch % pip3 install tiktoken
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
error: externally-managed-environment
× This environment is externally managed
╰─> To install Python packages system-wide, try brew install
xyz, where xyz is the package you are trying to
install.
If you wish to install a Python library that isn't in Homebrew,
use a virtual environment:
python3 -m venv path/to/venv
source path/to/venv/bin/activate
python3 -m pip install xyz
If you wish to install a Python application that isn't in Homebrew,
it may be easiest to use 'pipx install xyz', which will manage a
virtual environment for you. You can install pipx with
brew install pipx
You may restore the old behavior of pip by passing
the '--break-system-packages' flag to pip, or by adding
'break-system-packages = true' to your pip.conf file. The latter
will permanently disable this error.
If you disable this error, we STRONGLY recommend that you additionally
pass the '--user' flag to pip, or set 'user = true' in your pip.conf
file. Failure to do this can result in a broken Homebrew installation.
Read more about this behavior here: <https://peps.python.org/pep-0668/>
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.
hint: See PEP 668 for the detailed specification.
saint@MacBookPro LLMs-from-scratch %
saint@MacBookPro LLMs-from-scratch % pip3 install tiktoken --break-system-packages
Collecting tiktoken
Downloading tiktoken-0.8.0-cp313-cp313-macosx_11_0_arm64.whl.metadata (6.6 kB)
Collecting regex>=2022.1.18 (from tiktoken)
Downloading regex-2024.11.6-cp313-cp313-macosx_11_0_arm64.whl.metadata (40 kB)
Collecting requests>=2.26.0 (from tiktoken)
Downloading requests-2.32.3-py3-none-any.whl.metadata (4.6 kB)
Collecting charset-normalizer<4,>=2 (from requests>=2.26.0->tiktoken)
Downloading charset_normalizer-3.4.1-cp313-cp313-macosx_10_13_universal2.whl.metadata (35 kB)
Collecting idna<4,>=2.5 (from requests>=2.26.0->tiktoken)
Downloading idna-3.10-py3-none-any.whl.metadata (10 kB)
Collecting urllib3<3,>=1.21.1 (from requests>=2.26.0->tiktoken)
Downloading urllib3-2.3.0-py3-none-any.whl.metadata (6.5 kB)
Collecting certifi>=2017.4.17 (from requests>=2.26.0->tiktoken)
Downloading certifi-2025.1.31-py3-none-any.whl.metadata (2.5 kB)
Downloading tiktoken-0.8.0-cp313-cp313-macosx_11_0_arm64.whl (982 kB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 982.8/982.8 kB 24.2 MB/s eta 0:00:00
Downloading regex-2024.11.6-cp313-cp313-macosx_11_0_arm64.whl (284 kB)
Downloading requests-2.32.3-py3-none-any.whl (64 kB)
Downloading certifi-2025.1.31-py3-none-any.whl (166 kB)
Downloading charset_normalizer-3.4.1-cp313-cp313-macosx_10_13_universal2.whl (195 kB)
Downloading idna-3.10-py3-none-any.whl (70 kB)
Downloading urllib3-2.3.0-py3-none-any.whl (128 kB)
Installing collected packages: urllib3, regex, idna, charset-normalizer, certifi, requests, tiktoken
Successfully installed certifi-2025.1.31 charset-normalizer-3.4.1 idna-3.10 regex-2024.11.6 requests-2.32.3 tiktoken-0.8.0 urllib3-2.3.0
[notice] A new release of pip is available: 24.3.1 -> 25.0
[notice] To update, run: python3.13 -m pip install --upgrade pip
saint@MacBookPro LLMs-from-scratch %
Finally, my machine is in a state that can run the code in the Jupyter notebook! This is such a brittle environment. I need to switch to a managed environment to avoid this type of mess.
