- We now serialize the file descriptor table when spawning / executing
processes on Windows. This means you can now inherit more stuff than
just standard i/o. It's needed by bash, which duplicates the console
to file descriptor #255. We also now do a better job serializing the
environment variables, so you're less likely to encounter E2BIG when
using your bash shell. We also no longer coerce environ to uppercase
- execve() on Windows now remotely controls its parent process to make
them spawn a replacement for itself. Then it'll be able to terminate
immediately once the spawn succeeds, without having to linger around
for the lifetime as a shell process for proxying the exit code. When
process worker thread running in the parent sees the child die, it's
given a handle to the new child, to replace it in the process table.
- execve() and posix_spawn() on Windows will now provide CreateProcess
an explicit handle list. This allows us to remove handle locks which
enables better fork/spawn concurrency, with seriously correct thread
safety. Other codebases like Go use the same technique. On the other
hand fork() still favors the conventional WIN32 inheritence approach
which can be a little bit messy, but is *controlled* by guaranteeing
perfectly clean slates at both the spawning and execution boundaries
- sigset_t is now 64 bits. Having it be 128 bits was a mistake because
there's no reason to use that and it's only supported by FreeBSD. By
using the system word size, signal mask manipulation on Windows goes
very fast. Furthermore @asyncsignalsafe funcs have been rewritten on
Windows to take advantage of signal masking, now that it's much more
pleasant to use.
- All the overlapped i/o code on Windows has been rewritten for pretty
good signal and cancelation safety. We're now able to ensure overlap
data structures are cleaned up so long as you don't longjmp() out of
out of a signal handler that interrupted an i/o operation. Latencies
are also improved thanks to the removal of lots of "busy wait" code.
Waits should be optimal for everything except poll(), which shall be
the last and final demon we slay in the win32 i/o horror show.
- getrusage() on Windows is now able to report RUSAGE_CHILDREN as well
as RUSAGE_SELF, thanks to aggregation in the process manager thread.
This way complex runtime features (e.g. ftrace, symbol tables) can
always yoink zipos support. This is important now that apelink.com
automates embedding symbol tables for multiple cpus.
- Found some bugs in LLVM compiler-rt library
- The useless LIBC_STUBS package is now deleted
- Improve the overflow checking story even further
- Get chibicc tests working in MODE=dbg mode again
- The libc/isystem/ headers now have correctly named guards
- Work towards improving non-optimized build support
- Introduce MODE=zero which is -O0 without ASAN/UBSAN
- Use system GCC when ~/.cosmo.mk has USE_SYSTEM_TOOLCHAIN=1
- Have package.com check .privileged code doesn't call non-privileged
This change implements a new approach to function call logging, that's
based on the GCC flag: -fpatchable-function-entry. Read the commentary
in build/config.mk to learn how it works.
llama.com can now load weights that use the new file format which was
introduced a few weeks ago. Note that, unlike llama.cpp, we will keep
support for old file formats in our tool so you don't need to convert
your weights when the upstream project makes breaking changes. Please
note that using ggjt v3 does make avx2 inference go 5% faster for me.
- Fix UX issues with llama.com
- Do housekeeping on libm code
- Add more vectorization to GGML
- Get GGJT quantizer programs working well
- Have the quantizer keep the output layer as f16c
- Prefetching improves performance 15% if you use fewer threads
This change makes quantized models (e.g. q4_0) go 10% faster on Macs
however doesn't offer much improvement for Intel PC hardware.
This change syncs llama.cpp 699b1ad7fe6f7b9e41d3cb41e61a8cc3ea5fc6b5
which recently made a breaking change to nearly all its file formats
without any migration. Since that'll break hundreds upon hundreds of
models on websites like HuggingFace llama.com will support both file
formats because llama.com will never ever break the GGJT file format
Example use case for JSON completion:
$ m=opt
$ make -j16 m=$m o/$m/third_party/ggml/llama.com
$ o/$m/third_party/ggml/llama.com -m llama.bin -p '{"key": "life", "val": ' -r '}'
42}
This provides better control. More sophisticated facilities for
controlling text generation will be provided soon enough.
