- 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
- Get mprotect_test working on aarch64
- Get completion working on python.com repl again
- Improve quality of printvideo.com and printimage.com
- Fix bug in openpty() so examples/script.c works again
This change greatly reduces the number of modules that need to be
compiled. The only issue right now is that sometimes when viewing
symbol table entries, the aliased symbol is chosen.
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.
- Now using 10x better GCD semaphores
- We now generate Linux-like thread ids
- We now use fast system clock / sleep libraries
- The APE M1 loader now generates Linux-like stacks
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.
This change progresses our AARCH64 support:
- The AARCH64 build and tests are now passing
- Add 128-bit floating-point support to printf()
- Fix clone() so it initializes cosmo's x28 TLS register
- Fix TLS memory layout issue with aarch64 _Alignas vars
- Revamp microbenchmarking tools so they work on aarch64
- Make some subtle improvements to aarch64 crash reporting
- Make kisdangerous() memory checks more accurate on aarch64
- Remove sys_open() since it's not available on Linux AARCH64
This change makes general improvements to Cosmo and Redbean:
- Introduce GetHostIsa() function in Redbean
- You can now feature check using pledge(0, 0)
- You can now feature check using unveil("",0)
- Refactor some more x86-specific asm comments
- Refactor and write docs for some libm functions
- Make the mmap() API behave more similar to Linux
- Fix WIFSIGNALED() which wrongly returned true for zero
- Rename some obscure cosmo keywords from noFOO to dontFOO
There's a new program named ape/ape-m1.c which will be used to build an
embeddable binary that can load ape and elf executables. The support is
mostly working so far, but still chasing down ABI issues.
- 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
- Perform some housekeeping on scalar math function code
- Import ARM's Optimized Routines for SIMD string processing
- Upgrade to latest Chromium zlib and enable more SIMD optimizations
- Introduce epoll_pwait()
- Rewrite -ftrapv and ffs() libraries in C code
- Use more FreeBSD code in math function library
- Get significantly more tests passing on qemu-aarch64
- Fix many Musl long double functions that were broken on AARCH64
make -j8 o//third_party/radpajama/radpajama.com
make -j8 o//third_party/radpajama/radpajama-chat.com
This change gets the radpajama.mk config working. This package depends
on THIRD_PARTY_GGML but it's configured to call ggjt_v1(), so that the
library will provide the old quantizers. The ggml_quantize_chunk() API
will now dispatch to older quantizers based on the configured version.
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
- Utilities like pledge.com now build
- kprintf() will no longer balk at 48-bit addresses
- There's a new aarch64-dbg build mode that should work
- gc() and defer() are mostly pacified; avoid using them on aarch64
- THIRD_PART_STB now has Arm Neon intrinsics for fast image handling
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.
