so far my from-scratch mat*mats are still slower than just running more
invocations of the existing Metal ported mat*vec shaders - it should be
theoretically possible to make a mat*mat that's faster (for actual
mat*mat cases) than an optimal mat*vec, but it will need to be at
*least* as fast as the mat*vec op and then take special care to be
cache-friendly and save memory bandwidth, as the # of compute ops is the
same
* disable llama.cpp logging unless GPT4ALL_VERBOSE_LLAMACPP envvar is
nonempty
* make verbose flag for retrieve_model default false (but also be
overridable via gpt4all constructor)
should be able to run a basic test:
```python
import gpt4all
model = gpt4all.GPT4All('/Users/aaron/Downloads/rift-coder-v0-7b-q4_0.gguf')
print(model.generate('def fib(n):'))
```
and see no non-model output when successful
Use the $<CONFIG> generator expression instead of CMAKE_BUILD_TYPE. This
is needed because Visual Studio is a multi-configuration generator, so
we do not know what the build type will be until `cmake --build` is
called.
Fixes#1470
- Use F16 KV cache
- Store transposed V in the cache
- Avoid unnecessary Q copy
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
ggml upstream commit 0265f0813492602fec0e1159fe61de1bf0ccaf78
* Handle edge cases when generating embeddings
* Improve Python handling & add llmodel_c.h note
- In the Python bindings fail fast with a ValueError when text is empty
- Advice other bindings authors to do likewise in llmodel_c.h
* Makefiles, black, isort
* Black and isort
* unit tests and generation method
* chat context provider
* context does not reset
* Current state
* Fixup
* Python bindings with unit tests
* GPT4All Python Bindings: chat contexts, tests
* New python bindings and backend fixes
* Black and Isort
* Documentation error
* preserved n_predict for backwords compat with langchain
---------
Co-authored-by: Adam Treat <treat.adam@gmail.com>
* backend/gptj: use scratch buffers
reduces total memory required and makes eval buf not grow with n_past
* backend/mpt: use scratch bufs
* fix format-related compile warnings
* backend: factor out common structs in model code
prepping to hack on these by hopefully making there be fewer places to fix the same bug
rename
* use common buffer wrapper instead of manual malloc
* fix replit compile warnings
most of these can just shortcut out of the model loading logic llama is a bit worse to deal with because we submodule it so I have to at least parse the hparams, and then I just use the size on disk as an estimate for the mem size (which seems reasonable since we mmap() the llama files anyway)
use RTLD_LOCAL so that symbols are *only* exposed via dlsym
without this all symbols exported by the libs are available for symbol
resolution, resulting in different lib versions potentially resolving
*each other's* symbols, causing incredibly cursed behavior such as
https://gist.github.com/apage43/085c1ff69f6dd05387793ebc301840f6
* Initial Library Loader
* Load library as part of Model factory
* Dynamically search and find the dlls
* Update tests to use locally built runtimes
* Fix dylib loading, add macos runtime support for sample/tests
* Bypass automatic loading by default.
* Only set CMAKE_OSX_ARCHITECTURES if not already set, allow cross-compile
* Switch Loading again
* Update build scripts for mac/linux
* Update bindings to support newest breaking changes
* Fix build
* Use llmodel for Windows
* Actually, it does need to be libllmodel
* Name
* Remove TFMs, bypass loading by default
* Fix script
* Delete mac script
---------
Co-authored-by: Tim Miller <innerlogic4321@ghmail.com>
fixes a definite use-after-free and likely avoids some other
potential ones - std::string will convert to a std::string_view
automatically but as soon as the std::string in question goes out of
scope it is already freed and the string_view is pointing at freed
memory - this is *mostly* fine if its returning a reference to the
tokenizer's internal vocab table but it's, imo, too easy to return a
reference to a dynamically constructed string with this as replit is
doing (and unfortunately needs to do to convert the internal whitespace
replacement symbol back to a space)
- bug introduced in 0cb2b86730
- currently getting: `warning C5102: ignoring invalid command-line macro definition '/arch:AVX2'`
- solution is to use `_options(...)` not `_definitions(...)`
no effect, but avoids a *potential* bug later if we use
actualVocabSize - which is for when a model has a larger
embedding tensor/# of output logits than actually trained token
to allow room for adding extras in finetuning - presently all of our
models have had "placeholder" tokens in the vocab so this hasn't broken
anything, but if the sizes did differ we want the equivalent of
`logits[actualVocabSize:]` (the start point is unchanged), not
`logits[-actualVocabSize:]` (this.)
* porting over replit code model to gpt4all
* replaced memory with kv_self struct
* continuing debug
* welp it built but lot of sus things
* working model loading and somewhat working generate.. need to format response?
* revert back to semi working version
* finally got rid of weird formatting
* figured out problem is with python bindings - this is good to go for testing
* addressing PR feedback
* output refactor
* fixed prompt reponse collection
* cleanup
* addressing PR comments
* building replit backend with new ggmlver code
* chatllm replit and clean python files
* cleanup
* updated replit to match new llmodel api
* match llmodel api and change size_t to Token
* resolve PR comments
* replit model commit comment
Major change to the backend that allows for pluggable versions of llama.cpp/ggml. This was squashed merged from dlopen_backend_5 where the history is preserved.
Improves output quality by making these tokenizers more closely
match the behavior of the huggingface `tokenizers` based BPE
tokenizers these models were trained with.
Featuring:
* Fixed unicode handling (via ICU)
* Fixed BPE token merge handling
* Complete added vocabulary handling
