* Reduce memory usage and allocate enough memory for large contexts
* Simpler scratch buffer usage
* Reenable BLAS for quantized mul_mat
* Fix number of layers in 30B and 65B
* Fix KV cache size for F32
changes to EOS behavior in interactive and reverse prompt handling broke instruct mode by erroneously injecting instruct mode's reverse prompt and an extra newline.
* Support calling mlock() on loaded model data on Linux and macOS
This is enabled by a new --mlock command line option.
Using mlock() disables swapping and memory compression for the model
data. Doing so can be useful on systems where the model takes up a
large fraction of system RAM. In my experience, macOS is quite eager to
start compressing llama.cpp's memory, which then makes it halt for a few
seconds while it decompresses, even with a model that uses "only" 25GB
out of 32GB.
Of course, this comes at the cost of forcing the system to swap or
compress other processes' memory instead, so it needs to be used with
care and shouldn't be enabled by default.
In theory it should be possible to support this on Windows as well using
VirtualLock(), but I'm not much of a Windows user.
* Update llama.cpp
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* working but ugly
* add arg flag, not working on embedding mode
* typo
* Working! Thanks to @nullhook
* make params argument instead of hardcoded boolean. remove useless time check
* start doing the instructions but not finished. This probably doesnt compile
* Embeddings extraction support
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Improve interactive mode's coherence after EOS
Aims to improve coherence and ability to resume the interactive session when the user is given input back after an end of text token is reached.
Not sure what token 13 is or why it seems to help. See conversation for examples.
* Make newline token a constant
* dynamically determine newline token
* relocate previous newline token const
* cleanup whitespace
* print a new line on end of text in interactive
this may need to be looked into further when not using a reverse prompt
* only print manual newline with reverse prompt
fix formatting of reverse prompts so they don't end up at the end of the current line while not introducing unnecessary new lines otherwise
* alternate approach to replace end of text tokens
* Inject the reverse prompt again after eos in interactive mode
* tokenize reverse prompt when needed
makes this PR compatible with https://github.com/ggerganov/llama.cpp/pull/330
* tokenize and inject only first reverse prompt
thanks to tjohnman
* tokenize first reverse prompt once
* add newline token
* add newline token
* tokenize/inject reverse prompt for refactor
this doesn't seem right though
* tokenize nothing for antiprompt if no reverse
* Update main.cpp
* Update main.cpp
* tokenize and inject reverse prompt as needed
this doesn't seem to work if the reverse prompt is tokenized outside earlier on
* not needed
* remove newline token
* remove newline token
* tokenize newline token
* add space to comment
* Update main.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Slaren <2141330+slaren@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Don't force immediate interactive without -i
Sometimes we might want to use a reverse prompt but we want to let the
model generate tokens right after the initial prompt. So we don't force
user input mode if the -i flag wasn't specified and instead let it run
until we encounter the reverse prompt.
This gives use some more flexibility, since it doesn't force the user to
enter a newline if they want to let the model generate text right after
the initial prompt and only be asked for input if the reverse prompt is
encountered.
The `--interactive-first` flag is reintroduced to force the old
behavior. `-r` behaves like `-i` plus introduces a reverse prompt (it
can be specified more than once).
* Update help output.
---------
Co-authored-by: Johnman <tjohnman@github>
* Major refactoring - introduce C-style API
* Clean up
* Add <cassert>
* Add <iterator>
* Add <algorithm> ....
* Fix timing reporting and accumulation
* Measure eval time only for single-token calls
* Change llama_tokenize return meaning
* Improve performance by changing std::map to std::unordered_map and std::map<id, token> id_to_token; to std::vector<token> id_to_token;
* fix last commit on gpt_vocab_init add vocab.id_to_token.resize(vocab.token_to_id.size());
* Removed include <map>
* Nest struct token score inside gpt_vocab
* renamed token to tok
* [WIP, broken] Importer for GPTQ quantized LLaMA models
Based on: https://github.com/qwopqwop200/GPTQ-for-LLaMa
Current status: Something is busted. The output starts out decent, but
quickly degrades into gibberish. This doesn't happen with either the
original GPTQ-for-LLaMa using the same weights, or llama.cpp when using
weights quantized by its own quantizer. Is there a bug in the
conversion script that somehow only comes into play with a large context
size?
I did notice one potential issue. It's clearly not the main cause of
the gibberish, since it doesn't happen when using q4_1 weights quantized
by llama.cpp itself, but it seems concerning. When doing a matrix
multiplication of f16 * f32 => f32 or q4_1 * f32 => f32, at least when
the multiplication is not done with BLAS, the intermediate results are
stored in the smaller format rather than f32. This seems like an
unnecessary waste of precision, especially in the q4_1 case.
I was originally hoping to validate the results by matching the Python
implementation's output exactly, but precision and non-associativity
issues make this very difficult, including when performing matrix
multiplications and, especially, computing norms.
Anyway, design details:
The models being imported store per-layer weights in essentially q4_1
format, although the addend and scale are shared across an entire row
rather than every group of 32 weights. This script duplicates the
addend and scale to match ggml's expectations, at the cost of wasting
some memory.
However, there are two differences which I accommodated changing the
output format (and adding corresponding support to main.cpp) rather than
having the script match the existing one:
- The tok_embeddings and output weights (i.e. the weights that aren't
per-layer) are f16 instead of q4_1. They could be converted to q4_1,
and the impact of the loss of precision would probably be low, but
this would rule out exactly matching the Python implementation's
output for validation.
- There is no sharding, since the input doesn't have it, and for a
CPU-only implementation it seems more useful to avoid having to deal
with multiple files.
The new format is differentiated from existing q4_1 format by changing
the 'f16' header flag to a new value, 4. That said, I think a cleaner
approach would be to change main.cpp to support loading each tensor with
an arbitrary sharding configuration and type rather than hardcoding
specific combinations of types. So far I've wasted too much time
debugging to try implementing this...
* Add missing permutation. Now it works.
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Compute perplexity over prompt
* More accurate perplexity calculation - over all logits in the context window (so 512x more tokens!)
* Output all perplexitiies
* Add timing/ETA
* Enable ANSI colors on Windows 10+
On older versions function will silently fail without any ill effects
* Do not call SetConsoleMode if the mode is already set
* Update main.cpp
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Add test-tokenizer-0 to do a few tokenizations - feel free to expand
* Added option to convert-pth-to-ggml.py script to dump just the vocabulary
* Added ./models/ggml-vocab.bin containing just LLaMA vocab data (used for tests)
* Added utility to load vocabulary file from previous point (temporary implementation)
* Avoid using std::string_view and drop back to C++11 (hope I didn't break something)
* Rename gpt_vocab -> llama_vocab
* All CMake binaries go into ./bin/ now
* potential out of bounds read
* fix quantize
* style
* Update convert-pth-to-ggml.py
* mild cleanup
* don't need the space-prefixing here rn since main.cpp already does it
* new file magic + version header field
* readme notice
* missing newlines
Co-authored-by: slaren <2141330+slaren@users.noreply.github.com>
* fix coloring of last `n_batch` of prompt, and refactor line input
* forgot the newline that needs to be sent to the model
* (per #283) try to force flush of color reset in SIGINT handler
* Use F16 for memory_k and memory_v
* add command line switch to use f16 instead of f32 for memory k+v
---------
Co-authored-by: Ty Everett <ty@tyweb.us>
* Implement non-greedy tokenizer that tries to maximize token lengths
* Insert single space in front of the prompt
- this is to match original llama tokenizer behavior
---------
Co-authored-by: Jakub Horak <jakub.horak@ibawizard.net>
* added ctx_size parameter
* added it in more places
* Apply suggestions from code review
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* fixed color reset on exit
* added sigint handler for ansi_color_reset
* Update main.cpp
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
