Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Implement multimodal models (LLaVA) #3436

Merged
merged 36 commits into from Oct 12, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
36 commits
Select commit Hold shift + click to select a range
59aa1ac
WIP: start implementing LLaVA
monatis Oct 2, 2023
0f0e7c6
rm scratch buf for now, will revert after cleanup
monatis Oct 2, 2023
7e9120f
LLaVA image encoder is working. will combine with llama
monatis Oct 2, 2023
d37ed47
Add llava inference code, but it's buggy. debugging
monatis Oct 3, 2023
8690f42
LLaVA is working e2e, needs to optimize memory allocation + cleanup
monatis Oct 7, 2023
94eeac3
Use ggml_allocr + rm unnecessary code
monatis Oct 8, 2023
0c2bd79
fix: crlf -> lf
monatis Oct 8, 2023
204d08b
fix: new line at EoF
monatis Oct 8, 2023
95da79e
fix: trailing whitespace
monatis Oct 8, 2023
2a04d0b
Merge branch 'master' into llava
monatis Oct 8, 2023
444dbce
Add readme
monatis Oct 9, 2023
8af7e21
Update readme
monatis Oct 9, 2023
54495c9
Some cleanup
monatis Oct 9, 2023
9b0ec4d
Are you happy editorconfig?
monatis Oct 9, 2023
8278a73
rm unused batch image preprocessing
monatis Oct 9, 2023
d78e816
rm unused import
monatis Oct 9, 2023
4759bfd
fix: rm designated initializers
monatis Oct 9, 2023
325d240
introduce pad-to-square mode for non-square images
monatis Oct 9, 2023
d75a031
are you happy editorconfig?
monatis Oct 9, 2023
ae01c85
gitignore /llava
monatis Oct 9, 2023
5009ae9
Handle cases where image file does not exist
monatis Oct 9, 2023
96171de
add llava target to Makefile
monatis Oct 9, 2023
d640aae
add support for 13b model variant
monatis Oct 10, 2023
587bde8
Maybe seed is unlucky?
monatis Oct 11, 2023
f1564bb
Merge branch 'master' into llava
monatis Oct 11, 2023
ab21587
Check if apples are compared to apples
monatis Oct 11, 2023
0409ae0
are you happy editorconfig?
monatis Oct 11, 2023
f0f7834
Use temperature = 0.1 by default
monatis Oct 11, 2023
2bc1710
command line: use gpt_params_parse()
monatis Oct 11, 2023
1403d87
Merge master and fix conflicts
monatis Oct 11, 2023
dc913ea
minor
monatis Oct 12, 2023
56ccf97
handle default n_predict
monatis Oct 12, 2023
e9534ea
fix typo
monatis Oct 12, 2023
346e3c1
Merge branch 'master' into llava
ggerganov Oct 12, 2023
4bc5c9c
llava : code formatting, rename files, fix compile warnings
ggerganov Oct 12, 2023
0bd7e69
do not use Wno-cast-qual for MSVC
monatis Oct 12, 2023
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
1 change: 1 addition & 0 deletions .gitignore
Expand Up @@ -44,6 +44,7 @@ models-mnt
/infill
/libllama.so
/llama-bench
/llava
/main
/metal
/perplexity
Expand Down
5 changes: 4 additions & 1 deletion Makefile
@@ -1,7 +1,7 @@
# Define the default target now so that it is always the first target
BUILD_TARGETS = \
main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
simple batched batched-bench save-load-state server embd-input-test gguf llama-bench baby-llama beam-search \
simple batched batched-bench save-load-state server embd-input-test gguf llama-bench llava baby-llama beam-search \
speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o

# Binaries only useful for tests
Expand Down Expand Up @@ -627,6 +627,9 @@ convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggm
llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -Wno-cast-qual

baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

Expand Down
14 changes: 14 additions & 0 deletions common/common.cpp
Expand Up @@ -384,6 +384,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
break;
}
params.lora_base = argv[i];
} else if (arg == "--mmproj") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.mmproj = argv[i];
} else if (arg == "--image") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.image = argv[i];
} else if (arg == "-i" || arg == "--interactive") {
params.interactive = true;
} else if (arg == "--embedding") {
Expand Down Expand Up @@ -703,6 +715,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" -np N, --parallel N number of parallel sequences to decode (default: %d)\n", params.n_parallel);
printf(" -ns N, --sequences N number of sequences to decode (default: %d)\n", params.n_sequences);
printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
printf(" --image IMAGE_FILE path to an image file. use with multimodal models\n");
if (llama_mlock_supported()) {
printf(" --mlock force system to keep model in RAM rather than swapping or compressing\n");
}
Expand Down
4 changes: 4 additions & 0 deletions common/common.h
Expand Up @@ -104,6 +104,10 @@ struct gpt_params {
bool numa = false; // attempt optimizations that help on some NUMA systems
bool verbose_prompt = false; // print prompt tokens before generation
bool infill = false; // use infill mode

// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector
std::string image = ""; // path to an image file
};

bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
Expand Down