Flat fix

[peardox]

Fixingcommit conflicts

[peardox]

@danbev your #3090 and others are now in master so will this get thru CI now?

I just merged and everything looks good

[peardox]

In CI Ruby don't like my -DBINDINGS_FLAT (which also creates a GGML_BINDINGS_FLAT)

The whole point of the switch is that it makes the additions / changes transparent to other systems

[KitaitiMakoto]

@peardox Hi, I sent a pull request to fix Ruby problem to your branch: WhisperBooks#1

[peardox]

Removed superfluous line in ggml-backend.h
Added documentation via comments at end of whisper.cpp explaining placement
Renamed BINDING_FLAT -> WHISPER_BINDINGS_FLAT (in multiple files)
Reduced included in whisper-flat.h
Added params.use_gpu trigger to server + wchess examples if WHISPER_BINDING_FLAT
Checked both with and without -DWHISPER_BINDING_FLAT library exports to ensure extra exposed functions only if enabled
Tested both with and without -DWHISPER_BINDING_FLAT (7 examples modify execution paths with WHISPER_BINDING_FLAT enabled)

[peardox]

@danbev Made changes - can you stick it thru CI just to check pls

[peardox]

Arggh - ruby failed again
I altered @KitaitiMakoto fix to suit new naming of WHISPER_BINDINGS_FLAT

[KitaitiMakoto]

Seems a backup file ( ggml/include/ggml-backend.h.bak ) is commited and causes Ruby bindings' test failure.

[peardox]

Seems a backup file ( ggml/include/ggml-backend.h.bak ) is commited and causes Ruby bindings' test failure.

Dunno where that came from :(

[ggerganov]

What is the goal of this change?

Note that this is not suitable to merge - it's probably better to keep in a separate fork.

[peardox]

What is the goal of this change?

First it should be mentioned that WHISPER_BINDINGS_FLAT is specifically designed to only do anything if passed to the build command.
It is entirely conditional, if built without WHISPER_BINDINGS_FLAT the code doesn't exist

The main purpose is to make -DBUILD_SHARED_LIBS=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON work
If built with those switches no examples will work as the backends are not available when the model is loaded (try it).
This results in null pointers being used to obtain device informtion.

If devices are loaded via e.g. whisper_flat_backend_load_all() or ggml_backend_try_load_best("cpu", nullptr) before the model is loaded this situation is resolved and examples all work.
However, currently, whisper_load_backends also tries to load backends.
This causes double loading of backends or in the case of only loading e.g. cpu forcing the load of all others as well (which may be undesirable)

The rest of the stuff is for non C++ language support.
C++ specific structures such as std::vector are used freely throughout which causes issues for other languages.
In the case of whisper_init_from_file_with_params state is assigned but is unavailable to non-C++ languages - this is resolved by whisper_flat_get_state_from_context as whisper_state is required extensively
whisper_flat_get_activity_with_state uses a whisper_state to obtain timing (not available as data otherwise)
The whisper_flat...backend functions all deal with getting information about the backends assigned to the current model (so you can tell which ones are actually being used) - this is also unavailable otherwise