ci: [SYCL] Use main GPU and enable sysman

[qnixsynapse]

Set environmental variables for using the main GPU for now and enable SYSMAN for correct GPU memory reporting.

[ggerganov]

If you add the keyword ggml-ci somewhere in the commit message, it will trigger the CI on this branch, so we can see if it works before merging.

[qnixsynapse]

Ah. It is broken!

[ggerganov]

It could be related to the shapes of the specific model. Here is how to get and convert the model:

llama.cpp/ci/run.sh

Lines 411 to 421 in 48d7021

	# pythia_1.4b
	function gg_run_pythia_1_4b {
	cd ${SRC}
	gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
	gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
	gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
	gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
	gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin

llama.cpp/ci/run.sh

Lines 435 to 437 in 48d7021

	python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf

[qnixsynapse]

@Alcpz @ggerganov Confirmed Q4_0 broken by 08d5986

To quickly test, run llama-cli with this specific model with GGML_SYCL_DISABLE_OPT=1.

[Alcpz]

Thanks for pinpointing this. We may want to disable the GGML_SYCL_DISABLE_OPT path by default until it is fixed.
@qnixsynapse It's probably worth it to start a discussion in the SYCL discussion thread so everyone using the backend is aware.

[qnixsynapse]

We may want to disable the GGML_SYCL_DISABLE_OPT path by default until it is fixed.

Probably in a separate PR. This PR should be merged IMO(It enables the necessary flags).

I will enable SYCL CI on whisper.cpp and ggml repos tomorrow.