ggml : add option for controlling work distribution across threads

[ggerganov]

See ggerganov/llama.cpp#1507

And comment: ggerganov/llama.cpp#1507 (comment)

I guess we can extend ggml to be able to choose work chunk distribution method - either at compile time, or via a context parameter. We can factor out the range selections from the ggml forward implementations to make implementation more concise and extensible in the future

---

Another thing to be investigated is the usage of sched_yield() and potentially making it user configurable:

ggerganov/whisper.cpp@09a6325

[IsaacDynamo]

Making this configurable would also be nice for the cuBLAS backend. When the whole model fits on the GPU, increasing the number of threads doesn't improve token/sec eval time.

But it does increase the CPU load on the system due to the busy loop. Even with n_thread = 1 , I suspect that a lot of CPU cycles are wasted in the busy loop.

So a yield flag would be a great addition to give the user control.

A busy-loop with a fallback to a yield might also be a good 'automatic' solution, that could be used as default.