If I get some model in gguf format (say, glm-air-q8), use same type for kv cache (f16) and run it:

• on M2 Ultra
• on CUDA
• on CPU only

Will the types used for compute and activations be the same? Will we first de-quantize weights and cache to the same type (f16/f32/whatever), will we use same types for operations (operate on f16, accumulate to f32), etc.?

Or, will it depend on platform/kernel implementation on that platform? If it is different, what's the right place (pointers to code?) to learn these differences?

Thank you!