Generation using GPU offloading is much slower than without

[Barafu]

Recently, generating a text with large preexisting context has become very slow when using GPU offloading. I tried llama-cpp-python versions 0.1.61 and 0.1.57 - I get the same behavior.

I am running an Oobabooga installation on Windows 11. The machine is AMD3950X, 32 gb RAM, 3070Ti (8 Gb VRAM). When I try to extend an already long text, with GPU offloading used, I get these numbers:

llama_print_timings:        load time = 36434.52 ms
llama_print_timings:      sample time =    40.21 ms /    65 runs   (    0.62 ms per token)
llama_print_timings: prompt eval time = 146712.89 ms /  1294 tokens (  113.38 ms per token)
llama_print_timings:        eval time = 15896.38 ms /    64 runs   (  248.38 ms per token)
llama_print_timings:       total time = 163419.18 ms
Output generated in 163.82 seconds (0.39 tokens/s, 64 tokens, context 1623, seed 1809972029)
Llama.generate: prefix-match hit

note the prompt eval time - two minutes. But without the offloading I get:

llama_print_timings:        load time =  5647.99 ms
llama_print_timings:      sample time =    66.99 ms /   111 runs   (    0.60 ms per token)
llama_print_timings: prompt eval time = 10762.28 ms /  1245 tokens (    8.64 ms per token)
llama_print_timings:        eval time = 34006.30 ms /   110 runs   (  309.15 ms per token)
llama_print_timings:       total time = 46578.66 ms
Output generated in 46.96 seconds (2.34 tokens/s, 110 tokens, context 1576, seed 874103700)
Llama.generate: prefix-match hit

Those runs use the same model (13Bq5_1), same prompt, similar context. The only difference is the argument --n-gpu-layers 26 in the first case and none in the second.

I did check the loading logs for lines like

llama_model_load_internal: offloading 16 layers to GPU
llama_model_load_internal: total VRAM used: 4143 MB

to confirm that it loads as I intended.

[RiyanParvez]

Same here, much much slower without gpu offloading in my case its close to 80ish ms per token, but with off loading its 700ish ms per token...

[RiyanParvez]

And it takes more time to load the model too.

[TonyWeimmer40]

Experiencing the same exact issue with the debug version CUBLAS. For context the release version gives buggy output. #1735

The buggy Release version with the gibberish output is far faster for me, but the properly working Debug version is extremely slow like observed here.

UPDATE: Using latest build made both of these issues seemingly go away on release version, debug is still slow I don't know why. Anyway it is far faster now, if anyone else is having trouble I recommend to build Release x64 from source and use the latest version (as of now it is 74a69d2).

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.