search

airockchip / rknn-llm

Other
402 stars 33 forks source link

E RKNN: [10:27:30.619] failed to allocate handle, ret: -1, errno: 14, errstr: Bad address Segmentation fault (core dumped) #4

Open fydeos-alex opened 7 months ago

fydeos-alex commented 7 months ago

This error occurred when I ran the Phi-2 model on rk3588 8G npu. I ran the Qwen1.8B successfully on it, but the phi didn't work on it. I am not sure whether the error was caused by the chip memory. other info:

system: openfyde
chip: Rockchip rk3588
RAM: 8G
ERROR: E RKNN: [10:27:30.619] failed to allocate handle, ret: -1, errno: 14, errstr: Bad address Segmentation fault (core dumped)

BTW, the model load speed was awful, what can I do to improve the experience?

sdrzmgy commented 7 months ago

Hello, have you built the kernal with the provided rknpu files? I came into trouble when I built the kernal for "implicit declaration of function 'vm_flags_set'" and "implicit declaration of function 'vm_flags_clear'"

fydeos-alex commented 7 months ago

I just followed the doc and ran the model of qwen, but it didn't work for phi-2 to run on the rk3588 8G npu. And you can try the newest openfyde version which has already updated the kernel.

sdrzmgy commented 7 months ago

I just followed the doc and ran the model of qwen, but it didn't work for phi-2 to run on the rk3588 8G npu. And you can try the newest openfyde version which has already updated the kernel.

thanks a lot! may you have a nice day!

Pelochus commented 7 months ago

Found the same problem. I was able to run Qwen 1.8B on Orange Pi 5 (RK3588S) 4GB, but can't run Phi-2. @fydeos-alex What options did you insert to the llm.build() function? I used the default:

ret = llm.build(do_quantization=True, optimization_level=1, quantized_dtype='w8a8', target_platform='rk3588')

fydeos-alex commented 7 months ago

Your memory is too small. The phi-2 model will cost more than 4GB when it is running in the INT8 quantization. I think this may also relate to your operating system memory strategy. Hope this will help.

Found the same problem. I was able to run Qwen 1.8B on Orange Pi 5 (RK3588S) 4GB, but can't run Phi-2. @fydeos-alex What options did you insert to the llm.build() function? I used the default:

ret = llm.build(do_quantization=True, optimization_level=1, quantized_dtype='w8a8', target_platform='rk3588')

Pelochus commented 6 months ago

Doesn't seem too logical for me, if Qwen 1.8B runs on my board and this (Phi-2) is barely bigger...

Besides, I was able to run a non-optimised Phi-2 on the same board through Ollama and OpenWebUI (which uses more RAM due to the UI and being a non-optimised version). Not only that, but your 8GB RAM can't also run it.

I believe this is some kind of bug. Were you able to run it in the end @fydeos-alex ?

fydeos-alex commented 6 months ago

Phi-2 didn't run successfully on my openfyde system in the end, and I am still waiting for more low-level support from RK Offical. BTW, Ollama loads Phi-2 in INT4 quantized, while rkllm only can convert to INT8, which means the rk model size is twice the Ollama. @Pelochus

Pelochus commented 6 months ago

Then that makes more sense. Still you should be able to run it on your 8GB model anyway.

Let's see if Rockchip releases a new RKLLM version with more LLMs support, fixed and updated Phi-2 and INT4 optimisation...

A bit off topic, but have you been able to run Llama 2 or TinyLlama?

@fydeos-alex

fydeos-alex commented 6 months ago

No, Llama 2 is too big for me, you know, I also need to run the openfyde system which already costs some of my runtime memory. I haven't tried TinyLlama yet and don't know whether the chip supports it or not. Please let me know if it works well. Qwen is enough for me now. 😇

Pelochus commented 6 months ago

I've been able to convert TinyLlama, however it doesn't work for me. It makes sense that Llama 2 doesn't work on 8GB though, it would be nice to see if someone with 32G or 16G can try to run it. I can't even convert it to RKLLM format due to the amount of RAM it requires to do so...

fydeos-alex commented 6 months ago

Nice try, let's wait for RockChip to continue their work.

noah003 commented 6 months ago

the same problem

Pelochus commented 6 months ago

@noah003 how much RAM, which SBC? Orange Pi 5?

Pelochus commented 6 months ago

You guys might find this useful:

https://www.reddit.com/r/RockchipNPU/comments/1c2m69i/comment/kzhnwo8/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

Edit: More useful links

https://www.reddit.com/r/RockchipNPU/s/F2P3KzBzlg

noah003 commented 6 months ago

@noah003 how much RAM, which SBC? Orange Pi 5?

16GB, Orange Pi 5