Android port error

[NoNamedCat]

$ ./llama -m vicuna.bin main: seed = 1681462772 llama.cpp: loading model from vicuna.bin llama_model_load_internal: format = ggjt v1 (latest) llama_model_load_internal: n_vocab = 32001 llama_model_load_internal: n_ctx = 512 llama_model_load_internal: n_embd = 4096 llama_model_load_internal: n_mult = 256 llama_model_load_internal: n_head = 32 llama_model_load_internal: n_layer = 32 llama_model_load_internal: n_rot = 128 llama_model_load_internal: ftype = 2 (mostly Q4_0) llama_model_load_internal: n_ff = 11008 llama_model_load_internal: n_parts = 1 llama_model_load_internal: model size = 7B llama_model_load_internal: ggml ctx size = 59.11 KB llama_model_load_internal: mem required = 5809.33 MB (+ 1026.00 MB per state) afdsan: attempted to close file descriptor 3, expected to be unowned, actually owned by FILE* 0x7e73c0a018 Aborted

Don't know what is happening

I'm having the same problem. The only method that works for me is to use an outdated 7b llama model.

Here's what happens when I try and run anything different:

~/llama.cpp $ ./main
main: seed = 1681658496
llama.cpp: loading model from models/llama-7B/ggml-model.bin
llama_model_load_internal: format = ggjt v1 (latest)
llama_model_load_internal: n_vocab = 32000
llama_model_load_internal: n_ctx = 512
llama_model_load_internal: n_embd = 4096
llama_model_load_internal: n_mult = 256
llama_model_load_internal: n_head = 32
llama_model_load_internal: n_layer = 32
llama_model_load_internal: n_rot = 128
llama_model_load_internal: ftype = 2 (mostly Q4_0)
llama_model_load_internal: n_ff = 11008
llama_model_load_internal: n_parts = 1
llama_model_load_internal: model size = 7B
llama_model_load_internal: ggml ctx size = 59.11 KB
llama_model_load_internal: mem required = 5809.32 MB (+ 1026.00 MB per state)
fdsan: attempted to close file descriptor 3, expected to be unowned, actually owned by FILE* 0x758a3a7888
Aborted

[slaren]

The only method that works for me is to use an outdated 7b llama model.

Seems like this could be problem with mmap. Can you see if using --no-mmap works with new models?

The only method that works for me is to use an outdated 7b llama model.

Seems like this could be problem with mmap. Can you see if using --no-mmap works with new models?

Wow, I searched for days and you pinned down the problem right away.

--no-mmap allowed my newer 7b llama model to run. Thank you.

[slaren]

I suspect this is the problem:
https://github.com/ggerganov/llama.cpp/blob/3173a62eb9f90b94fb3184131032c1c8b7aa8d86/llama_util.h#L173-L179

@comex can you confirm is this is an error? Looks like the fd shouldn't be closed since it is owned by the FILE*.

[BartlomiejLewandowski]

I can confirm the same issue on android, --no-map is a valid work around.

[slaren]

Can you check if removing the line with close(fd) in llama_util.h line 179 fixes the issue?

Can you check if removing the line with close(fd) in llama_util.h line 179 fixes the issue?

Hi, removing the lines close(fd) from llama_util.h allowed my newer models to run without the -no-mmap.

Thank you.

[NoNamedCat]

For me is working now removing that line from the code. Now we can close this?