why LLAMA_MAX_SEQ limit to 64?

[yysj-zq]

#https://github.com/ggml-org/llama.cpp/blob/5fc7856815920f828f9e90cb759ac82c7f0c1ea5/src/llama-cparams.h#L7

As the title suggests. can anyone answer it?

[ggerganov]

See the answer here: 4f81b33#commitcomment-159972791

[yysj-zq]

i saw it, thank you.

[yysj-zq]

and here is another question could you answer it? what happened in llama.cpp from commit 557924f to commit 860a9e4(Tue Nov 19 2024 to Tue Jun 17 2025)? Originally, 100 parallel were possible, but now even 50 can cause oom(same model\ ctx_per_seq\ etc.)