server : remove self-extend features #9860
Conversation
| if (!params.ctx_shift) { | ||
| // this check is redundant (for good) | ||
| // we should never get here, because generation should already stopped in process_token() | ||
| slot.release(); | ||
| send_error(slot, "context shift is disabled", ERROR_TYPE_SERVER); | ||
| continue; | ||
| } |
There was a problem hiding this comment.
@ngxson I think the comment is not entirely correct because in process_token() we check agains the training context length (n_ctx_train), while the slot's context slot.n_ctx could be smaller. What do you think?
There was a problem hiding this comment.
Hmm no, I did add a check against slot.n_ctx. Is this what you're looking for?
llama.cpp/examples/server/server.cpp
Lines 1184 to 1191 in 11ac980
There was a problem hiding this comment.
Ah, I missed that, thanks.
Shouldn't we check this actually:
if (slot.n_prompt_tokens + slot.n_decoded >= n_ctx) {Hmm, or maybe:
if (slot.n_past + slot.n_decoded >= n_ctx) {Anyway, I will figure it out as I'm looking into this logic currently.
There was a problem hiding this comment.
Ah yeah I misunderstood n_decoded. Yeah, maybe we even need (int) system_tokens.size() + slot.n_prompt_tokens because system_tokens is already in KV cache before the first decode.
Thanks for looking into this.
There was a problem hiding this comment.
No sorry I haven't see #9811
ggml-ci
ggerganov
force-pushed
the
gg/server-remove-self-extend
branch
from
e5f74fe
to
8a1f439
Compare
* server : remove self-extend ggml-ci * server : fix context limit check to use slot.n_past ggml-ci
* server : remove self-extend ggml-ci * server : fix context limit check to use slot.n_past ggml-ci
target #9857, fix #9859
Drop support for the self-extend related arguments: