LLama.cpp Got Screwd

https://github.com/ggml-org/llama.cpp/discussions/25482

by trilogic • 14 days ago

11 5 14 days ago

trilogic2• 14 days ago

We are heading towards a closed and monitored system, better said a proprietary one soon. Get the last genuine llama.cpp build before to late b9925.

The CLI now asking for access to load the model?

Here the change that screw everything:

b9927 @github-actions github-actions released this 17 hours ago b9927 c264f65 Details cli : move to HTTP-based implementation (#24948)

cli: move to HTTP-based implementation

wip

working

remote server ok

cli support router mode

Co-authored-by: Piotr Wilkin ilintar@gmail.com

case: router with only one model

Apply suggestions from code review

Co-authored-by: Piotr Wilkin (ilintar) piotr.wilkin@syndatis.com

remove outdated comment

use destructor instead

add ftype

cli-view --> cli-ui

pimpl

no more json in header

nits fixes

also show model aliases

Adios llama.cpp.

rimworld• 14 days ago

only a matter of time, thx for heads up

magicalhippo2• 14 days ago

Here[1] is the offending pull request. The llama.cpp devs are trying to refactor the code, so the "engine" bits are shared between the current server and client applications.

Georgi himself noted it should be possible to replace the transport layer from HTTP to say an internal message queue, and the main developer of the PR said the HTTP bit is already abstracted and intended to include a non-HTTP transport example, so seems clear to me it's not intended to be HTTP-only.

[1] https://github.com/ggml-org/llama.cpp/pull/24948

trilogic2• 14 days ago

How is adding compulsory HTTP to CLI necessary for refactoring the code?

You realize that now the CLI needs http permissions to load a model locally. Still confused why, the server already does that by all means.

First they changed the remote downloading weights in server. Then they approved and merged claude code snippets. Now mandatory monitoring.

Everyone has a price

magicalhippo2• 14 days ago

If you check out the preceeding PR[1], linked from the merged PR, the use-case is if you got a llama-server running but is restricted to console, or you just want to debug, so you want to connect to it using llama-cli.

So it's easy to see how HTTP became the primary transport, because it facilitated the motivating use-case.

[1] https://github.com/ggml-org/llama.cpp/pull/21674

LLama.cpp Got Screwd

Related Stories