Eval bug: context length incorrectly capped in server for yarn extendable context models

[steampunque]

Name and Version

latest codebase at b8850

Operating systems

Linux

GGML backends

CUDA

Hardware

9900k/4070

Models

Qwen 3 8B

Problem description & steps to reproduce

If yarn is used to extend context length (many qwen models such as Qwen3 use yarn extension), the extension is blocked by this logic in server-context.cpp which incorrectly limits the context length to the base training context:

   int n_ctx_slot = llama_n_ctx_seq(ctx);
    if (n_ctx_slot > n_ctx_train) {
        SRV_WRN("the slot context (%d) exceeds the training context of the model (%d) - capping\n", n_ctx_slot, n_ctx_train);
        n_ctx_slot = n_ctx_train;
    }

Instead just a warning should be output as follows:

   int n_ctx_slot = llama_n_ctx_seq(ctx);
    if (n_ctx_slot > n_ctx_train) {
        SRV_WRN("the slot context (%d) exceeds the training context of the model (%d)\n", n_ctx_slot, n_ctx_train);
    }

First Bad Commit

Unknown, whenever the cap was added.

Relevant log output

Logs