Eval bug: reasoning off gives reasoning medium for gpt-oss

[thomas-0816]

Name and Version

./llama-cli --version
load_backend: loaded RPC backend from /home/tb/code/llama-b8287/libggml-rpc.so
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon 780M Graphics (RADV PHOENIX) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
load_backend: loaded Vulkan backend from /home/tb/code/llama-b8287/libggml-vulkan.so
load_backend: loaded CPU backend from /home/tb/code/llama-b8287/libggml-cpu-zen4.so
version: 8287 (acb7c7906)
built with GNU 11.4.0 for Linux x86_64

Using --reasoning off for gpt-oss gives "Reasoning: medium", maybe it should give "Reasoning: low" ?

Using --reasoning-budget 0 for gpt-oss gives "Reasoning: medium", maybe it should give "Reasoning: low" ?

Using --chat-template-kwargs '{"reasoning_effort": "low"}' for gpt-oss gives "Reasoning: low" (ok).

./llama-server -hf unsloth/gpt-oss-20b-GGUF:F16 --threads -1 --parallel 1 --ctx-size 16384 --temp 1.0 --min-p 0.0 --top-p 1.0 --top-k 0 --n_predict 4096 --reasoning off --direct-io 2>&1 | grep -i reasoning
Reasoning: medium

maybe related to #20297 @pwilkin

Operating systems

Linux

GGML backends

Vulkan

Hardware

AMD Ryzen 7 PRO 7840U

Models

No response

Problem description & steps to reproduce

use "--reasoning off" with "unsloth/gpt-oss-20b-GGUF:F16"

First Bad Commit

No response

Relevant log output

Logs

[pwilkin]

@aldehir While you're refactoring the GPT OSS template, care to relay --reasoning off to also set reasoning_effort to none?

[aldehir]

Not supported. Low is the lowest it goes.

[aviallon]

@aldehir doesn't it have 'minimal'?

[aldehir]

@aldehir doesn't it have 'minimal'?

You might be thinking of it's step-cousin, Solar Open 100B, which got lost in the purgemerge. GPT-OSS only has low, medium, and high.