llamacpp: HTTP 400 on valid unanchored JSON Schema patterns in tool definitions

Abierto
#1,100 1 comentario 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 1 día

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
3/5
Tiempo estimado
Medio día
Aptitud para principiantes
45/100
Tipo de issue
Error
Claridad
Bastante claro
Estado de actividad
Activo
Stack tecnológico
docker, go
Área
backend

Línea de trabajo

The fix is a version bump of the llama.cpp backend pin, LLAMA_SERVER_VERSION (currently b9879), so find where that variable is set in the repo. Confirm the upstream change in ggml-org/llama.cpp#26939 is included in the new build, then run the curl reproduction with the ^ses pattern and check for HTTP 200. Done means that request succeeds and the existing tool-call tests still pass.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

Summary

Docker Model Runner's llama.cpp backend rejects any chat completion request whose tool JSON Schema contains a valid pattern that is not fully anchored with ^/$. The whole request fails with HTTP 400 before inference starts.

Upstream llama.cpp fixed this class of failure in ggml-org/llama.cpp#26939 (merged 2026-08-20), but DMR still pins LLAMA_SERVER_VERSION=b9879, which predates the fix.

Real-world impact: OpenAI-compatible clients that emit prefix patterns such as ^ses cannot use grammar-constrained tool calling with DMR.

Environment

  • Docker Desktop 29.8.2 (client 29.9.1), Windows 11 10.0.26300
  • Docker Model Runner server v1.2.8
  • Model: docker.io/ai/gpt-oss:20b
  • Backend fingerprint reported by DMR: b1-72874f5 (LLAMA_SERVER_VERSION=b9879 in the repo)

Reproduction

Minimal, standalone request (no OpenCode or other client involved):

curl -sS http://localhost:12434/engines/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{
    "model": "docker.io/ai/gpt-oss:20b",
    "messages": [{"role": "user", "content": "hi"}],
    "max_tokens": 5,
    "tools": [{
      "type": "function",
      "function": {
        "name": "demo",
        "parameters": {
          "type": "object",
          "properties": {
            "sessionID": {"type": "string", "pattern": "^ses"}
          },
          "required": ["sessionID"]
        }
      }
    }]
  }'

Result: HTTP 400

{
  "error": {
    "code": 400,
    "message": "JSON schema conversion failed:\nPattern must start with '^' and end with '$'\nPattern must start with '^' and end with '$'",
    "type": "invalid_request_error"
  }
}

A/B test, changing only the pattern:

Pattern Result
^ses HTTP 400 (schema conversion failure)
^ses.*$ HTTP 200, completion succeeds

Expected behavior

A valid JSON Schema pattern should not fail the entire request. Unsupported regex constructs should degrade gracefully (for example, fall back to an unconstrained string) as upstream llama.cpp now does, rather than returning HTTP 400.

Suggested fix

Bump LLAMA_SERVER_VERSION from b9879 to a llama.cpp build that contains ggml-org/llama.cpp#26939, or cherry-pick that change.

Additional context

  • Concrete trigger: OpenCode v2 declares the built-in subagent tool parameter sessionID as {"type":"string","pattern":"^ses"} (session IDs start with ses). The same request works against Ollama, which tolerates unanchored patterns.
  • Client-side workaround: rewrite patterns to be fully anchored (^ses → ^ses.*$) before dispatch.
  • Related: docker/model-runner#534 ("Failed to parse grammar - opencode", closed).
  • Upstream PR: https://github.com/ggml-org/llama.cpp/pull/26939
Lenguaje dominante
Go
Estrellas
662
Forks
156
Merge medio
15 h 26 min
PR fusionados (30 d)
17

Preparar el entorno

  • Incluye un Dockerfile o un archivo de Docker Compose
  • Sin plantilla de pull request
  • Sin guía de contribución

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de docker/model-runner

Todos los issues de docker/model-runner

Issues similares

Más issues de Go

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.