[JonSnow1807]

Overview

Fixes #30835

An HTTP request containing an unsupported grammar_triggers[].type value can reach the sampling code and trigger GGML_ASSERT, terminating the entire llama-server process rather than rejecting the malformed request.

This PR validates grammar trigger types at the request-parsing boundary, before invalid enum values can reach sampler initialization. Unsupported types are rejected with a client error while existing behavior for valid trigger types is preserved.

Additional information

Build: cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CURL=ON && ninja -C build llama-server (macOS arm64), model ggml-org/test-model-stories260K.

Before the fix (master 9febe63):

POST /completion {"prompt":"hi","n_predict":1}                                                -> 200
POST /completion {"prompt":"hi","n_predict":1,"grammar_triggers":[{"type":99,"value":"x"}]}  -> connection dropped
log: common/sampling.cpp:258: GGML_ASSERT(false && "unknown trigger type") failed, process exited

After the fix the same request returns 400 with Field 'grammar_triggers': invalid grammar trigger type: 99 and a following valid /completion returns 200.

New tests in tools/server/tests/unit/test_completion.py:

• test_completion_invalid_grammar_trigger_type[-1|4|99]: negative and out-of-range values get a 400 and the server still answers a subsequent valid request.
• test_completion_valid_grammar_trigger_types: all four valid types (0..3) in one lazy-grammar request are accepted.

cd tools/server/tests
PYTEST_WORKERS=1 ./tests.sh unit/test_completion.py -k grammar_trigger
4 passed, 43 deselected in 5.73s

With tools/server/server-common.h reverted the invalid-type test fails with http.client.RemoteDisconnected (server aborted). The full unit/test_completion.py file: 45 passed, 1 skipped, 1 failed (test_completion_stream_with_openai_library_stops, a @slow test that needs a Phi-3.5 download and fails identically on unmodified master).

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Claude Code assisted with investigating the root cause, preparing the implementation, and developing regression tests. I reviewed the changes and am responsible for their correctness.

[ggml-gh-bot]

Hi @JonSnow1807, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.