dcdpr / jp / 24202651754

67%

Build Type

push

github

Committed by web-flow

Commit Message

feat(llm, cerebras): Add Cerebras LLM provider (#531)

Adds support for the Cerebras inference API as a new LLM provider,
giving users access to fast inference on models like `llama3.1-8b`,
`qwen-3-235b-a22b-instruct-2507`, `gpt-oss-120b`, and `zai-glm-4.7`.

The provider is OpenAI-compatible and reuses the SSE streaming types
from the `llamacpp` provider (now `pub(crate)`), extended with a serde
`alias = "reasoning"` on the `reasoning_content` field so that
Cerebras's `parsed` reasoning format is handled without duplication.

Reasoning is supported for `gpt-oss-120b` (low/medium/high effort) and
`zai-glm-4.7` (reasoning on by default; can be disabled with
`reasoning_effort: none`). All four known models are mapped to their
context-window and output-token limits.

Structured output is supported via a `transform_schema` step that adapts
the generic JSON Schema produced by `jp` to Cerebras's strict subset:
`additionalProperties: false` is enforced on all objects, all object
properties are added to `required`, and unsupported constraints
(`minItems`, `maxItems`, `pattern`, `format`) are stripped and appended
to the nearest `description` field as a soft hint to the model.

Signed-off-by: Jean Mertz <git@jeanmertz.com>
Co-authored-by: rgrant <rgrant@contract.design>

Coverage Stats

380 of 420 new or added lines in 6 files covered. (90.48%)

1 existing line in 1 file now uncovered.

20711 of 31934 relevant lines covered (64.86%)

301.51 hits per line