Jesse Gross
147509c0c5
The MLX runner accepted the API's format field but did not enforce it:
requests asking for JSON or a JSON Schema got unconstrained text, and
clients had no way to tell.
Enforce format with xgrammar: each sampling step masks the logits to
the tokens the grammar allows, so every emitted token and the end of
generation are valid under the constraint. Sampling, penalties, and
logprobs see the constrained distribution, and "json" yields a JSON
object, as the API documents and the llama-server path already
enforces. Only sampling waits on the mask; the forward pass is
dispatched before it, so constrained decoding stays pipelined.
The grammar engine is a dynamic library alongside MLX; when it is
missing, plain inference is unaffected and structured requests fail
with an explicit error. Constrained requests decode without
speculative decoding for now.
Decoding 256 tokens of a book-list schema on qwen3.8:27b-mlx (M5 Max,
seed 42, thinking off); pre-decode is the request time spent before
the first token:
unconstrained ~65 tok/s pre-decode ~70 ms
unconstrained, no draft ~32 tok/s pre-decode ~70 ms
JSON schema ~32 tok/s pre-decode ~70 ms
Schema and draft-less decoding are equal to within 0.1 tok/s in
paired adjacent requests, and a cold grammar compile adds nothing
measurable to pre-decode. The gap to unconstrained decoding is the
disabled draft model.
Fixes #16563
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Integration Tests
This directory contains integration tests to exercise Ollama end-to-end to verify behavior
By default, these tests are disabled so go test ./... will exercise only unit tests. To run integration tests, pass the integration tag and one of the scoped tags:
go test -tags=integration,fast -v -count 1 ./integration/
go test -tags=integration,release -v -count 1 -timeout 30m ./integration/
go test -tags=integration,library -v -count 1 -timeout 120m ./integration/
Tags:
fast: quick runner/model smoke coverage.release: release regression coverage.library: broad library coverage requiring about 2.5 TiB of disk space.
Scope wiring and model selections live in integration/reg_fast_test.go, integration/reg_release_test.go, and integration/reg_library_test.go.
The integration tests have 2 modes of operating.
- By default, on Unix systems, they will start the server on a random port, run the tests, and then shutdown the server. On Windows you must ALWAYS run the server on OLLAMA_HOST for the tests to work.
- If
OLLAMA_TEST_EXISTINGis set to a non-empty string, the tests will run against an existing running server, which can be remote based on yourOLLAMA_HOSTenvironment variable
Set OLLAMA_TEST_LOG_SERVER=1 to print the managed server log after each test
run, even when the tests pass. This only applies when the integration test
harness starts the server.
Important
Before running the tests locally without the "test existing" setting, compile ollama from the top of the source tree
go build .in addition to GPU support with cmake if applicable on your platform. The integration tests expect to find an ollama binary at the top of the tree.
Testing a New Model
When implementing new model architecture, use OLLAMA_TEST_MODEL to run the
integration suite against your model with either the fast or release coverage.
# Build the binary first
go build .
# Run integration tests against it
OLLAMA_TEST_MODEL=mymodel go test -tags=integration,fast -v -count 1 ./integration/