mlxrunner: add logprobs support

Match the ollamarunner and OpenAI semantics: raw, full-vocab log-softmax
with the top-K ranked by probability. Skipped on the GPU when the request
doesn't ask for logprobs so decode doesn't pay for it otherwise.
This commit is contained in:
Jesse Gross
2026-04-16 13:06:17 -07:00
parent 5d1021603a
commit 24e038d56a
11 changed files with 463 additions and 167 deletions

View File

@@ -406,10 +406,6 @@ func TestAPIShowModel(t *testing.T) {
}
func TestAPIGenerateLogprobs(t *testing.T) {
if testModel != "" {
// Logprobs requires runner support (e.g. llama.cpp has it, MLX does not).
t.Skip("logprobs not supported by all runners")
}
ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
defer cancel()
@@ -523,10 +519,6 @@ func TestAPIGenerateLogprobs(t *testing.T) {
}
func TestAPIChatLogprobs(t *testing.T) {
if testModel != "" {
// Logprobs requires runner support (e.g. llama.cpp has it, MLX does not).
t.Skip("logprobs not supported by all runners")
}
ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
defer cancel()