fix: change logprobs type from bool to int in Completions endpoint (#5343)

Fixes #5253

The `logprobs` field in `OpenAICompletionRequestWithExtraBody` was typed
as `bool` but the OpenAI Completions API expects an integer (0–5)
indicating how many top log probabilities to include per output token.
This caused a 400 error when callers passed an int:

```
openai.BadRequestError: ... 'Input should be a valid boolean, unable to interpret input'
```

Changes:
- `models.py`: `logprobs: bool | None` → `logprobs: int | None` with
`ge=0, le=5`
- `llama-stack-spec.yaml`: matching schema update (`boolean` → `integer`
with min/max)

The Chat Completions model (`OpenAIChatCompletionRequestWithExtraBody`)
is unchanged — its `logprobs` is correctly a boolean there.

---------

Signed-off-by: Extra Small <littleshuai.bot@gmail.com>
Signed-off-by: Sébastien Han <seb@redhat.com>
Co-authored-by: Sébastien Han <seb@redhat.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: 3 people

client-sdks/stainless/openapi.yml

@@ -5008,9 +5008,11 @@ components:
           description: The logit bias to use.
         logprobs:
           anyOf:
-          - type: boolean
+          - type: integer
+            maximum: 5.0
+            minimum: 0.0
           - type: 'null'
-          description: The log probabilities to use.
+          description: Include the log probabilities on the logprobs most likely output tokens.
         max_tokens:
           anyOf:
           - type: integer

docs/static/deprecated-ogx-spec.yaml

@@ -1311,9 +1311,11 @@ components:
           description: The logit bias to use.
         logprobs:
           anyOf:
-          - type: boolean
+          - type: integer
+            maximum: 5.0
+            minimum: 0.0
           - type: 'null'
-          description: The log probabilities to use.
+          description: Include the log probabilities on the logprobs most likely output tokens.
         max_tokens:
           anyOf:
           - type: integer

docs/static/experimental-ogx-spec.yaml

@@ -1752,9 +1752,11 @@ components:
           description: The logit bias to use.
         logprobs:
           anyOf:
-          - type: boolean
+          - type: integer
+            maximum: 5.0
+            minimum: 0.0
           - type: 'null'
-          description: The log probabilities to use.
+          description: Include the log probabilities on the logprobs most likely output tokens.
         max_tokens:
           anyOf:
           - type: integer

docs/static/ogx-spec.yaml

@@ -4565,9 +4565,11 @@ components:
           description: The logit bias to use.
         logprobs:
           anyOf:
-          - type: boolean
+          - type: integer
+            maximum: 5.0
+            minimum: 0.0
           - type: 'null'
-          description: The log probabilities to use.
+          description: Include the log probabilities on the logprobs most likely output tokens.
         max_tokens:
           anyOf:
           - type: integer

docs/static/stainless-ogx-spec.yaml

@@ -5008,9 +5008,11 @@ components:
           description: The logit bias to use.
         logprobs:
           anyOf:
-          - type: boolean
+          - type: integer
+            maximum: 5.0
+            minimum: 0.0
           - type: 'null'
-          description: The log probabilities to use.
+          description: Include the log probabilities on the logprobs most likely output tokens.
         max_tokens:
           anyOf:
           - type: integer

src/ogx/providers/remote/inference/vertexai/vertexai.py

@@ -132,7 +132,7 @@ def from_openai_params(cls, params: OpenAICompletionRequestWithExtraBody) -> Gem
             candidate_count=params.n,
             max_output_tokens=params.max_tokens,
             stop_sequences=stop_sequences,
-            response_logprobs=params.logprobs or None,
+            response_logprobs=params.logprobs is not None and params.logprobs > 0,
         )
 
 

src/ogx_api/inference/models.py

@@ -948,7 +948,13 @@ class OpenAICompletionRequestWithExtraBody(BaseModel, extra="allow"):
         default=None, ge=-2.0, le=2.0, description="The penalty for repeated tokens."
     )
     logit_bias: dict[str, float] | None = Field(default=None, description="The logit bias to use.")
-    logprobs: bool | None = Field(default=None, description="The log probabilities to use.")
+    logprobs: int | None = Field(
+        default=None,
+        ge=0,
+        le=5,
+        strict=True,
+        description="Include the log probabilities on the logprobs most likely output tokens.",
+    )
     max_tokens: int | None = Field(default=None, ge=1, description="The maximum number of tokens to generate.")
     n: int | None = Field(default=None, ge=1, description="The number of completions to generate.")
     presence_penalty: float | None = Field(

tests/unit/providers/inference/vertexai/test_adapter_completion.py

@@ -164,25 +164,25 @@ async def test_logprobs_param_sets_response_logprobs(self, make_completion_adapt
         params = OpenAICompletionRequestWithExtraBody(
             model="google/gemini-2.5-flash",
             prompt="hi",
-            logprobs=True,
+            logprobs=5,
         )
         await adapter.openai_completion(params)
         call_kwargs = fake_client.aio.models.generate_content.call_args.kwargs
         config = call_kwargs["config"]
         assert config.response_logprobs is True
 
-    async def test_logprobs_false_sets_response_logprobs_none(self, make_completion_adapter):
-        """Test that logprobs false sets response logprobs none."""
+    async def test_logprobs_zero_sets_response_logprobs_false(self, make_completion_adapter):
+        """Test that logprobs zero sets response logprobs false."""
         adapter, fake_client = make_completion_adapter()
         params = OpenAICompletionRequestWithExtraBody(
             model="google/gemini-2.5-flash",
             prompt="hi",
-            logprobs=False,
+            logprobs=0,
         )
         await adapter.openai_completion(params)
         call_kwargs = fake_client.aio.models.generate_content.call_args.kwargs
         config = call_kwargs["config"]
-        assert getattr(config, "response_logprobs", None) is None
+        assert config.response_logprobs is False
 
     async def test_stream_raises_not_implemented_removed(self, monkeypatch):
         """Test that stream raises not implemented removed."""