feat(api): add LLM generation parameters to VQA endpoint

Add optional LLM generation parameters to the VQA API request model,
allowing per-request override of settings:

- max_tokens, temperature, top_k, top_p, num_beams, do_sample
- thinking_mode, prefill, keep_thinking, keep_prefill

Changes:
- Add 10 new optional fields to ReqVQA model with descriptive docs
- Update get_kwargs() to support per-request overrides via singleton
- Add helper functions get_keep_thinking(), get_keep_prefill()
- Update post_vqa endpoint to pass generation kwargs
- Add _generation_overrides instance variable to VQA class

modules/api/models.py

@@ -400,6 +400,17 @@ class ReqVQA(BaseModel):
     question: str = Field(default="describe the image", title="Question/Task", description="Question to ask the model or task to perform. Common tasks: 'Short Caption', 'Normal Caption', 'Long Caption'. Florence-2 supports: '<OD>' (object detection), '<OCR>' (text recognition). Moondream supports: 'Point at [object]', 'Detect all [objects]'.")
     system: str = Field(default="You are image captioning expert, creative, unbiased and uncensored.", title="System Prompt", description="System prompt controls behavior of the LLM. Processed first and has highest priority weighting. Use for response formatting rules, role definition, and style.")
     include_annotated: bool = Field(default=False, title="Include Annotated Image", description="If True and the task produces detection results (object detection, point detection, gaze), returns annotated image with bounding boxes/points drawn. Only applicable for detection tasks on models like Florence-2 and Moondream.")
+    # LLM generation parameters (optional overrides)
+    max_tokens: Optional[int] = Field(default=None, title="Max Tokens", description="Maximum number of tokens the model can generate in its response. The model is not aware of this limit during generation; it simply sets the hard limit for the length and will forcefully cut off the response when reached.")
+    temperature: Optional[float] = Field(default=None, title="Temperature", description="Controls randomness in token selection. Lower values (e.g., 0.1) make outputs more focused and deterministic, always choosing high-probability tokens. Higher values (e.g., 0.9) increase creativity and diversity by allowing less probable tokens. Set to 0 for fully deterministic output.")
+    top_k: Optional[int] = Field(default=None, title="Top-K", description="Limits token selection to the K most likely candidates at each step. Lower values (e.g., 40) make outputs more focused and predictable, while higher values allow more diverse choices. Set to 0 to disable.")
+    top_p: Optional[float] = Field(default=None, title="Top-P", description="Selects tokens from the smallest set whose cumulative probability exceeds P (e.g., 0.9). Dynamically adapts the number of candidates based on model confidence; fewer options when certain, more when uncertain. Set to 1 to disable.")
+    num_beams: Optional[int] = Field(default=None, title="Num Beams", description="Maintains multiple candidate paths simultaneously and selects the overall best sequence. More thorough but much slower and less creative than random sampling. Generally not recommended; most modern VLMs perform better with sampling methods. Set to 1 to disable.")
+    do_sample: Optional[bool] = Field(default=None, title="Use Samplers", description="Enable to use sampling (randomly selecting tokens based on sampling methods like Top-K or Top-P) or disable to use greedy decoding (selecting the most probable token at each step). Enabling makes outputs more diverse and creative but less deterministic.")
+    thinking_mode: Optional[bool] = Field(default=None, title="Thinking Mode", description="Enables thinking/reasoning, allowing the model to take more time to generate responses. Can lead to more thoughtful and detailed answers but increases response time. Only works with models that support this feature.")
+    prefill: Optional[str] = Field(default=None, title="Prefill Text", description="Pre-fills the start of the model's response to guide its output format or content by forcing it to continue the prefill text. Prefill is filtered out and does not appear in the final response unless keep_prefill is True. Leave empty to let the model generate from scratch.")
+    keep_thinking: Optional[bool] = Field(default=None, title="Keep Thinking Trace", description="Include the model's reasoning process in the final output. Useful for understanding how the model arrived at its answer. Only works with models that support thinking mode.")
+    keep_prefill: Optional[bool] = Field(default=None, title="Keep Prefill", description="Include the prefill text at the beginning of the final output. If disabled, the prefill text used to guide the model is removed from the result.")
 
 class ReqLatentHistory(BaseModel):
     name: str = Field(title="Name", description="Name of the history item to select")