ollama - 💡(How to fix) Fix Add enable_thinking parameter to disable CoT/Reasoning generation [2 comments, 2 participants]

Fix Plan

To address the issue, we will implement a native boolean flag enable_thinking in the API and model file. This flag will allow users to explicitly disable Chain-of-Thought (CoT) reasoning for models like Qwen 3.5 and GPT-OSS.

Implementation Steps

• Add a new parameter enable_thinking to the API with a default value of True.
• Update the model file to accept the enable_thinking parameter and toggle reasoning generation accordingly.
• Modify the inference logic to bypass reasoning generation when enable_thinking is False.

Example Code

# API update
def generate_text(prompt, enable_thinking=True):
    # ...

# Model file update
class Qwen35Model:
    def __init__(self, enable_thinking=True):
        self.enable_thinking = enable_thinking

    def generate_text(self, prompt):
        if not self.enable_thinking:
            # Bypass reasoning generation
            return self.direct_generate_text(prompt)
        # ...

# Inference logic update
def inference(model, prompt):
    if not model.enable_thinking:
        # Direct generation
        return model.direct_generate_text(prompt)
    # ...

Verification

To verify the fix, test the API and model with the enable_thinking flag set to False. Measure the performance and response quality to ensure that the fix addresses the current issues.

Extra Tips

• Document the enable_thinking parameter in the API and model documentation to ensure users understand its purpose and usage.
• Consider adding additional logging or monitoring to track the usage and effectiveness of the enable_thinking flag.