fix: use max_new_tokens instead of max_length in set_model

[AjayBandiwaddar]

Problem

set_model() in app/ai.py uses max_length=1024 when
initializing the pipeline, while __init__() correctly uses
max_new_tokens=1024.

These two parameters behave differently:

• max_length limits the total tokens including the input prompt
• max_new_tokens limits only the generated output tokens

This inconsistency means that after calling /change-model,
responses may be significantly shorter than expected because
input tokens consume part of the max_length budget.

This was missed in commit 659ff99 which updated __init__
from max_length to max_new_tokens but did not update
set_model.

Fix

Changed max_length=1024 to max_new_tokens=1024 in
set_model() to match the behavior in __init__().

Testing

The fix is verified by code inspection — set_model() now
matches the pipeline initialization parameters used in
__init__(). Both now use max_new_tokens=1024 consistently.

[AjayBandiwaddar]

Problem

set_model() in app/ai.py uses max_length=1024 when initializing the pipeline, while __init__() correctly uses max_new_tokens=1024.

These two parameters behave differently:

• max_length limits the total tokens including the input prompt
• max_new_tokens limits only the generated output tokens

This inconsistency means that after calling /change-model, responses may be significantly shorter than expected because input tokens consume part of the max_length budget.

This was missed in commit 659ff99 which updated __init__ from max_length to max_new_tokens but did not update set_model.

Fix

Changed max_length=1024 to max_new_tokens=1024 in set_model() to match the behavior in __init__().

Testing

The fix is verified by code inspection — set_model() now matches the pipeline initialization parameters used in __init__(). Both now use max_new_tokens=1024 consistently.

Problem

set_model() in app/ai.py uses max_length=1024 when initializing the pipeline, while __init__() correctly uses max_new_tokens=1024.

These two parameters behave differently:

• max_length limits the total tokens including the input prompt
• max_new_tokens limits only the generated output tokens

This inconsistency means that after calling /change-model, responses may be significantly shorter than expected because input tokens consume part of the max_length budget.

This was missed in commit 659ff99 which updated __init__ from max_length to max_new_tokens but did not update set_model.

Fix

Changed max_length=1024 to max_new_tokens=1024 in set_model() to match the behavior in __init__().

Testing

The fix is verified by code inspection — set_model() now matches the pipeline initialization parameters used in __init__(). Both now use max_new_tokens=1024 consistently.

Note: This PR was developed with AI assistance (Claude). As per Sugar Labs contributing guidelines, I'm disclosing this. The AI helped me structure the fix, but I verified, and tested each change myself.