perf: Significant litellm import bottleneck in MTP runtime

[nk243547]

Feature Description

Examples

# Current (problematic):
import litellm  # - Around 2s delay at module load

def call_model(...):
    response = litellm.completion(...)  # Already loaded

# Proposed (lazy loading):
def get_litellm():
    import litellm  # - Only imported when actually needed
    return litellm

def call_model(...):
    litellm = get_litellm()  # - Import happens here
    response = litellm.completion(...)
 -->