fix(minillm): qwen2 teacher student model unmatched model vocab size

[anjilab]

Problem Statement

When distilling from teacher model Qwen/Qwen2.5-7B-Instruct to student model Qwen/Qwen2.5-3B, mismatched model vocab size issue.

RuntimeError: The size of tensor a (152064) must match the size of tensor b (151936) at non-singleton dimension 2

Description of Change

Though both model has same tokenizer vocab size of 151665, their model vocab size (embedding) are different where Qwen/Qwen2.5-7B-Instruct has 152064 and Qwen/Qwen2.5-3B has 151936. It is safe to truncate the extra padded tokens from the teacher logits (as no real token ID points to them).

Added a conditional fix to handle Qwen2 model types where vocab size mismatch may occur. The change ensures that teacher logits are truncated to match the student’s vocabulary size before loss computation.