tests/unittest/_torch/modules/tests_lora_modules/test_qwen3_sanity.py (1)

128-149: Minor: Inner loop break doesn't exit outer loop.

When any_differ is set in the inner logprobs comparison loop (line 146), the code breaks out of the inner loop but continues iterating through remaining (lora_out, base_out) pairs unnecessarily. This is functionally correct but slightly inefficient.

♻️ Optional: Use labeled loop pattern or helper

 def _assert_lora_changes_output(out_lora, out_base):
     """Assert that LoRA produces at least one different output (tokens or logprobs)."""
     any_differ = False
     for lora_out, base_out in zip(out_lora, out_base, strict=True):
+        if any_differ:
+            break
         lora_ids = lora_out.outputs[0].token_ids
         base_ids = base_out.outputs[0].token_ids
         if lora_ids != base_ids:
             any_differ = True
             break

         # Even if tokens match, logprobs should differ
         lp_lora = lora_out.outputs[0].logprobs
         lp_base = base_out.outputs[0].logprobs
         if lp_lora and lp_base:
             for lp_w, lp_wo in zip(lp_lora, lp_base, strict=True):
                 val_w = next(iter(lp_w.values())).logprob
                 val_wo = next(iter(lp_wo.values())).logprob
                 if abs(val_w - val_wo) > 1e-6:
                     any_differ = True
                     break
+            if any_differ:
+                break

     assert any_differ, "LoRA outputs identical to base model (same tokens AND same logprobs)"

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@tests/unittest/_torch/modules/tests_lora_modules/test_qwen3_sanity.py` around
lines 128 - 149, The helper _assert_lora_changes_output currently sets
any_differ inside the inner logprobs loop then breaks only that inner loop,
causing the outer for-loop over (lora_out, base_out) to continue unnecessarily;
to fix, make the function return immediately when a difference is detected (or
set any_differ and break the outer loop explicitly) so we stop processing
further pairs as soon as a differing token or logprob is found—update the inner
loop where val_w vs val_wo triggers any_differ to either call return or break
out of both loops (e.g., by returning from _assert_lora_changes_output) to
short-circuit the check.