TST Bump absolute tolerance for test (#1891)

The test test_4bit_lora_mixed_adapter_batches_lora allclose can fail on
some systems, even though it passes on others (like CI). Increase the
tolerance slightly to get rid of this.

tests/test_common_gpu.py

@@ -801,7 +801,7 @@ def test_4bit_lora_mixed_adapter_batches_lora(self):
  with torch.inference_mode():
  out_adapter1 = model(**inputs).logits
 
- atol, rtol = 1e-5, 1e-5
+ atol, rtol = 3e-5, 1e-5
  # sanity check, outputs have the right shape and are not the same
  assert len(out_base) >= 3
  assert len(out_base) == len(out_adapter0) == len(out_adapter1)