(cls)
| 45 | |
| 46 | @classmethod |
| 47 | def setUpClass(cls): |
| 48 | try: |
| 49 | llm = LLM( |
| 50 | model=MODEL_NAME, |
| 51 | max_num_batched_tokens=4096, |
| 52 | tensor_parallel_size=1, |
| 53 | lm_head_fp32=True, |
| 54 | engine_worker_queue_port=int(os.getenv("FD_ENGINE_QUEUE_PORT")), |
| 55 | cache_queue_port=int(os.getenv("FD_CACHE_QUEUE_PORT")), |
| 56 | ) |
| 57 | cls.llm = weakref.proxy(llm) |
| 58 | except Exception as e: |
| 59 | print(f"Setting up LLM failed: {e}") |
| 60 | raise unittest.SkipTest(f"LLM initialization failed: {e}") |
| 61 | |
| 62 | @classmethod |
| 63 | def tearDownClass(cls): |