Revert "Merge branch 'main' of github.com:ROCm/vllm" (vllm-project#72)

This reverts commit 2a3cbf9, reversing
changes made to 367aa5a.

tests/metrics/test_metrics.py

@@ -26,7 +26,6 @@ def test_metric_counter_prompt_tokens(
     vllm_model = vllm_runner(model,
                              dtype=dtype,
                              disable_log_stats=False,
-                             enforce_eager=True,
                              gpu_memory_utilization=0.4)
     tokenizer = vllm_model.model.get_tokenizer()
     prompt_token_counts = [len(tokenizer.encode(p)) for p in example_prompts]
@@ -60,7 +59,6 @@ def test_metric_counter_generation_tokens(
     vllm_model = vllm_runner(model,
                              dtype=dtype,
                              disable_log_stats=False,
-                             enforce_eager=True,
                              gpu_memory_utilization=0.4)
     vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
     tokenizer = vllm_model.model.get_tokenizer()
@@ -90,7 +88,6 @@ def test_metric_set_tag_model_name(vllm_runner, model: str, dtype: str,
     vllm_model = vllm_runner(model,
                              dtype=dtype,
                              disable_log_stats=False,
-                             enforce_eager=True,
                              gpu_memory_utilization=0.3,
                              served_model_name=served_model_name)
     stat_logger = vllm_model.model.llm_engine.stat_logger
@@ -108,7 +105,7 @@ def test_metric_set_tag_model_name(vllm_runner, model: str, dtype: str,
             f"{served_model_name[0]!r}\n"
             f"actual: {metrics_tag_content!r}")
 
-"""
+
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [4])
@@ -121,11 +118,11 @@ async def test_async_engine_log_metrics_regression(
     max_tokens: int,
     disable_log_stats: bool,
 ) -> None:
-    " ""
+    """
     Regression test ensuring async engine generates metrics
     when disable_log_stats=False
     (see: https://github.com/vllm-project/vllm/pull/4150#pullrequestreview-2008176678)
-    " ""
+    """
     engine_args = AsyncEngineArgs(model=model,
                                   dtype=dtype,
                                   disable_log_stats=disable_log_stats)
@@ -144,7 +141,6 @@ async def test_async_engine_log_metrics_regression(
                    len(example_prompts))
 
 
-
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [4])
@@ -162,15 +158,15 @@ def test_engine_log_metrics_regression(
     engine = LLMEngine.from_engine_args(engine_args)
     for i, prompt in enumerate(example_prompts):
         engine.add_request(
-            f\"request-id-{i}\",
+            f"request-id-{i}",
             prompt,
             SamplingParams(max_tokens=max_tokens),
         )
     while engine.has_unfinished_requests():
         engine.step()
 
     assert_metrics(engine, disable_log_stats, len(example_prompts))
-"""
+
 
 def assert_metrics(engine: LLMEngine, disable_log_stats: bool,
                    num_requests: int) -> None:

tests/worker/test_model_runner.py

@@ -137,7 +137,7 @@ def test_prepare_prompt(batch_size):
                             dtype=actual.dtype)
     torch.testing.assert_close(actual, expected)
 
-"""
+
 @pytest.mark.parametrize("batch_size", list(range(1, 257)))
 def test_prepare_decode_cuda_graph(batch_size):
     model_runner = _create_model_runner(
@@ -246,7 +246,7 @@ def test_prepare_decode_cuda_graph(batch_size):
                             device=actual.device,
                             dtype=actual.dtype)
     torch.testing.assert_close(actual, expected)
-"""
+
 
 def test_empty_seq_group():
     """Verify prepare prompt and decode returns empty output."""

tests/worker/test_swap.py

@@ -5,7 +5,7 @@
 from vllm.utils import get_distributed_init_method, get_ip, get_open_port
 from vllm.worker.worker import Worker
 
-"""
+
 def test_swap() -> None:
     # Configure the engine.
     engine_args = EngineArgs(model="facebook/opt-125m",
@@ -87,4 +87,3 @@ def test_swap() -> None:
         for src, dst in execute_model_req.blocks_to_swap_in:
             assert allclose(gpu_key_cache[dst], cpu_key_cache[src])
             assert allclose(gpu_value_cache[dst], cpu_value_cache[src])
-"""