cherry-pick vllm-project#1675 vllm-project#1589 vllm-project#1501 vllm-project#1420 vllm-project#1328 from v0.9.1-dev to main

Signed-off-by: 22dimensions <waitingwind@foxmail.com>

Author: shikang-hangzhou

examples/offline_data_parallel.py

@@ -54,17 +54,16 @@
                     --master-port=13345
 """
 
-import os
-from time import sleep
 import contextlib
 import gc
+import os
+from time import sleep
 
 import torch
-
 from vllm import LLM, SamplingParams
-from vllm.utils import get_open_port
 from vllm.distributed.parallel_state import (  # noqa E402
     destroy_distributed_environment, destroy_model_parallel)
+from vllm.utils import get_open_port
 
 os.environ["VLLM_USE_MODELSCOPE"] = "True"
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

tests/e2e/singlecard/test_aclgraph.py

@@ -20,6 +20,8 @@
 Run `pytest tests/compile/test_aclgraph.py`.
 """
 
+import os
+
 import pytest
 import torch
 from vllm import LLM, SamplingParams
@@ -84,3 +86,16 @@ def test_deepseek_raises_error(monkeypatch: pytest.MonkeyPatch) -> None:
                        max_model_len=1024,
                        enforce_eager=False)
         assert "ACL Graph does not support deepseek" in str(excinfo.value)
+
+
+@pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "0",
+                    reason="aclgraph only support on v1")
+@pytest.mark.parametrize("model", MODELS)
+def test_ray_backend_sets_no_compilation(
+        model: str, monkeypatch: pytest.MonkeyPatch) -> None:
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+        runner = VllmRunner(model,
+                            enforce_eager=False,
+                            distributed_executor_backend="ray")
+        assert runner.model.llm_engine.vllm_config.compilation_config.level == 0

vllm_ascend/ascend_config.py

@@ -76,6 +76,31 @@ def __init__(self, torchair_graph_config):
             raise ValueError(
                 "graph_batch_sizes_init is only valid when graph_batch_sizes is empty"
             )
+        if not self.enabled:
+            if self.use_cached_graph:
+                raise RuntimeError(
+                    "use_cached_graph is valid only when Torchair graph mode is enabled"
+                )
+            if self.graph_batch_sizes:
+                raise RuntimeError(
+                    "graph_batch_sizes is valid only when Torchair graph mode is enabled"
+                )
+            if self.graph_batch_sizes_init:
+                raise RuntimeError(
+                    "graph_batch_sizes_init is valid only when Torchair graph mode is enabled"
+                )
+            if self.enable_multistream_mla:
+                raise RuntimeError(
+                    "enable_multistream_mla is valid only when Torchair graph mode is enabled"
+                )
+            if self.enable_multistream_moe:
+                raise RuntimeError(
+                    "enable_multistream_moe is valid only when Torchair graph mode is enabled"
+                )
+            if self.enable_kv_nz:
+                raise RuntimeError(
+                    "enable_kv_nz is valid only when Torchair graph mode is enabled"
+                )
 
 
 class AscendSchedulerConfig: