[Misc] Refactor additional_config

.github/workflows/vllm_ascend_test.yaml

@@ -112,7 +112,13 @@ jobs:
             pytest -sv tests/singlecard/test_scheduler.py
             # guided decoding doesn't work, fix it later
             # pytest -sv tests/singlecard/test_guided_decoding.py.py
-            pytest -sv tests/singlecard/ --ignore=tests/singlecard/test_offline_inference.py --ignore=tests/singlecard/test_scheduler.py --ignore=tests/singlecard/test_guided_decoding.py
+            # test_ascend_config.py should be ran separately because it will regenerate the global config many times.
+            pytest -sv tests/singlecard/test_ascend_config.py
+            pytest -sv tests/singlecard/ \
+            --ignore=tests/singlecard/test_offline_inference.py \
+            --ignore=tests/singlecard/test_scheduler.py \
+            --ignore=tests/singlecard/test_guided_decoding.py \
+            --ignore=tests/singlecard/test_ascend_config.py
           else
             pytest -sv tests/multicard/test_ilama_lora_tp2.py
             VLLM_USE_MODELSCOPE=True pytest -sv tests/multicard/ --ignore=tests/multicard/test_ilama_lora_tp2.py
@@ -128,11 +134,14 @@ jobs:
             # guided decoding doesn't work, fix it later
             # pytest -sv tests/singlecard/test_guided_decoding.py.py
             pytest -sv tests/singlecard/test_camem.py
+            # test_ascend_config.py should be ran separately because it will regenerate the global config many times.
+            pytest -sv tests/singlecard/test_ascend_config.py
             pytest -sv tests/singlecard/ \
               --ignore=tests/singlecard/test_offline_inference.py \
               --ignore=tests/singlecard/test_scheduler.py \
               --ignore=tests/singlecard/test_guided_decoding.py \
-              --ignore=tests/singlecard/test_camem.py
+              --ignore=tests/singlecard/test_camem.py \
+              --ignore=tests/singlecard/test_ascend_config.py
           else
             pytest -sv tests/multicard/test_ilama_lora_tp2.py
             # Fixme: run VLLM_USE_MODELSCOPE=True pytest -sv tests/multicard/test_offline_inference_distributed.py will raise error.

docs/source/index.md

@@ -46,6 +46,7 @@ faqs
 user_guide/suppoted_features
 user_guide/supported_models
 user_guide/env_vars
+user_guide/additional_config
 user_guide/release_notes
 :::
 

docs/source/user_guide/additional_config.md

@@ -0,0 +1,70 @@
+# Additional Configuration
+
+addintional configuration is a mechanism provided by vLLM to allow plugins to control inner behavior by their own. vLLM Ascend uses this mechanism to make the project more flexible.
+
+## How to use
+
+With either online mode or offline mode, users can use additional configuration. Take Qwen3 as an example:
+
+**Online mode**:
+
+```bash
+vllm serve Qwen/Qwen3-8B --additional-config='{"config_key":"config_value"}'
+```
+
+**Offline mode**:
+
+```python
+from vllm import LLM
+
+LLM(model="Qwen/Qwen3-8B", additional_config={"config_key":"config_value"})
+```
+
+### Configuration options
+
+The following table lists the additional configuration options available in vLLM Ascend:
+
+| Name | Type | Default | Description |
+| ---- | ---- | ------- | ----------- |
+| `torchair_graph_config` | dict | `{}` | The config options for torchair graph mode |
+| `ascend_scheduler_config` | dict | `{}` | The config options for ascend scheduler  |
+| `expert_tensor_parallel_size` | str | `1` | Expert tensor parallel size the model to use. |
+
+The details of each config option are as follows:
+
+**torchair_graph_config**
+
+| Name | Type | Default | Description |
+| ---- | ---- | ------- | ----------- |
+| `enabled` | bool | `False` | Whether to enable torchair graph mode |
+| `use_cached_graph` | bool | `False` | Whether to use cached graph |
+| `graph_batch_sizes` | list[int] | `[]` | The batch size for torchair graph cache |
+| `graph_batch_sizes_init` | bool | `False` | Init graph batch size dynamically if `graph_batch_sizes` is empty |
+
+**ascend_scheduler_config**
+
+| Name | Type | Default | Description |
+| ---- | ---- | ------- | ----------- |
+| `enabled` | bool | `False` | Whether to enable ascend scheduler for V1 engine|
+
+ascend_scheduler_config also support the options from [vllm scheduler config](https://docs.vllm.ai/en/stable/api/vllm/config.html#vllm.config.SchedulerConfig). For example, you  can add `chunked_prefill_enabled: true` to ascend_scheduler_config as well.
+
+### Example
+
+A full example of additional configuration is as follows:
+
+```
+{
+    "torchair_graph_config": {
+        "enabled": true,
+        "use_cached_graph": true,
+        "graph_batch_sizes": [1, 2, 4, 8],
+        "graph_batch_sizes_init": true
+    },
+    "ascend_scheduler_config": {
+        "enabled": true,
+        "chunked_prefill_enabled": true,
+    },
+    "expert_tensor_parallel_size": 1
+}
+```

examples/dp_offline/data_parallel.py

@@ -62,7 +62,9 @@ def main():
               max_num_seqs=num_seqs,
               additional_config={
                   'expert_tensor_parallel_size': etp_size,
-                  'enable_graph_mode': False,
+                  'torchair_graph_config': {
+                      'enabled': False,
+                  },
               })
 
     outputs = llm.generate(prompts, sampling_params)

tests/long_term/spec_decode/e2e/conftest.py

@@ -167,17 +167,17 @@ def run_equality_correctness_test(
 
     # TODO current torchair graph mode needs clean torchair cache.
     # if do not clean, it will raise error
-    additional_config = common_llm_kwargs.get("additional_config")
-    enable_graph_mode = additional_config.get(
-        "enable_graph_mode") if additional_config else False
+    torchair_graph_enabled = common_llm_kwargs.get(
+        "additional_config", {}).get("torchair_graph_config",
+                                     {}).get("enabled", False)
 
     with vllm_runner(**org_args) as vllm_model:
-        if enable_graph_mode:
+        if torchair_graph_enabled:
             _clean_torchair_cache()
         org_outputs = vllm_model.generate_w_logprobs(prompts, sampling_params)
 
     with vllm_runner(**sd_args) as vllm_model:
-        if enable_graph_mode:
+        if torchair_graph_enabled:
             _clean_torchair_cache()
         if ensure_all_accepted or expected_acceptance_rate is not None:
             # Force log interval to be 0 to catch all metrics.

tests/long_term/spec_decode/e2e/test_mtp_correctness.py

@@ -218,7 +218,9 @@ def test_mtp_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
     "common_llm_kwargs",
     [{
         "additional_config": {
-            'enable_graph_mode': True,
+            'torchair_graph_config': {
+                "enabled": True,
+            },
         },
 
         # Print spec metrics.
@@ -262,7 +264,9 @@ def test_mtp_e2e_greedy_correctness_torchair_graph(
     "common_llm_kwargs",
     [{
         "additional_config": {
-            'enable_graph_mode': True,
+            'torchair_graph_config': {
+                "enabled": True,
+            },
         },
 
         # Print spec metrics.

tests/multicard/test_dynamic_npugraph_batchsize.py

@@ -18,8 +18,6 @@
 import torch
 from vllm import LLM, SamplingParams
 
-from vllm_ascend.utils import vllm_version_is
-
 MODELS = [
     "Qwen/Qwen2.5-0.5B-Instruct",
 ]
@@ -32,9 +30,6 @@
 ]
 
 
-@pytest.mark.skipif(
-    (vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1")),
-    reason="aclgraph not supported in v0.8.5 and v0.8.5.post1")
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
 @pytest.mark.parametrize("max_tokens", [64])

tests/multicard/test_offline_inference_distributed.py

@@ -31,9 +31,7 @@
 
 def test_models_distributed_QwQ():
     example_prompts = [
-        "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
-        "Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
-        "Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
+        "Hello, my name is",
     ]
     dtype = "half"
     max_tokens = 5
@@ -48,9 +46,7 @@ def test_models_distributed_QwQ():
 
 def test_models_distributed_DeepSeek():
     example_prompts = [
-        "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
-        "Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
-        "Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
+        "Hello, my name is",
     ]
     dtype = "half"
     max_tokens = 5

tests/compile/test_aclgraph.py → tests/singlecard/test_aclgraph.py

@@ -28,16 +28,12 @@
 
 from tests.conftest import VllmRunner
 from tests.model_utils import check_outputs_equal
-from vllm_ascend.utils import vllm_version_is
 
 MODELS = ["Qwen/Qwen2.5-0.5B-Instruct"]
 
 
 @pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "0",
                     reason="aclgraph only support on v1")
-@pytest.mark.skipif(
-    (vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1")),
-    reason="aclgraph not supported in v0.8.5 and v0.8.5.post1")
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("max_tokens", [32])
 def test_models(
@@ -88,9 +84,6 @@ def test_models(
 
 @pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "0",
                     reason="aclgraph only support on v1")
-@pytest.mark.skipif(
-    (vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1")),
-    reason="aclgraph not supported in v0.8.5 and v0.8.5.post1")
 def test_deepseek_raises_error(monkeypatch: pytest.MonkeyPatch) -> None:
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_MODELSCOPE", "True")

tests/singlecard/test_ascend_config.py

@@ -0,0 +1,118 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import pytest
+
+from tests.conftest import VllmRunner
+from vllm_ascend.ascend_config import clear_ascend_config, get_ascend_config
+
+
+def _clean_up_ascend_config(func):
+
+    def wrapper(*args, **kwargs):
+        clear_ascend_config()
+        func(*args, **kwargs)
+        clear_ascend_config()
+
+    return wrapper
+
+
+@_clean_up_ascend_config
+def test_run_without_ascend_config():
+    with VllmRunner("facebook/opt-125m"):
+        ascend_config = get_ascend_config()
+
+        assert not ascend_config.torchair_graph_config.enabled
+        assert not ascend_config.torchair_graph_config.use_cached_graph
+        assert ascend_config.torchair_graph_config.graph_batch_sizes == []
+        assert not ascend_config.torchair_graph_config.graph_batch_sizes_init
+        assert not ascend_config.ascend_scheduler_config.enabled
+        assert ascend_config.expert_tensor_parallel_size == 1
+
+
+@_clean_up_ascend_config
+def test_run_with_ascend_config():
+    input_additional_config = {
+        "torchair_graph_config": {
+            # torchair graph only works with deepseek. The e2e test should be added
+            # in multicard test with deepseek models.
+            "enabled": False,
+            "use_cached_graph": True,
+            "graph_batch_sizes": [1, 2, 4, 8],
+            "graph_batch_sizes_init": False,
+        },
+        "ascend_scheduler_config": {
+            "enabled": True,
+            "enable_chunked_prefill": True,
+        },
+        "expert_tensor_parallel_size": 1
+    }
+    with VllmRunner("facebook/opt-125m",
+                    additional_config=input_additional_config):
+        ascend_config = get_ascend_config()
+
+        assert not ascend_config.torchair_graph_config.enabled
+        assert ascend_config.torchair_graph_config.use_cached_graph
+        assert ascend_config.torchair_graph_config.graph_batch_sizes == [
+            1, 2, 4, 8
+        ]
+        assert not ascend_config.torchair_graph_config.graph_batch_sizes_init
+        assert ascend_config.ascend_scheduler_config.enabled
+        assert ascend_config.ascend_scheduler_config.enable_chunked_prefill
+        assert ascend_config.expert_tensor_parallel_size == 1
+
+
+@_clean_up_ascend_config
+def test_ascend_config_init_error():
+    # ascend_config should be initialized first
+    with pytest.raises(RuntimeError):
+        _ = get_ascend_config()
+
+
+@_clean_up_ascend_config
+def test_ascend_config_load_error():
+    # graph_batch_sizes should be list.
+    with pytest.raises(TypeError):
+        input_additional_config_fake_1 = {
+            "torchair_graph_config": {
+                "graph_batch_sizes": "fake_size",
+            },
+        }
+        with VllmRunner("facebook/opt-125m",
+                        additional_config=input_additional_config_fake_1):
+            pass
+
+    # graph_batch_sizes_init should not be True when graph_batch_sizes is not empty.
+    with pytest.raises(ValueError):
+        input_additional_config_fake_2 = {
+            "torchair_graph_config": {
+                "graph_batch_sizes": [1, 2, 4, 8],
+                "graph_batch_sizes_init": True,
+            },
+        }
+        with VllmRunner("facebook/opt-125m",
+                        additional_config=input_additional_config_fake_2):
+            pass
+
+    # torchair graph only works with deepseek.
+    with pytest.raises(NotImplementedError):
+        input_additional_config_fake_2 = {
+            "torchair_graph_config": {
+                "enabled": True,
+            },
+        }
+        with VllmRunner("facebook/opt-125m",
+                        additional_config=input_additional_config_fake_2):
+            pass