Merge branch 'vllm-project:main' into main

Author: zhaozx-cn

docs/source/user_guide/configuration/additional_config.md

@@ -48,6 +48,7 @@ The details of each config option are as follows:
 | `enable_multistream_mla`| bool | `False` | Whether to put vector ops of MLA to another stream. This option only takes effects on models using MLA (e.g., DeepSeek). |
 | `enable_multistream_moe`| bool | `False` | Whether to enable multistream shared expert. This option only takes effects on DeepSeek moe models. |
 | `enable_view_optimize` | bool | `True` | Whether to enable torchair view optimization |
+| `enable_frozen_parameter` | bool | `True` | Whether to fix the memory address of weights during inference to reduce the input address refresh time during graph execution. |
 | `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 |

tests/ut/test_ascend_config.py

@@ -53,6 +53,7 @@ def test_init_ascend_config_without_additional_config(self):
         self.assertFalse(torchair_graph_config.enable_multistream_mla)
         self.assertFalse(torchair_graph_config.enable_multistream_moe)
         self.assertTrue(torchair_graph_config.enable_view_optimize)
+        self.assertTrue(torchair_graph_config.enable_frozen_parameter)
         self.assertFalse(torchair_graph_config.enable_kv_nz)
 
         ascend_scheduler_config = ascend_config.ascend_scheduler_config
@@ -70,6 +71,7 @@ def test_init_ascend_config_with_additional_config(self):
                 "enable_multistream_mla": True,
                 "enable_multistream_moe": True,
                 "enable_view_optimize": True,
+                "enable_frozen_parameter": True,
                 "enable_kv_nz": True
             },
             "ascend_scheduler_config": {
@@ -89,6 +91,7 @@ def test_init_ascend_config_with_additional_config(self):
         self.assertTrue(torchair_graph_config.enable_multistream_mla)
         self.assertTrue(torchair_graph_config.enable_multistream_moe)
         self.assertTrue(torchair_graph_config.enable_view_optimize)
+        self.assertTrue(torchair_graph_config.enable_frozen_parameter)
         self.assertTrue(torchair_graph_config.enable_kv_nz)
 
         ascend_scheduler_config = ascend_config.ascend_scheduler_config

vllm_ascend/ascend_config.py

@@ -114,6 +114,8 @@ def __init__(self, torchair_graph_config):
             "enable_multistream_moe", False)
         self.enable_view_optimize = torchair_graph_config.get(
             "enable_view_optimize", True)
+        self.enable_frozen_parameter = torchair_graph_config.get(
+            "enable_frozen_parameter", True)
         self.enable_kv_nz = torchair_graph_config.get("enable_kv_nz", False)
 
         if not isinstance(self.graph_batch_sizes, list):

vllm_ascend/torchair/torchair_model_runner.py

@@ -359,7 +359,8 @@ def _get_torchair_lazy_compiled_model(self, batch_size: int):
         config = torchair.CompilerConfig()
         if get_ascend_config().torchair_graph_config.mode:
             config.mode = get_ascend_config().torchair_graph_config.mode
-        config.experimental_config.frozen_parameter = True
+        config.experimental_config.frozen_parameter = \
+        get_ascend_config().torchair_graph_config.enable_frozen_parameter
         # enabling tiling_schedule_optimize on 300I Duo has some bugs, so we have to
         # disable it on 300I Duo platform now.
         config.experimental_config.tiling_schedule_optimize = not is_310p()