Upgrade to new vllm commit

.github/workflows/format_pr_body.yaml

@@ -36,7 +36,7 @@ jobs:
 
       - name: Get vLLM version
         run: |
-          VLLM_COMMIT=17c540a993af88204ad1b78345c8a865cf58ce44
+          VLLM_COMMIT=c9461e05a4ed3557cfbf4b15ded1e26761cc39ca
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> $GITHUB_ENV
 
       - name: Checkout repository

.github/workflows/vllm_ascend_test.yaml

@@ -42,7 +42,7 @@ jobs:
   lint:
     uses: ./.github/workflows/pre-commit.yml
     with:
-      vllm: 17c540a993af88204ad1b78345c8a865cf58ce44
+      vllm: c9461e05a4ed3557cfbf4b15ded1e26761cc39ca
 
   changes:
     runs-on: ubuntu-latest
@@ -83,7 +83,7 @@ jobs:
         VLLM_USE_MODELSCOPE: True
     strategy:
       matrix:
-        vllm_version: [17c540a993af88204ad1b78345c8a865cf58ce44, v0.11.0]
+        vllm_version: [c9461e05a4ed3557cfbf4b15ded1e26761cc39ca, v0.11.0]
     steps:
       - name: Install packages
         run: |
@@ -140,7 +140,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [17c540a993af88204ad1b78345c8a865cf58ce44, v0.11.0]
+        vllm_version: [c9461e05a4ed3557cfbf4b15ded1e26761cc39ca, v0.11.0]
     # Note (yikun): If CI resource are limited we can split job into two chain jobs
     needs: [lint, changes]
     # only trigger e2e test after lint passed and the change is e2e related with pull request.

.github/workflows/vllm_ascend_test_full.yaml

@@ -69,7 +69,7 @@ jobs:
     name: e2e-full
     strategy:
       matrix:
-        vllm_version: [17c540a993af88204ad1b78345c8a865cf58ce44, v0.11.0]
+        vllm_version: [c9461e05a4ed3557cfbf4b15ded1e26761cc39ca, v0.11.0]
     needs: [changes]
     if: ${{ needs.changes.outputs.e2e_tracker == 'true' }}
     uses: ./.github/workflows/_e2e_test.yaml

examples/offline_data_parallel.py

@@ -63,7 +63,11 @@
 from vllm import LLM, SamplingParams
 from vllm.distributed.parallel_state import (  # noqa E402
     destroy_distributed_environment, destroy_model_parallel)
-from vllm.utils import get_open_port
+from vllm_ascend.utils import vllm_version_is
+if vllm_version_is("0.11.0"):
+    from vllm.utils import get_open_port
+else:
+    from vllm.utils.network_utils import get_open_port
 
 os.environ["VLLM_USE_MODELSCOPE"] = "True"
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

examples/offline_external_launcher.py

@@ -65,9 +65,15 @@
 import torch
 from vllm import LLM, SamplingParams
 from vllm.distributed.parallel_state import (  # noqa E402
-    destroy_distributed_environment, destroy_model_parallel, get_tp_group)
-from vllm.utils import get_open_port, GiB_bytes
+    destroy_distributed_environment, destroy_model_parallel, get_tp_group) 
 from safetensors.torch import load_file
+from vllm_ascend.utils import vllm_version_is
+if vllm_version_is("0.11.0"):
+    from vllm.utils import GiB_bytes, get_open_port
+
+else:
+    from vllm.utils.mem_constants import GiB_bytes
+    from vllm.utils.network_utils import get_open_port
 
 os.environ["VLLM_USE_MODELSCOPE"] = "True"
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

examples/offline_inference_sleep_mode_npu.py

@@ -20,7 +20,11 @@
 
 import torch
 from vllm import LLM, SamplingParams
-from vllm.utils import GiB_bytes
+from vllm_ascend.utils import vllm_version_is
+if vllm_version_is("0.11.0"):
+    from vllm.utils import GiB_bytes
+else:
+    from vllm.utils.mem_constants import GiB_bytes
 
 os.environ["VLLM_USE_MODELSCOPE"] = "True"
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

examples/offline_weight_load.py

@@ -66,8 +66,14 @@
 from vllm import LLM, SamplingParams
 from vllm.distributed.parallel_state import (  # noqa E402
     destroy_distributed_environment, destroy_model_parallel, get_tp_group)
-from vllm.utils import get_open_port, GiB_bytes
 from safetensors.torch import load_file
+from vllm_ascend.utils import vllm_version_is
+if vllm_version_is("0.11.0"):
+    from vllm.utils import GiB_bytes, get_open_port
+
+else:
+    from vllm.utils.mem_constants import GiB_bytes
+    from vllm.utils.network_utils import get_open_port
 
 os.environ["VLLM_USE_MODELSCOPE"] = "True"
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

tests/e2e/conftest.py

@@ -45,7 +45,6 @@
 from vllm.outputs import RequestOutput
 from vllm.platforms import current_platform
 from vllm.transformers_utils.utils import maybe_model_redirect
-from vllm.utils import get_open_port
 
 from tests.e2e.model_utils import (TokensTextLogprobs,
                                    TokensTextLogprobsPromptLogprobs)
@@ -55,6 +54,12 @@
 # we not explicitly patch here, some of them might be effectiveless
 # in pytest scenario
 from vllm_ascend.utils import adapt_patch  # noqa E402
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import get_open_port
+else:
+    from vllm.utils.network_utils import get_open_port
 
 adapt_patch(True)
 adapt_patch(False)

tests/e2e/multicard/test_single_request_aclgraph.py

@@ -19,9 +19,14 @@
 
 import openai
 import pytest
-from vllm.utils import get_open_port
 
 from tests.e2e.conftest import RemoteOpenAIServer
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import get_open_port
+else:
+    from vllm.utils.network_utils import get_open_port
 
 MODELS = [
     "Qwen/Qwen3-30B-A3B",

tests/e2e/nightly/models/test_qwen3_32b.py

@@ -18,10 +18,15 @@
 
 import openai
 import pytest
-from vllm.utils import get_open_port
 
 from tests.e2e.conftest import RemoteOpenAIServer
 from tools.aisbench import run_aisbench_cases
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import get_open_port
+else:
+    from vllm.utils.network_utils import get_open_port
 
 MODELS = [
     "Qwen/Qwen3-32B",

tests/e2e/singlecard/test_camem.py

@@ -21,11 +21,16 @@
 
 import torch
 from vllm import SamplingParams
-from vllm.utils import GiB_bytes
 
 from tests.e2e.conftest import VllmRunner
 from tests.e2e.utils import fork_new_process_for_each_test
 from vllm_ascend.device_allocator.camem import CaMemAllocator
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import GiB_bytes
+else:
+    from vllm.utils.mem_constants import GiB_bytes
 
 
 @fork_new_process_for_each_test

tests/ut/core/test_schedule_config.py

@@ -78,21 +78,6 @@ def test_not_implemented_policy(self):
             str(context.exception),
         )
 
-    def test_not_implemented_send_delta_data(self):
-        with self.assertRaises(NotImplementedError) as context:
-            AscendSchedulerConfig.initialize_from_config(
-                self.basic_scheduler_config,
-                AscendSchedulerConfig(
-                    send_delta_data=True,
-                    max_num_batched_tokens=2048,
-                    max_model_len=2048,
-                ),
-            )
-        self.assertIn(
-            "currently AscendScheduler doesn't support send_delta_data",
-            str(context.exception),
-        )
-
     def test_no_override(self):
         ascend_config = AscendSchedulerConfig.initialize_from_config(
             self.basic_scheduler_config, {})

tests/ut/core/test_scheduler.py

@@ -9,7 +9,6 @@
 from vllm.multimodal.inputs import (MultiModalFeatureSpec,
                                     MultiModalKwargsItem, PlaceholderRange)
 from vllm.sampling_params import SamplingParams
-from vllm.utils import sha256
 from vllm.v1.core.kv_cache_utils import (get_request_block_hasher,
                                          init_none_hash)
 from vllm.v1.core.sched.output import SchedulerOutput
@@ -24,6 +23,11 @@
 from vllm_ascend.core.scheduler_dynamic_batch import SchedulerDynamicBatch
 from vllm_ascend.utils import vllm_version_is
 
+if vllm_version_is("0.11.0"):
+    from vllm.utils import sha256
+else:
+    from vllm.utils.hashing import sha256
+
 EOS_TOKEN_ID = 50256
 MODEL = "Qwen3-0.6B"
 ENABLE_PREFIX_CACHING = None

tests/ut/kv_connector/test_mooncake_connector.py

@@ -12,7 +12,13 @@
 
 import msgspec
 import zmq
-from vllm.utils import make_zmq_path
+
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import make_zmq_path
+else:
+    from vllm.utils.network_utils import make_zmq_path
 
 fake_engine = types.ModuleType("mooncake.engine")
 fake_engine.TransferEngine = MagicMock()  # type: ignore[attr-defined]

tests/ut/kv_connector/utils.py

@@ -10,7 +10,6 @@
 from vllm import SamplingParams
 from vllm.config import (CacheConfig, DeviceConfig, KVTransferConfig,
                          ModelConfig, SchedulerConfig, VllmConfig)
-from vllm.utils import sha256
 from vllm.v1.core.kv_cache_utils import (get_request_block_hasher,
                                          init_none_hash)
 from vllm.v1.core.sched.scheduler import Scheduler
@@ -22,6 +21,11 @@
 
 from vllm_ascend.utils import vllm_version_is
 
+if vllm_version_is("0.11.0"):
+    from vllm.utils import sha256
+else:
+    from vllm.utils.hashing import sha256
+
 EOS_TOKEN_ID = 50256
 os.environ["VLLM_USE_V1"] = "1"
 

tests/ut/model_loader/netloader/test_netloader.py

@@ -22,6 +22,7 @@
 from torch import nn
 
 from vllm_ascend.model_loader.netloader.netloader import ModelNetLoaderElastic
+from vllm_ascend.utils import vllm_version_is
 
 
 class DummyDeviceConfig:
@@ -173,7 +174,11 @@ def __exit__(self, a, b, c):
         "vllm_ascend.model_loader.netloader.netloader.process_weights_after_loading",
         lambda *a, **k: None)
     # patch get_ip
-    monkeypatch.setattr("vllm.utils.get_ip", lambda: "127.0.0.1")
+    if vllm_version_is("0.11.0"):
+        monkeypatch.setattr("vllm.utils.get_ip", lambda: "127.0.0.1")
+    else:
+        monkeypatch.setattr("vllm.utils.network_utils.get_ip",
+                            lambda: "127.0.0.1")
     # patch find_free_port
     monkeypatch.setattr(
         "vllm_ascend.model_loader.netloader.netloader.find_free_port",

tests/ut/worker/test_input_batch.py

@@ -20,14 +20,19 @@
 import pytest
 import torch
 from vllm.sampling_params import SamplingParams
-from vllm.utils import make_tensor_with_pad
 from vllm.v1.pool.metadata import PoolingMetadata
 from vllm.v1.sample.logits_processor import LogitsProcessors
 from vllm.v1.sample.metadata import SamplingMetadata
 
+from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.block_table import BlockTable, MultiGroupBlockTable
 from vllm_ascend.worker.npu_input_batch import CachedRequestState, InputBatch
 
+if vllm_version_is("0.11.0"):
+    from vllm.utils import make_tensor_with_pad
+else:
+    from vllm.utils.torch_utils import make_tensor_with_pad
+
 VOCAB_SIZE = 1024
 NUM_OUTPUT_TOKENS = 20
 MAX_PROMPT_SIZE = 100

tests/ut/worker/test_worker_v1.py

@@ -5,6 +5,7 @@
 from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
 
 from tests.ut.base import TestBase
+from vllm_ascend.utils import vllm_version_is
 
 
 class TestNPUWorker(TestBase):
@@ -178,15 +179,26 @@ def test_init_npu_worker_with_custom_cache_dtype(
         # Create NPUWorker instance
         from vllm_ascend.worker.worker_v1 import NPUWorker
 
-        with patch("vllm.utils.STR_DTYPE_TO_TORCH_DTYPE",
-                   {"float32": torch.float32}):
-            worker = NPUWorker(
-                vllm_config=self.vllm_config_mock,
-                local_rank=self.local_rank,
-                rank=self.rank,
-                distributed_init_method=self.distributed_init_method,
-                is_driver_worker=self.is_driver_worker,
-            )
+        if vllm_version_is("0.11.0"):
+            with patch("vllm.utils.STR_DTYPE_TO_TORCH_DTYPE",
+                       {"float32": torch.float32}):
+                worker = NPUWorker(
+                    vllm_config=self.vllm_config_mock,
+                    local_rank=self.local_rank,
+                    rank=self.rank,
+                    distributed_init_method=self.distributed_init_method,
+                    is_driver_worker=self.is_driver_worker,
+                )
+        else:
+            with patch("vllm.utils.torch_utils.STR_DTYPE_TO_TORCH_DTYPE",
+                       {"float32": torch.float32}):
+                worker = NPUWorker(
+                    vllm_config=self.vllm_config_mock,
+                    local_rank=self.local_rank,
+                    rank=self.rank,
+                    distributed_init_method=self.distributed_init_method,
+                    is_driver_worker=self.is_driver_worker,
+                )
 
         # Verify cache_dtype is set to custom value
         self.assertEqual(worker.cache_dtype, torch.float32)

vllm_ascend/core/schedule_config.py

@@ -99,9 +99,6 @@ def __post_init__(self, *args) -> None:
             raise NotImplementedError(
                 f"currently AscendScheduler only supports fcfs policy, got {self.policy}"
             )
-        if self.send_delta_data:
-            raise NotImplementedError(
-                "currently AscendScheduler doesn't support send_delta_data.")
         if getattr(self, "scheduler_delay_factor", 0) > 0:
             raise NotImplementedError(
                 "currently AscendScheduler doesn't support scheduler_delay_factor."

vllm_ascend/distributed/cpu_offload_manager/metadata.py

@@ -9,11 +9,18 @@
 import vllm.envs as envs
 import zmq
 from vllm.config import KVTransferConfig, VllmConfig
-from vllm.utils import get_dtype_size, logger, make_zmq_socket
+from vllm.utils import logger
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 from vllm_ascend.distributed.cpu_offload_manager.cpu_kv_cache_manager import \
     CPUKVCacheManager
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import get_dtype_size, make_zmq_socket
+else:
+    from vllm.utils.network_utils import make_zmq_socket
+    from vllm.utils.torch_utils import get_dtype_size
 
 
 @dataclass

vllm_ascend/distributed/llmdatadist_c_mgr_connector.py

@@ -25,19 +25,25 @@
 from vllm.distributed.parallel_state import (get_dcp_group, get_tp_group,
                                              get_world_group)
 from vllm.forward_context import ForwardContext
-from vllm.utils import get_ip, logger
+from vllm.utils import logger
 from vllm.v1.core.kv_cache_manager import KVCacheBlocks
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.request import Request, RequestStatus
 
 import vllm_ascend.envs as envs_ascend
 from vllm_ascend.utils import (AscendSocVersion, get_ascend_soc_version,
-                               prefill_context_parallel_enable)
+                               prefill_context_parallel_enable,
+                               vllm_version_is)
 
 if prefill_context_parallel_enable():
     from vllm.distributed.parallel_state import \
         get_prefill_context_model_parallel_rank
 
+if vllm_version_is("0.11.0"):
+    from vllm.utils import get_ip
+else:
+    from vllm.utils.network_utils import get_ip
+
 TORCH_DTYPE_TO_NPU_DTYPE = {
     torch.half: llm_datadist.DataType.DT_FLOAT16,
     torch.float16: llm_datadist.DataType.DT_FLOAT16,