[CI] upgrade vllm to 0.8.5 (#715)

1. Upgrade vllm to 0.8.5
2. Drop 0.8.4 support
3. Keep doc to 0.8.4rc2 until we release 0.8.5

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm_ascend/worker/draft_model_runner.py

@@ -28,7 +28,6 @@ from vllm.worker.model_runner_base import (ModelRunnerBase,
                                            ModelRunnerWrapperBase)
 
 from vllm_ascend.attention.attention import AscendMetadata
-from vllm_ascend.utils import vllm_version_is
 
 # A flag to enable debug prints for the updated input tensors
 # before each step.
@@ -287,17 +286,11 @@ class TP1DraftModelRunner(ModelRunnerWrapperBase):
             if not self.is_driver_worker:
                 return []
             # Sample the next token.
-            if vllm_version_is("0.8.4"):
-                output = self.model.sample(
-                    logits=logits,
-                    sampling_metadata=model_input.sampling_metadata,
-                )
-            else:
-                assert self.model_runner.sampler is not None
-                output = self.model_runner.sampler(
-                    logits=logits,
-                    sampling_metadata=model_input.sampling_metadata,
-                )
+            assert self.model_runner.sampler is not None
+            output = self.model_runner.sampler(
+                logits=logits,
+                sampling_metadata=model_input.sampling_metadata,
+            )
             outputs.append(output)
 
             if model_input.attn_metadata.num_prefills == 0 \

vllm_ascend/worker/model_runner.py

@@ -34,6 +34,7 @@ from vllm.attention.backends.utils import CommonAttentionState
 from vllm.config import VllmConfig
 from vllm.core.scheduler import SchedulerOutputs
 from vllm.distributed import get_pp_group
+from vllm.distributed.kv_transfer import get_kv_transfer_group
 from vllm.forward_context import set_forward_context
 from vllm.inputs import INPUT_REGISTRY, InputRegistry
 from vllm.logger import logger
@@ -42,7 +43,7 @@ from vllm.lora.request import LoRARequest
 from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
 from vllm.model_executor import SamplingMetadata, SamplingMetadataCache
 from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
-from vllm.model_executor.layers.sampler import SamplerOutput
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
 from vllm.model_executor.models import supports_lora, supports_multimodal
@@ -63,13 +64,6 @@ from vllm.worker.model_runner_base import (
     _init_attn_metadata_from_tensor_dict,
     _init_sampling_metadata_from_tensor_dict)
 
-from vllm_ascend.utils import vllm_version_is
-
-if vllm_version_is("0.8.4"):
-    from vllm.distributed import get_kv_transfer_group
-else:
-    from vllm.distributed.kv_transfer import get_kv_transfer_group
-
 if TYPE_CHECKING:
     from vllm.attention.backends.abstract import AttentionBackend
 
@@ -935,12 +929,7 @@ class NPUModelRunnerBase(ModelRunnerBase[TModelInputForNPU]):
         self.sampling_metadata_cache: SamplingMetadataCache = \
               SamplingMetadataCache() \
                 if self.parallel_config.pipeline_parallel_size == 1 else None
-
-        if vllm_version_is("0.8.4"):
-            self.sampler = None
-        else:
-            from vllm.model_executor.layers.sampler import get_sampler
-            self.sampler = get_sampler()
+        self.sampler = get_sampler()
 
     def get_model(self) -> nn.Module:
         return self.model
@@ -1409,17 +1398,10 @@ class NPUModelRunner(NPUModelRunnerBase[ModelInputForNPUWithSamplingMetadata]):
             model_input.async_callback()
 
         # Sample the next token.
-        if vllm_version_is("0.8.4"):
-            output = self.model.sample(
-                logits=logits,
-                sampling_metadata=model_input.sampling_metadata,
-            )
-        else:
-            assert self.sampler is not None
-            output = self.sampler(
-                logits=logits,
-                sampling_metadata=model_input.sampling_metadata,
-            )
+        output = self.sampler(
+            logits=logits,
+            sampling_metadata=model_input.sampling_metadata,
+        )
         if (self.observability_config is not None
                 and self.observability_config.collect_model_forward_time
                 and output is not None):

vllm_ascend/worker/model_runner_v1.py

@@ -47,13 +47,13 @@ from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
 from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                         KVCacheSpec)
 from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT, ModelRunnerOutput
+from vllm.v1.sample.sampler import Sampler
 from vllm.v1.utils import bind_kv_cache
 from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
 
 from vllm_ascend.attention.attention import AttentionMaskBuilder
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
 from vllm_ascend.platform import NPUPlatform
-from vllm_ascend.utils import vllm_version_is
 
 if TYPE_CHECKING:
     import xgrammar as xgr  # type: ignore[import-untyped]
@@ -291,11 +291,7 @@ class NPUModelRunner:
         self.attn_mask_builder = AttentionMaskBuilder.initialize_from_len(
             self.attn_mask_len, self.dtype)
 
-        if vllm_version_is("0.8.4"):
-            self.sampler = None
-        else:
-            from vllm.v1.sample.sampler import Sampler
-            self.sampler = Sampler()
+        self.sampler = Sampler()
 
     def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
         """Update the cached states and the persistent batch with the scheduler
@@ -346,34 +342,19 @@ class NPUModelRunner:
                 generator.manual_seed(sampling_params.seed)
             else:
                 generator = None
-            if vllm_version_is("0.8.4"):
-                self.requests[req_id] = CachedRequestState(
-                    req_id=req_id,
-                    prompt_token_ids=new_req_data.prompt_token_ids,
-                    prompt=new_req_data.prompt,
-                    mm_inputs=new_req_data.mm_inputs,
-                    mm_positions=new_req_data.mm_positions,
-                    sampling_params=sampling_params,
-                    generator=generator,
-                    block_ids=new_req_data.block_ids,
-                    num_computed_tokens=new_req_data.num_computed_tokens,
-                    output_token_ids=[],
-                    lora_request=new_req_data.lora_request,
-                )
-            else:
-                # the prompt removed by: https://github.com/vllm-project/vllm/pull/17214
-                self.requests[req_id] = CachedRequestState(
-                    req_id=req_id,
-                    prompt_token_ids=new_req_data.prompt_token_ids,
-                    mm_inputs=new_req_data.mm_inputs,
-                    mm_positions=new_req_data.mm_positions,
-                    sampling_params=sampling_params,
-                    generator=generator,
-                    block_ids=new_req_data.block_ids,
-                    num_computed_tokens=new_req_data.num_computed_tokens,
-                    output_token_ids=[],
-                    lora_request=new_req_data.lora_request,
-                )
+
+            self.requests[req_id] = CachedRequestState(
+                req_id=req_id,
+                prompt_token_ids=new_req_data.prompt_token_ids,
+                mm_inputs=new_req_data.mm_inputs,
+                mm_positions=new_req_data.mm_positions,
+                sampling_params=sampling_params,
+                generator=generator,
+                block_ids=new_req_data.block_ids,
+                num_computed_tokens=new_req_data.num_computed_tokens,
+                output_token_ids=[],
+                lora_request=new_req_data.lora_request,
+            )
 
             req_ids_to_add.append(req_id)
 
@@ -666,17 +647,10 @@ class NPUModelRunner:
 
         # Sample the next token and get logprobs if needed.
         sampling_metadata = self.input_batch.sampling_metadata
-        if vllm_version_is("0.8.4"):
-            sampler_output = self.model.sample(
-                logits=logits,
-                sampling_metadata=sampling_metadata,
-            )
-        else:
-            assert self.sampler is not None
-            sampler_output = self.sampler(
-                logits=logits,
-                sampling_metadata=sampling_metadata,
-            )
+        sampler_output = self.sampler(
+            logits=logits,
+            sampling_metadata=sampling_metadata,
+        )
 
         # TODO(woosuk): The following loop can be slow since it iterates over
         # the requests one by one. Optimize.

vllm_ascend/worker/multi_step_runner.py

@@ -23,7 +23,6 @@ from vllm.worker.multi_step_model_runner import (ModelOutput,
                                                  PythonizationCache,
                                                  StatefulModelInput)
 
-from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.model_runner import (
     ModelInputForNPUWithSamplingMetadata, NPUModelRunnerBase)
 
@@ -318,13 +317,7 @@ class MultiStepModelNPURunner(NPUModelRunnerBase[StatefulModelInputForNPU]):
                     dtype=torch.long,
                     device="cpu",
                     pin_memory=True)
-
-            if vllm_version_is("0.8.4"):
-                self._base_model_runner.model.sampler.include_gpu_probs_tensor = (
-                    True)
-            else:
-                assert self._base_model_runner.sampler is not None
-                self._base_model_runner.sampler.include_gpu_probs_tensor = True
+            self._base_model_runner.sampler.include_gpu_probs_tensor = True
             if frozen_model_input.sampling_metadata:
                 frozen_model_input.sampling_metadata.skip_sampler_cpu_output = (
                     True)

vllm_ascend/worker/worker.py

@@ -28,6 +28,7 @@ from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment,
                               set_custom_all_reduce)
+from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
 from vllm.logger import logger
 from vllm.lora.request import LoRARequest
 from vllm.model_executor import set_random_seed
@@ -46,15 +47,10 @@ from vllm.worker.worker_base import (LocalOrDistributedWorkerBase, WorkerBase,
 from vllm_ascend.device_allocator.camem import CaMemAllocator
 from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
 from vllm_ascend.platform import NPUPlatform
-from vllm_ascend.utils import try_register_lib, vllm_version_is
+from vllm_ascend.utils import try_register_lib
 from vllm_ascend.worker.model_runner import NPUModelRunner
 from vllm_ascend.worker.pooling_model_runner import NPUPoolingModelRunner
 
-if vllm_version_is("0.8.4"):
-    from vllm.distributed import ensure_kv_transfer_initialized
-else:
-    from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
-
 
 class NPUWorker(LocalOrDistributedWorkerBase):
     """A worker class that executes (a partition of) the model on a NPU.

vllm_ascend/worker/worker_v1.py

@@ -29,6 +29,7 @@ from vllm.config import VllmConfig
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment,
                               set_custom_all_reduce)
+from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
 from vllm.logger import logger
 from vllm.model_executor import set_random_seed
 from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
@@ -41,14 +42,9 @@ from vllm.v1.worker.worker_base import WorkerBase
 
 from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
 from vllm_ascend.platform import NPUPlatform
-from vllm_ascend.utils import try_register_lib, vllm_version_is
+from vllm_ascend.utils import try_register_lib
 from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
 
-if vllm_version_is("0.8.4"):
-    from vllm.distributed import ensure_kv_transfer_initialized
-else:
-    from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
-
 
 class NPUWorker(WorkerBase):