[CI] Fix sample backward compatibility problem (#648)

vllm-project/vllm@b411418
this vllm commit change the sample usage. This PR adapt the change for
main and make sure it works for 0.8.4 as well.

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

Author: wangxiyuan

vllm_ascend/worker/draft_model_runner.py

@@ -28,6 +28,7 @@
                                            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.
@@ -286,10 +287,17 @@ def execute_model(
             if not self.is_driver_worker:
                 return []
             # Sample the next token.
-            output = self.model.sample(
-                logits=logits,
-                sampling_metadata=model_input.sampling_metadata,
-            )
+            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,
+                )
             outputs.append(output)
 
             if model_input.attn_metadata.num_prefills == 0 \

vllm_ascend/worker/model_runner.py

@@ -937,6 +937,12 @@ def __init__(
               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()
+
     def get_model(self) -> nn.Module:
         return self.model
 
@@ -1404,10 +1410,17 @@ def execute_model(
             model_input.async_callback()
 
         # Sample the next token.
-        output: SamplerOutput = self.model.sample(
-            logits=logits,
-            sampling_metadata=model_input.sampling_metadata,
-        )
+        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,
+            )
         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

@@ -53,6 +53,7 @@
 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]
@@ -290,6 +291,12 @@ def __init__(self, vllm_config: VllmConfig, device: torch.device):
         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()
+
     def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
         """Update the cached states and the persistent batch with the scheduler
         output.
@@ -645,10 +652,17 @@ def execute_model(
 
         # Sample the next token and get logprobs if needed.
         sampling_metadata = self.input_batch.sampling_metadata
-        sampler_output = self.model.sample(
-            logits=logits,
-            sampling_metadata=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,
+            )
 
         # TODO(woosuk): The following loop can be slow since it iterates over
         # the requests one by one. Optimize.