async_output_proc: Add virtual engine support

vllm/engine/llm_engine.py

@@ -1,7 +1,7 @@
 import time
 from collections import deque
 from contextlib import contextmanager
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from typing import (TYPE_CHECKING, Any, ClassVar, Deque, Dict, Iterable, List,
                     Mapping, Optional)
 from typing import Sequence as GenericSequence
@@ -40,7 +40,8 @@
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import (EmbeddingSequenceGroupOutput, ExecuteModelRequest,
                            SamplerOutput, Sequence, SequenceGroup,
-                           SequenceGroupMetadata, SequenceStatus)
+                           SequenceGroupMetadata, SequenceStatus,
+                           AsyncCallbackData)
 from vllm.tracing import (SpanAttributes, SpanKind, extract_trace_context,
                           init_tracer)
 from vllm.transformers_utils.config import try_get_generation_config
@@ -88,6 +89,19 @@ class SchedulerOutputState:
     last_output: Optional[SamplerOutput] = None
 
 
+@dataclass
+class SchedulerContext:
+    output_queue: Deque[Tuple[List[SamplerOutput],
+                              List[Tuple[ScheduledSequenceGroup,
+                                         SequenceGroupMetadata]],
+                              SchedulerOutputs]] = field(
+                                  default_factory=lambda: deque())
+
+    request_outputs: List[Union[RequestOutput,
+                                EmbeddingRequestOutput]] = field(
+                                    default_factory=lambda: [])
+
+
 class LLMEngine:
     """An LLM engine that receives requests and generates texts.
 
@@ -406,12 +420,17 @@ def get_tokenizer_for_seq(sequence: Sequence) -> AnyTokenizer:
             for _ in range(self.parallel_config.pipeline_parallel_size)
         ]
 
-        # Async output processing pointers
-        self.output_queue: Deque[Tuple[List[SamplerOutput],
-                                       List[SequenceGroupMetadata],
-                                       SchedulerOutputs]] = deque()
-        self.request_outputs: List[Union[RequestOutput,
-                                         EmbeddingRequestOutput]] = []
+        self.scheduler_contexts = [
+            SchedulerContext()
+            for _ in range(self.parallel_config.pipeline_parallel_size)
+        ]
+
+        self.async_callback_data = [
+            AsyncCallbackData(self._process_model_outputs, {
+                "virtual_engine": v_id,
+                "is_async": True,
+            }) for v_id in range(self.parallel_config.pipeline_parallel_size)
+        ]
 
     def _initialize_kv_caches(self) -> None:
         """Initialize the KV cache in the worker(s).
@@ -1214,32 +1233,28 @@ def _process_sequence_group_outputs(
 
         return
 
-    def _process_model_outputs(self,
-                               is_async: bool,
-                               clear_outputs: bool = True) -> None:
+    def _process_model_outputs(self, virtual_engine: int,
+                               is_async: bool) -> None:
         """Apply the model output to the sequences in the scheduled seq groups.
 
+        virtual_engine: The engine id to operate on
         is_async: Indicates whether this postprocessor runs in 
             parallel with the GPU forward pass and is processing 
             tokens from the previous step. If this is true, then
             no tokens need to be appended since it is already done
             externally (before the next schedule() call)
-        clear_outputs: Sometimes existing outputs need to be combined 
-            with outputs of this call. This happens for postprocessor
-            draining at the final stage (like when sequences are finished)
         
         Returns RequestOutputs that can be returned to the client.
         """
         now = time.time()
 
-        if clear_outputs:
-            self.request_outputs.clear()
+        ctx: SchedulerContext = self.scheduler_contexts[virtual_engine]
 
-        if len(self.output_queue) == 0:
+        if len(ctx.output_queue) == 0:
             return None
 
         (outputs, seq_group_metadata_list,
-         scheduler_outputs) = self.output_queue.popleft()
+         scheduler_outputs) = ctx.output_queue.popleft()
 
         # Sanity check
         assert len(seq_group_metadata_list) == len(
@@ -1314,11 +1329,11 @@ def _process_model_outputs(self,
             if (seq_group.is_finished()
                     if self.step_return_finished_only else True):
                 request_output = RequestOutputFactory.create(seq_group)
-                self.request_outputs.append(request_output)
+                ctx.request_outputs.append(request_output)
 
         for seq_group in scheduler_outputs.ignored_seq_groups:
             request_output = RequestOutputFactory.create(seq_group)
-            self.request_outputs.append(request_output)
+            ctx.request_outputs.append(request_output)
 
         if is_async:
             # Log stats.
@@ -1414,29 +1429,41 @@ def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
                 "Pipeline parallelism is only supported through AsyncLLMEngine "
                 "as performance will be severely degraded otherwise.")
 
+        virtual_engine = 0
+
         # These are cached outputs from previous iterations. None if on first
         # iteration
-        cached_outputs = self.cached_scheduler_outputs[0]
+        cached_outputs = self.cached_scheduler_outputs[virtual_engine]
         seq_group_metadata_list = cached_outputs.seq_group_metadata_list
         scheduler_outputs = cached_outputs.scheduler_outputs
         allow_async_output_proc = cached_outputs.allow_async_output_proc
 
+        ctx = self.scheduler_contexts[virtual_engine]
+
         # Skip the scheduler if there are any remaining steps in the seq groups.
         # This ensures that the scheduler is only called again when the current
         # batch has completed.
         if not self._has_remaining_steps(seq_group_metadata_list):
+
+            # Clear outputs on scheduler iteration start
+            ctx.request_outputs.clear()
+
+            # Schedule iteration
             (seq_group_metadata_list, scheduler_outputs,
-             allow_async_output_proc) = self.scheduler[0].schedule()
+             allow_async_output_proc
+             ) = self.scheduler[virtual_engine].schedule()
 
-            if not allow_async_output_proc and len(self.output_queue) > 0:
-                self._process_model_outputs(is_async=True)
+            # Maybe switch from async mode to sync mode
+            if not allow_async_output_proc and len(ctx.output_queue) > 0:
+                self._process_model_outputs(virtual_engine=virtual_engine,
+                                            is_async=True)
 
             if (self.scheduler_config.is_multi_step
                     and scheduler_outputs.num_lookahead_slots > 0):
                 # cache the scheduler outputs for the next iteration if we have
                 # lookahead slots
                 self._cache_scheduler_outputs_for_multi_step(
-                    0, seq_group_metadata_list, scheduler_outputs,
+                    virtual_engine, seq_group_metadata_list, scheduler_outputs,
                     allow_async_output_proc)
 
         assert seq_group_metadata_list is not None
@@ -1447,14 +1474,14 @@ def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
 
         if not scheduler_outputs.is_empty():
             finished_requests_ids = self.scheduler[
-                0].get_and_reset_finished_requests_ids()
+                virtual_engine].get_and_reset_finished_requests_ids()
 
             # Check if we have a cached last_output from the previous iteration.
             # For supporting PP this is probably the best way to pass the
             # sampled_token_ids, as a separate broadcast over all the PP stages
             # will cause one virtual engine's microbatch to block the pipeline.
             last_sampled_token_ids = \
-                self._get_last_sampled_token_ids(0)
+                self._get_last_sampled_token_ids(virtual_engine)
 
             execute_model_req = ExecuteModelRequest(
                 seq_group_metadata_list=seq_group_metadata_list,
@@ -1469,20 +1496,24 @@ def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
                 last_sampled_token_ids=last_sampled_token_ids)
 
             if allow_async_output_proc:
-                execute_model_req.output_proc_callback_fn = \
-                    self._process_model_outputs
+                execute_model_req.async_callback = self.async_callback_data[
+                    virtual_engine]
 
             output = self.model_executor.execute_model(
                 execute_model_req=execute_model_req)
 
-            # we need to do this here so that last step's sampled_token_ids can
+            # We need to do this here so that last step's sampled_token_ids can
             # be passed to the next iteration for PP.
             if self.scheduler_config.is_multi_step:
-                self._update_cached_scheduler_output(0, output)
+                self._update_cached_scheduler_output(virtual_engine, output)
         else:
-            if len(self.output_queue) > 0:
+            # Nothing scheduled => If there is pending async postprocessor,
+            # then finish it here.
+            if len(ctx.output_queue) > 0:
                 assert not self.scheduler_config.is_multi_step
-                self._process_model_outputs(is_async=True)
+                self._process_model_outputs(virtual_engine=virtual_engine,
+                                            is_async=True)
+            # No outputs in this case
             output = []
 
         # Finish the current step for all the sequence groups.
@@ -1497,7 +1528,7 @@ def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
 
             # Add results to the output_queue
             # (for async or non-async postprocessing)
-            self.output_queue.append(
+            ctx.output_queue.append(
                 (output, seq_group_metadata_list, scheduler_outputs))
 
             if output and allow_async_output_proc:
@@ -1508,23 +1539,27 @@ def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
                     output[0], seq_group_metadata_list,
                     scheduler_outputs.scheduled_seq_groups)
 
+            # Check if need to run the usual non-async path
             if not allow_async_output_proc:
-                self._process_model_outputs(is_async=False)
+                self._process_model_outputs(virtual_engine=virtual_engine,
+                                            is_async=False)
 
                 # Log stats.
                 self.do_log_stats(scheduler_outputs, output)
 
                 # Tracing
                 self.do_tracing(scheduler_outputs)
         else:
+            # Multi-step case
             self.request_outputs = []
 
         if not self.has_unfinished_requests():
-            # Drain async postprocessor
-            if len(self.output_queue) > 0:
+            # Drain async postprocessor (if exists)
+            if len(ctx.output_queue) > 0:
                 assert not self.scheduler_config.is_multi_step
-                self._process_model_outputs(is_async=True, clear_outputs=False)
-            assert len(self.output_queue) == 0
+                self._process_model_outputs(virtual_engine=virtual_engine,
+                                            is_async=True)
+            assert len(ctx.output_queue) == 0
 
             # Stop the execute model loop in parallel workers until there are
             # more requests to process. This avoids waiting indefinitely in
@@ -1533,7 +1568,7 @@ def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
             # queued control plane messages, such as add/remove lora adapters.
             self.model_executor.stop_remote_worker_execution_loop()
 
-        return self.request_outputs
+        return ctx.request_outputs
 
     def _has_remaining_steps(
         self, seq_group_metadata_list: Optional[List[SequenceGroupMetadata]]

vllm/sequence.py

@@ -811,6 +811,9 @@ def remove(self, seq_id: int) -> None:
         self.is_single_seq = len(self.seqs) == 1
 
     def is_finished(self) -> bool:
+        if self.is_single_seq:
+            return self.seqs[0].is_finished()
+
         return all(seq.is_finished() for seq in self.seqs)
 
     def is_prefill(self) -> bool:
@@ -1259,6 +1262,12 @@ def expand_with_bonus_tokens(
             [self.hidden_states, self.second_last_token_hidden_states])[index]
 
 
+@dataclass
+class AsyncCallbackData:
+    func: Callable
+    kw_args: Dict[str, Any]
+
+
 class ExecuteModelRequest(
         msgspec.Struct,
         array_like=True,  # type: ignore[call-arg]
@@ -1290,8 +1299,8 @@ class ExecuteModelRequest(
     finished_requests_ids: List[str] = msgspec.field(default_factory=list)
     # The last sampled token ids for multi step decoding.
     last_sampled_token_ids: Optional[torch.Tensor] = None
-    # Async postprocessor
-    output_proc_callback_fn: Optional[Callable] = None
+    # Async callback
+    async_callback: Optional[AsyncCallbackData] = None
 
     @property
     def is_first_multi_step(self) -> bool:
@@ -1338,4 +1347,4 @@ def clone(
             finished_requests_ids=self.finished_requests_ids,
             last_sampled_token_ids=self.last_sampled_token_ids.clone()
             if self.last_sampled_token_ids is not None else None,
-            output_proc_callback_fn=self.output_proc_callback_fn)
+            async_callback=self.async_callback)

vllm/worker/model_runner.py

@@ -42,7 +42,7 @@
     LRUCacheWorkerPromptAdapterManager)
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import (IntermediateTensors, SamplerOutput,
-                           SequenceGroupMetadata)
+                           SequenceGroupMetadata, AsyncCallbackData)
 from vllm.utils import (CudaMemoryProfiler, PyObjectCache, async_tensor_h2d,
                         flatten_2d_lists, is_hip, is_pin_memory_available)
 from vllm.worker.model_runner_base import (
@@ -90,7 +90,7 @@ class ModelInputForGPU(ModelRunnerInputBase):
     request_ids_to_seq_ids: Optional[Dict[str, List[int]]] = None
     finished_requests_ids: Optional[List[str]] = None
     virtual_engine: int = 0
-    output_proc_callback_fn: Optional[Callable] = None
+    async_callback: Optional[AsyncCallbackData] = None
 
     def as_broadcastable_tensor_dict(self) -> Dict[str, Any]:
         tensor_dict = {
@@ -1456,9 +1456,11 @@ def execute_model(
         if not self.is_driver_worker:
             return []
 
-        if model_input.output_proc_callback_fn is not None:
-            model_input.output_proc_callback_fn(is_async=True)
-
+        if model_input.async_callback is not None:
+            func = model_input.async_callback.func
+            kw_args = model_input.async_callback.kw_args
+            func(**kw_args)
+
         # Sample the next token.
         output: SamplerOutput = self.model.sample(
             logits=logits,

vllm/worker/worker_base.py

@@ -263,11 +263,10 @@ def _get_driver_input_and_broadcast(
             broadcast_data.update(kwargs)
             broadcast_tensor_dict(broadcast_data, src=0)
 
-        if execute_model_req.output_proc_callback_fn:
+        if execute_model_req.async_callback:
             model_input = dataclasses.replace(  # type: ignore
                 model_input,
-                output_proc_callback_fn=execute_model_req.
-                output_proc_callback_fn)
+                async_callback=execute_model_req.async_callback)
 
         return model_input, worker_input, kwargs