first pass at n>1

Signed-off-by: Andrew Feldman <[email protected]>

Author: afeldman-nm

vllm/v1/engine/llm_engine.py

@@ -1,5 +1,5 @@
 from typing import Dict, List, Mapping, Optional, Type, Union
-
+from dataclasses import dataclass
 from typing_extensions import TypeVar
 
 from vllm.config import VllmConfig
@@ -26,6 +26,42 @@
 
 _G = TypeVar("_G", bound=BaseTokenizerGroup, default=BaseTokenizerGroup)
 
+def _none_safe_min(x,y):
+    if x is None:
+        return y
+    if y is None:
+        return x
+    return min(x,y)
+
+def _none_safe_max(x,y):
+    if x is None:
+        return y
+    if y is None:
+        return x
+    return max(x,y)
+
+def _none_safe_sum(x,y):
+    if x is None:
+        return y
+    if y is None:
+        return x
+    return x+y
+
+@dataclass
+class ParallelSampleChildRequestInfo:
+    """Info for aggregating parallel sampling child requests under parent"""
+    parent_req_id: str
+    index: int
+
+@dataclass
+class ParallelSampleParentRequestInfo:
+    """Parallel sampling parent request info"""
+    n: int
+    n_finished: int = 0
+
+    def num_child_requests_remaining(self):
+        assert self.n >= self.n_finished
+        return self.n - self.n_finished
 
 class LLMEngine:
     """Legacy LLMEngine for backwards compatibility."""
@@ -46,6 +82,14 @@ def __init__(
         # TODO: Can we avoid this?
         self.model_config = vllm_config.model_config
 
+        # Parallel sampling metadata
+        # - Metadata for aggregating the child requests associated with a parent request
+        self.child_req_id_to_parent_req_info: Dict[
+            str, ParallelSampleChildRequestInfo] = {}
+        # - Parent request metadata i.e. degree of parallelism and other characteristics
+        self.parent_req_id_info: Dict[str,
+                                      ParallelSampleParentRequestInfo] = {}
+
         # Tokenizer (+ ensure liveness if running in another process).
         self.tokenizer = init_tokenizer_from_configs(
             model_config=vllm_config.model_config,
@@ -117,8 +161,52 @@ def _get_executor_cls(cls, vllm_config: VllmConfig) -> Type[Executor]:
 
         return executor_class
 
-    def get_num_unfinished_requests(self) -> int:
+    def _get_num_core_unfinished_requests(self) -> int:
+        """Total number of unfinished requests in engine core
+        
+        Does not account for parallel sampling, i.e. a request
+        with `n=3` contributes `(3-n_complete)` to the total
+        (the parent request
+        does not count); an unfinished request with `n=1`
+        contributes 1 to the total.
+
+        Returns:
+          Total requests in engine core
+        """
         return self.detokenizer.get_num_unfinished_requests()
+    
+    def _get_num_parallel_sampling_parent_unfinished_requests(self) -> int:
+        """Total number of requests with parallel sampling
+        
+        i.e. an unfinished request with `n=<blah>` counts as 1,
+        all other requests count a 0.
+
+        Returns:
+          Number of parallel sampling parent requests
+        """
+        return len(self.parent_req_id_info)
+
+    def _get_num_parallel_sampling_child_unfinished_requests(self) -> int:
+        """Total number of parallel sampling child requests.
+        
+        i.e. an unfinished request with `n>1` counts as `(n-n_complete)`,
+        all other requests count as 0.
+
+        Returns:
+          Number of parallel sampling child requests
+        """
+        return sum([preq_info.num_child_requests_remaining() 
+                    for (_,preq_info) in self.parent_req_id_info.items()])
+
+    def get_num_unfinished_requests(self) -> int:
+        """Number of unfinished requests.
+        
+        Each request submitted by the user counts as 1; the child requests
+        spawned by parallel sampling requests are not reflected in this count.
+        """
+        return (self._get_num_core_unfinished_requests() - 
+                self._get_num_parallel_sampling_child_unfinished_requests() +
+                self._get_num_parallel_sampling_parent_unfinished_requests())
 
     def has_unfinished_requests(self) -> bool:
         return self.detokenizer.has_unfinished_requests()
@@ -127,11 +215,78 @@ def has_unfinished_requests(self) -> bool:
     def validate_outputs(cls, outputs, output_type):
         return outputs
 
+    def _forget_parallel_sample_child_request_and_maybe_parent(
+        self,
+        child_request_id:str,
+    ) -> None:
+        """Forget child request parallel sampling metadata, & its' parent's metadata if necessary.
+        
+        Parent request parallel sampling metadata is forgotten once all child requests have finished.
+
+        Args:
+          child_request_id: id of finished child request
+        """
+        # Forget child request metadata
+        parent_req_id=self.child_req_id_to_parent_req_info[child_request_id].parent_req_id
+        self.child_req_id_to_parent_req_info.pop(child_request_id, None)
+        # Track parent request's remaining child requests & erase parent request metadata
+        # if there are no remaining child requests
+        self.parent_req_id_info[parent_req_id].n_finished+=1
+        if self.parent_req_id_info[parent_req_id].num_child_requests_remaining() == 0:
+            self.parent_req_id_info.pop(parent_req_id, None)
+
+    def _maybe_forget_parallel_sample_child_requests(
+            self, possible_child_request_ids: List[str]) -> None:
+        """When a request aborts, if it is a child of a parallel sampling request,
+        forget its parallel sampling metadata. Apply this to a list of possible child
+        request ids. If the request is not associated with parallel sampling, this
+        method has no effect on it.
+        
+        Args:
+          request_ids: list of request ids to possibly forget parallel sampling metadata for
+        """
+        for possible_child_req_id in possible_child_request_ids:
+            # Check if request is a parallel sampling child request
+            if possible_child_req_id in self.child_req_id_to_parent_req_info:
+                # If so, forget child request parallel sampling metadata
+                self._forget_parallel_sample_child_request_and_maybe_parent(possible_child_req_id)
+            
+
     def abort_request(self, request_ids: List[str]) -> None:
         """Remove request_ids from EngineCore and Detokenizer."""
 
         self.engine_core.abort_requests(request_ids)
         self.detokenizer.abort_requests(request_ids)
+        self._maybe_forget_parallel_sample_child_requests(request_ids)
+
+    def _register_parallel_sampling_parent_request(
+        self,
+        parent_req_id: str,
+        parallel_sample_parent_req_info: ParallelSampleParentRequestInfo,
+    ) -> None:
+        """Register the attributes associated with a parallel sampling request (i.e. the parent request)"""
+        self.parent_req_id_info[
+            parent_req_id] = parallel_sample_parent_req_info
+
+    def _register_parallel_sampling_child_request(
+        self,
+        parallel_sample_child_req_info: ParallelSampleChildRequestInfo,
+    ) -> str:
+        """Register the association of a parallel sampling child req with its parent req.
+        
+        Generates a child request id
+
+        Side effect: internal mapping from child req id -> parent req info structure
+
+        Returns:
+          Child request id
+        """
+        parent_req_id = parallel_sample_child_req_info.parent_req_id
+        index = parallel_sample_child_req_info.index
+        child_req_id = f"{parent_req_id}_parallel_sample_{index}"
+        self.child_req_id_to_parent_req_info[
+            child_req_id] = parallel_sample_child_req_info
+        return child_req_id
 
     def add_request(
         self,
@@ -144,6 +299,29 @@ def add_request(
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
     ) -> None:
+        if isinstance(params, SamplingParams) and params.n > 1:
+            # Register parallel sampling request
+            n = params.n
+            self._register_parallel_sampling_parent_request(
+                request_id, ParallelSampleParentRequestInfo(n))
+            params.n = 1  # Engine core cannot see `n`
+            for ndx in range(n):
+                # Register child request with parent
+                child_req_id = self._register_parallel_sampling_child_request(
+                    ParallelSampleChildRequestInfo(request_id, ndx))
+                # Recurse to add child request; `n=1` prevents further recursion
+                self.add_request(
+                    request_id=child_req_id,
+                    prompt=prompt,
+                    params=params,
+                    arrival_time=arrival_time,
+                    lora_request=lora_request,
+                    trace_headers=trace_headers,
+                    prompt_adapter_request=prompt_adapter_request,
+                    priority=priority,
+                )
+            # The top-level add_request call is done
+            return
 
         # 1) Process raw inputs into the request.
         detokenizer_req, engine_core_req = self.processor.process_inputs(
@@ -156,6 +334,80 @@ def add_request(
         # 3) Add the request to EngineCore.
         self.engine_core.add_request(engine_core_req)
 
+    def _is_parallel_sampling_child_request(
+            self,
+            possible_child_request_id:str,
+    ) -> bool:
+        return possible_child_request_id in self.child_req_id_to_parent_req_info
+
+    def _maybe_get_parallel_sampling_child_request_info(
+            self,
+            possible_child_request_id: str,
+    ) -> Optional[ParallelSampleChildRequestInfo]:
+        return self.child_req_id_to_parent_req_info.get(possible_child_request_id,None)
+
+    def _merge_parallel_sampling_child_request_output_in_place(
+        self,
+        parent_req_output: RequestOutput,
+        child_req_output: RequestOutput,
+    ) -> None:
+        # Parent is finished when all children are finished
+        parent_req_output.finished=parent_req_output.finished and child_req_output.finished
+        p_met=parent_req_output.metrics
+        c_met=child_req_output.metrics
+        if p_met is None:
+            # If current parent request metrics are `None`, update with this child's metrics
+            # (which may also be None)
+            parent_req_output.metrics=c_met
+        elif c_met is not None:
+            # Only merge in child request output metrics if the child request output metrics
+            # are not `None` 
+            p_met.last_token_time=max(p_met.last_token_time,c_met.last_token_time)
+            p_met.first_scheduled_time=_none_safe_min(p_met.first_scheduled_time,
+                                                      c_met.first_scheduled_time)
+            p_met.first_token_time=_none_safe_min(p_met.first_token_time,c_met.first_token_time)
+            p_met.time_in_queue=_none_safe_sum(p_met.time_in_queue,c_met.time_in_queue)
+            p_met.finished_time=_none_safe_max(p_met.finished_time,c_met.finished_time)
+            p_met.last_token_time=max(p_met.last_token_time,c_met.last_token_time)
+            p_met.model_execute_time=_none_safe_sum(p_met.model_execute_time,c_met.model_execute_time)
+            p_met.model_forward_time=_none_safe_sum(p_met.model_forward_time,c_met.model_forward_time)
+            p_met.scheduler_time=_none_safe_sum(p_met.scheduler_time,c_met.scheduler_time)
+            p_met.time_in_queue=_none_safe_sum(p_met.time_in_queue,c_met.time_in_queue)
+        parent_req_output.outputs.extend(child_req_output.outputs)
+        parent_req_output.num_cached_tokens=_none_safe_sum(parent_req_output.num_cached_tokens,
+                                                           child_req_output.num_cached_tokens)
+
+    def _maybe_aggregate_parallel_sampling_child_requests(
+            self,
+            request_outputs: List[RequestOutput],
+    ) -> List[RequestOutput]:
+        agg_request_outputs: List[RequestOutput]=[]
+        parent_req_id_to_idx: Dict[str,int]={}
+        for req_output in request_outputs:
+            possible_child_req_id=req_output.request_id
+            maybe_child_req_info = self._maybe_get_parallel_sampling_child_request_info(possible_child_req_id)
+            if maybe_child_req_info:
+                parent_req_id=maybe_child_req_info.parent_req_id
+                if parent_req_id not in parent_req_id_to_idx:
+                    # For a particular parent id, this is the first child request output we have seen.
+                    # Repurpose the child request output structure to be the parent request output structure
+                    req_output.request_id=parent_req_id
+                    agg_request_outputs.append(req_output)
+                    # Remember where the parent request output structure resides in the output list
+                    parent_req_id_to_idx[parent_req_id]=len(agg_request_outputs)-1
+                else:
+                    # Merge this child request output into the growing request output data structure associated
+                    # with its parent.
+                    parent_req_output=agg_request_outputs[parent_req_id_to_idx[parent_req_id]]
+                    self._merge_parallel_sampling_child_request_output_in_place(parent_req_output,req_output)
+            else:
+                # Not a parallel sampling request; don't touch it
+                agg_request_outputs.append(req_output)
+        return agg_request_outputs
+        
+        
+
+
     def step(self) -> List[RequestOutput]:
 
         # 1) Get EngineCoreOutput from the EngineCore.
@@ -169,7 +421,9 @@ def step(self) -> List[RequestOutput]:
         if requests_to_abort:
             self.abort_request(requests_to_abort)
 
-        return request_outputs
+        # 4) If necessary, aggregate outputs for parallel sampling child requests
+        #    to be associated with parent request
+        return self._maybe_aggregate_parallel_sampling_child_requests(request_outputs)
 
     # TODO(rob): Can we get rid of these?