Allow AsyncLLMEngine.generate to target a specific DP rank (#19102)
Signed-off-by: Jon Swenson <jmswen@gmail.com>
This commit is contained in:
@@ -55,6 +55,7 @@ class EngineCoreRequest(
|
||||
arrival_time: float
|
||||
lora_request: Optional[LoRARequest]
|
||||
cache_salt: Optional[str]
|
||||
data_parallel_rank: Optional[int]
|
||||
|
||||
# Index of the client, used to ensure outputs are sent back to the same
|
||||
# client for this request when scaling out the front-end.
|
||||
|
||||
@@ -229,6 +229,7 @@ class AsyncLLM(EngineClient):
|
||||
trace_headers: Optional[Mapping[str, str]] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
priority: int = 0,
|
||||
data_parallel_rank: Optional[int] = None,
|
||||
) -> RequestOutputCollector:
|
||||
"""Add new request to the AsyncLLM."""
|
||||
|
||||
@@ -245,7 +246,7 @@ class AsyncLLM(EngineClient):
|
||||
prompt_str, request = self.processor.process_inputs(
|
||||
request_id, prompt, params, arrival_time, lora_request,
|
||||
tokenization_kwargs, trace_headers, prompt_adapter_request,
|
||||
priority)
|
||||
priority, data_parallel_rank)
|
||||
|
||||
if params.n == 1:
|
||||
await self._add_request(request, prompt_str, None, 0, queue)
|
||||
@@ -291,6 +292,7 @@ class AsyncLLM(EngineClient):
|
||||
trace_headers: Optional[Mapping[str, str]] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
priority: int = 0,
|
||||
data_parallel_rank: Optional[int] = None,
|
||||
) -> AsyncGenerator[RequestOutput, None]:
|
||||
"""
|
||||
Main function called by the API server to kick off a request
|
||||
@@ -321,6 +323,7 @@ class AsyncLLM(EngineClient):
|
||||
trace_headers=trace_headers,
|
||||
prompt_adapter_request=prompt_adapter_request,
|
||||
priority=priority,
|
||||
data_parallel_rank=data_parallel_rank,
|
||||
)
|
||||
|
||||
# The output_handler task pushes items into the queue.
|
||||
|
||||
@@ -982,7 +982,16 @@ class DPAsyncMPClient(AsyncMPClient):
|
||||
resources.stats_update_task = asyncio.create_task(
|
||||
run_engine_stats_update_task())
|
||||
|
||||
def get_core_engine_for_request(self) -> CoreEngine:
|
||||
def get_core_engine_for_request(self,
|
||||
dp_rank: Optional[int] = None
|
||||
) -> CoreEngine:
|
||||
if dp_rank is not None:
|
||||
# engines are already in rank order
|
||||
if dp_rank < 0 or dp_rank >= len(self.core_engines):
|
||||
raise ValueError(f"Requested DP rank {dp_rank} is out of "
|
||||
f"range [0, {len(self.core_engines)})")
|
||||
return self.core_engines[dp_rank]
|
||||
|
||||
if not self.lb_engines:
|
||||
return self.core_engines[0]
|
||||
# TODO use P2C alg for larger DP sizes
|
||||
@@ -1018,7 +1027,8 @@ class DPAsyncMPClient(AsyncMPClient):
|
||||
request.current_wave = self.current_wave
|
||||
request.client_index = self.client_index
|
||||
|
||||
chosen_engine = self.get_core_engine_for_request()
|
||||
chosen_engine = self.get_core_engine_for_request(
|
||||
request.data_parallel_rank)
|
||||
self.reqs_in_flight[request.request_id] = chosen_engine
|
||||
|
||||
to_await = self._send_input(EngineCoreRequestType.ADD, request,
|
||||
|
||||
@@ -212,6 +212,7 @@ class Processor:
|
||||
trace_headers: Optional[Mapping[str, str]] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
priority: int = 0,
|
||||
data_parallel_rank: Optional[int] = None,
|
||||
) -> tuple[Optional[str], EngineCoreRequest]:
|
||||
|
||||
# TODO(woosuk): Support pooling models.
|
||||
@@ -328,6 +329,7 @@ class Processor:
|
||||
arrival_time=arrival_time,
|
||||
lora_request=lora_request,
|
||||
cache_salt=decoder_inputs.get("cache_salt"),
|
||||
data_parallel_rank=data_parallel_rank,
|
||||
)
|
||||
|
||||
def _validate_model_inputs(self,
|
||||
|
||||
Reference in New Issue
Block a user