ef6abf3dc7
Adds ALL files needed for Dockerfile build:
- qwen3_6_scripts/ (baseline patches + our optimizations)
- vllm/ (full vllm package)
- paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
- Dockerfile + computility-run.yaml
Our optimizations vs baseline:
1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
Triton try/fallback, V2 heuristic, threshold 32K→64K
2. paged_attention_v2_pytorch.py: fills NotImplementedError,
single-bmm Phase 1 (195 launches → 3)
3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
max-num-batched-tokens 8192→16384
This repo can now be submitted to dev.modelhub.org.cn as-is.
136 lines
3.7 KiB
Python
136 lines
3.7 KiB
Python
from dataclasses import dataclass
|
|
from enum import Enum
|
|
from typing import List, Mapping, Optional, Union, overload
|
|
|
|
from vllm import PoolingParams
|
|
from vllm.inputs import PromptType
|
|
from vllm.lora.request import LoRARequest
|
|
from vllm.outputs import RequestOutput
|
|
from vllm.prompt_adapter.request import PromptAdapterRequest
|
|
from vllm.sampling_params import SamplingParams
|
|
from vllm.utils import deprecate_kwargs
|
|
|
|
VLLM_RPC_SUCCESS_STR = "SUCCESS"
|
|
|
|
IPC_INPUT_EXT = "_input_socket"
|
|
IPC_OUTPUT_EXT = "_output_socket"
|
|
IPC_HEALTH_EXT = "_health_socket"
|
|
IPC_DATA_EXT = "_data_socket"
|
|
|
|
|
|
class MQEngineDeadError(RuntimeError):
|
|
pass
|
|
|
|
|
|
@dataclass
|
|
class RPCProcessRequest:
|
|
prompt: PromptType
|
|
params: Union[SamplingParams, PoolingParams]
|
|
request_id: str
|
|
lora_request: Optional[LoRARequest] = None
|
|
trace_headers: Optional[Mapping[str, str]] = None
|
|
prompt_adapter_request: Optional[PromptAdapterRequest] = None
|
|
priority: int = 0
|
|
|
|
@overload # DEPRECATED
|
|
def __init__(
|
|
self,
|
|
*,
|
|
inputs: PromptType,
|
|
params: Union[SamplingParams, PoolingParams],
|
|
request_id: str,
|
|
lora_request: Optional[LoRARequest] = None,
|
|
trace_headers: Optional[Mapping[str, str]] = None,
|
|
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
|
priority: int = 0,
|
|
) -> None:
|
|
...
|
|
|
|
@overload
|
|
def __init__(
|
|
self,
|
|
prompt: PromptType,
|
|
params: Union[SamplingParams, PoolingParams],
|
|
request_id: str,
|
|
lora_request: Optional[LoRARequest] = None,
|
|
trace_headers: Optional[Mapping[str, str]] = None,
|
|
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
|
priority: int = 0,
|
|
) -> None:
|
|
...
|
|
|
|
@deprecate_kwargs(
|
|
"inputs",
|
|
additional_message="Please use the 'prompt' parameter instead.",
|
|
)
|
|
def __init__(
|
|
self,
|
|
prompt: Optional[PromptType] = None,
|
|
params: Optional[Union[SamplingParams, PoolingParams]] = None,
|
|
request_id: Optional[str] = None,
|
|
lora_request: Optional[LoRARequest] = None,
|
|
trace_headers: Optional[Mapping[str, str]] = None,
|
|
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
|
priority: int = 0,
|
|
*,
|
|
inputs: Optional[PromptType] = None, # DEPRECATED
|
|
) -> None:
|
|
if inputs is not None:
|
|
prompt = inputs
|
|
assert (prompt is not None and params is not None
|
|
and request_id is not None)
|
|
|
|
super().__init__()
|
|
|
|
self.prompt = prompt
|
|
self.params = params
|
|
self.request_id = request_id
|
|
self.lora_request = lora_request
|
|
self.trace_headers = trace_headers
|
|
self.prompt_adapter_request = prompt_adapter_request
|
|
self.priority = priority
|
|
|
|
|
|
@dataclass
|
|
class RPCError:
|
|
request_id: Optional[str]
|
|
is_engine_errored: bool
|
|
exception: BaseException
|
|
|
|
|
|
@dataclass
|
|
class RPCAbortRequest:
|
|
request_id: str
|
|
|
|
|
|
class RPCStartupRequest(Enum):
|
|
IS_SERVER_READY = 1
|
|
|
|
|
|
@dataclass
|
|
class RPCStartupResponse:
|
|
tracing_enabled: bool
|
|
|
|
|
|
class RPCUProfileRequest(Enum):
|
|
START_PROFILE = 1
|
|
STOP_PROFILE = 2
|
|
|
|
|
|
RPC_REQUEST_T = Union[RPCProcessRequest, RPCAbortRequest, RPCStartupRequest,
|
|
RPCUProfileRequest]
|
|
|
|
REQUEST_OUTPUTS_T = Union[List[RequestOutput], RPCError]
|
|
|
|
|
|
def ENGINE_DEAD_ERROR(
|
|
error: Optional[BaseException] = None) -> MQEngineDeadError:
|
|
if error is None:
|
|
return MQEngineDeadError(
|
|
"Engine loop is not running. Inspect the stacktrace to "
|
|
"find the original error")
|
|
|
|
return MQEngineDeadError(
|
|
"Engine loop is not running. Inspect the stacktrace to "
|
|
f"find the original error: {repr(error)}.")
|