[Core] Add span metrics for model_forward, scheduler and sampler time (#7089)

This commit is contained in:
Mahesh Keralapura
2024-08-09 13:55:13 -07:00
committed by GitHub
parent 70d268a399
commit 933790c209
17 changed files with 189 additions and 21 deletions
+5 -3
View File
@@ -4,8 +4,8 @@ from typing import Any, Dict, List, Optional, Tuple, Type
import torch
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
ModelConfig, MultiModalConfig, ParallelConfig,
PromptAdapterConfig, SchedulerConfig)
ModelConfig, MultiModalConfig, ObservabilityConfig,
ParallelConfig, PromptAdapterConfig, SchedulerConfig)
from vllm.logger import init_logger
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.multimodal import MultiModalInputs
@@ -45,6 +45,7 @@ class EmbeddingModelRunner(
is_driver_worker: bool = False,
prompt_adapter_config: Optional[PromptAdapterConfig] = None,
multimodal_config: Optional[MultiModalConfig] = None,
observability_config: Optional[ObservabilityConfig] = None,
):
super().__init__(model_config,
parallel_config,
@@ -56,7 +57,8 @@ class EmbeddingModelRunner(
kv_cache_dtype=kv_cache_dtype,
is_driver_worker=is_driver_worker,
prompt_adapter_config=prompt_adapter_config,
multimodal_config=multimodal_config)
multimodal_config=multimodal_config,
observability_config=observability_config)
@torch.inference_mode()
def execute_model(