[Renderer] Move InputPreprocessor into Renderer (1/2) (#34510)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Cyrus Leung
2026-02-14 10:14:21 -08:00
committed by GitHub
parent b3c14229b0
commit 73391a1baa
39 changed files with 456 additions and 458 deletions
+90 -6
View File
@@ -18,7 +18,7 @@ import dataclasses
import json
import time
from datetime import datetime
from typing import Any
from typing import TYPE_CHECKING, Any
import numpy as np
@@ -28,9 +28,6 @@ from vllm.benchmarks.datasets import (
)
from vllm.benchmarks.throughput import get_requests
from vllm.engine.arg_utils import EngineArgs
from vllm.multimodal.processing.context import (
get_timing_stats_from_engine_client,
)
from vllm.utils.gc_utils import freeze_gc_heap
from vllm.utils.import_utils import PlaceholderModule
@@ -39,16 +36,103 @@ try:
except ImportError:
pd = PlaceholderModule("pandas")
if TYPE_CHECKING: # Avoid having to mock during docs build
from vllm.v1.engine.llm_engine import LLMEngine
else:
LLMEngine = object
def get_timing_stats_from_engine(llm_engine: LLMEngine) -> dict[str, dict[str, float]]:
"""
Get all multimodal timing stats from the LLM engine.
Collects both preprocessing stats (HF processor, hashing, cache lookup,
prompt update) and encoder forward pass timing, merged by request_id.
Args:
llm_engine: The LLM engine (has input_processor and workers).
Returns:
Dictionary mapping request_id to merged stats dict containing
both preprocessing and encoder timing metrics.
Example:
{
'request-123': {
'hf_processor_time': 0.45,
'hashing_time': 0.02,
'cache_lookup_time': 0.01,
'prompt_update_time': 0.03,
'preprocessor_total_time': 0.51,
'encoder_forward_time': 0.23,
'num_encoder_calls': 1
}
}
"""
observability_config = llm_engine.vllm_config.observability_config
if not observability_config or not observability_config.enable_mm_processor_stats:
return {}
renderer = llm_engine.renderer
mm_processor = renderer.get_mm_processor()
preprocessing_stats = mm_processor.info.ctx.get_all_timing_stats()
encoder_stats = dict[str, dict[str, float]]()
for worker_stats in llm_engine.collective_rpc("get_encoder_timing_stats"):
if not worker_stats:
continue
for request_id, stats_dict in worker_stats.items():
if request_id not in encoder_stats:
encoder_stats[request_id] = dict(stats_dict)
else:
# Aggregate timing metrics across workers
current_time = encoder_stats[request_id].get(
"encoder_forward_time", 0.0
)
new_time = stats_dict.get("encoder_forward_time", 0.0)
encoder_stats[request_id]["encoder_forward_time"] = max(
current_time, new_time
)
current_calls = encoder_stats[request_id].get("num_encoder_calls", 0)
new_calls = stats_dict.get("num_encoder_calls", 0)
encoder_stats[request_id]["num_encoder_calls"] = max(
current_calls, new_calls
)
merged_stats = dict[str, dict[str, float]]()
for request_id, prep_dict in preprocessing_stats.items():
merged_stats[request_id] = dict(prep_dict)
for request_id, enc_dict in encoder_stats.items():
if request_id in merged_stats:
merged_stats[request_id].update(enc_dict)
continue
# In V1 engine, the request_id in encoder_stats has a suffix
# appended to the original request_id (which is used in
# preprocessing_stats).
# We try to strip the suffix to find the matching request.
possible_original_id = request_id.rpartition("-")[0]
if possible_original_id and possible_original_id in merged_stats:
merged_stats[possible_original_id].update(enc_dict)
else:
merged_stats[request_id] = dict(enc_dict)
return merged_stats
def collect_mm_processor_stats(
llm_engine: Any,
llm_engine: LLMEngine,
num_warmup_reqs: int = 0,
) -> dict[str, list[float]]:
"""
Collect multimodal processor timing stats.
Returns a dictionary mapping stage names to lists of timing values (in seconds).
"""
all_stats = get_timing_stats_from_engine_client(llm_engine)
all_stats = get_timing_stats_from_engine(llm_engine)
stat_keys = [
"hf_processor_time",