Filter non-int token ids in benchmark and observe decode-side bootstrap/alloc metrics (#24684)
This commit is contained in:
@@ -66,8 +66,12 @@ def compute_random_lens(full_len: int, range_ratio: float, num: int) -> List[int
|
|||||||
|
|
||||||
@lru_cache(maxsize=1)
|
@lru_cache(maxsize=1)
|
||||||
def get_available_tokens(tokenizer):
|
def get_available_tokens(tokenizer):
|
||||||
"""Get all available token ids from the tokenizer vocabulary."""
|
"""Get valid token ids from the tokenizer vocabulary."""
|
||||||
return list(tokenizer.get_vocab().values())
|
return [
|
||||||
|
token_id
|
||||||
|
for token_id in tokenizer.get_vocab().values()
|
||||||
|
if isinstance(token_id, int)
|
||||||
|
]
|
||||||
|
|
||||||
|
|
||||||
def gen_prompt(tokenizer, token_num):
|
def gen_prompt(tokenizer, token_num):
|
||||||
|
|||||||
@@ -936,6 +936,16 @@ class SchedulerReqTimeStats(ReqTimeStatsBase):
|
|||||||
)
|
)
|
||||||
self.trace_slice(stage, self.decode_prealloc_queue_entry_time, ts)
|
self.trace_slice(stage, self.decode_prealloc_queue_entry_time, ts)
|
||||||
|
|
||||||
|
if self.enable_metrics and self.bootstrap_done_time > 0:
|
||||||
|
bootstrap_ms = (
|
||||||
|
self.bootstrap_done_time - self.decode_prealloc_queue_entry_time
|
||||||
|
) * 1000
|
||||||
|
alloc_ms = (ts - self.bootstrap_done_time) * 1000
|
||||||
|
self.metrics_collector.observe_kv_transfer_bootstrap(
|
||||||
|
bootstrap_ms=bootstrap_ms,
|
||||||
|
alloc_ms=alloc_ms,
|
||||||
|
)
|
||||||
|
|
||||||
def set_bootstrap_done_time(self, ts=None):
|
def set_bootstrap_done_time(self, ts=None):
|
||||||
ts = ts or time.perf_counter()
|
ts = ts or time.perf_counter()
|
||||||
if self.bootstrap_done_time == 0.0:
|
if self.bootstrap_done_time == 0.0:
|
||||||
|
|||||||
@@ -1,10 +1,10 @@
|
|||||||
"""
|
"""
|
||||||
Usage:
|
Usage:
|
||||||
# replay from a folder
|
# replay from a folder
|
||||||
python3 replay_request_dump.py --file-number 100 --parallel 512 --input-folder /data/lianmin/sglang_request_dump/grok-mini-0220-engine-5756f8f94-28bm6/
|
python3 replay_request_dump.py --file-number 100 --parallel 512 --input-folder /data/lianmin/sglang_request_dump/engine-34xd1/
|
||||||
|
|
||||||
# replay from a single file
|
# replay from a single file
|
||||||
python3 replay_request_dump.py --parallel 512 --input-file /data/sglang_crash_dump/memx-cti-34-sr1.xpop.twttr.net/crash_dump_2025-06-04_20-13-18.pkl
|
python3 replay_request_dump.py --parallel 512 --input-file /data/sglang_crash_dump/crash_dump_2025-06-04_20-13-18.pkl
|
||||||
"""
|
"""
|
||||||
|
|
||||||
import argparse
|
import argparse
|
||||||
|
|||||||
Reference in New Issue
Block a user