minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Tiny add more information in retract logging. (#15694)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-12-24 00:43:36 +08:00
committed by GitHub
parent 762846531f
commit 705287b2e5
1 changed files with 18 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler.py
+18 -6
View File
@@ -2055,13 +2055,19 @@ class Scheduler(
return batch return batch
# Check if decode out of memory # Check if decode out of memory
if not batch.check_decode_mem(self.decode_mem_cache_buf_multiplier) or ( if (
TEST_RETRACT and self.forward_ct % TEST_RETRACT_INTERVAL == 0 kv_full_retract_flag := not batch.check_decode_mem(
): self.decode_mem_cache_buf_multiplier
)
) or (TEST_RETRACT and self.forward_ct % TEST_RETRACT_INTERVAL == 0):
old_available_tokens = self.token_to_kv_pool_allocator.available_size()
old_ratio = self.new_token_ratio old_ratio = self.new_token_ratio
retracted_reqs, new_token_ratio, reqs_to_abort = batch.retract_decode( retracted_reqs, new_token_ratio, reqs_to_abort = batch.retract_decode(
self.server_args, self.decode_mem_cache_buf_multiplier self.server_args, self.decode_mem_cache_buf_multiplier
) )
new_available_tokens = self.token_to_kv_pool_allocator.available_size()
new_token_gained = new_available_tokens - old_available_tokens
self.num_retracted_reqs = len(retracted_reqs) self.num_retracted_reqs = len(retracted_reqs)
if self.enable_metrics and (x := len(retracted_reqs)) > 0: if self.enable_metrics and (x := len(retracted_reqs)) > 0:
self.metrics_collector.increment_num_retracted_reqs(x) self.metrics_collector.increment_num_retracted_reqs(x)
@@ -2072,11 +2078,17 @@ class Scheduler(
AbortReq(abort_message=abort_reason.message, rid=req.rid), req AbortReq(abort_message=abort_reason.message, rid=req.rid), req
) )
logger.info( msg_prefix = (
"KV cache pool is full. Retract requests. " "KV cache pool is full. Retract requests. "
f"#retracted_reqs: {len(retracted_reqs)}, " if kv_full_retract_flag
f"#new_token_ratio: {old_ratio:.4f} -> {new_token_ratio:.4f}" else "Testing retraction. "
) )
msg_details = f"#retracted_reqs: {len(retracted_reqs)}, #new_tokens_gained: {new_token_gained}"
if kv_full_retract_flag:
msg_details += (
f", #new_token_ratio: {old_ratio:.4f} -> {new_token_ratio:.4f}"
)
logger.warning(msg_prefix + msg_details)
for req in retracted_reqs: for req in retracted_reqs:
self._add_request_to_queue(req, is_retracted=True) self._add_request_to_queue(req, is_retracted=True)