Fix get_load API (#13991)

This commit is contained in:
Liangsheng Yin
2025-11-26 22:07:51 +08:00
committed by GitHub
parent b704b0a922
commit eff6a07c8f
5 changed files with 41 additions and 81 deletions
+7 -17
View File
@@ -171,6 +171,10 @@ class DecodeRequest:
waiting_for_input: bool = False
metadata_buffer_index: int = -1
@property
def seqlen(self) -> int:
return self.req.seqlen
class DecodePreallocQueue:
"""
@@ -299,11 +303,6 @@ class DecodePreallocQueue:
if is_retracted:
self.retracted_queue.append(req)
self.scheduler.num_waiting_tokens += (
len(req.origin_input_ids)
+ len(req.output_ids)
+ self.num_reserved_decode_tokens
)
else:
if req.bootstrap_host == FAKE_BOOTSTRAP_HOST:
kv_receiver_class = get_kv_class(
@@ -326,7 +325,6 @@ class DecodePreallocQueue:
self.queue.append(
DecodeRequest(req=req, kv_receiver=kv_receiver, waiting_for_input=False)
)
self.scheduler.num_waiting_tokens += len(req.origin_input_ids)
def _check_if_req_exceed_kv_capacity(self, req: Req) -> bool:
if len(req.origin_input_ids) > self.max_total_num_tokens:
@@ -367,8 +365,6 @@ class DecodePreallocQueue:
req.is_retracted = False
self._pre_alloc(req)
allocatable_tokens -= required_tokens_for_request
# update counter for resumed request
self.scheduler.num_waiting_tokens -= required_tokens_for_request
# load from cpu, release the cpu copy
req.load_kv_cache(self.req_to_token_pool, self.token_to_kv_pool_allocator)
@@ -544,15 +540,9 @@ class DecodePreallocQueue:
RequestStage.DECODE_BOOTSTRAP, decode_req.req.rid, auto_next_anon=True
)
preallocated_queue = []
for i, entry in enumerate(self.queue):
if i not in indices_to_remove:
preallocated_queue.append(entry)
else:
# update counter for allocated or failed request
self.scheduler.num_waiting_tokens -= len(entry.req.origin_input_ids)
self.queue = preallocated_queue
self.queue = [
entry for i, entry in enumerate(self.queue) if i not in indices_to_remove
]
return preallocated_reqs
@@ -200,7 +200,6 @@ class PrefillBootstrapQueue:
self._process_req(req)
req.add_latency(RequestStage.PREFILL_PREPARE)
self.queue.append(req)
self.scheduler.num_waiting_tokens += len(req.origin_input_ids)
trace_slice_end(RequestStage.PREFILL_PREPARE, req.rid, auto_next_anon=True)
def extend(self, reqs: List[Req], num_kv_heads: int) -> None:
@@ -271,8 +270,6 @@ class PrefillBootstrapQueue:
self.scheduler.stream_output([req], req.return_logprob)
indices_to_remove.add(i)
failed_reqs.append(req)
# update counter for failed bootstrap request
self.scheduler.num_waiting_tokens -= len(req.origin_input_ids)
if self.scheduler.enable_metrics:
self.scheduler.metrics_collector.increment_bootstrap_failed_reqs()
continue