EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[RL] fix skip_server_warmup and rl health_generate logic (#8757)

Browse Source
This commit is contained in:
Zilin Zhu
2025-08-08 19:34:38 +08:00
committed by GitHub
parent a947154286
commit dd650e0e21
3 changed files with 12 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/entrypoints/http_server.py
View File

@@ -1172,6 +1172,8 @@ def _wait_and_warmup(
pipe_finish_writer, pipe_finish_writer,
): ):
return return
else:
_global_state.tokenizer_manager.server_status = ServerStatus.Up
logger.info("The server is fired up and ready to roll!") logger.info("The server is fired up and ready to roll!")

5
python/sglang/srt/managers/scheduler.py
View File

@@ -473,6 +473,7 @@ class Scheduler(
self.memory_saver_adapter = TorchMemorySaverAdapter.create( self.memory_saver_adapter = TorchMemorySaverAdapter.create(
enable=server_args.enable_memory_saver enable=server_args.enable_memory_saver
) )
self.offload_tags = set()
self.init_profier() self.init_profier()
self.recv_skipper = SchedulerRecvSkipper.maybe_create(server_args) self.recv_skipper = SchedulerRecvSkipper.maybe_create(server_args)
@@ -1040,7 +1041,9 @@ class Scheduler(
for recv_req in recv_reqs: for recv_req in recv_reqs:
# If it is a health check generation request and there are running requests, ignore it. # If it is a health check generation request and there are running requests, ignore it.
if is_health_check_generate_req(recv_req) and ( if is_health_check_generate_req(recv_req) and (
self.chunked_req is not None or not self.running_batch.is_empty() self.chunked_req is not None
or not self.running_batch.is_empty()
or len(self.offload_tags) > 0
): ):
self.return_health_check_ct += 1 self.return_health_check_ct += 1
continue continue

6
python/sglang/srt/managers/scheduler_update_weights_mixin.py
View File

@@ -78,6 +78,9 @@ class SchedulerUpdateWeightsMixin:
if tags is None or len(tags) == 0: if tags is None or len(tags) == 0:
tags = [GPU_MEMORY_TYPE_WEIGHTS, GPU_MEMORY_TYPE_KV_CACHE] tags = [GPU_MEMORY_TYPE_WEIGHTS, GPU_MEMORY_TYPE_KV_CACHE]
for tag in tags:
self.offload_tags.add(tag)
if GPU_MEMORY_TYPE_KV_CACHE in tags: if GPU_MEMORY_TYPE_KV_CACHE in tags:
self.memory_saver_adapter.pause(GPU_MEMORY_TYPE_KV_CACHE) self.memory_saver_adapter.pause(GPU_MEMORY_TYPE_KV_CACHE)
self.flush_cache() self.flush_cache()
@@ -97,6 +100,9 @@ class SchedulerUpdateWeightsMixin:
if tags is None or len(tags) == 0: if tags is None or len(tags) == 0:
tags = [GPU_MEMORY_TYPE_WEIGHTS, GPU_MEMORY_TYPE_KV_CACHE] tags = [GPU_MEMORY_TYPE_WEIGHTS, GPU_MEMORY_TYPE_KV_CACHE]
for tag in tags:
self.offload_tags.remove(tag)
if GPU_MEMORY_TYPE_WEIGHTS in tags: if GPU_MEMORY_TYPE_WEIGHTS in tags:
self.memory_saver_adapter.resume(GPU_MEMORY_TYPE_WEIGHTS) self.memory_saver_adapter.resume(GPU_MEMORY_TYPE_WEIGHTS)
torch.distributed.barrier(self.tp_cpu_group) torch.distributed.barrier(self.tp_cpu_group)