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

clean up step function (#635)

Browse Source
This commit is contained in:
Ying Sheng
2024-07-16 20:15:24 -07:00
committed by GitHub
parent 5f90e0769c
commit a470e60c97
1 changed files with 33 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
python/sglang/srt/managers/controller/tp_worker.py
View File

@@ -228,23 +228,7 @@ class ModelTpServer:
# Print stats # Print stats
if self.tp_rank == 0 and self.decode_forward_ct % 40 == 0: if self.tp_rank == 0 and self.decode_forward_ct % 40 == 0:
num_used = self.max_total_num_tokens - ( self.print_stats()
self.token_to_kv_pool.available_size()
+ self.tree_cache.evictable_size()
)
throughput = self.num_generated_tokens / (
time.time() - self.last_stats_tic
)
self.num_generated_tokens = 0
self.last_stats_tic = time.time()
logger.info(
f"[gpu_id={self.gpu_id}] Decode batch. "
f"#running-req: {len(self.running_batch.reqs)}, "
f"#token: {num_used}, "
f"token usage: {num_used / self.max_total_num_tokens:.2f}, "
f"gen throughput (token/s): {throughput:.2f}, "
f"#queue-req: {len(self.forward_queue)}"
)
if self.running_batch.is_empty(): if self.running_batch.is_empty():
self.running_batch = None self.running_batch = None
@@ -253,17 +237,38 @@ class ModelTpServer:
if self.out_pyobjs and self.running_batch.has_stream(): if self.out_pyobjs and self.running_batch.has_stream():
break break
else: else:
# Check the available size self.check_memory()
available_size = (
self.token_to_kv_pool.available_size() def print_stats(self):
+ self.tree_cache.evictable_size() num_used = self.max_total_num_tokens - (
) self.token_to_kv_pool.available_size()
if available_size != self.max_total_num_tokens: + self.tree_cache.evictable_size()
warnings.warn( )
"Warning: " throughput = self.num_generated_tokens / (
f"available_size={available_size}, max_total_num_tokens={self.max_total_num_tokens}\n" time.time() - self.last_stats_tic
"KV cache pool leak detected!" )
) self.num_generated_tokens = 0
self.last_stats_tic = time.time()
logger.info(
f"[gpu_id={self.gpu_id}] Decode batch. "
f"#running-req: {len(self.running_batch.reqs)}, "
f"#token: {num_used}, "
f"token usage: {num_used / self.max_total_num_tokens:.2f}, "
f"gen throughput (token/s): {throughput:.2f}, "
f"#queue-req: {len(self.forward_queue)}"
)
def check_memory(self):
available_size = (
self.token_to_kv_pool.available_size()
+ self.tree_cache.evictable_size()
)
if available_size != self.max_total_num_tokens:
warnings.warn(
"Warning: "
f"available_size={available_size}, max_total_num_tokens={self.max_total_num_tokens}\n"
"KV cache pool leak detected!"
)
def handle_generate_request( def handle_generate_request(
self, self,