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

[DLLM] Fix dLLM regression (#15371)

Browse Source
This commit is contained in:
Zehuan Li
2025-12-19 10:55:46 +08:00
committed by GitHub
parent e88e75a9df
commit f6c9db4bc4
2 changed files with 6 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler.py
+5 -7
View File
@@ -1834,13 +1834,11 @@ class Scheduler(
# Determine chunked_prefill_size for this batch # Determine chunked_prefill_size for this batch
chunked_prefill_size = self.chunked_prefill_size chunked_prefill_size = self.chunked_prefill_size
if self.chunked_req is not None: if self.chunked_req is not None and self.enable_dynamic_chunking:
self.chunked_req.init_next_round_input() history_len = len(self.chunked_req.prefix_indices)
if self.enable_dynamic_chunking: dynamic_size = self.predict_next_chunk_size(history_len)
history_len = len(self.chunked_req.prefix_indices) if dynamic_size is not None:
dynamic_size = self.predict_next_chunk_size(history_len) chunked_prefill_size = dynamic_size
if dynamic_size is not None:
chunked_prefill_size = dynamic_size
# Prefill policy # Prefill policy
adder = PrefillAdder( adder = PrefillAdder(
test/srt/dllm/test_llada2_mini.py
+1 -1
View File
@@ -57,7 +57,7 @@ class TestLLaDA2Mini(CustomTestCase):
metrics = run_eval_few_shot_gsm8k(args) metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}") print(f"{metrics=}")
self.assertGreater(metrics["accuracy"], 0.6) self.assertGreater(metrics["accuracy"], 0.88)
self.assertGreater(metrics["output_throughput"], 150) self.assertGreater(metrics["output_throughput"], 150)
def test_bs_1_speed(self): def test_bs_1_speed(self):