minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

Fix BailingMoeV3 reading enable_dp_lm_head off live topology instead of config (#36584)

Browse Source
This commit is contained in:
Alison Shao
2026-08-26 18:41:39 -07:00
committed by GitHub
parent a8b72c4c34
commit 15688fea7d
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/bailing_moe_v3.py
+1 -1
View File
@@ -1333,7 +1333,7 @@ class BailingMoeV3ForCausalLM(nn.Module):
# in the logits processor. Accuracy-neutral on ling-v3. # in the logits processor. Accuracy-neutral on ling-v3.
params_dtype=torch.bfloat16, params_dtype=torch.bfloat16,
quant_config=quant_config, quant_config=quant_config,
use_attn_tp_group=get_parallel().enable_dp_lm_head, use_attn_tp_group=get_parallel().config.enable_dp_lm_head,
) )
) )
self.logits_processor = LogitsProcessor(config) self.logits_processor = LogitsProcessor(config)
test/registered/unit/models/test_shared_experts_fusion_gates.py
+1 -1
View File
@@ -335,7 +335,7 @@ class TestBailingMoeV3Gate(_FusionGateCase):
parallel = SimpleNamespace( parallel = SimpleNamespace(
tp_size=1, tp_size=1,
moe_ep_size=1, moe_ep_size=1,
enable_dp_lm_head=False, config=SimpleNamespace(enable_dp_lm_head=False),
) )
with ( with (
unittest.mock.patch.object( unittest.mock.patch.object(