minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Disagg] Fix MegaMoE topk_ids dtype mismatch and FakeKVManager missing kv_args (#25380)

Browse Source 
Co-authored-by: JoeLee314 <liqichao@baidu.com>
This commit is contained in:
JoeLee314
2026-05-15 22:46:44 -07:00
committed by GitHub
co-authored by JoeLee314
parent daade9cc00
commit bda01d2435
2 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/disaggregation/fake/conn.py
+1
View File
@@ -28,6 +28,7 @@ class FakeKVManager(BaseKVManager):
is_mla_backend: Optional[bool] = False, is_mla_backend: Optional[bool] = False,
): ):
super().__init__(args, disaggregation_mode, server_args, is_mla_backend) super().__init__(args, disaggregation_mode, server_args, is_mla_backend)
self.kv_args = args
self.req_to_decode_prefix_len = {} self.req_to_decode_prefix_len = {}
def register_to_bootstrap(self): def register_to_bootstrap(self):
python/sglang/srt/layers/moe/mega_moe.py
+2 -2
View File
@@ -206,8 +206,8 @@ def _run_mega_routed(
) )
if num_tokens > 0: if num_tokens > 0:
topk_ids_in = topk_ids topk_ids_in = topk_ids.to(torch.int32)
topk_weights_in = topk_weights topk_weights_in = topk_weights.to(torch.float32)
else: else:
topk_ids_in = hidden_states.new_empty((0, top_k), dtype=torch.int32) topk_ids_in = hidden_states.new_empty((0, top_k), dtype=torch.int32)
topk_weights_in = hidden_states.new_empty((0, top_k), dtype=torch.float32) topk_weights_in = hidden_states.new_empty((0, top_k), dtype=torch.float32)