A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Fix] Remove divisibility requirement between num_kv_heads and tp_size in bailing_moe (#26876)
Signed-off-by: vito.yy <vito.yy@antgroup.com>
ant-yy committed
Oct 15, 2025 at 16:44 UTC
5c3bae1a6a73aad8c0883a097079448b506fbfcc
1 file changed
+1
-2
vllm/model_executor/models/bailing_moe.py
+1
-2
@@ -86,13 +86,12 @@ class BailingAttention(nn.Module):
86
tp_size = get_tensor_model_parallel_world_size()
87
88
assert self.total_num_heads % tp_size == 0
89
- assert self.total_kv_heads % tp_size == 0
89
assert self.total_num_heads >= self.total_kv_heads
90
91
self.num_heads = self.total_num_heads // tp_size
92
self.head_dim = config.head_dim or (self.hidden_size // self.total_num_heads)
93
self.q_size_per_rank = self.head_dim * self.num_heads
95
- self.num_kv_heads = self.total_kv_heads // tp_size
94
+ self.num_kv_heads = max(1, self.total_kv_heads // tp_size)
95
self.kv_size_per_rank = self.num_kv_heads * self.head_dim
96
self.scale = self.head_dim**-0.5
97
self.use_qk_norm = getattr(config, "use_qk_norm", False)