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 13, 14:30 UTC
[Misc] Add some comments in qwen3-next (#28267)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Nov 9, 2025 at 15:59 UTC
7ae5a5fb11151e029609009b7950cc46ff097407
1 file changed
+2
vllm/model_executor/models/qwen3_next.py
+2
@@ -462,6 +462,8 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
462
# ============================================================
463
# Part 2: Core Attention (Custom Op)
464
# ============================================================
465
+ # Note: we should not use torch.empty here like other attention backends,
466
+ # see discussions in https://github.com/vllm-project/vllm/pull/28182
467
core_attn_out = torch.zeros(
468
(num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim),
469
dtype=hidden_states.dtype,