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 6, 15:45 UTC
[BUGFIX][Mamba] Use uint64 for address in KVBlockZeroer (#37197)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Mar 17, 2026 at 04:39 UTC
d157216093ac50603bab5c2236437cdc68512f6d
1 file changed
+1
-1
vllm/v1/worker/utils.py
+1
-1
@@ -180,7 +180,7 @@ class KVBlockZeroer:
180
)
181
self._ids_gpu = torch.empty(self._id_cap, dtype=torch.int64, device=self.device)
182
self._meta = (
183
- torch.tensor(seg_addrs, dtype=torch.int64, device=self.device),
183
+ torch.tensor(seg_addrs, dtype=torch.uint64, device=self.device),
184
page_size_el,
185
blk_size,
186
len(seg_addrs),