[Misc] Use meta tensor for KV cache stride calculation (#47316)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Lucas Wilkinson committed
Jul 10, 2026 at 23:24 UTC
04d553f390fd37e09ab111936ef1592881299957
1 file changed
+1
-1
vllm/v1/worker/gpu/attn_utils.py
+1
-1
@@ -238,7 +238,7 @@ def _reshape_attention_kv_cache(
238
page_stride = kv_cache_spec.page_size_bytes // dtype_size
239
240
num_blocks_dim = inv_order[0]
241
- strides = list(torch.empty(permuted_kv_cache_shape).stride())
241
+ strides = list(torch.empty(permuted_kv_cache_shape, device="meta").stride())
242
strides[num_blocks_dim] = page_stride
243
244
kv_cache = torch.as_strided(