Use explicit uint64 dtype for Tensor data_ptr() to avoid overflow (#11994)
This commit is contained in:
@@ -89,6 +89,7 @@ def write_cache_indices(
|
||||
prefix_pointers = torch.tensor(
|
||||
[t.data_ptr() for t in prefix_tensors],
|
||||
device=req_to_token_pool.device,
|
||||
dtype=torch.uint64,
|
||||
)
|
||||
# TODO: some tensors can be reused for ForwardBatchInfo (e.g., extend_lens, cumsum_start)
|
||||
write_req_to_token_pool_triton[(req_pool_indices_tensor.shape[0],)](
|
||||
|
||||
Reference in New Issue
Block a user