[Bugfix][XPU] store the pointer raw bit pattern instead of its numeric value (#54514)
Signed-off-by: Lai, Yejing <yejing.lai@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Yejing Lai committed
Sep 24, 2026 at 17:05 UTC
14f98cbe5d447d452bb9f670fa82a6e65e3f00e6
1 file changed
+4
-4
vllm/v1/worker/gpu/model_states/prompt_embeds.py
+4
-4
@@ -37,8 +37,8 @@ class PromptEmbedsState:
37
38
# Indexed by req_state index. Stale entries after removal are
39
# harmless: add_request rewrites all fields for every index it claims.
40
- self.embeds_ptrs = UvaBackedTensor(max_num_reqs, dtype=torch.int64)
41
- self.mask_ptrs = UvaBackedTensor(max_num_reqs, dtype=torch.int64)
40
+ self.embeds_ptrs = UvaBackedTensor(max_num_reqs, dtype=torch.uint64)
41
+ self.mask_ptrs = UvaBackedTensor(max_num_reqs, dtype=torch.uint64)
42
self.embeds_lens = UvaBackedTensor(max_num_reqs, dtype=torch.int32)
43
44
def add_request(self, req_index: int, new_req_data: NewRequestData) -> None:
@@ -100,8 +100,8 @@ class PromptEmbedsState:
100
def _apply_prompt_embeds_kernel(
101
inputs_embeds_ptr,
102
inputs_embeds_stride,
103
- embeds_ptrs_ptr, # int64 [max_num_reqs], device pointers (0-len = unused)
104
- mask_ptrs_ptr, # int64 [max_num_reqs], 0 = no is-token-ids mask
103
+ embeds_ptrs_ptr, # uint64 [max_num_reqs], device pointers (0-len = unused)
104
+ mask_ptrs_ptr, # uint64 [max_num_reqs], 0 = no is-token-ids mask
105
embeds_lens_ptr, # int32 [max_num_reqs]
106
idx_mapping_ptr,
107
query_start_loc_ptr,