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, 10:45 UTC
[Model][Qwen3VL] Simplify `get_mrope_input_positions` using numpy (#28302)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Nov 12, 2025 at 02:55 UTC
cbb799e314178abf2fa99e7306087198c5f43580
1 file changed
+11
-33
vllm/model_executor/models/qwen3_vl.py
+11
-33
@@ -1432,13 +1432,11 @@ class Qwen3VLForConditionalGeneration(
1432
vision_start_token_id = hf_config.vision_start_token_id
1433
spatial_merge_size = hf_config.vision_config.spatial_merge_size
1434
1435
- input_tokens_tensor = torch.tensor(input_tokens)
1436
- vision_start_indices = torch.argwhere(
1437
- input_tokens_tensor == vision_start_token_id
1438
- ).squeeze(1)
1439
- vision_tokens = input_tokens_tensor[vision_start_indices + 1]
1440
- image_nums = (vision_tokens == image_token_id).sum()
1441
- video_nums = (vision_tokens == video_token_id).sum()
1435
+ input_tokens_array = np.array(input_tokens)
1436
+ vision_start_mask = input_tokens_array == vision_start_token_id
1437
+ vision_tokens = input_tokens_array[vision_start_mask.nonzero()[0] + 1]
1438
+ image_nums = np.count_nonzero(vision_tokens == image_token_id)
1439
+ video_nums = np.count_nonzero(vision_tokens == video_token_id)
1440
llm_pos_ids_list: list = []
1441
1442
st = 0
@@ -1474,43 +1472,23 @@ class Qwen3VLForConditionalGeneration(
1472
1473
st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
1474
llm_pos_ids_list.append(
1477
- torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
1475
+ np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
1476
)
1477
1480
- t_index = (
1481
- torch.arange(llm_grid_t)
1482
- .view(-1, 1)
1483
- .expand(-1, llm_grid_h * llm_grid_w)
1484
- .flatten()
1485
- )
1486
- h_index = (
1487
- torch.arange(llm_grid_h)
1488
- .view(1, -1, 1)
1489
- .expand(llm_grid_t, -1, llm_grid_w)
1490
- .flatten()
1491
- )
1492
- w_index = (
1493
- torch.arange(llm_grid_w)
1494
- .view(1, 1, -1)
1495
- .expand(llm_grid_t, llm_grid_h, -1)
1496
- .flatten()
1497
- )
1498
- llm_pos_ids_list.append(
1499
- torch.stack([t_index, h_index, w_index]) + text_len + st_idx
1500
- )
1478
+ grid_indices = np.indices((llm_grid_t, llm_grid_h, llm_grid_w))
1479
+ llm_pos_ids_list.append(grid_indices.reshape(3, -1) + text_len + st_idx)
1480
st = ed + llm_grid_t * llm_grid_h * llm_grid_w
1481
1482
if st < len(input_tokens):
1483
st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
1484
text_len = len(input_tokens) - st
1485
llm_pos_ids_list.append(
1507
- torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
1486
+ np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
1487
)
1488
1510
- llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
1489
+ llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1)
1490
mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()
1512
-
1513
- return llm_positions, mrope_position_delta
1491
+ return torch.from_numpy(llm_positions), mrope_position_delta
1492
1493
def get_language_model(self) -> torch.nn.Module:
1494
return self.language_model