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 14, 05:30 UTC
[Perf] Further optimization for Qwen3-VL `fast_pos_embed_interpolate` (#25347)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Sep 22, 2025 at 04:12 UTC
af7dfb0d1a95fb097a93de156ddd6c1eb5c72796
1 file changed
+32
-18
vllm/model_executor/models/qwen3_vl.py
+32
-18
@@ -405,25 +405,39 @@ class Qwen3_VisionTransformer(nn.Module):
405
dh = h_idxs - h_floor
406
dw = w_idxs - w_floor
407
408
- w00 = ((1 - dh)[:, None] * (1 - dw)[None, :]).reshape(-1)
409
- w01 = ((1 - dh)[:, None] * dw[None, :]).reshape(-1)
410
- w10 = (dh[:, None] * (1 - dw)[None, :]).reshape(-1)
411
- w11 = (dh[:, None] * dw[None, :]).reshape(-1)
412
-
413
- idx00 = (h_floor[:, None] * num_grid_per_side +
414
- w_floor[None, :]).reshape(-1)
415
- idx01 = (h_floor[:, None] * num_grid_per_side +
416
- w_ceil[None, :]).reshape(-1)
417
- idx10 = (h_ceil[:, None] * num_grid_per_side +
418
- w_floor[None, :]).reshape(-1)
419
- idx11 = (h_ceil[:, None] * num_grid_per_side +
420
- w_ceil[None, :]).reshape(-1)
421
-
422
- indices = torch.stack([idx00, idx01, idx10, idx11], dim=0)
408
+ # Create meshgrid view for all h, w vars
409
+ dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing='ij')
410
+ h_floor_grid, w_floor_grid = torch.meshgrid(h_floor,
411
+ w_floor,
412
+ indexing='ij')
413
+ h_ceil_grid, w_ceil_grid = torch.meshgrid(h_ceil,
414
+ w_ceil,
415
+ indexing='ij')
416
+ h_floor_grid_idx = h_floor_grid * num_grid_per_side
417
+ h_ceil_grid_idx = h_ceil_grid * num_grid_per_side
418
+
419
+ # original computation of weights
420
+ # w00 = (1 - dh_grid) * (1 - dw_grid)
421
+ # w01 = (1 - dh_grid) * dw_grid
422
+ # w10 = dh_grid * (1 - dw_grid)
423
+ # w11 = dh_grid * dw_grid
424
+ # we reuse w11 here to avoid duplicate
425
+ # dh_grid * dw_grid computation
426
+ w11 = dh_grid * dw_grid
427
+ w10 = dh_grid - w11
428
+ w01 = dw_grid - w11
429
+ w00 = 1 - dh_grid - dw_grid + w11
430
+
431
+ idx00 = h_floor_grid_idx + w_floor_grid
432
+ idx01 = h_floor_grid_idx + w_ceil_grid
433
+ idx10 = h_ceil_grid_idx + w_floor_grid
434
+ idx11 = h_ceil_grid_idx + w_ceil_grid
435
+
436
+ indices = torch.stack([idx00, idx01, idx10, idx11],
437
+ dim=0).reshape(4, -1)
438
weights = torch.stack([w00, w01, w10, w11],
424
- dim=0).to(dtype=self.dtype,
425
- device=self.device)
426
- weights = weights.unsqueeze(-1)
439
+ dim=0).reshape(4, -1, 1)
440
+ weights = weights.to(dtype=self.dtype, device=self.device)
441
442
embeds = self.pos_embed(indices)
443
weighted_embeds = embeds * weights