xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-10 01:07:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Models][Qwen3VL] Speedup fast_pos_embed_interpolate (#26647)

Browse Source 
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
This commit is contained in:
Lukas Geiger 2025-10-12 18:20:07 +01:00 committed by GitHub
parent 18ed7746ea
commit a6049be73c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
vllm/model_executor/models/qwen3_vl.py
View File

@ -467,8 +467,6 @@ class Qwen3_VisionTransformer(nn.Module):
dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing="ij") dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing="ij")
h_floor_grid, w_floor_grid = torch.meshgrid(h_floor, w_floor, indexing="ij") h_floor_grid, w_floor_grid = torch.meshgrid(h_floor, w_floor, indexing="ij")
h_ceil_grid, w_ceil_grid = torch.meshgrid(h_ceil, w_ceil, indexing="ij") h_ceil_grid, w_ceil_grid = torch.meshgrid(h_ceil, w_ceil, indexing="ij")
h_floor_grid_idx = h_floor_grid * num_grid_per_side
h_ceil_grid_idx = h_ceil_grid * num_grid_per_side
# original computation of weights # original computation of weights
# w00 = (1 - dh_grid) * (1 - dw_grid) # w00 = (1 - dh_grid) * (1 - dw_grid)
@ -480,30 +478,25 @@ class Qwen3_VisionTransformer(nn.Module):
w11 = dh_grid * dw_grid w11 = dh_grid * dw_grid
w10 = dh_grid - w11 w10 = dh_grid - w11
w01 = dw_grid - w11 w01 = dw_grid - w11
w00 = 1 - dh_grid - dw_grid + w11 w00 = 1 - dh_grid - w01
idx00 = h_floor_grid_idx + w_floor_grid h_grid = torch.stack([h_floor_grid, h_floor_grid, h_ceil_grid, h_ceil_grid])
idx01 = h_floor_grid_idx + w_ceil_grid w_grid = torch.stack([w_floor_grid, w_ceil_grid, w_floor_grid, w_ceil_grid])
idx10 = h_ceil_grid_idx + w_floor_grid h_grid_idx = h_grid * num_grid_per_side
idx11 = h_ceil_grid_idx + w_ceil_grid
indices = torch.stack([idx00, idx01, idx10, idx11], dim=0).reshape(4, -1) indices = (h_grid_idx + w_grid).reshape(4, -1)
weights = torch.stack([w00, w01, w10, w11], dim=0).reshape(4, -1, 1) weights = torch.stack([w00, w01, w10, w11], dim=0).reshape(4, -1, 1)
weights = weights.to( weights = weights.to(dtype=self.dtype)
dtype=self.dtype, device=self.device, non_blocking=True
)
embeds = self.pos_embed(indices) embeds = self.pos_embed(indices)
weighted_embeds = embeds * weights weighted_embeds = embeds * weights
p0, p1, p2, p3 = weighted_embeds.unbind(dim=0) combined = weighted_embeds.sum(dim=0)
combined = p0 + p1 + p2 + p3
combined = combined.view(h * w, hidden_dim) combined = combined.reshape(
repeated = combined.unsqueeze(0).expand(t, -1, -1).contiguous() h // m_size, m_size, w // m_size, m_size, hidden_dim
repeated = repeated.view(
t, h // m_size, m_size, w // m_size, m_size, hidden_dim
) )
repeated = repeated.permute(0, 1, 3, 2, 4, 5).reshape(-1, hidden_dim) combined = combined.permute(0, 2, 1, 3, 4).reshape(1, -1, hidden_dim)
repeated = combined.expand(t, -1, -1).reshape(-1, hidden_dim)
outputs.append(repeated) outputs.append(repeated)
return torch.cat(outputs, dim=0) return torch.cat(outputs, dim=0)