[Perf] Avoid more GPU<->CPU syncs in multimodal encoders (#52369)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-08-14 10:40:14 -07:00
committed by GitHub
parent 9b0ab5dd53
commit 3e3ceb1961
9 changed files with 81 additions and 72 deletions
+25 -17
View File
@@ -59,6 +59,7 @@ from vllm.multimodal.processing import (
)
from vllm.sequence import IntermediateTensors
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import async_tensor_h2d
from .interfaces import (
MultiModalEmbeddings,
@@ -608,7 +609,9 @@ class KeyeSiglipEncoder(nn.Module):
[height_position_ids, width_position_ids],
dim=-1,
)
max_grid_size = pids.max() + 1
# The ids are built from the grids above, so `h`/`w` bound them
# and the table size is known on the host.
max_grid_size = max(max(h, w) for _, h, w in flatten_image_grid_thw)
rope_emb_max_grid = self.rotary_pos_emb(max_grid_size)
rope_emb = rope_emb_max_grid[pids].flatten(1)
rope_emb = rope_emb.repeat(1, 2)
@@ -694,19 +697,19 @@ class KeyeSiglipVisionTransformer(nn.Module):
last_hidden_state = self.post_layernorm(last_hidden_state)
sample_hidden_state = list()
if cu_seqlens is None:
raise ValueError(
"cu_seqlens cannot be None for "
"SiglipVisionTransformer output processing."
)
for i in range(cu_seqlens.shape[0] - 1):
start = cu_seqlens[i]
end = cu_seqlens[i + 1]
tensor = last_hidden_state[:, start:end, :].squeeze(0)
sample_hidden_state.append(tensor)
return sample_hidden_state
# `cu_seqlens` is the running sum of the per-image `t * h * w`, so the
# split sizes are known on the host and slicing needs no device read.
split_sizes = [
int(np.prod(thw)) for thw in self.encoder.flatten_list(image_grid_thw)
]
return [
tensor.squeeze(0) for tensor in last_hidden_state.split(split_sizes, dim=1)
]
class KeyeSiglipVisionModel(nn.Module):
@@ -1307,13 +1310,17 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):
)
else:
pixel_values = image_input["pixel_values"].type(self.visual.dtype)
# These are all built on the host, so stage them over
# non-blocking rather than forcing a sync per transfer.
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(
pixel_values.device
pixel_values.device, non_blocking=True
)
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(
pixel_values.device
cu_seqlens = async_tensor_h2d(
cu_seqlens, dtype=torch.int32, device=pixel_values.device
)
sample_indices = torch.concat(sample_indices, dim=0).to(
pixel_values.device, non_blocking=True
)
sample_indices = torch.concat(sample_indices, dim=0).to(pixel_values.device)
image_embeds = self.visual(
pixel_values=pixel_values,
@@ -1357,14 +1364,15 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):
)
else:
pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
# These are host-built; stage them over non-blocking.
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(
pixel_values_videos.device
pixel_values_videos.device, non_blocking=True
)
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(
pixel_values_videos.device
cu_seqlens = async_tensor_h2d(
cu_seqlens, dtype=torch.int32, device=pixel_values_videos.device
)
sample_indices = torch.concat(sample_indices, dim=0).to(
pixel_values_videos.device
pixel_values_videos.device, non_blocking=True
)
video_embeds = self.visual(