[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
@@ -867,7 +867,7 @@ class DiffusionGemmaModelState(ModelState):
self.diffusion_states.add_request(req_index)
if not new_req_data.req_id.startswith("_warmup_"):
prompt_len = len(new_req_data.prompt_token_ids)
self.diffusion_states.prompt_len[req_index] = prompt_len
self.diffusion_states.prompt_len[req_index].fill_(prompt_len)
def remove_request(self, req_id: str) -> None:
idx = self._req_id_to_index.pop(req_id, None)
@@ -1276,7 +1276,10 @@ class DiffusionSampler:
# before canvas padding so phantom positions stay uniform.
if num_decode > 0:
top_k, top_p = self.sampling_states.get_top_k_top_p(
decode_slots.repeat_interleave(valid_canvas_len), decode_slots_np
decode_slots.repeat_interleave(
valid_canvas_len, output_size=int(valid_canvas_len_np.sum())
),
decode_slots_np,
)
if top_k is not None or top_p is not None:
logits = apply_top_k_top_p(logits.float(), top_k, top_p)
+9 -5
View File
@@ -72,6 +72,7 @@ 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 vllm.v1.attention.backends.registry import AttentionBackendEnum
from .ernie45_vl_moe import Ernie4_5_VLMoeForCausalLM
@@ -446,6 +447,9 @@ class Ernie4_5_VisionTransformer(nn.Module):
pos_ids = torch.cat(pos_ids, dim=0)
max_grid_size = grid_thw[:, 1:].max()
rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
# `pos_ids` is built on the host; stage it over non-blocking so the
# gather below doesn't index a device tensor with a CPU one.
pos_ids = pos_ids.to(rotary_pos_emb_full.device, non_blocking=True)
rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
return rotary_pos_emb
@@ -530,7 +534,7 @@ class Ernie4_5_VisionTransformer(nn.Module):
else:
max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
cu_seqlens = cu_seqlens.to(device)
cu_seqlens = cu_seqlens.to(device, non_blocking=True)
return {
"rotary_pos_emb": rotary_pos_emb,
@@ -806,8 +810,8 @@ class VariableResolutionResamplerModel(nn.Module):
b_offset + (temp_offset + 1) * spatial_size,
)
)
slice_offsets = torch.tensor(np.concatenate(slice_offsets, axis=-1)).to(
x.device
slice_offsets = async_tensor_h2d(
np.concatenate(slice_offsets, axis=-1), device=x.device
)
slice_offsets2 = []
@@ -823,8 +827,8 @@ class VariableResolutionResamplerModel(nn.Module):
b_offset + (temp_offset + 1) * spatial_size,
)
)
slice_offsets2 = torch.tensor(np.concatenate(slice_offsets2, axis=-1)).to(
x.device
slice_offsets2 = async_tensor_h2d(
np.concatenate(slice_offsets2, axis=-1), device=x.device
)
x_timestep_1 = torch.index_select(x, dim=0, index=slice_offsets)
+18 -37
View File
@@ -546,14 +546,6 @@ class Glm4vVisionEmbeddings(nn.Module):
0, hidden_size, device=device, dtype=pos_embed_weight.dtype
)
else:
# Convert inputs to tensors if needed
if isinstance(lengths, list):
lengths = async_tensor_h2d(lengths, device=device, dtype=torch.long)
if not isinstance(image_shapes, torch.Tensor):
image_shapes = async_tensor_h2d(
image_shapes, device=device, dtype=torch.long
)
# Prepare 2D position embedding
orig_size_sq = pos_embed_weight.shape[0]
orig_size = int(orig_size_sq**0.5)
@@ -564,32 +556,23 @@ class Glm4vVisionEmbeddings(nn.Module):
.to(device=device, dtype=torch.float32)
)
# Calculate target dimensions for each patch
# Add bounds checking for data parallel mode
if len(lengths) > image_shapes.shape[0]:
# In data parallel mode, some GPUs might not have all
# image shapes
# Use available image shapes, cycling if necessary
target_h_list = []
target_w_list = []
for i in range(len(lengths)):
# Cycle through available shapes
shape_idx = i % image_shapes.shape[0]
target_h_list.append(image_shapes[shape_idx, 1].repeat(lengths[i]))
target_w_list.append(image_shapes[shape_idx, 2].repeat(lengths[i]))
target_h = torch.cat(target_h_list).to(
device=device, dtype=torch.float32
)
target_w = torch.cat(target_w_list).to(
device=device, dtype=torch.float32
)
else:
target_h = torch.cat(
[image_shapes[i, 1].repeat(lengths[i]) for i in range(len(lengths))]
).to(device=device, dtype=torch.float32)
target_w = torch.cat(
[image_shapes[i, 2].repeat(lengths[i]) for i in range(len(lengths))]
).to(device=device, dtype=torch.float32)
# Calculate target dimensions for each patch. `lengths` and
# `image_shapes` are host data, so expand them with numpy and move
# the result across once rather than per-image.
# Shapes are cycled: in data parallel mode some GPUs might not
# have all image shapes.
shapes_np = np.asarray(image_shapes)
shape_idx = np.arange(len(lengths)) % shapes_np.shape[0]
target_h = async_tensor_h2d(
np.repeat(shapes_np[shape_idx, 1], lengths),
device=device,
dtype=torch.float32,
)
target_w = async_tensor_h2d(
np.repeat(shapes_np[shape_idx, 2], lengths),
device=device,
dtype=torch.float32,
)
# Normalize coordinates to [-1, 1] range for grid_sample
h_coords = h_coords.to(device=device, dtype=torch.float32)
@@ -818,9 +801,7 @@ class Glm4vVisionTransformer(nn.Module):
)
lengths = [h * w] * t
image_shapes = async_tensor_h2d(
[[t, h, w]], dtype=torch.long, device=device
)
image_shapes = [[t, h, w]]
# Build the coordinates on the host (cheap integer math) but move
# them across pinned + non-blocking, so the consumer's
+2 -2
View File
@@ -424,7 +424,7 @@ class InternS1MultiModalProcessor(BaseMultiModalProcessor[InternS1ProcessingInfo
pixel_values=MultiModalFieldConfig.flat_from_sizes(
"image", image_num_patches
),
image_num_patches=MultiModalFieldConfig.batched("image"),
image_num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
image_embeds=MultiModalFieldConfig.batched("image"),
image_token_id=MultiModalFieldConfig.shared(
"image", num_images, keep_on_cpu=True
@@ -432,7 +432,7 @@ class InternS1MultiModalProcessor(BaseMultiModalProcessor[InternS1ProcessingInfo
pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
"video", video_num_patches
),
video_num_patches=MultiModalFieldConfig.batched("video"),
video_num_patches=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
video_token_id=MultiModalFieldConfig.shared(
"video", num_videos, keep_on_cpu=True
),
+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(
+11 -5
View File
@@ -68,6 +68,7 @@ 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 vllm.v1.attention.backends.registry import AttentionBackendEnum
from .ernie45 import Ernie4_5ForCausalLM
@@ -800,16 +801,18 @@ class SiglipEncoder(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)
rotary_pos_emb = rope_emb_max_grid[pids].flatten(1)
if cu_seqlens is None:
raise ValueError("cu_seqlens cannot be None for SiglipEncoder.")
if not isinstance(cu_seqlens, torch.Tensor):
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32, device=device)
cu_seqlens = async_tensor_h2d(cu_seqlens, dtype=torch.int32, device=device)
else:
cu_seqlens = cu_seqlens.to(device=device)
cu_seqlens = cu_seqlens.to(device=device, non_blocking=True)
max_seqlen = None
if self.attn_backend in {
@@ -1131,10 +1134,13 @@ class PaddleOCRVLForConditionalGeneration(nn.Module, SupportsMultiModal, Support
siglip_position_ids.append(image_position_ids)
cu_seqlens.append(cu_seqlens[-1] + numel)
# Both are built on the host; stage them over non-blocking.
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(
pixel_values.device
pixel_values.device, non_blocking=True
)
cu_seqlens = async_tensor_h2d(
cu_seqlens, dtype=torch.int32, device=pixel_values.device
)
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(pixel_values.device)
vision_outputs = self.visual(
pixel_values=pixel_values,
+1 -1
View File
@@ -905,7 +905,7 @@ class Phi4MMMultiModalProcessor(BaseMultiModalProcessor[Phi4MMProcessingInfo]):
return dict(
input_image_embeds=MultiModalFieldConfig.batched("image"),
image_attention_mask=MultiModalFieldConfig.batched("image"),
image_sizes=MultiModalFieldConfig.batched("image"),
image_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
num_img_tokens=MultiModalFieldConfig.batched("image"),
input_audio_embeds=MultiModalFieldConfig.batched("audio"),
)
+5 -1
View File
@@ -244,7 +244,11 @@ class UltravoxMultiModalProcessor(BaseMultiModalProcessor[UltravoxProcessingInfo
# higher than the number of audio samples
audio_features=MultiModalFieldConfig.flat_from_sizes("audio", num_chunks),
audio_token_len=MultiModalFieldConfig.flat_from_sizes("audio", num_chunks),
audio_lens=MultiModalFieldConfig.flat_from_sizes("audio", num_chunks),
# Only ever used to derive the encoder attention metadata on the
# host, so keep it there.
audio_lens=MultiModalFieldConfig.flat_from_sizes(
"audio", num_chunks, keep_on_cpu=True
),
# num_chunks can convert audio_chunked to audio batch dimension
audio_num_chunks=MultiModalFieldConfig.batched("audio"),
audio_embeds=MultiModalFieldConfig.batched("audio"),
+5 -2
View File
@@ -27,11 +27,14 @@ def prepare_chunk_indices(cu_seqlens: torch.Tensor, chunk_size: int) -> torch.Te
for n in triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()
]
)
return torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1).to(cu_seqlens)
chunk_indices = torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1)
return chunk_indices.to(
device=cu_seqlens.device, dtype=cu_seqlens.dtype, non_blocking=True
)
@tensor_cache
def prepare_chunk_offsets(cu_seqlens: torch.Tensor, chunk_size: int) -> torch.Tensor:
return torch.cat(
[cu_seqlens.new_tensor([0]), triton.cdiv(prepare_lens(cu_seqlens), chunk_size)]
[cu_seqlens.new_zeros(1), triton.cdiv(prepare_lens(cu_seqlens), chunk_size)]
).cumsum(-1)