Add torch compile for qwen3_vl encoder (#40116)

Signed-off-by: Tianyu Guo <guoty9@mail2.sysu.edu.cn>
Signed-off-by: Tianyu Guo <guoty@inferact.ai>
This commit is contained in:
Tianyu Guo
2026-08-08 08:23:18 +00:00
committed by GitHub
parent 643c125fab
commit 653ebb52df
+22 -1
View File
@@ -48,7 +48,10 @@ from transformers.models.qwen3_vl.video_processing_qwen3_vl import (
)
from transformers.video_utils import VideoMetadata
from vllm.compilation.decorators import support_torch_compile
from vllm.compilation.decorators import (
should_torch_compile_mm_encoder,
support_torch_compile,
)
from vllm.config import VllmConfig
from vllm.config.multimodal import (
BaseDummyOptions,
@@ -356,6 +359,13 @@ def pos_embed_interpolate_native(
return repeated.to(dtype=dtype)
@support_torch_compile(
dynamic_arg_dims={
"x": 0,
},
enable_if=should_torch_compile_mm_encoder,
is_encoder=True,
)
class Qwen3_VisionPatchEmbed(nn.Module):
def __init__(
self,
@@ -422,6 +432,17 @@ class Qwen3_VisionMLP(nn.Module):
return mlp_output
@support_torch_compile(
dynamic_arg_dims={
"x": 0,
"cu_seqlens": 0,
"rotary_pos_emb_cos": 0,
"rotary_pos_emb_sin": 0,
"sequence_lengths": 0,
},
enable_if=should_torch_compile_mm_encoder,
is_encoder=True,
)
class Qwen3_VisionBlock(nn.Module):
def __init__(
self,