Fix issues from #28242 (#28257)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

vllm/config/compilation.py

@@ -251,13 +251,6 @@ class CompilationConfig:
     disabled when running with Inductor: mode>=VLLM_COMPILE and use_inductor=True.
     Inductor generates (fused) Triton kernels for disabled custom ops."""
     splitting_ops: list[str] | None = None
-
-    """
-    Provide control over whether to compile the multimodal encoder 
-    such as Qwen2_5_vl 
-    """
-    compile_mm_encoder: bool = True
-
     """A list of ops to exclude from cudagraphs, used in piecewise compilation.
 
     The behavior depends on use_inductor_graph_partition:
@@ -275,6 +268,9 @@ class CompilationConfig:
 
     If None, defaults to attention ops for piecewise cudagraphs.
     If empty list [], no ops are excluded (suitable for full cudagraphs)."""
+    compile_mm_encoder: bool = True
+    """Whether or not to compile the multimodal encoder.
+    Currently, this only works for `Qwen2_5_vl`."""
 
     # Inductor capture
     use_inductor: bool | None = None

vllm/model_executor/models/qwen2_5_vl.py

@@ -67,9 +67,7 @@ from vllm.model_executor.layers.linear import (
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.module_mapping import MultiModelKeys
-from vllm.model_executor.models.transformers.utils import (
-    should_torch_compile_mm_vit,
-)
+from vllm.model_executor.models.vision import should_torch_compile_mm_vit
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.evs import (
     compute_mrope_for_media,

vllm/model_executor/models/transformers/utils.py

@@ -205,14 +205,3 @@ def can_enable_torch_compile(vllm_config: "VllmConfig") -> bool:
     # Dynamic rope scaling is not compatible with torch.compile
     rope_scaling: dict = getattr(text_config, "rope_scaling", None) or {}
     return rope_scaling.get("rope_type") != "dynamic"
-
-
-def should_torch_compile_mm_vit(vllm_config: "VllmConfig") -> bool:
-    """
-    Callable to be passed to `@support_torch_compile`'s `enable_if` argument.
-
-    Defaults to `True` but is disabled in the following situations:
-
-    - The model uses dynamic rope scaling.
-    """
-    return vllm_config.compilation_config.compile_mm_encoder

vllm/model_executor/models/vision.py

@@ -11,6 +11,7 @@ import torch
 from transformers import PretrainedConfig
 
 from vllm.attention.backends.registry import _Backend
+from vllm.config import VllmConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -100,6 +101,11 @@ def get_vit_attn_backend(
     return current_platform.get_vit_attn_backend(head_size, dtype)
 
 
+def should_torch_compile_mm_vit(vllm_config: VllmConfig) -> bool:
+    """Callable to be passed to `@support_torch_compile`'s `enable_if` argument."""
+    return vllm_config.compilation_config.compile_mm_encoder
+
+
 VisionFeatureSelectStrategyStr = Literal["class", "default", "full"]
 
 VisionFeatureSelectStrategy: TypeAlias = (