diff --git a/tests/distributed/test_weight_transfer.py b/tests/distributed/test_weight_transfer.py
index 715a8069d7..8a13b24dc5 100644
--- a/tests/distributed/test_weight_transfer.py
+++ b/tests/distributed/test_weight_transfer.py
@@ -37,23 +37,18 @@ from vllm.platforms import current_platform
 from vllm.utils.network_utils import get_open_port
 
 
-def _weight_transfer_ray_env_vars() -> dict[str, str]:
-    if not current_platform.is_rocm():
-        return {}
-
-    return {
-        "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
-        "RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES": "1",
-        "RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES": "1",
-    }
-
-
 def _init_ray_for_weight_transfer() -> None:
     if ray.is_initialized():
         return
     ray.init(
         ignore_reinit_error=True,
-        runtime_env={"env_vars": _weight_transfer_ray_env_vars()},
+        runtime_env={
+            "env_vars": {
+                "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
+                "RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES": "1",
+                "RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES": "1",
+            }
+        },
     )
 
 
@@ -66,7 +61,7 @@ def _get_ray_assigned_device() -> torch.device:
 
 def _set_ray_assigned_device() -> torch.device:
     device = _get_ray_assigned_device()
-    torch.accelerator.set_device_index(device)
+    current_platform.set_device(device)
     return device
 
 
diff --git a/tests/v1/e2e/general/test_mamba_prefix_cache.py b/tests/v1/e2e/general/test_mamba_prefix_cache.py
index 4644a6cc7e..dd9efc6696 100644
--- a/tests/v1/e2e/general/test_mamba_prefix_cache.py
+++ b/tests/v1/e2e/general/test_mamba_prefix_cache.py
@@ -403,6 +403,7 @@ def _run_ref_mamba_state_worker():
         GPUModelRunner._sample = fake_sample_fn
         engine = LLM(
             model=MODEL,
+            load_format="dummy",
             block_size=BLOCK_SIZE,
             hf_overrides={"num_hidden_layers": NUM_HIDDEN_LAYERS},
             seed=42,
@@ -750,6 +751,7 @@ def test_mamba_prefix_cache_mrv1(monkeypatch: pytest.MonkeyPatch):
 
     engine = LLM(
         model=MODEL,
+        load_format="dummy",
         enable_prefix_caching=True,
         block_size=BLOCK_SIZE,
         mamba_cache_mode="align",
diff --git a/vllm/model_executor/layers/fused_moe/routed_experts.py b/vllm/model_executor/layers/fused_moe/routed_experts.py
index 1b1deb250c..ce87d0b1b3 100644
--- a/vllm/model_executor/layers/fused_moe/routed_experts.py
+++ b/vllm/model_executor/layers/fused_moe/routed_experts.py
@@ -907,7 +907,7 @@ class RoutedExperts(PluggableLayer):
                 # Unified loading logic for fused and non-fused experts
                 loaded_experts = experts_shard.unbind()
                 for expert_id, loaded_expert in enumerate(loaded_experts, start=start):
-                    success = self.weight_loader(
+                    success = param.weight_loader(
                         param=param,
                         loaded_weight=loaded_expert,
                         weight_name=weight_name,
diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
index 211c88129c..9530ac58b6 100644
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -1445,6 +1445,8 @@ class DeepseekV2Model(nn.Module):
                 hidden_states, residual = combined_states.split(
                     [self.hidden_size, self.hidden_size], dim=-1
                 )
+                # fused_add_rms_norm requires a contiguous residual
+                residual = residual.contiguous()
             if idx in self.aux_hidden_state_layers:
                 aux_hidden_state = hidden_states + residual
                 if aux_hidden_state.shape[0] != positions.shape[0]:
@@ -1469,6 +1471,8 @@ class DeepseekV2Model(nn.Module):
             hidden_states, residual = combined_states.split(
                 [self.hidden_size, self.hidden_size], dim=-1
             )
+            # fused_add_rms_norm requires a contiguous residual
+            residual = residual.contiguous()
 
         if self.end_layer in self.aux_hidden_state_layers:
             aux_hidden_states.append(hidden_states + residual)
diff --git a/vllm/model_executor/models/gemma3_mm.py b/vllm/model_executor/models/gemma3_mm.py
index 9e58438f4c..0d9f8f1418 100644
--- a/vllm/model_executor/models/gemma3_mm.py
+++ b/vllm/model_executor/models/gemma3_mm.py
@@ -761,6 +761,7 @@ class Gemma3ForConditionalGeneration(
         max_frames_per_batch: int,
         device: torch.device,
         dtype: torch.dtype,
+        path: str = "default",
     ):
         from vllm.v1.worker.encoder_cudagraph_defs import (
             EncoderCudaGraphCaptureInputs,
@@ -792,6 +793,7 @@ class Gemma3ForConditionalGeneration(
         mm_kwargs: dict[str, Any],
         max_batch_size: int,
         max_frames_per_batch: int,
+        path: str = "default",
     ):
         from vllm.v1.worker.encoder_cudagraph_defs import (
             EncoderCudaGraphReplayBuffers,
@@ -804,6 +806,7 @@ class Gemma3ForConditionalGeneration(
     def encoder_cudagraph_forward(
         self,
         values: dict[str, torch.Tensor],
+        path: str = "default",
     ) -> torch.Tensor:
         pixel_values = values["pixel_values"]
         image_features = self.vision_tower(pixel_values)