[Bugfix] Disable moe inplace for torch >= 2.9 (#26497)

Signed-off-by: Bill Nell <bnell@redhat.com>

vllm/model_executor/layers/fused_moe/fused_marlin_moe.py

@@ -14,7 +14,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import moe_align_block_size
 from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
     TopKWeightAndReduceNoOP,
 )
-from vllm.model_executor.layers.fused_moe.utils import _resize_cache
+from vllm.model_executor.layers.fused_moe.utils import _resize_cache, disable_inplace
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     marlin_make_workspace_new,
     marlin_moe_intermediate_size,
@@ -235,7 +235,11 @@ def fused_marlin_moe(
     ).view(-1, topk, K)
 
     if output is None:
-        output = hidden_states if inplace else torch.empty_like(hidden_states)
+        if inplace and not disable_inplace():
+            output = hidden_states
+        else:
+            output = torch.empty_like(hidden_states)
+
     return torch.sum(intermediate_cache3.view(-1, topk, K), dim=1, out=output)
 
 

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -39,6 +39,7 @@ from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
 from vllm.model_executor.layers.fused_moe.utils import (
     _resize_cache,
     activation_without_mul,
+    disable_inplace,
     moe_kernel_quantize_input,
 )
 from vllm.model_executor.layers.quantization.utils.mxfp4_utils import dequant_mxfp4
@@ -1516,7 +1517,7 @@ def torch_vllm_outplace_fused_experts(**kwargs) -> torch.Tensor:
 
 
 def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]:
-    if inplace:
+    if inplace and not disable_inplace():
         return torch_vllm_inplace_fused_experts
     return torch_vllm_outplace_fused_experts
 
@@ -1766,7 +1767,10 @@ def fused_experts_impl(
     else:
         raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
 
-    out_hidden_states = hidden_states if inplace else torch.empty_like(hidden_states)
+    if inplace and not disable_inplace():
+        out_hidden_states = hidden_states
+    else:
+        out_hidden_states = torch.empty_like(hidden_states)
 
     if ocp_mx_scheme is not None:
         # TODO: On platforms for which `current_platform.supports_mx()` is True

vllm/model_executor/layers/fused_moe/modular_kernel.py

@@ -13,6 +13,7 @@ from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
 from vllm.model_executor.layers.fused_moe.utils import (
     _resize_cache,
     count_expert_num_tokens,
+    disable_inplace,
 )
 from vllm.utils import cdiv
 from vllm.v1.worker.ubatching import (
@@ -1139,7 +1140,7 @@ class FusedMoEModularKernel(torch.nn.Module):
         - torch.Tensor: The output tensor after applying the MoE layer.
         """
 
-        if inplace and self.shared_experts is None:
+        if inplace and self.shared_experts is None and not disable_inplace():
             output = hidden_states
         else:
             output = torch.zeros_like(hidden_states)

vllm/model_executor/layers/fused_moe/utils.py

@@ -23,7 +23,7 @@ from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
     mxfp8_e4m3_quantize,
 )
 from vllm.triton_utils import tl, triton
-from vllm.utils import cdiv
+from vllm.utils import cdiv, is_torch_equal_or_newer
 from vllm.utils.flashinfer import flashinfer_fp4_quantize
 
 
@@ -321,3 +321,10 @@ def _validate_scale_shape(
 
 def activation_without_mul(activation: str) -> str:
     return activation + "_no_mul"
+
+
+# Torch custom ops can't deal with outputs aliasing inputs so we need to
+# disable inplace for torch >= 2.9.
+# See https://github.com/vllm-project/vllm/issues/26378
+def disable_inplace() -> bool:
+    return is_torch_equal_or_newer("2.9")