[Bugfix] Fix FP8 MoE EP Weight Loading for ModelOpt Llama4 (#32886)

Signed-off-by: baonudesifeizhai <baonudesifeizhai@gmail.com>

vllm/model_executor/models/llama4.py

@@ -51,6 +51,8 @@ from vllm.model_executor.model_loader.weight_utils import (
 )
 from vllm.model_executor.models.interfaces import MixtureOfExperts
 from vllm.model_executor.models.utils import sequence_parallel_chunk
+from vllm.platforms import current_platform
+from vllm.utils.torch_utils import is_torch_equal_or_newer
 
 from .llama import LlamaForCausalLM, LlamaMLP, LlamaModel
 from .utils import (
@@ -504,7 +506,25 @@ class Llama4Model(LlamaModel):
                         .flatten()
                         .to(new_loaded_weight.device)
                     )
-                    new_loaded_weight = new_loaded_weight[local_expert_indices]
+                    # Workaround for FP8 CPU indexing on older PyTorch:
+                    # https://github.com/vllm-project/vllm/issues/32862
+                    is_fp8_dtype = new_loaded_weight.dtype == (
+                        current_platform.fp8_dtype()
+                    ) or (
+                        new_loaded_weight.dtype.is_floating_point
+                        and new_loaded_weight.element_size() == 1
+                    )
+                    if (
+                        new_loaded_weight.device.type == "cpu"
+                        and is_fp8_dtype
+                        and not is_torch_equal_or_newer("2.11.0")
+                    ):
+                        # PyTorch < 2.11 doesn't support CPU float8 indexing.
+                        new_loaded_weight = new_loaded_weight.to(torch.float16)[
+                            local_expert_indices
+                        ].to(new_loaded_weight.dtype)
+                    else:
+                        new_loaded_weight = new_loaded_weight[local_expert_indices]
                     expert_id = local_expert_indices[0].item()
             else:
                 # TODO: add EP support for non fused weights