In _load_batched_moe_fp8_tensor, when param.shape[0] == loaded_weight.shape[0], calling weight_loader with the 3D loaded_weight and expert_id=0 will cause a shape mismatch and crash in vLLM. vLLM's weight_loader for MoE expects a 2D tensor for a single expert and copies it to param[expert_id]. We should remove this branch and always use the loop to unbind and load each expert individually. This is robust and works correctly in all environments (including single-GPU or non-expert-parallel setups).

    # Retain the compact wire format, but let the loader map global expert IDs one view at a time.
    loaded_any = False
    for expert_id, expert_weight in enumerate(loaded_weight.unbind(0)):
        loaded_any = (
            bool(
                weight_loader(
                    param,
                    expert_weight,
                    target_name,
                    shard_id=shard_id,
                    expert_id=expert_id,
                    return_success=True,
                )
            )
            or loaded_any
        )
    if not loaded_any:
        raise ValueError(f"No local expert accepted batched MoE tensor {wire_name!r}")
    return True

Checking for specific Blackwell compute capabilities (10, 0) and (10, 3) is fragile and might break on other Blackwell GPU variants (e.g., GB200 or future revisions with minor version changes like 10.1 or 10.2). It is safer and more future-proof to check if the major version of the compute capability is 10.