feat(CRITICAL): 从 GitHub 扫描搬运 ixformer SDK + xllm 完整 GDN/MoE 代码

来源:
  1. Chranos/ixformer (GitHub) → ixformer_sdk/ (230 files, 70K lines)
     - inference/functions/vllm.py: vllm_moe_topk_softmax 完整实现 (2033 lines)
     - inference/functions/moe.py: MoE ops 完整实现 (1380 lines)
     - contrib/vllm_flash_attn/: FA2 Python 接口 (1018 lines)
     - contrib/tgi/fused_moe.py: TGI fused MoE (429 lines)
     - csrc/include/ixformer/: C++ kernel headers + cmake

  2. Deep-Spark/xllm (GitHub) → upstream_ref/xllm_latest/ (+15 files)
     - npu_torch/qwen3_5_decoder_layer_impl.cpp/.h
     - npu_torch/qwen3_5_gated_delta_net.cpp/.h
     - npu_torch/qwen3_next_*.cpp/.h (6 files)
     - npu_torch/attention.cpp/.h + fused_moe.cpp/.h + CMakeLists.txt
     - models/llm/qwen3_5.h + qwen3_5_mtp.h + qwen3_next.h
     - models/vlm/qwen3_5.h

调用链完整性:
  ixformer_sdk/inference/functions/vllm.py
    → ops.infer.moe_topk_softmax() (C++ 层)
    → 这就是 base 镜像 libixformer.so 里的实现

  upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
    → ixformer::infer::topk_softmax() (直接 C++ 调用)
    → ixformer::infer::group_gemm() → 完整 7-step MoE pipeline

ixformer_sdk/train/functions/gelu.py

@@ -0,0 +1,47 @@
+from typing import List, Union
+
+import ixformer._C as ops
+import torch
+from torch.autograd.function import Function, FunctionCtx
+
+__all__ = [
+    "gelu",
+]
+
+
+class GeluFunction(Function):
+    @staticmethod
+    def forward(
+        ctx, input: torch.Tensor, in_place: bool = False, training: bool = False
+    ):
+        if training:
+            if in_place:
+                ctx.save_for_backward(input.clone())
+            else:
+                ctx.save_for_backward(input)
+        if in_place:
+            return ops.train.gelu_forward(input, input)
+        else:
+            return ops.train.gelu_forward(input)
+
+    @staticmethod
+    def backward(ctx: FunctionCtx, grad_outputs):
+        input = ctx.saved_tensors[0]
+        grad_input = ops.train.gelu_backward(input, grad_outputs)
+        return grad_input, None, None
+
+
+def gelu(
+    input: torch.Tensor, in_place: bool = False, training: bool = False
+) -> torch.Tensor:
+    """
+    等价实现：
+    torch.nn.functional.gelu
+
+    Args:
+        input: dtype:[torch.float, torch.half, torch.bfloat16]
+        in place: bool. Whether to operate directly on the original input data.
+    Returns:
+        output: dtype:[torch.float, torch.half, torch.bfloat16]
+    """
+    return GeluFunction.apply(input, in_place, training)