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/inference/functions/add.py

@@ -0,0 +1,46 @@
+import ixformer._C as ops
+import torch
+
+__all__ = [
+    "ref_add",
+    "add",
+]
+
+
+def ref_add(input: torch.Tensor, other: torch.Tensor, out: torch.Tensor = None):
+    return torch.add(input, other, out=out)
+
+
+def add(input: torch.Tensor, other: torch.Tensor, out: torch.Tensor = None):
+    """
+    out = input + other
+    Support elementwise addition, but broadcasting is not supported yet.
+    Note: The dtype of input and other needs to be the same.
+    Args:
+        input:              (...)        torch.float32, torch.float16, torch.bfloat16
+        other:              (...)        same as input
+        out:                (...)        same as input
+    Returns:
+        out:                (...)        same as input
+    """
+    if input.dtype not in [torch.float16, torch.float32, torch.bfloat16]:
+        return torch.add(input, other, out=out)
+    if not input.is_contiguous() or not other.is_contiguous():
+        return torch.add(input, other, out=out)
+    if out is not None and not out.is_contiguous():
+        return torch.add(input, other, out=out)
+
+    if input.dtype != other.dtype:
+        return torch.add(input, other, out=out)
+    if out is not None and out.dtype != input.dtype:
+        return torch.add(input, other, out=out)
+
+    assert input.shape == other.shape, (f"broadcasting is not supported yet."
+        "input is {input.shape}, other is {other.shape}")
+    
+    if out is None:
+        out = torch.empty_like(input)
+
+    ops.infer.add(input, other, out)
+
+    return out