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/csrc/include/ixformer/kernels/error.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include <stdexcept>
+
+#include "status.h"
+
+namespace ixformer::kernels {
+
+class KernelError : public std::runtime_error {
+public:
+    template<class ERROR_STR>
+    KernelError(KernelStatus error, const ERROR_STR str) : error_{error}, std::runtime_error(str) {}
+
+    KernelStatus status() {
+        return error_;
+    }
+
+private:
+    KernelStatus error_;
+};
+
+}// namespace ixformer::kernels

ixformer_sdk/csrc/include/ixformer/kernels/status.h

@@ -0,0 +1,20 @@
+#pragma once
+
+#include <string>
+
+namespace ixformer::kernels {
+
+enum KernelStatus {
+    kernelSuccess,
+    kernelFail,
+    kernelCudaError,
+    kernelInvalidArgument,
+    kernelCuinferError,
+    kernelUnsupported,
+};
+
+
+std::string to_string(KernelStatus status);
+
+
+}// namespace ixformer::kernels

ixformer_sdk/csrc/include/ixformer/kernels/tensor.h

@@ -0,0 +1,92 @@
+#pragma once
+
+#include <string>
+
+namespace ixformer::kernels {
+
+const uint8_t MAX_TENSOR_NDIM = 8;
+
+// align with at::ScalarType
+enum DType {
+    Byte = 0,
+    Char = 1,
+    Short = 2,
+    Int = 3,
+    Long = 4,
+    Half = 5,
+    Float = 6,
+    Double = 7,
+    ComplexHalf = 8,
+    ComplexFloat = 9,
+    ComplexDoubl = 10,
+    Bool = 11,
+    QInt8 = 12,
+    QUInt8 = 13,
+    QInt32 = 14,
+    BFloat16 = 15,
+    QUInt4x2 = 16,
+    QUInt2x4 = 17,
+    Bits1x8 = 18,
+    Bits2x4 = 19,
+    Bits4x2 = 20,
+    Bits8 = 21,
+    Bits16 = 22,
+    Float8_e5m2 = 23,
+    Float8_e4m3fn = 24,
+    Undefined = 25,
+    NumOptions = 26
+};
+
+struct TensorDesc {
+
+public:
+    // delete default constructor
+    TensorDesc() = delete;
+    // All information must be (should be) prepared when constructing a TensorDesc object.
+    TensorDesc(DType scalar_type, void *data_ptr, int64_t numel, int64_t dim, const int64_t *size, const int64_t *stride, bool is_contiguous, bool is_cuda)
+        : dtype(scalar_type), ptr(data_ptr), nnumel(numel), ndim(dim), sizes(size), strides(stride), contiguous(is_contiguous), cuda(is_cuda) {}
+
+    inline DType scalar_type() const {
+        return dtype;
+    }
+
+    inline void *data_ptr() const {
+        return ptr;
+    }
+
+    inline int64_t numel() const {
+        return nnumel;
+    }
+
+    inline int64_t dim() const {
+        return ndim;
+    }
+
+    inline int64_t size(int64_t dim) const {
+        return dim < 0 ? sizes[ndim - dim] : sizes[dim];
+    }
+
+    inline int64_t stride(int64_t dim) const {
+        return dim < 0 ? strides[ndim - dim] : strides[dim];
+    }
+
+    inline bool is_contiguous() const {
+        return contiguous;
+    }
+
+    inline bool is_cuda() const {
+        return cuda;
+    }
+
+private:
+    void *ptr{nullptr};
+    DType dtype;
+    int64_t nnumel{0};
+    int64_t ndim{0};
+    const int64_t *sizes{nullptr};
+    const int64_t *strides{nullptr};
+    bool contiguous{false};
+    bool cuda{false};
+};
+
+}// namespace ixformer::kernels