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/matmul.py

@@ -0,0 +1,108 @@
+import ixformer._C as ops
+import torch
+from torch.autograd.function import Function, FunctionCtx
+
+__all__ = ["matmul", "MatmulFunction"]
+
+
+class MatmulFunction(Function):
+    @staticmethod
+    def forward(
+        ctx: FunctionCtx,
+        input: torch.Tensor,
+        other: torch.Tensor,
+        out: torch.Tensor = None,
+        transa: bool = False,
+        transb: bool = False,
+        alpha: float = 1.0,
+        beta: float = 0.0,
+    ):
+        ctx.save_for_backward(input, other)
+        ctx.params = (transa, transb, alpha, beta)
+        if out is None:
+            return ops.train.matmul(
+                input, other, transa=transa, transb=transb, alpha=alpha, beta=beta
+            )
+        else:
+            return ops.train.matmul(
+                input,
+                other,
+                out=out,
+                transa=transa,
+                transb=transb,
+                alpha=alpha,
+                beta=beta,
+            )
+
+    @staticmethod
+    def backward(ctx: FunctionCtx, dy):
+        input, other = ctx.saved_tensors
+        transa, transb, alpha, beta = ctx.params
+
+        if beta in [1, None]:
+            raise RuntimeError("Backward don't support beta == 1.0f")
+
+        if not transa and not transb:
+            dx = matmul(dy, other, transb=True, alpha=alpha)
+            do = matmul(input, dy, transa=True, alpha=alpha)
+            return dx, do, None, None, None, None, None
+
+        if transa and not transb:
+            dx = matmul(other, dy, transb=True, alpha=alpha)
+            do = matmul(input, dy, alpha=alpha)
+            return dx, do, None, None, None, None, None
+
+        if not transa and transb:
+            dx = matmul(dy, other, alpha=alpha)
+            do = matmul(dy, input, transa=True, alpha=alpha)
+            return dx, do, None, None, None, None, None
+
+        if transa and transb:
+            dx = matmul(other, dy, transa=True, transb=True, alpha=alpha)
+            do = matmul(dy, input, transa=True, transb=True, alpha=alpha)
+            return dx, do, None, None, None, None, None
+
+
+def matmul(
+    input: torch.Tensor,
+    other: torch.Tensor,
+    *,
+    out: torch.Tensor = None,
+    transa: bool = False,
+    transb: bool = False,
+    alpha: float = 1.0,
+    beta: float = 0.0
+) -> torch.Tensor:
+    """
+    等价实现:
+    def pt_matmul(a, b, transa, transb, alpha):
+        if transa:
+            dims = list(range(a.ndim))
+            dims[-1], dims[-2] = dims[-2], dims[-1]
+            a = a.permute(*dims).contiguous()
+
+        if transb:
+            dims = list(range(b.ndim))
+            dims[-1], dims[-2] = dims[-2], dims[-1]
+            b = b.permute(*dims).contiguous()
+
+        return alpha * torch.matmul(a, b)
+     Arguments:
+        input:
+            当transa为False  shape : [...,m,k]       dtype: torch.half
+            当transa为True  shape :  [...,k,m]       dtype: torch.half
+        other:
+            当transb为False  shape : [...,k,n]       dtype: torch.half
+            当transb为True  shape :  [...,n,k]       dtype: torch.half
+    Return:
+        output:   [...m,n]                         dtype: [torch.half]
+
+    """
+    if not input.is_contiguous():
+        input = input.contiguous()
+
+    if not other.is_contiguous():
+        if not other.transpose(-2, -1).is_contiguous():
+            other = other.contiguous()
+
+    return MatmulFunction.apply(input, other, out, transa, transb, alpha, beta)