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/linalg.py

@@ -0,0 +1,50 @@
+from typing import Union
+
+import ixformer._C as ops
+import torch
+
+__all__ = ["solve", "ref_slove"]
+
+
+def ref_slove(
+    A: torch.Tensor, B: torch.Tensor, *, left: bool = True, out: torch.Tensor = None
+):
+    out = torch.linalg.solve(A, B, left=left)
+    return out
+
+
+def solve(
+    A: torch.Tensor, B: torch.Tensor, *, left: bool = True, out: torch.Tensor = None
+):
+    """
+    Args:
+        A:                  (..., n, n)                                                         torch.float
+        B:                  (..., n) or (..., n, k) or (n,...) or (n, k) or (n)                 torch.float
+        left:                                                                                   bool
+                whether to solve the system AX=B or XA=B. Default: True, 目前只支持left =True
+        out:                (..., n, k)                                                         torch.float
+    Returns:
+        out:                (..., n, k)                                                         torch.float
+    """
+
+    n = A.shape[-1]
+    batch_count = A.numel() // (n * n)
+    if B.dim() == 1:
+        k = 1
+    elif B.dim() == 2:
+        if A.dim() > 2 and B.shape == (batch_count, n):
+            k = 1
+        else:
+            nid = 0 if left else 1
+            k = B.shape[nid ^ 1]
+    else:
+        k = B.size(B.dim() - 1 if left else B.dim() - 2)
+
+    if n <= 64 and k <= 64 and left:
+        return ops.infer.solve(A, B, left)
+    else:
+        device = A.device
+        cpu_A = A.cpu()
+        cpu_B = B.cpu()
+        cpu_res = ref_slove(A=cpu_A, B=cpu_B, left=left)
+        return cpu_res.to(device)