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/bnb_rowcol_absmax.py

@@ -0,0 +1,52 @@
+from typing import List, Union
+
+import ixformer._C as ops
+import torch
+
+__all__ = ["bnb_rowcol_absmax", "ref_bnb_rowcol_absmax"]
+
+
+# input  : input   shape : [row, col]
+# threshold : abs of element exceeds threshold will be ignored
+# type
+#   0 : row absmax
+def ref_bnb_rowcol_absmax(
+    input: torch.Tensor,
+    training: bool = False,
+    threshold: float = 0.0,
+    type: int = 0,
+):
+    input = input.float()
+    if threshold ==0.0:
+        threshold = float('inf')
+    mask = (torch.abs(input) < threshold)
+    masked_input = mask * input
+    masked_input = masked_input.half()
+    if type == 0:
+        out = torch.amax(torch.abs(masked_input), dim=1)
+        
+    else:
+        out = torch.amax(torch.abs(masked_input), dim=0)
+    return out
+
+
+def bnb_rowcol_absmax(
+    input: torch.Tensor,
+    training: bool = False,
+    threshold: float = 0.0,
+    type: int = 0,
+) -> torch.Tensor:
+    
+    """
+    Args:
+        input:              (row, col)        torch.half
+                目前col值必须满足col%2==0
+        training:                             bool
+        threshold:                            float
+                abs of element exceeds threshold will be ignored
+        type:                                  int
+                row absmax, 目前只支持type=0
+    Returns:
+        Tensor:             (row)              torch.half
+    """
+    return ops.infer.bnb_rowcol_absmax(input, threshold, type)