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_dequant.py

@@ -0,0 +1,55 @@
+from typing import List, Union
+
+import ixformer._C as ops
+import torch
+from torch.autograd.function import Function, FunctionCtx
+
+__all__ = [
+    "bnb_dequant",
+    "ref_bnb_dequant",
+]
+
+
+def ref_bnb_dequant(
+    qA: torch.Tensor,
+    SA: torch.Tensor,
+    training: bool = False,
+    scale: float = 127.0,
+    dequant_type: int = 0,
+):
+    A = torch.empty(qA.shape, dtype = SA.dtype, device = SA.device)
+    if dequant_type == 0:
+        for i in range(qA.size(0)): 
+            A[i:] = qA[i:] * (SA[i].to(torch.float) / scale).to(SA.dtype)
+    else:
+        for i in range(qA.size(1)):
+            A[:,i] =  qA[:,i] * (SA[i].to(torch.float) / scale).to(SA.dtype)
+            
+    return A
+
+
+def bnb_dequant(
+    qA: torch.Tensor,
+    SA: torch.Tensor,
+    training: bool = False,
+    scale: float = 127.0,
+    dequant_type: int = 0,
+) -> torch.Tensor:
+    
+    """
+    Args:
+        qA:                  (row, col)            torch.int8
+                dequant input
+        SA:                  (row) or (col)        torch.half
+                scale vector
+        training:                                  bool
+        scale:                                     float
+        dequnt_type:                               int
+                0 : every row shared a scale, SA shape : [row]
+                1 : every col shared a scale, SA shape : [col]
+    Returns:
+        Tensor:              (row, col)            torch.half
+                dequant output    
+        
+    """
+    return ops.infer.bnb_dequant(qA, SA, scale, dequant_type)