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_quant.py

@@ -0,0 +1,58 @@
+from typing import List, Union
+
+import ixformer._C as ops
+import torch
+
+__all__ = ["bnb_quant", "ref_bnb_quant"]
+
+
+# A : input   shape : [row, col]
+# SA : scale vector
+# quant_type
+#   0 : every row shared a scale, SA shape : [row]
+#   1 : every col shared a scale, SA shape : [col]
+def ref_bnb_quant(
+    A: torch.Tensor,
+    SA: torch.Tensor,
+    training: bool = False,
+    scale: float = 127.0,
+    quant_type: int = 0,
+):
+    qA = torch.empty(A.shape, device = SA.device)
+    if quant_type == 0:
+        for i in range(A.size(0)):
+            qA[i:] = torch.round(A[i:] * (scale / SA[i].to(torch.float)))
+    else:
+        for i in range(A.size(1)):
+            qA[:,i] =  torch.round(A[:,i] * (scale / SA[i].to(torch.float)))
+
+    qA_clamped = torch.clamp(qA, min=-128, max=127)
+    qA = qA_clamped.to(torch.int8)
+    return qA
+
+
+def bnb_quant(
+    A: torch.Tensor,
+    SA: torch.Tensor,
+    training: bool = False,
+    scale: float = 127.0,
+    quant_type: int = 0,
+) -> torch.Tensor:
+    
+    """
+    Args:
+        A:                   (row, col)            torch.half
+                quant input
+        SA:                  (row) or (col)        torch.half
+                scale vector
+        training:                                  bool
+        scale:                                     float
+        qunt_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.int8
+                quant output    
+        
+    """
+    return ops.infer.bnb_quant(A, SA, scale, quant_type)