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_mm_dequant.py

@@ -0,0 +1,73 @@
+from typing import List, Union
+
+import ixformer._C as ops
+import torch
+from torch.autograd.function import Function, FunctionCtx
+
+__all__ = ["bnb_mm_dequant"]
+
+
+# A : quant input              shape : [row, col]  shape : torch.int
+def bnb_mm_dequant(
+    A: torch.Tensor,
+    quant_state: tuple,
+    row_stats: torch.Tensor,
+    col_stats: torch.Tensor,
+    bias: torch.Tensor = None,
+    add_bias: bool = False,
+    training: bool = False,
+) -> torch.Tensor:
+    
+    """
+    Args:
+        A:                  (row, col)            torch.int8
+        quant_state:                              tuple
+        row_stats:          (row)                 torch.float
+        col_stats:          (col)                 torch.float
+        bias:               (col)                 torch.half
+        add_bias:                                 bool
+        training:                                 bool
+    Returns:
+        Tensor:             (row, col)            torch.half
+        
+    """
+    
+    assert A.dtype == torch.int
+    if bias is not None:
+        add_bias = True
+        print("bias.dtype:", bias.dtype)
+        assert bias.dtype == torch.half
+    else:
+        bias = A
+    out_shape = quant_state[0]
+    if len(out_shape) == 3:
+        out_shape = (out_shape[0] * out_shape[1], out_shape[2])
+    out = torch.full(size=out_shape, fill_value=0, dtype=torch.half, device=A.device)
+    new_row_stats = torch.full(
+        size=(out_shape[0],), fill_value=0, dtype=torch.float, device=A.device
+    )
+    new_col_stats = torch.full(
+        size=(out_shape[1],), fill_value=0, dtype=torch.float, device=A.device
+    )
+
+    assert (
+        new_row_stats.shape[0] == row_stats.shape[0]
+    ), f"{new_row_stats.shape} vs {row_stats.shape}"
+    assert (
+        new_col_stats.shape[0] == col_stats.shape[0]
+    ), f"{new_col_stats.shape} vs {col_stats.shape}"
+    numRows = out_shape[0]
+    numCols = out_shape[1]
+    ops.infer.bnb_mm_dequant(
+        A,
+        row_stats,
+        col_stats,
+        out,
+        new_row_stats,
+        new_col_stats,
+        numRows,
+        numCols,
+        add_bias,
+        bias,
+    )
+    return out