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/gemv.py

@@ -0,0 +1,48 @@
+import os
+from typing import Union
+
+import ixformer._C as ops
+import torch
+
+__all__ = ["gemv", "ref_gemv"]
+
+
+def ref_gemv(x: torch.Tensor, A: torch.Tensor, gemv_max_batch: int = 1):
+    output = torch.nn.functional.linear(x, A)
+    return output
+
+
+def gemv_conditions(input, weight, gemv_max_batch):
+    # gemv 使用的条件 input:[m,k] weight:[n,k]
+    # 1. m<=gemv_max_batch
+    # 2. k%2==0 n%2==0
+    # 3. bias is None
+    input = input.view(-1, input.shape[-1])
+    weight = weight.view(-1, weight.shape[-1])
+    m = input.shape[0]
+    k = input.shape[1]
+    n = weight.shape[0]
+    if m <= gemv_max_batch and k % 2 == 0 and n % 2 == 0:
+        return True
+    return False
+
+
+def gemv(x: torch.Tensor, A: torch.Tensor, gemv_max_batch: int = 1):
+    
+    """
+    Args:
+        x:                  (..., k)                torch.float16, torch.bfloat16
+        A:                  (n,k)                   torch.float16, torch.bfloat16, torch.float32
+        gemv_max_batch:                             int
+                        用于是否满足gemv使用条件的判断,目前只支持到1   
+    Returns:
+        Tensor:             (..., n)                torch.float16, torch.bfloat16
+    """
+    disable_infer_gemm_ex = os.getenv("DISABLE_INFER_GEMM_EX", "0")
+    use_gemv = gemv_conditions(x, A, gemv_max_batch) and disable_infer_gemm_ex != "1"
+    assert use_gemv == True
+    output_shape = list(x.shape)
+    output_shape[-1] = A.shape[0]
+    output = x.new_empty(output_shape)
+    output = ops.infer.linear_ex(x, A, None, output)
+    return output