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/act_and_mul.py

@@ -0,0 +1,88 @@
+from typing import List, Union
+
+import ixformer._C as ops
+import torch
+import torch.nn.functional as NNF
+
+__all__ = ["ref_silu_and_mul", "ref_gelu_and_mul", "ref_gelu_tanh_and_mul",
+           "silu_and_mul", "gelu_and_mul", "gelu_tanh_and_mul"]
+
+
+def ref_silu_and_mul(input: "torch.Tensor") -> torch.Tensor:
+    x1, x2 = input.chunk(chunks=2, dim=-1)
+    res = NNF.silu(x1) * x2
+    return res
+
+
+def ref_gelu_and_mul(input: "torch.Tensor", gate_first=True) -> torch.Tensor:
+    x1, x2 = input.chunk(chunks=2, dim=-1)
+    if gate_first:
+        res = NNF.gelu(x1) * x2
+    else:
+        res = NNF.gelu(x2) * x1
+    return res
+
+
+def ref_gelu_tanh_and_mul(input: "torch.Tensor") -> torch.Tensor:
+    x1, x2 = input.chunk(chunks=2, dim=-1)
+    res = NNF.gelu(x1) * x2
+    return res
+
+
+def silu_and_mul(input: torch.Tensor, output: torch.Tensor = None):
+    
+    """
+    Args:
+        input:              (..., 2*hidden_size)    torch.float16, torch.bfloat16, torch.float32
+        output:             (..., hidden_size)      torch.float16, torch.bfloat16, torch.float32
+        
+    Returns:
+        output:             (..., hidden_size)      torch.float16, torch.bfloat16, torch.float32
+    """
+    if output is None:
+        output_shape = list(input.shape)
+        output_shape[-1] = output_shape[-1] // 2
+        output = input.new_empty(output_shape)
+    
+    ops.infer.silu_and_mul(input, output)
+    
+    return output
+
+
+def gelu_and_mul(input: "torch.Tensor", output: torch.Tensor = None, gate_first=True):
+    
+    """
+    Args:
+        input:              (..., 2*hidden_size)    torch.float16, torch.bfloat16, torch.float32
+        output:             (..., hidden_size)      torch.float16, torch.bfloat16, torch.float32
+        gate_first:                                 bool
+    Returns:
+        output:             (..., hidden_size)      torch.float16, torch.bfloat16, torch.float32
+    """
+    if output is None:
+        output_shape = list(input.shape)
+        output_shape[-1] = output_shape[-1] // 2
+        output = input.new_empty(output_shape)
+
+    ops.infer.gelu_and_mul(input, output, gate_first)
+
+    return output
+
+
+def gelu_tanh_and_mul(input: torch.Tensor, output: torch.Tensor = None):
+    
+    """
+    Args:
+        input:              (..., 2*hidden_size)    torch.float16, torch.bfloat16, torch.float32
+        output:             (..., hidden_size)      torch.float16, torch.bfloat16, torch.float32
+    Returns:
+        output:             (..., hidden_size)      torch.float16, torch.bfloat16, torch.float32
+    """
+    if output is None:
+        output_shape = list(input.shape)
+        output_shape[-1] = output_shape[-1] // 2
+        output = input.new_empty(output_shape)
+
+    ops.infer.gelu_tanh_and_mul(input, output)
+
+    return output