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/utils/benchmark/cuda_benchmark.py

@@ -0,0 +1,69 @@
+import time
+from collections import namedtuple, OrderedDict
+from typing import List, Dict, Any
+
+import tabulate
+import torch
+import torch.distributed as dist
+
+DeviceTime = namedtuple("DeviceTime", ["cpu", "gpu"])
+
+
+class Functor:
+
+    def __init__(self, fn, *args, **kwargs):
+        self.fn = fn
+        self.args = args
+        self.kwargs = kwargs
+
+    def __call__(self):
+        return self.fn(*self.args, **self.kwargs)
+
+
+def cuda_timeit(fn: Functor, dist_barrier=False) -> DeviceTime:
+    torch.cuda.synchronize()
+
+    start = torch.cuda.Event(enable_timing=True)
+    stop = torch.cuda.Event(enable_timing=True)
+    start.record(torch.cuda.current_stream())
+
+    t0 = time.time()
+    fn()
+    t1 = time.time()
+
+    stop.record(torch.cuda.current_stream())
+
+    torch.cuda.synchronize()
+    if dist_barrier:
+        dist.barrier()
+
+    gpu_time = start.elapsed_time(stop)
+    cpu_time = t1 - t0
+    return DeviceTime(cpu_time, gpu_time)
+
+
+def cuda_benchmark(fn: Functor, num_repeated=10, num_warmup=1, dist_barrier=False) -> DeviceTime:
+    [fn() for _ in range(num_warmup)]
+    times = [cuda_timeit(fn, dist_barrier=dist_barrier) for _ in range(num_repeated)]
+    times.sort(key=lambda t: t.gpu)
+
+    if num_repeated >= 10:
+        times = times[3:-3]
+
+    avg_gpu_time = sum([t.gpu for t in times]) / len(times)
+    avg_cpu_time = sum([t.cpu for t in times]) / len(times)
+
+    return DeviceTime(avg_cpu_time * 1000, avg_gpu_time)
+
+
+def show_benchmark_results(times: List[DeviceTime], extra_info: Dict[Any, List]=None):
+    data = extra_info or OrderedDict()
+
+    if len(times) != 0:
+        cpu_times = [round(t.cpu, 6) for t in times]
+        gpu_times = [round(t.gpu, 6) for t in times]
+
+        data["CPU Time(ms)"] = cpu_times
+        data["GPU Time(ms)"] = gpu_times
+
+    print(tabulate.tabulate(extra_info, headers=data.keys()))