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/train/speedformer/layers/gpt2/attention.py

@@ -0,0 +1,45 @@
+import torch
+import os
+from einops import rearrange
+from flash_attn import flash_attn_varlen_func
+
+
+@staticmethod
+def replace_flash_attn_forward(self, q, k, v, attention_mask, query_length, dropout=0.0, softmax_scale=None):
+
+    # flash-attn(ixdnn)存在gpt2(118M,338M,738M) shape没适配，只能采用普通版本
+    assert os.getenv('ENABLE_FLASH_ATTENTION_WITH_IXDNN', "1") == '0', "flash-attn should not be use ixdnn version, please set variables" \
+        " in shell \"export ENABLE_FLASH_ATTENTION_WITH_IXDNN=0 \" "
+    assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
+    assert all((i.is_cuda for i in (q, k, v)))
+
+    batch_size, seqlen_q = q.shape[0], q.shape[1]
+    seqlen_k = k.shape[1]
+
+    q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
+    cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32,
+                                device=q.device)
+
+    if query_length != 1:
+        # during training q,k,v always have same seqlen
+        assert seqlen_k == seqlen_q
+
+        is_causal = self.is_causal
+        cu_seqlens_k = cu_seqlens_q
+        dropout_p = dropout
+    else:
+        # turn off FA causal mask after first inference autoregressive iteration
+        # only on first autoregressive step q,k,v have same seqlen
+        is_causal = seqlen_q == seqlen_k
+        cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32,
+                                    device=q.device)
+        dropout_p = 0
+
+    output = flash_attn_varlen_func(
+        q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k,
+        dropout_p,
+        softmax_scale=softmax_scale, causal=is_causal
+    )
+    # print(f"{output}")
+    output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
+    return output