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/t5.py

@@ -0,0 +1,97 @@
+from typing import Union
+
+import ixformer._C as ops
+import torch
+
+__all__ = [
+    "t5_split_qkv",
+    "t5_split_qkv_update_kv_cache",
+    "ref_t5_split_qkv_update_kv_cache",
+    "ref_t5_split_qkv",
+]
+
+
+def reshape_query(query, head_num, head_dim):
+    batch_size, seq_len, _ = query.shape
+    query = query.view(batch_size, seq_len, head_num, head_dim)
+    query = query.transpose(1, 2)
+    return query
+
+
+def ref_t5_split_qkv(qkv: "torch.Tensor", head_num: int, head_dim: int):
+    assert qkv.size(-1) == head_dim * head_num * 3
+    batch_size, seq_len, _ = qkv.shape
+    q, k, v = torch.chunk(qkv, 3, dim=-1)
+    q = reshape_query(q, head_num, head_dim)
+    k = reshape_query(k, head_num, head_dim)
+    v = reshape_query(v, head_num, head_dim)
+    return q, k, v
+
+
+def t5_split_qkv(qkv: "torch.Tensor", head_num: int, head_dim: int):
+    
+    """
+    Args:
+        qkv:                (batch_size, seq_len, head_dim * head_num * 3)        torch.half, torch.bfloat16
+        head_num:                                                                 int
+        head_dim:                                                                 int
+    Returns:
+        q:                  (batch_size, head_num, seq_len, head_dim)             torch.half, torch.bfloat16
+        k:                  (batch_size, head_num, seq_len, head_dim)             torch.half, torch.bfloat16  
+        v:                  (batch_size, head_num, seq_len, head_dim)             torch.half, torch.bfloat16                  
+    """
+    batch_size, seq_len, _ = qkv.shape
+    q = qkv.new_empty([batch_size, head_num, seq_len, head_dim])
+    k = qkv.new_empty([batch_size, head_num, seq_len, head_dim])
+    v = qkv.new_empty([batch_size, head_num, seq_len, head_dim])
+    ops.infer.t5_split_qkv(qkv, q, k, v, head_num, head_dim)
+    return q, k, v
+
+
+def ref_t5_split_qkv_update_kv_cache(
+    qkv: "torch.Tensor",
+    past_key: "torch.Tensor",
+    past_value: "torch.Tensor",
+    head_num: int,
+    head_dim: int,
+):
+    assert qkv.size(-1) == head_dim * head_num * 3
+    batch_size, seq_len, _ = qkv.shape
+    q, k, v = torch.chunk(qkv, 3, dim=-1)
+    q = reshape_query(q, head_num, head_dim)
+    k = reshape_query(k, head_num, head_dim)
+    v = reshape_query(v, head_num, head_dim)
+    k = torch.cat([past_key, k], dim=2)
+    v = torch.cat([past_value, v], dim=2)
+    return q, k, v
+
+
+def t5_split_qkv_update_kv_cache(
+    qkv: "torch.Tensor",
+    past_key: "torch.Tensor",
+    past_value: "torch.Tensor",
+    head_num: int,
+    head_dim: int,
+):
+    
+    """
+    Args:
+        qkv:                (batch_size, 1 , head_dim * head_num * 3)               torch.half, torch.bfloat16
+        past_key:           (batch_size, head_num, seq_len - 1, head_dim)           torch.half, torch.bfloat16
+        past_value:         (batch_size, head_num, seq_len - 1, head_dim)           torch.half, torch.bfloat16
+        head_num:                                                                   int
+        head_dim:                                                                   int
+    Returns:
+        q:                  (batch_size, head_num, 1, head_dim)                     torch.half, torch.bfloat16
+        k:                  (batch_size, head_num, seq_len, head_dim)               torch.half, torch.bfloat16  
+        v:                  (batch_size, head_num, seq_len, head_dim)               torch.half, torch.bfloat16                  
+    """
+    batch_size, _, past_seq_len, _ = list(past_key.shape)
+    seq_len = past_seq_len + 1
+    q = qkv.new_empty([batch_size, head_num, 1, head_dim])
+    k = qkv.new_empty([batch_size, head_num, seq_len, head_dim])
+    v = qkv.new_empty([batch_size, head_num, seq_len, head_dim])
+    ops.infer.t5_split_qkv_update_kv_cache(
+        qkv, past_key, past_value, q, k, v, head_num, head_dim
+    )
+    return q, k, v