代码/code/bench.py

@@ -291,8 +291,8 @@ def _parse_args():
                     help="逗号分隔的 keep_fp32_modules，如 linear,rep_encoder.input_norm")
     ap.add_argument("--feasign-none", action="store_true",
                     help="不截断特征（max_feasign_per_slot=None）")
-    ap.add_argument("--attn", choices=["auto", "flex", "sdpa"], default=None,
-                    help="注意力实现：flex=块对角FlexAttention, sdpa=稠密(原), auto=SM80自动")
+    ap.add_argument("--attn", choices=["sdpa", "flex", "varlen"], default=None,
+                    help="注意力：sdpa=稠密(原), flex=FlexAttention, varlen=嵌套张量变长flash")
     ap.add_argument("--moe", choices=["dense", "loop"], default=None,
                     help="MoE实现：dense=向量化(新), loop=逐expert循环(原)")
     ap.add_argument("--compile", action="store_true", help="开启 torch.compile")
@@ -322,7 +322,7 @@ if __name__ == "__main__":
     if a.keep is not None:
         cfg["keep_fp32_modules"] = tuple(x for x in a.keep.split(",") if x)
     if a.attn is not None:
-        cfg["use_flex_attn"] = {"auto": "auto", "flex": True, "sdpa": False}[a.attn]
+        cfg["attn"] = a.attn
     if a.moe is not None:
         cfg["vectorize_moe"] = (a.moe == "dense")
     if a.compile:

代码/code/infer.py

@@ -40,28 +40,27 @@ CONFIG = {
     "signid_mode": "clamp",   # "clamp" 或 "modulo"：处理超界 sign id 的方式
     "sync_timing": False,     # bench 里设 True，做 torch.cuda.synchronize 真实计时
     "filter_test_users": True,  # 只处理含测试样本的用户（跳过会被丢弃的用户，省算力）
-    # 实测：FlexAttention + 稠密MoE 在本模型上反而慢 5-6 倍（模型是开销瓶颈非算力瓶颈），
-    # 故默认回到已验证最快的 sdpa + loop；flex/dense 仅作 bench 对照选项。
-    "use_flex_attn": False,   # "auto"(SM80+用flex,否则SDPA) / True / False
+    # 实测(A800)：sdpa+loop=15.1s 最快；flex/dense/compile/小batch 都更慢。
+    # attn: "sdpa"(稠密mask,默认/已验证) / "flex"(FlexAttention,慢) / "varlen"(嵌套张量变长flash)
+    "attn": "sdpa",
     "vectorize_moe": False,   # True=稠密向量化MoE；False=原逐expert循环（默认，已验证更快）
-    "compile": False,         # 是否 torch.compile
+    "compile": False,         # 是否 torch.compile（实测慢5×，勿开）
 }
 
 
-def _use_flex(device):
-    """决定是否用 FlexAttention：auto 模式下仅在 SM80+（Ampere/A800）启用。"""
-    mode = CONFIG.get("use_flex_attn", "auto")
-    if not _HAS_FLEX or mode is False:
-        return False
-    if mode is True:
-        return True
-    if device is not None and device.type == "cuda":
-        try:
-            major, _ = torch.cuda.get_device_capability(device)
-            return major >= 8
-        except Exception:
-            return False
-    return False
+def _resolve_attn(device):
+    """解析实际使用的注意力实现。flex 需 SM80+ 且可用，否则回退 sdpa。"""
+    attn = CONFIG.get("attn", "sdpa")
+    if attn == "flex":
+        if not _HAS_FLEX:
+            return "sdpa"
+        if device is not None and device.type == "cuda":
+            try:
+                if torch.cuda.get_device_capability(device)[0] < 8:
+                    return "sdpa"
+            except Exception:
+                return "sdpa"
+    return attn
 
 
 def _force_fp32_io(module):
@@ -353,12 +352,35 @@ class RepEncoder(nn.Module):
         return rep_emb
 
 
+def _varlen_attention(q, k, v, user_offsets):
+    """嵌套张量变长 flash 注意力：每个用户当独立序列、is_causal 块对角因果。
+    一个内核处理一 batch 内所有用户，无稠密 mask、无 padding 浪费、开销低。
+    q,k,v: [1, H, S, Dh]；user_offsets: [B+1]（S 上的用户边界）。返回 [1, H, S, Dh]。
+    """
+    _, H, S, Dh = q.shape
+    offs = user_offsets.to(torch.int64)
+    # [1,H,S,Dh] -> [S,H,Dh]
+    qv = q.squeeze(0).transpose(0, 1).contiguous()
+    kv = k.squeeze(0).transpose(0, 1).contiguous()
+    vv = v.squeeze(0).transpose(0, 1).contiguous()
+    # 按用户边界做 jagged 嵌套张量：[B, ragged, H, Dh] -> [B, H, ragged, Dh]
+    qn = torch.nested.nested_tensor_from_jagged(qv, offsets=offs).transpose(1, 2)
+    kn = torch.nested.nested_tensor_from_jagged(kv, offsets=offs).transpose(1, 2)
+    vn = torch.nested.nested_tensor_from_jagged(vv, offsets=offs).transpose(1, 2)
+    out = F.scaled_dot_product_attention(qn, kn, vn, is_causal=True)  # [B,H,ragged,Dh]
+    out = out.transpose(1, 2).values()  # [S, H, Dh]
+    return out.transpose(0, 1).unsqueeze(0).contiguous()  # [1, H, S, Dh]
+
+
 def scaled_dot_product(q, k, v, extension):
     """注意力分发：
-    - 若 extension 带 block_mask → FlexAttention 块对角因果（每用户只在自己序列内
-      做因果注意力，避免对 ~14000 长拼接序列做 O(S²) 稠密注意力，计算量砍数十倍）。
-    - 否则 → 标准 SDPA（稠密 mask，数学等价、用于回退/对照）。
+    - varlen_offsets → 嵌套张量变长 flash（每用户独立序列、块对角因果，开销低）。
+    - block_mask     → FlexAttention 块对角因果。
+    - mask（默认）   → 标准 SDPA 稠密 mask（数学等价、已验证最快）。
     """
+    if extension is not None and extension.get("varlen_offsets") is not None:
+        return _varlen_attention(q, k, v, extension["varlen_offsets"])
+
     if extension is not None and extension.get("block_mask") is not None:
         return flex_attention(q, k, v, block_mask=extension["block_mask"])
 
@@ -562,7 +584,10 @@ class CTRModel(nn.Module):
     def forward(self, batch):
         seq_input = self.rep_encoder(batch)
         user_offsets = batch["user_offsets"]
-        if _use_flex(seq_input.device):
+        attn = _resolve_attn(seq_input.device)
+        if attn == "varlen":
+            extension = {"varlen_offsets": user_offsets}
+        elif attn == "flex":
             S = seq_input.shape[0]  # rep_encoder 输出 [S, D]，S=总 token 数
             extension = {"block_mask": self.build_block_mask(user_offsets, S)}
         else:
@@ -652,8 +677,7 @@ def load_model(ckpt_path, device='cuda:0'):
     model.to(dev)
     model.eval()
 
-    use_flex = _use_flex(dev)
-    print(f"[INFO] attention={'FlexAttention(block-causal)' if use_flex else 'SDPA(dense)'}, "
+    print(f"[INFO] attention={_resolve_attn(dev)}, "
           f"moe={'dense' if CONFIG.get('vectorize_moe', True) else 'loop'}")
 
     if CONFIG.get("compile", False):

代码/code/tests/test_equiv.py

@@ -64,11 +64,32 @@ def test_moe_dense_matches_loop():
     print(f"[PASS] MoE 稠密向量化 == 逐expert循环 (max err={err:.2e}, dev={dev})")
 
 
+def test_varlen_matches_dense_attention():
+    if not torch.cuda.is_available():
+        print("[SKIP] varlen 等价测试（需 CUDA）")
+        return
+    torch.manual_seed(0)
+    dev = "cuda"
+    H, Dh = 8, 64
+    offs = _offsets([10, 25, 7, 40, 18], dev)
+    S = int(offs[-1])
+    q = torch.randn(1, H, S, Dh, device=dev, dtype=torch.float16)
+    k = torch.randn(1, H, S, Dh, device=dev, dtype=torch.float16)
+    v = torch.randn(1, H, S, Dh, device=dev, dtype=torch.float16)
+    with torch.no_grad():
+        dense = infer.scaled_dot_product(q, k, v, {"mask": _dense_causal_mask(offs)[None, None]})
+        varlen = infer.scaled_dot_product(q, k, v, {"varlen_offsets": offs})
+    err = (dense.float() - varlen.float()).abs().max().item()
+    assert torch.allclose(dense.float(), varlen.float(), atol=2e-2, rtol=2e-2), \
+        f"varlen 不等价 max err={err:.3e}"
+    print(f"[PASS] varlen(嵌套张量) == 稠密SDPA (max err={err:.2e})")
+
+
 def test_flex_matches_dense_attention():
     ok = (torch.cuda.is_available() and infer._HAS_FLEX
           and torch.cuda.get_device_capability()[0] >= 8)
     if not ok:
-        print("[SKIP] FlexAttention 等价测试（需 CUDA SM80+，当前环境不满足）")
+        print("[SKIP] FlexAttention 等价测试（需 CUDA SM80+）")
         return
     torch.manual_seed(0)
     dev = "cuda"
@@ -88,5 +109,6 @@ def test_flex_matches_dense_attention():
 
 if __name__ == "__main__":
     test_moe_dense_matches_loop()
+    test_varlen_matches_dense_attention()
     test_flex_matches_dense_attention()
     print("[DONE] 等价测试结束")