From aa8b9f5c56821c7f5ba8c82d3b51634b8fbd83fd Mon Sep 17 00:00:00 2001
From: HaoHaoXueXiHxy <20211005@bjtu.edu.cn>
Date: Wed, 3 Dec 2025 12:11:55 +0800
Subject: [PATCH 1/2] =?UTF-8?q?2025=E5=B9=B4=E6=98=87=E8=85=BEAI=E5=88=9B?=
 =?UTF-8?q?=E6=96=B0=E5=A4=A7=E8=B5=9B-=E6=98=87=E6=80=9D=E6=A8=A1?=
 =?UTF-8?q?=E5=9E=8B=E5=BC=80=E5=8F=91=E6=8C=91=E6=88=98=E8=B5=9B=EF=BC=88?=
 =?UTF-8?q?S1=E8=B5=9B=E5=AD=A3)--MoE=E8=B5=9B=E9=A2=98--=E6=88=91?=
 =?UTF-8?q?=E6=83=B3=E5=A5=BD=E5=A5=BD=E7=9D=A1=E4=B8=80=E8=A7=89=E9=98=9F?=
 =?UTF-8?q?=E6=8F=90=E4=BA=A4?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../README.md"                                | 262 ++++++++++++++++++
 .../patches.zip"                              | Bin 0 -> 30103 bytes
 2 files changed, 262 insertions(+)
 create mode 100644 "2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md"
 create mode 100644 "2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/patches.zip"

diff --git "a/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md" "b/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md"
new file mode 100644
index 00000000..0b0ad724
--- /dev/null
+++ "b/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md"
@@ -0,0 +1,262 @@
+#### 1. MoE模块前向优化
+在moe模块可以拆分 prefill（长序列预填充）和 decode（单 token 解码）阶段逻辑，能够解决解决 “两阶段结果不一致” 和 “推理效率低” 问题。
+
+DeepseekMoE模块
+显式拆分prefill/decode阶段
+```python
+def forward(self, hidden_states):
+    if orig_shape[1] == 1:  # decode阶段（单token）
+        y = self.moe_infer_decode(...).view(*orig_shape)
+    else:  # prefill阶段（长序列）
+        y = self.moe_infer_prefill(...).view(*orig_shape)
+```
+通过`sequence_length == 1`显式区分两阶段，分别调用专属推理函数，避免共享逻辑导致的结果偏差。
+
+decode阶段（单token推理）：极简逐专家计算
+```python
+@no_grad()
+def moe_infer_decode(self, x, flat_expert_indices, flat_expert_weights):
+    expert_cache = ops.zeros_like(x)
+    for i in range(self.num_experts_per_tok):
+        expert_id = flat_expert_indices[i].item()  # 取第i个选中的专家ID
+        weight = flat_expert_weights[i].item()    # 取对应权重
+        expert = self.experts[expert_id]          # 选中专家网络
+        expert_out = expert(x)                    # 专家处理单token输入
+        expert_cache += expert_out * weight       # 加权累加输出
+    return expert_cache
+```
+单token场景下，直接遍历门控网络选中的`num_experts_per_tok`个专家，逐个计算并加权累加；
+无冗余操作：单token无需批量索引/掩码，直接遍历激活专家，计算效率最大化；
+  `@no_grad()`：关闭梯度计算，避免推理阶段显存占用；
+权重直接乘：跳过复杂的掩码/索引，保证计算语义简单可追溯。
+
+prefill阶段（长序列推理）：批量高效计算
+```python
+@no_grad()
+def moe_infer_prefill(self, x, flat_expert_indices, flat_expert_weights):
+    expert_cache = ops.zeros_like(x)
+    idxs = flat_expert_indices.argsort()  # 按专家ID排序，便于批量处理
+    tokens_per_expert = flat_expert_indices.bincount().cumsum(0)  # 每个专家的token数量
+    token_idxs = idxs // self.num_experts_per_tok  # 还原token在原序列中的索引
+    
+    for i, end_idx in enumerate(tokens_per_expert):
+        start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+        if start_idx == end_idx: continue  # 无token的专家跳过
+        
+        expert = self.experts[i]
+        length = (end_idx - start_idx).item()
+        # 1. 切片获取当前专家处理的token索引（MindSpore原生操作）
+        exp_token_idx = ops.narrow(token_idxs, 0, start_idx, length)
+        # 2. 批量提取token的输入特征
+        expert_tokens = F.embedding(exp_token_idx, x)
+        # 3. 专家计算 + 权重加权
+        expert_out = expert(expert_tokens)
+        expert_out = expert_out.mul(F.embedding(ops.narrow(idxs, 0, start_idx, length), flat_expert_weights))
+        # 4. 批量累加结果到缓存（替代循环赋值）
+        expert_cache = mindspore.mint.scatter_add(
+            expert_cache, 0, 
+            exp_token_idx.view(-1, 1).tile((1, x.shape[-1])),  # 扩展索引到特征维度
+            expert_out
+        )
+    return expert_cache
+```
+长序列场景下，先按专家ID排序token，再批量处理每个专家的所有token，避免逐token循环；
+
+排序+批量处理：将同专家的token归拢，一次调用专家网络处理所有token，计算效率提升数倍；
+`ops.narrow`替代Python切片：适配MindSpore张量操作，避免维度不匹配；
+`scatter_add`批量累加：替代逐token赋值，减少显存交互开销；
+跳过空专家：无token的专家直接跳过，减少无效循环。
+
+---
+
+Qwen2MoeSparseMoeBlock模块
+（deepseek模块的实现参考了培训给出的一些示例，基于这个经验，在qwen2_moe模块也进行了拆分获取了收益）
+
+与DeepseekMoE的“完全拆分函数”不同，Qwen2-MoE在同一个`forward`函数内通过分支区分阶段（代码写的没那么漂亮，主打一个效率）：
+```python
+def forward(self, hidden_states):
+    if sequence_length == 1:  # decode阶段
+        # 仅循环活跃专家的精简逻辑
+        ...
+    else:  # prefill阶段
+        # 遍历所有专家的标准逻辑
+        ...
+```
+
+decode阶段（单token）：活跃专家筛选+复用prefill核心逻辑
+```python
+if sequence_length == 1:
+    # 1. 生成专家掩码（与prefill格式完全一致）
+    expert_mask = nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+    # 2. 筛选活跃专家（仅处理被选中的专家，减少循环次数）
+    expert_usage = ops.sum(expert_mask, dim=(1, 2))  # 统计每个专家被选中的次数
+    active_experts = ops.nonzero(expert_usage > 0, as_tuple=False).squeeze(-1)
+    # 3. 遍历活跃专家（复用prefill的index_add逻辑）
+    for expert_idx_tensor in active_experts:
+        expert_idx = int(expert_idx_tensor.asnumpy().item())
+        expert_layer = self.experts[expert_idx]
+        idx, top_x = ops.nonzero(expert_mask[expert_idx], as_tuple=True)
+        if 0 not in idx.shape:
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+            # 关键：与prefill使用相同的index_add累加，保证结果一致
+            final_hidden_states = final_hidden_states.index_add(
+                0, top_x.int(), current_hidden_states.to(hidden_states.dtype)
+            )
+```
+- **核心优化**：
+  - 掩码格式对齐：decode阶段生成与prefill完全一致的`expert_mask`，消除掩码维度/排列差异导致的结果偏差；
+  - 活跃专家筛选：仅循环被选中的专家（而非所有专家），单token推理循环次数从`num_experts`降至`top_k`（通常2），效率提升；
+  - 累加逻辑一致：复用`index_add`而非直接赋值，保证两阶段输出的数学语义完全一致。
+
+prefill阶段（长序列）：标准掩码+全专家遍历
+```python
+else:
+    # 1. 生成专家掩码：(num_experts, top_k, batch*seq_len)
+    expert_mask = nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+    # 2. 遍历所有专家（保证长序列处理兼容性）
+    for expert_idx in range(self.num_experts):
+        expert_layer = self.experts[expert_idx]
+        idx, top_x = ops.nonzero(expert_mask[expert_idx], as_tuple=True)
+        if 0 not in idx.shape:
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+            # 核心：index_add累加，保证批量token结果正确
+            final_hidden_states = final_hidden_states.index_add(
+                0, top_x.int(), current_hidden_states.to(hidden_states.dtype)
+            )
+```
+- **核心保留**：
+  - 全专家遍历：长序列场景下，确保所有可能被选中的专家都被处理，避免遗漏；
+  - 掩码索引：通过`nonzero`提取每个专家处理的token索引，批量计算后用`index_add`累加，保证结果与原始逻辑一致。
+
+共享专家（Shared Expert）：两阶段统一逻辑
+```python
+shared_expert_output = self.shared_expert(hidden_states)
+shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output
+final_hidden_states = final_hidden_states + shared_expert_output
+```
+- 共享专家是所有输入都会经过的“基础专家”，用于兜底计算；
+- 门控权重：通过`sigmoid`激活的门控网络（`shared_expert_gate`）控制共享专家的输出强度，两阶段逻辑完全一致，避免额外偏差。
+
+这样把prefill和decode分开来写带来了在decode阶段有达到秒级的收益。这样做完总分提升到了160+。
+
+
+#### 2. 培训中提到的使用ops算子替换索引操作
+
+```python
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    # x1 = x[..., : x.shape[-1] // 2]
+    # x2 = x[..., x.shape[-1] // 2 :]
+    x1, x2 = ops.split(x, x.shape[-1] // 2, dim = -1)
+    return ops.cat((-x2, x1), dim=-1)
+    
+    
+...
+
+    # cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+    # sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+    cos = F.embedding(position_ids, cos).unsqueeze(unsqueeze_dim)
+    sin = F.embedding(position_ids, sin).unsqueeze(unsqueeze_dim)
+
+...
+    
+    hidden_states_expand = ops.unsqueeze(hidden_states, 2)
+    hidden_states = hidden_states_expand.broadcast_to((batch, num_key_value_heads, n_rep, slen, head_dim))
+
+...
+            # self.cos_cached[:seq_len].to(dtype=x.dtype),
+            # self.sin_cached[:seq_len].to(dtype=x.dtype),
+            ops.narrow(self.cos_cached, 0, 0, seq_len).to(dtype=x.dtype),
+            ops.narrow(self.sin_cached, 0, 0, seq_len).to(dtype=x.dtype),
+            
+...
+
+            attention_mask_expanded = ops.unsqueeze(ops.unsqueeze(attention_mask, dim=1), dim=2)
+            padding_mask = ops.narrow(causal_mask, -1, 0, mask_length) + attention_mask_expanded
+
+```
+使用ops替换掉tensor索引的一些操作能够带来一些细微的收益（几百ms），有的替换甚至没有收益（也可能是那块代码没执行）。
+
+#### 3. 使用StaticCache，并开启JIT优化
+
+使用StaticCache需要先在cache_utils.py文件中修改一下StaticCache类的update函数中的某一部分（被注释掉的是原来的）：
+```python
+            else:
+                # use index_add for mindspore since tensor slice is too slow and no implementation of index_copy
+                # k_out = ops.index_add(k_out, 2, cache_position.int(), key_states)
+                # v_out = ops.index_add(v_out, 2, cache_position.int(), value_states)
+                k_out.index_add_(2, cache_position.int(), key_states)
+                v_out.index_add_(2, cache_position.int(), value_states)
+```
+
+由于默认是DynamicCache，因此需要在utils.py中generate接口处修改一些地方，支持StaticCache的创建使用
+```python
+        else:
+            model_type = getattr(self.config, 'model_type', '')
+            supports_cache_position = model_type in ['qwen2_moe']
+            # print('StaticCache')
+            if (
+                hasattr(self, '_supports_static_cache')
+                and self._supports_static_cache
+                and not requires_cross_attention_cache
+                and supports_cache_position  
+            ):
+                if hasattr(self.config, "_pre_quantization_dtype"):
+                    cache_dtype = self.config._pre_quantization_dtype
+                else:
+                    cache_dtype = self.dtype
+                # print('StaticCache')
+                model_kwargs[cache_name] = self._get_cache(
+                    cache_implementation="static",
+                    max_batch_size=batch_size,
+                    max_cache_len=max_cache_length,
+                    model_kwargs=model_kwargs,
+                )
+            else:
+                num_hidden_layers = self.config.get_text_config().num_hidden_layers
+                model_kwargs[cache_name] = (
+                    DynamicCache(num_hidden_layers)
+                    if not requires_cross_attention_cache
+                    else EncoderDecoderCache(DynamicCache(num_hidden_layers), DynamicCache(num_hidden_layers))
+                )
+```
+这里StaticCache只限制了在qwen2_moe中使用，因为我尝试在deepseek中使用最后好像出现了mismatch，可能精度上有一些损失。
+
+最后JIT优化使用在了RMSNorm部分，还在Qwen2MoeModel的_update_causal_mask部分加了JIT，都带来了收益
+```python
+    @mindspore.jit
+    def forward(self, hidden_states):
+        # if use_pyboost():
+        #     return F.rms_norm(hidden_states, self.weight, self.variance_epsilon)
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(mindspore.float32)
+        variance = ops.mean(hidden_states.pow(2), -1, keepdim=True)
+        hidden_states = hidden_states * ops.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+        
+    ...  
+    
+    @mindspore.jit
+    def _update_causal_mask(
+        self,
+        attention_mask: mindspore.Tensor,
+        input_tensor: mindspore.Tensor,
+        cache_position: mindspore.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+```
+
+最后JIT优化牺牲了一些prefill的时延，但是对decode的时延带来了很大的提升，所以最后总分也提升到了280+。
+
+---
+#### 最终评测结果
+
+|   评测指标    |     平均得分     |
+| :-------: | :----------: |
+|   峰值显存    |     100      |
+| Prefill时延 |   62.3694    |
+| Decode时延  |   696.3557   |
+|  **总分**   | **286.2417** |
diff --git "a/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/patches.zip" "b/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/patches.zip"
new file mode 100644
index 0000000000000000000000000000000000000000..cd92eb435e67c529bdaa1ce68df8587df05a4326
GIT binary patch
literal 30103
zcmagFQ?M{huq1eF+qP}nwr$(CZQHibcWm3X?fvJ*#_rC(Om{RMI`XlytFlr-8W;o$
z;D3#bS_|#}x%}T71ON_zgQ1JDg{d>WstP0ku!1DKrO|)O)dLy;5ab*f006|~zqbni
zA)x$sLi`UxB6Fz6F**RioE-oF!T$$=fq{XEmeJPI+4g@y(z;uC()|zk|AYL$(EoqP
zH#)Y?TVhGSb^46vz8p+MhDk}LlBlIk*~6*5&CZ?9Rd&@XR9HmeNNNNF00nwZ-m0^9
zHWQI1U#oK09CC;6Nj2`t5}ps6akaz#Vg#k~ztC&E%m8Kp2}LGck@f$aGEq2|cK(|`
zF2)1Kak}KGL2@sQy7;v*G3iw+haR%{<tcHC#Kj|xS$tB|#NtpRn-rXe+Q%fe35SYG
z6fzC|B`Ops;BkPNkdIAvL;ZW~7k9Dg>1h+q(o@c|J6XPdWgGV_TDDfGSDk_AuJLjS
zYo}S9Oj3G{=m7#35R|E8QAfzc#3mICY7{UzHnZP>YCVulFPHXu-Up^iTs%y&QLaO^
zS9+#&{T%H)3<+~*;1W9xRcRX>*PLA^t2X6wxNi=k_0tPZOUyFr)-=iIFXBRB;S=);
z<&n+HMRB*7ZLS)*u*ilp9P&#oL$2LYvTqKli(%3iJbzmaW>bTxrh1HgcQ2w=sgthR
zV#q_%MyL3amNfuDRP=58h)T>dmr0X27HMkEB`z*87-`bDC|ffyrh@fDAm>Vcg!#A_
z6<DoU6KCeI@#o>*iW@TY#b5>yx<o@qshbX^YKtvara5uTqf#Xr6ftoz1)*L?>;TnQ
ziDyr9$=?!|H0mJShhGwb1Eo=w-E^Lg8;e(BC!jfrhJWVdU>UeoEn;T#u9J2{EIl#O
zGg@)WS!J4N>9Ons)UZ=ih-^e%38Wuq<BmfwWpESE@Uqa0?_4hD^O8BDJ)BZhvSWp5
z;bIRi)k298vMxqIutsa+l$q0%OwH3~jEoRYD+bOObeb$O$<qppE|#cZqSJB@tSOmK
zDX`TG&)YJuC=G<=MBu9#qjG^Jz`}#mf=2ATOp;|~C)NcPM*BA4yj02(=b~zm6e?L!
z^o=@nC<y8?2{&9?)I(;ClX<-IwCH6;swvhx=5tFYOiyq+N#H}a=92^lEKDSlJTqwH
z{iCW(WCzO^w=`%j&64nwB3U^Gz+v;xV!>J~G5-`5GlYA~j1ShwKMO&h22+8z<|+`U
zUX7p#1!#MyY6Y^w>m^uFT7KHUZfd1})z!D~x9xKOY|ai$|AOPe!j4<t_48pp-~Ba{
zDZZ-J-S|3B_BgKeI~slOes9s$x8Uu*wAtO%=|n&r=xoX-d1!X+I<UTS6h+fr#wOe|
z2p(pugxXeTMXRiOomcgFr<NmtcK*H>R;&6<I+91kwa@TuYg_5;+g5kCf9{J{yW8FO
z{+9dw`d@y{L}HPHBFt1MxY+IfWaP<`RtIggb&cpqm!dU+w<}_X$QWx)?~|`{Ani0=
z*=oMF-wJH~T}yrKP8L>K{|>dTCUO6k+BebX+Gca@botNxpUPib+rRPYKF(G#apz%T
zR4+1?vtCXGA_Q`HCh^WtJGh%GA}|YG%gLNL%!L%j4L5UcMFcT4=M!Xhfm2y1=VfVj
z9Ir=YC=}uxl7j<zb{@zuIs4`%f74{iYoN#z(~h^<9UD+MPQdXWXqWsPDqj|~_fxsy
z>p#!<!+MuCdkAPo+TiiQcDV!=MXnJVx#M@Z_0G8GZsc2T|LsE#U*1+5aAVPKwRpHl
za-=di5N4w;fYvN|y;<XF4(M$Q?7UY-Uydt2wndN+vyG(bI?1sp%yNETKEdd$RI;0^
zdkZ94VsrtwNMlklm^${bl)<N4WvS7`T@f3XYQbkP#p<9NWU}QHP9+wLI|00oaO(&?
zQv?{RzCGgw60cvOZb1=Jm}uw70BjzdM21Y3-#@Z=hM5>0PaF(6c|lBI+5s?zYIMl7
z8*;07$9l3i%#LO5kH=*fxnmI8CJh)fOh|DhJ;|?B1PjAOV7x-<l57_*pmYJSD?pUe
zNXE%(4wGd{*hdc1YoZPe&qk;8Msjwk#>Ou@Or^WR8v$hH+t*{z-$E6aI-Wwk^Wk*`
z3CrGRgF-3RD~35u7snw?)dsxoG=*L<8D6M51Q};p&ZX&gpRT-=!r%|8CecIXPM_1Q
z1~vr0<=|su5rc@mS+v3+!@&0;{}-)8Fd^y8kYWN5T&WST3vhCMW-NoiGVTK8Q+qiv
zq9b()`9U>V$NJ3&@plOx7r;t>iX7qx$Fl2W#1=Fi0z!-zp=bnksn>u>{BW2t_n<N>
z;c||3fwSE@PzLHjoZnu^O6})mo<kM9h8oScnJ~m*Vy0^@)3L3p7-;38HAPgyB31;l
z%}w~H%H)i$VWukBuaa~)O*Qr>hm1QJT}~2(>NKtg)8AZVO!_#)akF%2J=G8i#~>gd
zc{ESD=F=F)j~-Pu-3~=$)1?}lnoX3w9dp1wkgQ`V@M8*usAIZacSa7)=z$hOcw-|`
z48;Mtneog~<VPJT$yBY@`mP)*is6(??kHTuJ--YvIkC$i<~4~UT87DRPv(uEEd+z`
zRJWEzm#NilpcP;S1_1J4lMG{f3bIqlY)mG-BtL-wvNfR)$6j!Z%OWM|Fs^4WU<b%C
zt2OD_aiA{RFXo#n>48Zm5jnT$*HL}N1mI2wKL6QZO8)DD-b4hUBBD$+rGgnfhDb3$
zUV>m8PT-&~{3ekFv=?oVJxU-BwNaleG!rwK5fGV)=Ij9iko~!GR6w$TZZHTPRb7_E
z+*Sf2iHo}#QW8H*Q+n@)sU9$m_I;x#Ke>Cg<~MBiY#t{V?t!$@8<&tMDqe6ZRCL;(
zSuq3MF@H<|C5AI?<yxtQIgBktd91qOVif**88_LBCex`>Di<X;f1?}2*C9C*tBb0}
z7qZ*<f$VfgTv@^mgL9RocNWcbJsw}MGAvP9X#v8;&BvsVeN`}f5$TuSSy}RF8v|%W
zfo>SP@1<z{80kQMmByS}lv`gZ!%Wj0gpPt-RXOl&Y#>L)4yE5N4?`Wk+ZRp8T7C#3
z1OJ=KWirBr=$I8OwdfDc->FRSo)43f15Hw7Hno%@WZNlhPm9d5CJe7J^JvM7WQ{&-
zlH{$BkXeuJj8K^Ol~Pg?;VoPEgsC;EOi=7N9#>+MjWykEJ@q$1aoSL+9Y1=2uia%~
zW`)Q$b+k>yYIe3ra8Z>nuO~>Y>A{i|*KGl2Afl9F^F0x~BCBHBPcH97rQy$Q(yKK)
zPK+roQnno|h{mk3iPD^5mdTnJI3tY>&93xu<jS6rl%6E}gktgxWR8n#(%L^i=jb7Q
zU9k`P*YR|F$bNE)?rM&Pd)N;r80c#eH1NmiRAJP8kNNv&4--v}-D0#H19c7ii2qtm
zkJ$I8>Q%INX())zCc&%L#6}R>6o=Ja$BL~mxFk~!$=Ro`xqQNwEGlY|ApZzL2WP9m
zaPb`^l)3QF1MiW3j?1l?(A!pKJ3=8!<D!HF<<o;nG6YYb3P}?}X9a4T2V@KH7I*wv
zjs(kZP{?=4O%*xS$%G@DG#oJ0+gU-9`x~u>P5gz5WQD8gij(@*^{>Z!cf+`CJcf8h
z$hE8jf1%|;s8(bwF6Jc<^X;&~-EKP?==R{CxKoh{<d6ATbI2Jdt1R@05M@8s&}f<P
z++&>$7?)hEanzjQR5iTs^jx7(MSlNn!9f-i%)tun6b>d-Ul^<wv-u!^EX6MN1ar%+
zHvkNO;rM3Nf_5;*e+-)3<`W1TElqd?@wy8ZH6EqtG}|(GZ2G*+PSpXh?568PYpF^?
z7W(;&vsf$iePRLqodyB`F<pzREynD?e}Nszu!SL3Rt@q2xW}2&gI(axDFL|@mr+8m
z;i^DC?ogBTcIE=_3xr{ph0vxkMo46FSyzc58mz=2XaCZ5@_`5s0s#=~Ow16JPKXuY
zpO~$b;(axN_D_~;=Jo`o1kn&D6I1PQ$QS{oa*TNgNUID{X?de6a?59~O58ET-*ncI
z)#AJvjzbvKDfry136h!m)J6uUGYGa`16K~&zAjIwwZ>_1qQ<K%CSL5RsG*|^hV{S0
za|snBw}&<IW4WRUNFpdfQV{?}Kss@K-qz8p6|0?TRV+bn*i^BSEuOeo%xa$U?AB?e
z057l!r9?uP)uz0JLG-ZLMHh}56xe9^pvf$FZJ~EHsI1BBxCtdEjx9F^h-!H=#0xfP
zktP|keh}ldsEu>D0R{OX7PY%+wwtnSLq^+xPW_<*RlJ&$kYQ%f*NpW2CQQ<}{h(co
zj49J)tbG89FVlfVozA0TP97$8EdHoF!f_a(pU9*7s{1B(K^f)xt3-_NIfNaU<4joL
z24bG5;)KZIA4Z3GCYZQ5&k^v5JO;cXBmDJ2n)mQ*IG#m-;Slm(gQdbgP<;Y1+~iAW
zl$$6m`qL1ZTN2}p5n{a*hl>aR%VsuxR(T=LS%pxuh%^ifZ`}1$35Yxdt!PUdu@-#p
znTT5HtnA~mX^?o8l*!&S<-sSB4j!8J=n3ukNuJ+Yr^;z#St4A@U;rsI<<rp-q!uH=
zvMr{}<A)WqmEQ6?9th^#N}vb{a&86|iO-u#;co*>?)x$R-;`Y{4L(MH*}B_CU$yy2
z&)Sa71^baV@aOH84f$Dj(DpuW`}oGJK_A&#0~WNuAvoYMhHI_gP6F2ZM>8D3zxjKQ
zL}Pze@8|iVcSvn0!a2c4=kf0Y(0{4p+tz-970;SPKHuo;VdGP+LPLQA37(%>XoFYf
zmYNb_;M`L@o-u@fAOT*ZzQReGHXCbGHu8bS0yRT(`+eW{3m&%;i7>GOGul{mq`ZG7
z!tOdy_<q)#sUI$H=F<WI;11sFV`+it^kWY-ARvESwOA3Jh?eq?_RA9|!woj<dNn;@
zVFa6wLq%Q0>z0B5MiF>T(=mnJ$zv`{qF#f7I(Zm`Ndvc4QO=*<Bx7=riOI%{yR_;X
zg#<JXN=}0)0d_Ednx0B)pTY<iNxDQry2ZZByRi~fMlfKi*!CJAZ=}hvP}9JMmBfe&
zW#H^A^I}zGL7Wej7+uA&E;7(?M3g4N#|kfL*eYDU%E6!L@GCBblWgF%9)-Z5*l1L@
zA2OT!S!7X;rH_zHoO{*@(f2Y_KA3kK)Vm37O+!UE)~m@l!O5Z?W)S>OONvpEXq1C8
zuzbpy;x54n(Y%?1sEfMgjnoYt8SzY5mamnVX?AHxzoF2c!CY<Gk@a2H!O{-{14)!Y
z<04BTSR{XDfI2O_LK~huyG00OWPy=_lBnTECG)}BC?uCOASpe=$qWqfmvx48Pr-U7
zj<W$X1P~YyGM?EW<Q;96Sg63J^^-28Vc7A?>j{D^fS|RufZvUpHR?V|!X#+XmE4$N
zPX<Gl_iEv=iIk1=Y7*+xj$3vJ-ho^-<;C>pVU=?kEGkihci+t8!*IX@rXZ@{okXid
zMz|_Ke-(Ij+K!N0kgMc)IQwUu$VlXEAsHhMa^HRCB92^qv4lk1@dvzOTS9kJ5?X0z
zXA!q)y=FrtlVVZHxFr3?uP4;*I1FMnAc8+CC2g7afSw3e#F83*64COD<NiuUinn^M
zzIZz27YeW9OV2T36Zs2=ysm?UlIr8E%`$Q$ZO@b}8XZUxF=^Y=Qvt7!qJHnde3gSr
z!!1DG6yP<H?jRqM!S574p=VvZdBLeS)L72bYD-e-HNuyX%CX>B8^%mfS2Q8%8E|D7
zdpK|Cy9hFdtP*-{=#;k!%HIv;dbMeaz6L<;8wT#sJX?TkXW;L;es-AQMPj<xLJ#~y
zyFnU>a(J~4QISqm9$*0RNLH4`OGDO^s<bJnbF$2r9vLjahAgV?xk5cfzNjjYd)YF8
zIv1Wj*s$Y=qc%Hi90O8;_KZmuvn)!y<s~iE!oK8{XszDD?5L^tRGgtUa-siGEYi=v
z%EE45ynPtzLEwQIhugscv*n8|^ktJmS+P?_JSGx(nwl7K1}`n#I;+d}Ih?sf^zz2p
zh*d5~7`I!yHhTZ&<_Agdqsijhl^gH;TfP7CoA>bk#?$=HR^Pkl*~y##Qagv5=W${C
zD-$Lv)VhAxJ^kLLKA&g^0~wC9eTJjr+SZ-wy|%utjo(-Czu&Xp9a#S>`Cs_G_w-f9
zn(69xr%#X>?uBb$26##LpcAPcUmpcJ&PTljp1nxxy8s?~-jHr|B=y^>7x3_K5>!b%
zWIt(_Pl&#gtxb76FX~MXr3`J)-!AdtT(@Og`JP%CFPoNJ^@|sHv$644NRzROS9qgp
zEm`pONerJCXuTurKf~}+<WkIm?RxKH%Ir}eD=nkB8~>)ppL^lI@a&uRyL**#KKOV@
zr4sWz)D@O-ymjOc?pKjdf(>5Z#QaQx-*|uf{JoH6upTXId%Hbuc`Ue*GZu-jVDI0N
z*8eh3H#WRmy^Zys<39htASuZ9H6I&w#;y$vX-oM)*r_KUdCEi(e3<<F(*Of5PVWR3
zT3{o`j9aa>xz@Vv1Ag=QuEwTrCqT8vr~98BRGq0=vYL~Ye#cr@*JKoZnZ;;DXgm0W
zkr*0EZ!#K2?3QYzd77|aGD$+ylge+Sr($kp(35a@6IEq0rDvBGsUl!wsyGbe8Divl
z7jz&4V@~3~vcT_4q=^t%a?S@2oFQawi&13CDCEf_Q;exbU6-)>dk~K(x%Ic5Hr_4S
zo(@zZ#)6I11c<zcf;l3$ngls=mi5Xmu)A0G{<k}QJO2+-o9r8J|Ghr<iM<YB?V6W^
ze0TL#Z$udcQ+c%!kz0+`IcibgZ`TB=uIAp?ibmiIZ<vBd>g*VQ!3rtpfbY)1mgG}i
zK)Iy1UxIipe#n3fY^kH6A!`^nK41@LE==RMB^DV?2YXxLkCpt<LZQYrS$x~~VfJ3T
zmy|P!Jd<>pMzuY*rNJ!Ko%TehQ9h5C=Gt_j9%E&C##eKT3(1JD0ovYPis#zG^!RZ8
zziIZZxBh)ele01D-^ri6ZW3@E+aBx1d{XBs+_heR@QdYs8Mln}`X`5k859=)^KExy
z=Hmr#?PQ)o?0ett^j+(IhMSbtL9Q^%w+An4th$HCsJq;qr8BL&@7N!cC0D(d{X3?c
zKepZ;Zogj1{`l<YpNjT3=ciup1K$rd7K&OZc}K3lY4`4Qc05~MYyNkPFu+I#(N=r=
zPy0Op`RBhTutqU~)tx1HH^1+E1`cEUkD2h_OxM1^jxix!0>9=v{JZz^c8?)Wvnr2R
z`n{(V%Rq}Q_*K)+hEsnBGw-Xa{#U#`upaaQ(rD$a=EeBEsVz?bwf=iN6|N_z+@(I>
zr9Cfn7D9BM%jKVT$mNf1<MnM0KV#)mDQup6<Vn%~sN}w^9rm@KC-c(1_27j63P9A~
ztnE|T2BLifEuCL?%+Cz`5}cRmoI!(3)UNSSH{k6{5n(eSCg_A#&?Q)DniRW=61|Kw
z^)0hM=k<QQ%^#T#jCy}UgWgAWSD)$IxCZtNFzC=-kJa$K5bJ<BmHt0yx2|@4bb{n<
ziVH|Qkz}ZJ69?}~Q|T(HIvo()zz>_5bKiG&_cAM%?}1y`xso@<gN^Kp5We<%0W&ac
zHoorFFRP#m&hR(!A0|)5(Jx;+fRfr`<q5UJR~RL?A~*Y@23_(m_2o$=^*)3zGRGoF
z=LF*<wf6WyT?CnqwE;yUSv<`r_C9-DKWivE-yvsRMBF&<J*gxQesekp)$(U`Dp_CA
zqC1BQcd;q?Zv%?_Z;Jh!lvF3r#UVRakIVN9;a)xY<x#H{<N4Q_MS`U?@Ws$;B>fP9
znk4hHh#%?hEEj87zXhz{=C#<`zvWjvSMUD5<PtpATW_n{?)U<2yS*QUe;3}h(pOvj
zI)ANR{!vx#B|xDSe-A&e<mFp?Joo%w=4avb+E?sf8mr}8nqQA&wJXs|Vz-Mtssf91
z&O*KAM)~PaHS7FCQy;n6`XQ}Ua}o{HJ4<VBcy$uc59<83y>F^lmu3DB&OMF4w`;X8
zu_!2S9uZJh`G34SpYP0i;?9;rc})d93PV|kld%3TKe^p|fgSJ7zY1Kn^)0))w)y%O
z0sa05^Uwo;Pc(SbAPr{bGz_byZ+ax<+z=Q)qJ_O0Undu0zS~T_D>7EQTCkgQ82XFQ
zW%|+S<pmxe{wK4!(HT4ZdyndKZlXs0YRvL;$VkAAXK5qxkND=h`5so8KOS})4KVdO
zXfIIy79K>eTLzp+eBpEcFvdm%`Q*T)wJs{uFfiZv|GRn+z!+r+LIwbUJPQDT{lBXR
zjI=JM&Mu6Mw5<PU;b59a*8WJs&HXP*>t#?SyGT@O6@4}M!nL~`d&$jvi<cW3157rN
zG}=1K<Y}nQjY-gH!uAabg!4GZS}s~N0O!%13Hez6xVMT(Q&g#Kn-^Ck!&2L1rCR;&
ztJmv9U9yEJ_``syXxW9_%%3F>Pop@{b6>8<(SM&~H|Vu|HjvYe#tA9(IA?SiC)v+`
z^p{wz-~P4-h=KWWGjhLw@9g>Q_v(xF<K@RIll$Oh?TKd==JF&A+pqO$apFA2N1itq
zCt>Lv4L}2Yn92OqVzqr-&wugmF6ZotUx7V4F}fKYKTZ$X+22EwDa)*EgL{4B;>C;G
z<Rn}VWB5z+Fyw+rN{vPGO?OOJiF@vwfc22IlFuLbFMfQt%yW&UBr^F!v7=rt%yWow
zV`~T%c4%E+6c2XMH-3?$^f}D`ZOFRYgJ=#%*3iuNa*P{GYlY`kbp*9cix{~)N?zV4
z%i&s8#Jb@=F(91cIeYx5A&|>@ejp$NJVxima|RXsGAd*aJ2@18KBYq%0r0quH4lkj
z#xh7nuC=#LZh9gCfM7EJxr>`SuU{iU?{0H0qPZV7{A?>`Gm!8(uOseEvxT1}W^h>-
zW;;}L+RvW_r%1rIeHYHbc#5_EH|v?~R3kru_uKFllE)F4Tz0|}tmqW)z$5l&EB7Yh
z^K3QlF$O1zLJpQc26(KLZ7UphYkpmupi1t&$uZ?qLJT9zQeab?e7rLEpAgGU)j(is
zi)pVpi998xIza<%8MP%2X-bLTxst*lL+F_0{t!ApsV2%``gt5jq4>Xu_iBy;YAiFE
zREZXZ`)vq_#p-&G;_HU>jTvI@?!qf4eb|KdT7rDBnqD$fKKv!L0(}tmFlEr$RX*76
zt=hq5klLi`5Hn9CBta4|1U>$`ecLjR;L%nX)T+J)8ZHbvnjwbr)nQDHXj+wC*lpm?
zT{fK5T^u<YTp!pk&bpT0Y^U!?|GRMJOci`*{=&)|GR3S-IWVuVBO5U%1VmZYwG=Su
zq1b<+{<iOxvcl-sjzDw~KO_|$76mui+~l2dvDR6Zxkd;LgtN^pOQJup#V|e&64due
zG(5zPK{B^IF}pD?&*9<}Mh`LHY;QZ*Ldt8Al#&4WQq!0Z=p@0tL9jqyNXl#a@B6$s
zbgo>T#UxO{tdNv11J~0g0dnX)FrXP0O$jsxc}(|m$5IUudWZxfabc9GG+A%~tOG?C
z3i)<JL?6xPIfDi|26(2@P9vzot-k6a!L<Zc20(&_IRk&hq;wwqWAGnki*(G{>iqW~
zh0R6$|5(CZHq&ny)CGyquoz!Y470`z@|Z6%k7)|zCLzWtL2X2VV?<_diHbojWbdT?
zh~Wh?(EKENqsB@;I!-kLLirLvZ2;e)YB_btHQ`tex6GJqpud%I&W{DnysZm@?;ats
z-bc`eIuT`~4?1y=?UM1D7Ijo)I&G*79y})unMN_7A%boK^tf-eXdWp>NG1!Z_-e#-
zoZvr&V@O~3(IKd6Y)#r#PikbT!0mFfaZmRqa$o^I7ENZTRv`U~L4<hd0Sw=SJr}Ac
z1vdX_1UBFE&WqQ2lPUt;Tfl{i-XO)R!$oUqjkvPMTsz&{#SUyk@>=<WG0*xGgeN<`
zP#%Xn#6hKkDgubnCV~iLa}+XO>u*3lJ|vrTZ^H{)D(p#8CvMWj+pKBBU`V(_oR@?B
zHVM{n<EH_HC=W+dgp;=k*JZTmCkibn=f1mU6>3bl2v79@O<GL@1Oy%x*~Gd`Gff8v
zR%R!cXB^B%CltVnp)dtf7T}OC)ZQ7kB{F|qBM3eTl>nMwdRU-`nj4Zi;VuoKeRZ5)
z|2E*x_;1&4t--dXGP`S--=^j8qRf6Gu6XKD;SJpEGlS5Rl#h+k-#i_`@8xcY)5q!I
zZ;9LM=j7w(Z;{i_&hF;;F<XaSZw>MDkmRP6%HbwDu&BiLqsdiwqWK|>funzL!)V7R
zlSs9^8oqmKz;8P6Y@c9U>!reFhLd3`2LU*drcBKOTYp++1rO<(Y39tnz7Z<vojOOn
zs(np;e`1-@6CkS)E3%R9w397Gb^^LsgAqOXwrp`7Lpzn5wz~b=V7!NNKF?R>&lJNB
z<fElHTnf3+^GJ0w+LQis<5}4vtjqECzBRdBlq0O*4XM{c*2}2B=5$rt$J^9*TC*uf
zcv5g7)_u%TQE)vV8Wm2H!nMhk;w9b#6hG;Zxk{^xnqCv(Fb-yz=zIze;Al_08{8xS
zq39T3jJRaXqV~zS*TcXWr`|D`h&Qke9swrc-!2VcDES9i>*Nke6M|tWdN9_NC5Inj
zks%a!6^0Sn_RJd}=C)rWSO$fNAR;O_e(p;aQR5mr3WK`8bf_Cz20N}hTmJwNv<S#x
z!vnU*=VN8ILX58NuN@gpXtuByK!E!M@=7Z2wAaY;df61TbRj{vF84SeW&vQG0C^4m
zj|Bn|J<!><nISi<ee-sr)7R(okljB1=4~(6&%@ctCcoGJ?Q&-8=&hRzo>RNYP_C|y
z2E0sAuNk6)^tjiEXoURX-04K0o;hg@16=sHq#01uV7ladlHyS<B3ZC3s6Z?B%7=0L
zl*WW)l_aO>JS?n0w8K>qB%)a3DrJsP_Mb#S^m+aXAXr}ktVX=nyJVK?2U|6y8BMcX
zGew+Y#m?}Y#Ub~KSFG(-ltXHp6v(ujvCmS#`SHgSSgp4EoVcab1aPX-#Df*{)Vd<I
z|IvIGH*e;dHn>@+dg94Hrg6?05)6v40t3v;ItkPqg1~^3Owwtj)Nwl9;F(7P-;S3-
zJ|mDdq$U0%OLl($2+A1nS25rO2Xi>i%+Y3B_}6imVcirEyyk}cQ);8xj$OI1?O8dv
zu&s@aG@>T_AhFH9@N2XF^0o?e;IJHS9|$Z=G`bXKE|@&6RwE$o*a4!XVGP63&MZD8
zc;?FX^yNX~Xz&Cxrl27kHsY1Zbgxpy2ypTZi%BVoWJ!PSBtr2O4~SX|1i@#3lqee)
z@*v&Z(`-0obcjY%q8U7fb6bY_mh6T=CXt#-QA2kg?t%l?HXG%sBnqP#xC{T(BO{xE
zeCoER=sMx$Z$Fqo+;-+vc{Y$aGVpC)*|PmSv8^t0(wrR)_tdBg*FJ!zZ99jQ2|Zci
z7?+FT=YKsl&SeZ4K5QC>6E3Sr>EbpGa#m?_>BY5RQL?AJF;OhvQ$)m(EA)s6sRMX0
zVJQqZjRF-&!1tt8I$joE>IUbg<pI>~T}nQHQ|z!<b~S9bi2S3*>6<vMAn9iNkp!xV
z)Sxnt_)xbCq06PcnFEQR<<6=#VGptqQ;}T6BF6>z5ZD?QBnZ6L)>%JQ)(OaGVM118
zEm;35s*%E}hdDsSvqcf29yS*X_%V#WJK4|S$>m{0ejEU$U|VGFUFe?Fmk0b?>~<cY
zVksu&WeT@onEaI?%>UzafMpa(C(gP2DW*yW6|<_IQ85&I+=w*-Jc93caCyll5-6o2
zz#GD3+2Yi7@M96h4(Pzht3`~X<O05Ow`;0;`}+p9ufm}-71HY#X5v_L{-SNxyhn}w
zaQytHI0ZpJ7){zzo-}o3Jr(Fxh;PvL*OXMh!kNS_hk)}y3r~d9vYKUhv(OL&de4U9
zIJB`ap@>!?HqOzmisr4nJF{^7q#U~xHNbLB;~LOz3l9eJM#2(Vn~2yS7s*~UThYqr
zp`50I1j*anWFex|d6r3J5YS2k4YG3!CBpUt;s*xB5a(dN5HRBw^rgx^S4yn=(Si^$
zjjh9tL|qec+72Q|QJY9`$i?<E68LCYHB!VT@ScZvj0Q*sDP}@(JM7JQpG1SSL^MM-
zQ8rbcClo|TGLM)yfo=^|9!`-_J<zp@7U&5%$*+!#hn~krrSKGE*L7y=B)L0djb$~+
zK!GQQ&|C!x=@NIY%Ktw8>0~~T$zkKI9zAGvs6#LsqN8=)obPnVdb6OUinbwg@$Diu
zD=uw?(uUccqW;SkUJUtg;bD#r46mQUQ(HxrM%3@{E{Zm+_^d7*Vm~4Nn(MQZ&Oil(
zz$|K%cZ2`TNNRwD-wR>t`iJ35BsE%B+&!j8N24j2&=chY<rtqz-~EQg3LQlWY7U7C
z8X02wmu=8wHrptYVa3+b%;4Zs+8Th4zaWg_UmmRFc;WXg3@vg*^zc|Vhefy|K@D02
z`Zy?72dE5=eJK2bz#53_yn}yit`;j#kcgt>IC<`j$Ov?W=xQqmp3^c($>OJLJR`_!
zH`(5jS4EkTxDYCKEm=Wa7R~=46?VvjtL`_zR3O95TPP^fHpB%RWt4A~D33)*h|2N%
zc-=K^i|IJ9L7+y?F*&1*@V9}IT~z}!#IOYSYjBEH#D7^M87dmm<r>xpv0xJMEZ#zW
z+_-Yz5aapoL7#efC<L{0&}=EE?XWf4O^ICWgKoF=F~SOyF+0&Y9p`#TzP^T$Dj#}P
zVXNQ?o26&t&Ex-3h1mk%>@I%@nAl^1`<3mmhcmMS2TO?)vsB+JY2D$CZfEOG;G^o`
zHP9(qGoYRVKwb?yHnbs+a@uq?Us%=oE>Fb53Bq`ACD7)g%y?{nr1C9~qE7Z)^9qa9
zq3Sjod(>XvvhX!ffuLm<qYhhMp9?WS5|CNz4VPIj0AxY6c^{?#k2zuzn*$LmNo6Ze
zf}9Z;z?%g5)0<q=mqG|^gpUIm;JQ@#b==Sl9+20Qo%(Ki`LwEo(W)w(AoSev0R_?b
zV1o4c1B5hO<J+>~TBG9=S`C{LrY)P8{zRPFLd48<RGMwuWH>6%8VGSSF5qh;u*GMU
zT^!$JUbbQ3jYZ$CL(bqg;B*p~<%^8%8C)*idsHB{#pb%S&Dd1Sb$F-g8fA^`TXl<Z
zCKz)ei1rVf6v5Y^=@cKip=+bq403%MHoHt;;gl=BmT|xo;L&@y@I_#XnrOCNB(j1=
zlZ(dn#MO1oyCU+|5!E2k7|2VBl_!ht8z5lq1xRcg+hX~37kkv1Y=1t>U1ef$x__jf
zDIV%?O>0PEx;C~K6$Aq#Nl*|3l&udIhYZ6$uM%Ky4^cL=><x*qYfD2zy471d@{m-M
zYM~nl!NWh$6gkrDmV{Qc{W_6Y>N@w%43IB0FnW7yv6ZzgwVa4}0z)2!(0voU2eEwn
z3t;`;UNJP)BabZ&Yp{^OY94}}=qUu$2f5Sww39@8j87y^_Il4j@F)F#<o|sx<^M*O
zhW-7}`+t1Q{eANL|CQ(ee#7(so$dX6r~kdz`~QvQ|K8NEkO0=5-;ELg#@{G$-`i7w
zD6L4yO#AY@N#5hZ##1*wQ(71f^=HJdc0)$X7Kqw!tc3bozGLcIvLZeHtei&3+Gh=k
zqbiZzcTpgA*oemH`|j(f{IFFtyHm`0(kaxCLBvftL%9<(_OBWImOJaF77AtZf&M4^
znY%`g7Y*yKnh0URLS?l%uMS`VzhlW_&QYVuXh^xW%I19)VxM?44>=s=<Y(+lb{(8n
zv*=s;-YSWo0w5*zB9b&L5gEu@$N=uo)LN;TmEf6OlC&7!!-Q~MCGD!m0tV4ti%6K4
z^LVI->a%Dng^LQ()Q-@z!FQ*!dRojD?sIa}k@LvxC{R3gq*R<NVOi8TL=vHYnYvj{
zCyzhqkc)0rN=7~lzcZ&?ZtW{~o5EDl*gP$&73-)qAE<p@f3ktJVBN}(b_}%kp%qmG
z_<@+``Ydi@V_K4=L=FHe3?{V%(=IhVScD+nV*DF*WVZ5AM0sH!>C^}sxk+yNj%R%p
z<!Pv5yr-g>22=|XHckwui*p4cjzz<KU7ZMyI&P%6P^oH$AR@=ssA(E-Y76>LxnA0`
zlBJI{VrXt<JrXwCto|<$X2?a&zbindq!$1ECBLXdMFCUOg;SGjp9tE9#<s5x=tvNB
z5EAwW)mY^HD(cW0iudU-h<esq;2Z!AYg>__S;J@IWY{hecdhRpCOUdQ5}NxmKhm9o
zbXROmN-v<$rN$dAcPed`cH%Q!*$Y_!E^u@t6-*aX3{ciQiB<ikx-+oF?`)aT;NeSF
zpy(PLB06LE5U*sEg?pOJc^OZhkRENi415-pffuvY<r*;7)Kuk?rkr>#8_%W6q)bSr
z6qr**JCL81--{p<IvUn%3<1Y*pv)z!S{=z?=TRn27#Oh}0n6L0Hsgr<jn^GPKnc%N
zXXdl7^fn9q%m+67L(+T~!0n+2)Wf&tXmu6%WJ(jYHtkPl#dwGoUmZkwx&U+46P-v*
zE|<H1HNnP0#V=(%Z6NEtek`J`DqE*mUL>3`v&n(i2ss=mEA8ys%~<O5Td`4z_vI}8
zSSSk_11>&3$;zdJmr0>>zO7PsbQfhl15TbIW)e2BnXG_i|F=ca94UwXBmFMFEK)mF
zD=|aY{JdN^H{MRb)$^93D+e7K)6|_ZJ|gx<j2GFb8Nn?M(?X%E_G^4n!j53ro#JI_
zTd;l9UP6cmXFLQ#q65d4%J9;qt@M$GuAkN_GoEVWhYqI(`MBI)3V015zY6F`J~BI3
zgE3?56N|M576iA!phj!d(IA8afUK@sd!TZ@^DfYN|CZ@BW)mn>*aRKxBWeW{Tx&Ub
z%@z?0rtpmme45v0S(rzoy3=%Ar_|+lP93K~GRxWsI<8GuWzXH<IBcgHM1BS4MqHZ%
zfIZh-imWz%1BF=q+8hIP>_QNC92$as?2UrNLnEn6{&0Dv_H+gpg)Wvx1t+H4Pgs+v
z`0WZklJ`#ma@CFI$X^mb2ilc6(-C8W%B>m99Du|RJX`v<VU;Rd<8j4siYy=-aK=s1
zn?Sb|32(ZHCbwmz2xDdy!=Rc0WR}M$hH%XT-3_;~4<+h!TK|N)a4M8Dobxb<%aPEe
z)+RUwgaHw`inmDvx{>v!Jrp~?aNVliGTYB*<+@3wphZXx_R{bMQ2B*RB_UofJA}o6
zbyx&<(6kh=Zwrr6{S9v&pFcq#RUQA5u|Gz(7gPzbw-SC&mcnhD!Z_554#+JsA|0e8
zOpXxDIGJ@%d+CCn1dLVbo`s^fFmFam^)gZ^7dhNRVP}y}`RSeZQBhm%chaO-6lQBi
z9Qya)d38nhC1Qn5XSQ`~XXIvnS%9uyWS`7!Xnh9kzFkijigS2+oZ#4l6HJ49=mJpF
zhD_=fUQMc0SpiTNr3sZUHRo2Bpv6*tnpek**$~YF%5g_<C;LTrnBhz<-^=#ktE^x)
z0sk&sizS&gNR!JHk@x*jPFk%1m`}<q-a<&Ta_=dm8t49ZcXPA*9shd2xAQme*Abdq
zPZ$mgJ%2BE>t3(F<L6AH?`v5vb_Ah{z$dLA<yU~$mK#uG;p@WHQL+8pZZGGLLBH4A
zuHVNk#HlO7F<<Z&_Fox$dFIk!<DHy-??)M*9L>%Fz=ovdf-%o9KX+i+kXq{YD4n(T
z43Wr*67ITEFY4Jt(L+PSqsBC9JySzBi?!z<l_HbKt3Xdkc^k%nl;&Y$)RCA=)fTYP
z;9><&U%r#YSKj2B?H854-KMr{e?epa<I0gYO*q@DWJw<O?-Zi5@VERK!PUe)X*4Ip
z*ZxTjwQQdaAX~~BbTmO3RSFck_lW8Mp&Bd_y-Q=?#-ZP>c|^VqrFG2!RF0-I2T&6V
zuSsOS^`bFN8rPZz7+1<pCeH(B{>A+eYv_;%Z`~_-eN+__KSZsRJ{8qsJG7r%!$(WO
zC<NMJh7-}=b}5j6*G#;rEZAA+caHhDXk@P`W9EG}aG}{%vyGtk6gYN;QA6I%70;w{
z?NqkQfTZEH4j=kj>Boiny0QOHoWk6kNGhCY9YmHtWa1W>IS+XC-WAY*DQ7Wj^z{nz
znfdx+tS_(a;XnBFdw$4gQjq%@0qw*Z=~mI=|7d`Zax)NUe(glprCrnN4w`lZv?Dd@
zN1CkBj8`M|r~|w~sb9!v3t|WVY%uw?Ol0e2<PzdC)qej&V9)4BDgpM}-g|8|hb<{|
z?}xhOosc0dk}e~<I%89k>?`HEgshnro-TbrV|X~gMx;ruw5i|sE5}aKH&u&3J_k~T
z%AoSA<+EtX*KV&cP}+Y62pt6E=V+q})-(<DH{NVG`P#Yh3Lragwbrh=MzW4~O%a2N
z%P$WK-qmt}fORi5T{*}c&oQLshJ}x2Q*B*=;>`(~ZOC(|*-g6T)#*k^=Y95fy8gi5
zEVHmM+~efgoJYglF7jR6tiSO=Rwahp&u7y-Ob~yduA<}aGp9p;N8(%H6O_&RvU^dn
z7xT21ez_GrmSfsI`u0!%BG_k;8v_nvjM~YvQD#+NyREoGYe1ML+Hsn$r*JQ+niE&q
zO`f|-wGr;$dDOwe_7#|Q<R-(DbOGgLDsU&eEf*XCQOet*#1Uhn1#f(JhLZoM1z)ah
z^33|(BRl(qE{@_f16u6v;jI#9Vzadqk(Qw)DjM%iv_s{2yhCS?ypUp_zL9lSA0%kY
z53<zy3u3g|``PJRTrR)g*9&|34`C)*u5(kK`UH7cSW||4PPIqo_u34hBn8()bow$f
zCxoCc8b}UA5v+8@!MOf#^hU?S-P`Fo&riHtG+=@CC<=st$gn<yD1ax6*VPc?_W7&)
z%IYA%kI{ZLggS?Z|IL};i?F9RrX23$8RTb3=pd$$TG${$-^k3?K9X8Pht3wve+`6v
zY<HHb^hx`Rq~{$pJ)w_Xtv@kQS!DdgBBS#v^B$;yOY)uciG?9G%v+UHBXC+5q@_<7
zO}jJH=1LU3LpD)Fo;#SSAYR>?sKEHuW@;3+Il2!CivD3QL-Wft#HzD@HN#gjg-RER
zZk`j^`wJThoaka@Ke3?fg>*DHub#_5+-@O1in;i*pZpa~a^sjOC>suXEEecr_l!Gc
zo`h{h9@Lk>DlbT1W27YGawdbNCAM64Nf}k;21NBX9&m~stBNToL`l`v_h*qL%}%~d
zpO8l&@$(!wakJa=)$Buip=Y#blsbieX#`~`;6Z<p>L4K0QGeqAsXUd6*9%0ZUaX~r
zjZPKsY-hf7yO!raInzt9<t5VQ$8q~vH<Pq0UntN{@<QJ!*OE~3GRG#YX*{KG#Zot~
zQy1n;%SYz)zdyHY=H5!%0F(M&%^CKljNa&0JzlktUL7?F>tCi$r4?RpC2&vy7h$(|
z#(SP{I6r0CyY_9i3yHpWFPM9RA$O<&a9rhafz8*P&)FHCZ*jo=?B}pUe)bLIvT#SG
zoJG+uSzysKXqG&$uO(o?YwJ!!5N+XNXty?Bh<Y{ed|luGNoJC}epV4=V+&*3cW`(7
zdlCc@^Hrd9yFIK2Sh*Y*vX73Qu<ITMPT=yt^sJkImt5*5##$vP$o1%66&|;_u%Oq{
zJW?v)B`A)7Biu0w;WV>?^ZIv4$M<h&QPjcqlf%bkLH2oN=Q+xBhdzD~mR(MqDidp|
zjB069hn7>@K(5Xd8pMDGJpgn86MsPxQv*VC05T~NTjm8{X}0bxTrmD>p<xI2C;%QE
z99E5k+sUfBf0xxSC>(_PA1_2ZT8=`4`fYkwrqur6e%19vjzv{kI1*;A2fq5EutzPL
zEb?jYtC|8<_13gR)@WVC<!tKB+6qOjt44}9Q%8+>O<R98lA4psx@RLSv5EZ)COH0d
zZS3|FIS)Q6q64_rb(V!ro^Kw*4JUx6DTh+YO^H>qm>e%^px;qm*V=5jP?5in?^{L-
zGo}(FMESa5WOFA{DQR6-^ECrR_daFL%C+XOi?fz)vzAfW-L%cEz$(vNbH8kf<8nB(
zK&^fa1jE{H>&V9VqpFptrBE<wCI|u{#}fljArg{TA3I~w5RxYm^+8SOeCK0DyDat(
zX>sozn+c7|#J-k-()M06F*!@7%C<AryJ1!y$6GFSDV?e)?5>NW9V46_6$y?pvZ53-
z)<~rjlLrN&@EJddGq;=&K!8|nqkvEAAIL#6nQB@l6%n=jC2Q`=l&WvdEaJfWBU{<W
zSW9pV8W$c=gEVE2FJEvo=#+;Df=c5s>3d4xdX9o~v;_9!N?w^Y{3F@6Jr`opZ&xP7
zJr;;8)el0cs?}D0LciY(r|QLJ4J`#3(8n0`&JrM9W{VQ`MvNZP7^?r;=IUw<e!f3X
zr+pD7FN~x!r@*8gBHy_S%CUd|<n_nlqUZteQ8Zt~7#3kb;<tGsv}B{SrY=x7FyRk3
z1bM!|4mD#(+!3y>hr43xvr4;S1$o)73CS2t#FB5IUEf__1;1=W%rkgc{>;H7sse$z
z-1H1DxznzIlx}I$k6ZHUmYb$XWEo}HWs5&-w#;M=7FpCS8j5e%7dzw09O%ttZD`$*
zESN75{)UqZb>7d?X;Pd5c0-2@<zMJpFb981G3rtuWkmmrYI*Y&Uiz%|CQZ`)<M+qi
zhlb@T`cv+z_*OG&tYzaS+!{0;b!m+XxnuopQ)XaQ)CD(Pj>Pjj>EQ(`)btTVDB>q(
zpOJ-lxy&ADmy~Q1;ATQ3v9ga3KeA~poZcmn`v}F<f|z6|d&kw_5;~y&!u;*R)DH&;
za#o(1Gc?qpwPIr4%n+5IOAiP{LjT5}wguA0F7H`T`HOklRv}Ow>0T6rH<{ggXf#Vg
zJPaiw?ZE7kfC`!|WN2;1<WgipY!vc+fbWwFDc4+tA+2?$T|@pD*$_F(Ddt&-FJP3S
z+6{Bt1Z+2)aw$7x$=vB=j2WE+&F-pzvq|`(d-2bXfHomWSQzMF#5JIfLOJ^3#EO$I
zl8cC?1x!OQ#F1lS6PI{B4RC1fk{oTlun1m^mCLaOmKa;jek~5%L-KsN?{Ob&fza4T
z2~oKSN}azKku2I~N7(w9YJjYm@Ls~JCNi!G?l2Zdq&vDY=*^zjO-#i`&p0uX;j!Yt
zSaHB<M+Bu8D$;<UtF*u}??$k-eSwuBTF4&><X{lQafw8}QyR*@D4~bmJ^`7`_OL;<
zX1~E5mdgjzt{C!wgZUUNzD%7BR>0MCN2)-@G9E6(&*lJvL9RF8fk)KM#cyM#jOhlr
ztwLI5uxmtIdW#`B_3R`9)f1q9Zkeyu>J<yCh0Eqt*No0_F|W?aHrhy4+0It%mpO-;
zJLQHeYbv=qCFV(&95$2%^bpk2jQ%X<4>4WKF&{SrO<TIFJ-X7urvu1GX2AKM>nG&X
z4(8r|5Q)w|dli$dOFg*9WK{Ks=}VRVttJzbN0m*Pf((s0DYuxg8cv<q8)P3V@%pS;
zqz)r7QRy@dd$O*XdDI_VTk*m4bZp(|datSfKG%@#T1!Hwb<~rhRHQ-<X^gAy;}ekk
z6O2z#IgEBGBmp@}(~L3YjF=_h$V+Lx>+Pe;KKz+Ej9dy4l@*OXNyS%Cpr|Cgj6WWc
z|C&2$o-EZV*4{#vP|P!%+_C^rTHTQHGq4a`!<OC$&;%;hgKM@OL?oQi=pgl>>B^=r
zy};b;dn^)wv)Y*t5J-D%k6MUl$$80|5$Op-U7z!MJU11~+u?iv=A8K^uxoDr=TG`b
zpNPt#GfLA8l6iYI>l)E#;fBa2F}n%zVZsw#4TidKL2Wj6Wtw}Gh=PDX`HSSKVY?Cn
zTp=OW7ifEZ2n|QT;LU?>P(yKAjHYDlHc~$9%*!HBd8;7;PkLoBP0|V9^DG%0TW(vR
z4Gl2*kUDv)YIv}UdJ5;)(KQE7QTsFm6jE>M!n1r#Wy#Np5tC~`2*}&pbbss{Y}uGq
zzYu}7Y>GV)Kn6s@eqcWs3baznyrtzFt~Fw+iSQ5`C9!ozockuWQe=lje54){;B@=d
zjL61sI)bdHL}?VI41&5d3W&l#-i?F76nXpCV?n8f6X9f~CkfWN4$K~rOz^x@P>$J0
z#UgLWQ<Re9V=gHrW<B$dOQWov@oRQbhtEWlW`*Z*vB(tfoU+v2*Kw9YSEydZsIJWH
zmjsoSRHaC(_VV&}nXVmG8a_a~FWep>;c}Twl_wsS@)lLNz5^F`o@0KNm^sS4dgYV1
z5E|W%0KMc2y`7qnR2v;dnPUo`PxBNWea8-czR+Y^-mOmDK|xJ|E(w1op4!rM#HM0p
zfP^&EsxV1QR)kk&5jj=zCPeu&1EKV}1zeIU96;&Pk)U*j_w0`2x2Jk61c-i$Yi>3q
zK`-%$sw=IrwbIwAZL{7|1SX9cPLLg+HjR|DaBT58$p2IN)`a{y_ySRadta?vV{6q2
zy4%gVX4md!VfA-qh7gpb9@VK|SulvrhD`7_Jz=1*uN(}NgLEhPh=)ar<eF`+@Rr+Q
zR=-9wcB^Q#ws5poBe-|DhLXKlVC@M)w_3Pf-0>GKqZ{^UeN|-pC;Hwk7I+ONi<T4m
zix`To!lOOuC<jn~K7wI9@E4N9T}D;P4gYa>#1^gtuC1+Mi4u@7xJFIW8IugzdI2fr
z;#E0Kpo93MzbZl|O3QM05}a_ZNC!PN(i;FSNpdOWN%UrSEv@F?VL?3>1_GpnG%gbC
z;_4A>R2$<^gM=x(M43Fs#IlG6OHA4F02N`_BC;(RR8<nC1PM>~gvp3<HCy&S!4tJ3
z_rPwdUMg>(odu5_J_|K4VEuB<gP^quebTVC*`pvvm2<5Y1lW&eMR)A~N-+sAR+(<X
z3+I@dnJ%oYP0P7@IAMbh^Eb2gy}vfgS8L2CmlRu>4z8@37j6M*Tf%h%_!ACavd*AI
z=nEBt&nawkKI5qGZ;WQr%8$cNphS77KUN;TQ~71I{i(1_)DD)J+eP&2A;Q$&_X97!
z3=l`{4hE2WJ$}?vJ>a3weNY!m<yk9Hv=M}?2TAq8Q8xufTlynvclY&PykW0?xumU`
z4D)ZFL(PI1Zry9VbMfZpYpkJxNoX{oK}b*$Gv|RLVc+q$oKwfV197nr;2AvkEwSiZ
zivLmdA;l%myAyMmP!mvGkR8Lhm$>*J`2XHmR%3Xz-h}}Gpc)47Ut8J#Yi$3|md;Ge
z$jHF)KeqJ$b7$ENueJS_SliwU_30{p9m-Uc33r@rP4I0SL|2Eb{he)KcYq;93rkaH
znuKaeQpY;{tQ`(_Eq>=rv*jkoM#)TD<xb1SaAWTKTtw1OrN7}Ej|A*Qp@?KmqXx|a
z7#gSZ>2&5(-q68Cl%CLqjG4m3aX7>5S(tpx@t7FXQ0Kz*`8|c4xf~i(G4aw8aq$xI
zc}!hlyaT4GT#<`c&Tlx_(uK@5O8}<$9VOTITR7Ek@aJx*=W{mnQ?sgQwx@TO7Mn7M
zzU)bha-%yca(r<-3Uj2`0{AB<0E0(R0CyrLK91ne#exghtD@^?JcoH)@dUj8*8}%m
zVd2q`GDWgvd>RTfJ9acI49u_sRVtU>;!}qO&YZCh8$+IPzp5>n^`2zv8Dj4*$RB*6
zpL)Cetg$VF_M&5RzH_wuH!>DpsNEqegFraG2xw~rY)W=0=b;Z0Y%YdKmDzq&<z*J)
zoji%6nFR)*!SpQ*iBV&Z7DYi|Am)n5<U>-`NvqLEho&M+ae=y2hfzrOqRX4F@jRlw
zfF%zT8!sD2l9X)xfexSYc%cjzFB6YU#1^#O|JB$#1!)q6TcTy#HoI)wwrzLWUAAqz
z%Xa-`+xpAsvb(J5GZQ!NIdkX4oy?cqFW<`88L@Z9xAt0#g+xW^1@Os<heM<Q#?;?O
zjru{4!gvg#rlH1wC(V%d7#(%M($yj1X9?9$VMc>$R;$KiK=+$%C0f+_roa6iFcr`z
zrZ@7YL)@Xw1?|7mzj6Q3QipGeFmlS>5lnRns>ZfH-IQ}@uc;=>odbz*iXnN1Avu;@
zVh-0?Y>G46@a_Aj1UZnxSwSvu;bDbLF}67>ajJhr^gLHp1(PA7d0vDecuRrgC1y{!
zj-QdF@XWg`IU(hb414vT2kj|5s>lq``8Nk^KWPRM1uk4|ZA1<5;iTf@k%^IQgK>=k
z%huwerApa}aa#zzH5?hJTYgO|%b8MwGjUhdD$Tm=jUFDOQD#i+QVaml{@mx(&i>qa
z5EhCaSrIe3%M=L_?r@#v<t^>;;ch^@Dio7JFj1zMM?2cykO3E+SgVQ~)sPuT8@F1s
zTh4_&+vsM3l8}$1P?kD4KEYzqfnvxFjGH&)2a1qv39w6c_(r<6GYBgBW+eJ+4-ljG
z4!yD+c~9G1r&}<p4IXL^(xkXR%XN=^C*oVVZAE21qm#nZorhv%rBI6~JX$HDCHEte
z<@RU1Ubd)0Xi^-}oQEedsJ9ZVX}$62t3*rq5QooCNlyMwKPVTB;2_r<O{4Dx`OZGY
zsXEm__s^hvYq=%SScz4t!U>2zJcF<bU~FJqA4}9^xDyoj4ctB#yE(PMJi&z=d9<3c
zDvg#XD`j6x{`^(&G4+VWTc6z=xr57XvBW*<p8LtCJcjya-dq?yZNFLqSXOdQKJAP`
zAi@cHO6@VT8MMu6jQuU67ms#mdwEq}{!(J=QNjB?l15=7P=#)>$r@#^02`frAw%^M
z(tt(sNo_87ueUUemn5}O%)F@KgiA-2&F7CaV@m0$jGvarmtoQ84=!|SWLn@Nu_E7P
zPCK~d>Luu_NUQ1W0VQ5)Phko^({@xOhR2B`oWSMy&}t1IDUcY5M%)6iN{p&0Zjn`Z
z{<jL%Q{)O)g3CPOAlk-OXSA~|9_*%DonchuPCc6@1zJ5O)g8Q>iVa54nJg%t{FEz)
zRj~5*m|RXy-f?gSqaXYrlh#ZVl;<S@7NdYu5R*!nhdfGf0;>^f%bU^8_>F4~dx#n;
zikC>xIk?h{%hg|9^9NM8kiXWltE(YWE2b1aB};tv;v<W-Js?>ayy6$v7%nx7LJ{6x
z=pW$<Fb07^NowA#EKziHnv@1-k~MZ%&ko&C9Zi-6eC8Eg*HsZ%n6s1sI2DLxU4vY;
zJMVK04w0_;)8<-~b)l8evoTszV}oDPHf@b2*`}Erqb3b;)O>oee@+`vB0*=WUf~NV
zRKmpd6*P<!PR}J6cdCC+tH~+y-HxUxM5^ulbYhB<2jl|!Uk#U9m7FCK9t0^zGST0u
zlT6KLY&#=6<ISgLLt9{1HBB&F2KlUs-vr!0r*&doe9FM~nT(Kzwc<K1c>oYpU`v`?
z5-9kny04&Wi5ei4Sv_iSpTv@Q!|%)riL*Shv9^#)#4E*>=`jy2==8j?qf)m+$6g9^
zCC#Bik@fu+&-B5OBU!NHq3P5j&<~C%Mg79zvk@O0<2sNvGF0G}7G9U;tZZ(@b{;Am
zWtoqxJx4#`n8{txkZCRIC=bJpz&OvaunydkkdmrWcicn3N6Rw!ojOEaHoHJ9;3b&%
zP1IqFHjZxM0R`eWib2pnGS^^n_Vz|J58swTX5Qmw^!lzD*4l`PcKi)@Ke}5xO&J91
z9GD2`1>W6n;B0`G=VXlxn{0t1TL~pog?hextM?D1k7JSD5|MGS2ms^sRQSH@T7sRg
z&acCe(YgBCJjYt>oYkFzS}mnoyXsJ#lPww;dSvNbMQ0LcY#$l%9%>R;_kAjgcpVgZ
z+rPr7dp)&>9in8J!xnMyM=Boih&&{9gxXr=%mXqr$zSZHC@(Uc6^AqH!(4dy0~bVX
zBnc--&Z?g=BOm&p@%p*2n&OoB8^yfT25wF3aql^;pY^p`$an<;R<TD`LbD&SCtLDR
z6t1O1V!s95I-JrrBYDgLU%icOkg?!qD_V@JZEL~{hrzn^#iLrl)s?i-JUQr`_lR~*
z<eM99zKfsI!(B7pv3OxQw+J$I<qJX1)W5b^c#x)8Mur_Gce{}_I4~i$Ov>N3=^pjw
zke}-HjQrGgwR!b*BXkHZ1H3!L|F|?jy6QA=-Yh#h7EO2idl6aNYpl<7MB13LG7N+d
zYchh+##L3rrlwEp^X6Cd;h@LBGbH1wrOH>y!>(+TfI*e#8S|S0BtU3KK)y@(iIpgh
z2=%y8e<4!iIMu_0Vw6mINs4kS^26~-{0*SXPKtSZ*2SE%7A;`g4U9_ADp9k^E)8)`
z&BSwrjO?Z`Oo-u?b6&Fsb-Q@@0Ur)>{L}8Dtko58cLW=$;PC#w`NU!Oz!RzcdC^C}
zSQ{=93l4c>H5hZ>*dLT~3T`k2ijI8aI{#s)IdO70aFO<B$13_3A7USLJwm<kAno=-
zNKaz|gM<;!aD=*}pIOXlKw2{f;$4;n^6$j#dAYfzp+s7#4ej~1iP_-Ac=tQCs#7z>
z6Fbr!#d;3=19j+AtbT*&G9@pR;&iIfSa>W+u=-CR&S3p6SrnN*$`G8N+Mb`%V6=~u
z7E?wH#665>1GmZn1dO`d;Id?5_CyoXur!(I>cP_HEV4W*VNb6m!&T^$*#K@mw`tPx
z9o7l!TJx&E0|<_j^h#E8T~p)`A~XIGLyAVqfZa**#Pt)j@cjtLdwXrG@3e|5`YEEM
ztAZAX8~qsh+qEgr4`9lUl#J0~r8A4`yj~_Xh%NDHQaZHUu3mvaTRQQo;}?v7f%_wH
z^J#X#Gr+-K-~4h~q>H5c!`x_9?9#8J!T0v<8{i-+>53t;_GX!U@Fgi4oP*dKBJEJ9
z@p~1|4FWXh7o5mKPDPe|Ti%$H!^9e6a&ub6FUCA6x1hv9pgU~QR-OD-p9?zJt+?#M
zpTlbj7m>s6OLM%^G7x=1-cABOh618af|{dBpo~c_t<JZgI&imzs4kd&CV6{^j4=q3
zz2B&9^;^I1F2&vbaSg!Ne`1SN7?K`LYx$~#+apZnWRiT3fQ?)XV%3k8;9ySm3;qai
zp<B1DIgNzH*>bWGDHM1Nk27q3l)Pybg^lw}pW#Nr#Qc0?#u=qoNS;K3%(5j6F8tUj
zfSnEf&~~;ca;K2no@esYDk5Hn3?>(u!6X%giK<8OyuccA;B2$<ndHd|G~H20(3~a9
znl(ylVo9!g_07SF0<hgBKkH!=(+Q0iHSpOEa1aBOw6!x}uQ_lk@u>t@vGfX61*CBH
zI{am1I|Yp147}xVPrt%L`0B?}dr|BH5>NzxJ1%TYY_vw6h&>|bsF1UxtF{d^Np}{;
z<OULc`W}82I{U<ol(cyAqkL0y`ZzMv-)#0s8Uli{GsQI{QBAuwtO-ZL>A<Pei~ojg
zxf+<fTTKh_>vyPre>lc%+dVp7H#smQATPfDtYcpI*@LtLI3YAZ#zW;{98e~A(iIPf
zyHyXRYRYoIE^DDslk)8zyqD<dzBqZzhwAw1*pP0^piiQ=yUQ55$U#x%^0*OU1KGzZ
z9c^Z~XhuYVyNJQGyCT~4GPwqj<AmxlUF*eE%gZtE2U!E|R+`^7OG5erellFLU_DLc
z2hag$sCQ=US&mR5z%T8(tI5vIU41<p+dP4~ofFHTz`yAg4?dd&&$kZ1D`&^9L*uKk
zR{!R^qlw%Oo^`LM{>^UC(39@a@!i%3fV+-m=+mPQe@iuR%_LCTR&-DjjFhQP*Dj-0
zYpsT^dKoGRug>9HdXa0bJ<w(y;y9`xpYxp(;`~>)kaz%4z)47+%ybbgIMWKRd(K=q
zPr&cn(%;q7aPo)S3;_Q<sxaN}xBqc#>&Mc-Dmf#FJ^tHl+1qH$+ctIVfN)3Ul#IL?
zNFj_Yn6ThR;p+C7xDHFWF%ASGaEq^Wr2MRM?{|rXiKKAqqKb_qJeQDom?Q|4i7h)=
zm!wJ=D#)hB9qznjTW!pd!o=~;R+`KCn}%|bIkn4O(&J*@I(+b!JEAo6Z=@>WO>oGm
zw+$#7$Hay=ri6<OVef5N6>`>-dX%KnS)YD>7<FV!dIwFKDm*OQS&{4LDXqJXdyyB9
z);MC282u#m5s3bbv(o;19MEuOlEb<OtowFxz3(36RLFw4iDMT>7w5UnUphuFil$71
z7@Oj$E*I9*={f5AlfL{YU39m1HRjJuurTO1e<eU&B>px9;#+!|%^=95M->7l8n8wi
z8Iw=;A7UAJr3?hc4GD235U8ajRldjPm0g_7JAho!!*s!mq_9V!?D$Jv{=X}&{U$3P
z2FXDyb{6xlKqd5kQN*sfz9!it4#QDM39)52<rqVDN*%KVd|_m^XIA2GB|IIcp1PEO
zB7lPu%<-5%HdUq@<Eg&5J*Hf-S-@IBxBwnk^$@rtL2%YXu=Q3i<cu9@XGn}G&IE4*
zZtAs>nK<eU!~d8%CN&Mv>QgQDWdDd(;6R0&NmT}F`nGO*wNBic`vOpsnqw&~nw)Tx
z{3~^utNC=PIA}P^dr)3PAuKsYPlU%=!6331PM?E=GR=?S@RxLZx7Gni4k%RZ{gxul
z@SlS3{}km{9J*<^`pAkql0^sWjArY5TM%*!YaWcm_5!9u!pZ|X%Wgk>J^)v5)J)B{
znOeML=y_B|QU1t5)Uc_a@&j74kkw~*xp3NizJpB28LL?x|8`~_;A^ync(BiSbw%df
zfoT7j-s~nf6$J^S8=TS6=haK77|K5`M>Y#NgT&eZQVX<UJWbJOHK<#FY}Fz8&1H`$
z4BI{{`-+TH;^@3FvJLk41sOgzDlhO7ZfE$ctd0uB7K4L>?4@57k84O=S#l{`J<ma<
zv)0Y;(8A9>rGTw}nckK27y6^)>6_(`3<KE`;S3onLjxXVLt#e6y<=-b;LOy*IlkQL
zqQ!Sl$O*<MVX_as11pNZu3L7{Es97oZc_H&M0aU4zBN}CmCRIsx||J>BxO4g)zP$-
zK@Zev;=I-ijt*q2>#at>OX-+tNn2oXTrUtxkWhU0(?H~+IG!eUZJ%xd)97yBW{c}9
z8hN;EIz^zT(ebKAZ>rQ~)J|5h2mvD4{P20Q%KaPK&N<1#TUt3LyMc@wGk6U?lpi~|
zHNIGR!eHg$Z}{VV#^zJ}MGdaUQ36b1+xzb)A0q^4aqjIdJA^1s4z+}HeBad}jr5K7
z-uErV>-yz3tqUfP+Br@n;$IE|Jvx1Dy1wpfDF-(PATgF>Gzoc$qH#5Ko8^RMcqJsj
z<ECD1XIb122?Row2)AfHKJZux3db-=<pka=fwgieMWG+yT1+|@Tsl6|6iHI5hJGSH
zxHNx8A5j85F7$*==AgEdG-6th))n=g!OGgYt+#<TTSnIW+pWE?AA2h)Z|5PaQ+Q40
z`K%|^)T>&%4M#seN0s(Yn@z}SsQxK0@7D?$$D<oR?suSQ>XE7re)&2DtEA>lP+D{5
z<WRPt1Wj@nAHD(J?llI?92$WnPyFY#^7v|0?2oN|9Gr~FxnGZCt$7@pREG-!`TSq_
zQry37NwQ-c3P<&R7A}H6P*f9C$tPlhEwT<sXzqF5?m-jFWFTW-6I?RE4r3V6Q5{>W
zuT%K$=H;*L$#H(W7x~|yR%ai|8z<0`O{#{M79XvKw|*QC#2A4XH0+l8Te?^Jn{_s}
zVJJ_yr*`*>Sa33&2MUiVJDXp_KAYK;Q8o_s*s3iuT?%%T+C@;y^=!<T+{92e4Gz>P
zWqFb3jG&Fz<&e@TY~`Oe8T2cqHuX%Ga;aPd8GTJqp6*@}2=H~wvsc^K8R>wO+3Uq(
zmKW4y%<0_WJ|yVptE@SKd%RK=8Yna3p=UIfytnl{t-hn#B+x4w4}3Zw#ZuX0g16F~
zIv=T%Egf=zD>qV>T=7M;DbH5|M`j<pcllDNro<U6fV$~Jef(|dOzp9025&UMLr%J2
zKHy{fZrUrTndlEJgy;}KXOWMV`-4*_tr_ih-3h~9bMQqaE{5l{n_u=*8F%&#JI(Mu
zi`Mk(VLYAZM%~4ZbB_3;r~T}6i%gH*@NwG~eQ@Upq~906X-7zSRk;fZmoMiF3$1Y7
zk2xxHS$>5X#m{2oxEfyUYiK^()v!a)$HAcb1zb4$U3I^GiGFVU?AC67MXPe92@i}+
zWqgc57Vf)Mt<Cu0$arFDwV$w-Y`}*)f1-d%wzCXcSt!e{UBO@0mu8(o_bAul?}%De
z<k%PY_<;$VL09f}$R4%hqvb^xr?tEA!L5)QV0Xntw6k05aAUrc{4iV_H`EPJpCZ3w
zSyn&ac=jHjgyuRayxq~mJ>9xxXm~%pT(*su1OIoWGtVP2fx_`zNczHv%;;bLj4tYe
z)$9j5vNxJ^X+@y8H<<LzndC;0zMxC{q51W-W9?d7U;6_iod<95ZS&X^>C}Kms=H^!
z*RiI<yQ?-am6Qk4+!Lt)C|JEY+F=M))z^3+GsIo<OUy4%#wZvLBiyu6bg26E2TV;~
zZf4{dmg>1}7(s<<i*J(pup}*6rH#fPK^mpk<ue<Thg7m56P>hYa!ySpRiT}aKt^-^
z!U-=|7MUU&vXz=?d#qUnUaJDaqGpV;f+K$}7TpYK=FA!jU2a+1xRZooQBNBSU$2{q
zi?w#<mhH5fcfX1f-e~k+B$_aPJSEl}vuaplms)y%rSP~vjy=}`+!MedEhP?cH;9Ao
za<C<dCgry@5G*CnDox2Cx*J5Gf<N(f5X%LAK_HQL>b~m_oaC_Rx;<dxSNLl0PS%k&
zzKeUKWJ6f0>pkYuhWO<b^0=R!HDK~2s30Af7$<BIynZ0OS|WVH;h{ZI9H$~&p+z0$
zQH7*oMlwf&9s%_iPsJ`DjRFN#EiHd|h5K8h@TOwSsJlX`J{;*;@svq7tkIn#s2I~+
zhkpKb{WXiWVJJp+Oxo3AuU_iY!C-P1?WDHet3PaB`s#>SqO&aylCuowO)dL7x!Twi
zs}o)%IsYe#RI|XP#K*JSKHP(JP~)r#w21-TYza5csx4AUaa{&#*&cfX#lER8o)2XL
zEO-!1NJACWt*|I>ZB@}Uw;<?vB(okBv9quf^S<wi`y^tV$r-A%l3*61d-Tywp7^~D
zWUR=N-NDKRk!nB1IGi^P=xmYc5%aqn`a2_S4o3~GHcBM0h^3)JRw7vN{<elrhp7kd
zFn_wEz%vX(C;)#HzWh~IUY5gof@fO<j8cEnsHd3cZ4R<AaPjXi<4ggbw}fntiQ$PA
zR!=?|I13(^Yc+>5P9-R=Dg)XIYb{P~Ps(mq>7(<XQj?e%zDxxKe?1l*@M#E%z3{c8
z@JH$0x*q0Vi`_G<+Uw&loI{e}n*vewgc9)g8-0fDY)oD;6n=h;XS}L|=TV=?$(!Am
z>AVRD*5LeDCn<NHo2%L@zQ3OLlC&aTBqu}wXsf&a5aHep^`ZEK5!LKfpe=KLLc@%H
zoLh{Qc2+!<hyu5O47XJOHQ;bRVe%Cwd@mvPM{^)R+qs;M-Y=rNmq^vk62wRAOYj|U
zA%alAnEb7#1p3DqZEyl_S0&yeNH6T+^C&_=z;D67`Tj)cD%=<At(Q7~)bUnWKi4}d
zwF1Yo@XeHW$`xQ45DCeOzadD$3`oEge+Q(Kxuk;&H{%rl`@PQhpy_UGV(h)yPA4iN
zj5~-+pPkqNKwh1S>p%0**bCL(3m(5P6I;%m>5AmxdV=(b1yG}=;bs?5Uh~i*;R`no
zEbmRI^J2aKZ+yr<5_QJ{9|VLq6a<9mzw#mM3?}C047N7^wD!#0|No50rT&#Ga5U??
zuFq8J&6v5=Shd((M$LPX_wM-ml3lBNY1Wz@Mw*dQoZ_EHuertBz!_!Z_S;o<>D_Uy
zBBRYBMOqH$$$GTW7j96p(6Ld+$%a8-45V;y@}-<r@IoX-!~5m+!<+Au&VA`ci%i)@
zZmou$MzgG4i%!KK`DWcL8^tuc){V2u$$!SJ4x{W9mg6>KR94mz>hVb~-a(&9l`fsE
z9%$qAi6`>VU;f)=eQo)C?K1u@KV4hxHtCesx(JOb;_3|B6A}c>-rRz6Sw7LS$mjSj
zgh6POG_a~x!YXN1>F3g>%&fXvGpx6bNaUD4{$56%NIzP;XJy8X9?7K}Ri>Vdj#5!m
zk4XcY@I9(smpv7LcDz4xx^@;YZPnKPGK>jsKGvpJcyhva`-_cVBG8&j8LTC=<v}@k
zOW~(_@XM~fSWOwS4Dz{1X*;v*+#Dm?D&aarJn>1ki_h&Q+Ir_>m<+#gNS4bi*{UQg
zYXu3lsF00WMp^hSkJ$L?HY$^mfP%Tv_Oa>U;M-}>P-W%nH~umungRUPy)HMKkIyGh
zo9xla%dtK0(y1r5n<c7Fl<nl#t0kM0!tt)rrjKWRTT)7yn4I2Hul#2_d^z)%Jt*ff
z{mlEwsF<jt)-@f#EIDWw{IOVal%)W)*Ly^_^h>jaG%t7DbeFZw=^rWIaCqhA>9+MY
zAF^$Y?7Y@SRAsxc8*ageC_k=GrrH#Zsmpv+=vm_QJE+g@<zK|Bf~w@pQ%L0EFDs1)
zAMCD^>LS{)_q?bsz3ND9q3yAM>JC|*4{MYmGe9pr)NQLLbk3_&N+6^iCI)V|HQI9E
zEpzJrp<<4wM@wkS_StUCS&i=fXB${?@+%}@70#zvDwjk}!^~mGO{<5J(;a+WG9#!C
z%P~qTwgPrC%pAH}#(s>N(XZ34bD8bD?$A5@?0s2s0@r@noXZ@+d$n%pXv>Pmci#*t
zOW0-|A&nMmG+M^;C5I4h3R-iJ-$23CF^b@fiwCSAo(7`lX9aln;hj4!;aIa%maqbu
z!-n~p^bu}6{g{k6oD)<yHoQs%Xr1jHpI2?RA-hRt4ywYaS?Wu!x~f;vV6#TKr_4>F
zSj`!@MMLCp4mMWuw#mjl9(hqLK4AtPlbHEktyZQKGv4^w9XpxVT@k+MeAB17vTX8x
zc}|88R1$yFs9j)^OJ~tW^E<8WLP7G%b&EB=_^@Uh_iPxyq3q2P=NvA_QhhyM{v({h
zWgp>ue!^1mlj-GBbYJ$0jP6G92Wc5#FcSfmAUuLV0tA<U_z$fb-|<6wr2%otm^>=P
zJ`2JD&DlG`CC~*V)+_u52L~rB3v2wyS=BW+9iwP$Y|PuOAcrmkL|Ws|)Jryb8+Z(c
zRvN2)kazaoQHt}ZkyC$@Fy%R{PQ`-ozjY5%g(3?@O_xhg6TH8Mv>(9%=6LPxbQC5p
z-NLi0*My^IUU<G87c2|MTJWQH_<M{O4mNC3%T1n|siVcLs|DHCjkNhyCw}}7-K7Fj
z<vu19;sy@GT%o>|iQ;AJZbjjw#2rm($9M*Ga#7n!M_0S8$Em6X4{a8bQS>RY^6cq@
zY)Z~y+TduW6<LEO`~U?r%1uV(5kC4-YNg!ip>mMEKS~CW;dozfnuUOM?ag+c*z&+B
z?_FH6U62LuM0vXO8x)~Z#P3{mRg99VvGM(i;(8w(N(j@xX@vHQJ*=`ZiKk832O|?v
zt9s0G{zkL<fo<!?MQg`Krx$8=gg(i#eI@ze8*Ckq;<QHM8*2K&d#t>uGT3!=79#ma
z4+Cz4+-Dtp2eu<qN=0OOBDGY4YzsUp&lLOvCkul-k$}^Oy3kZ?FUhfzDN4oR*}d$W
zn_Q_$EA~`!!=Sgx6lF|*v9!v0L#z$)R!HQ&sbO6jdTi1ecKR5^O+f`d8S0TYd7qM=
zja)KEwXiY20-UGqVJga8N#3HWRI_qX2rHIW*hb~+BLD9}dv`-s$c#2&n^jtIK>R|B
zZsCBbuHm+bz3E|>Zo?@Q&43J{Y3V+X0x^QlAp>_aHJxZC$I_4hsx87zJ-Y@>(7evH
znMfywKSXQ8ImRa_y3p1#8aOEF`3h}Iix$>gm%)#ZI|$pf{d&<04L6#55V@|pj2Blc
zM0Q4TQ1j%_AgWbdDXVTOsG6Wv6Rf+l`!ZwWaL$-2PCGeV@`-3<4=w>r%ciISKt^|J
zuW0KRc+vH{0lAsan7b{}0VIp8(X{iIiwHxS*><tYyL$BSb>TFJ&Kz82-nvt|Lu<E#
zjY^qzvvOMXv1(ax7eAX;_Egc77$!ROu67xtLS6a?1!ou<ImI}+^u^*L$Yyed;AzJl
zX|ts2Q}$+gd$<pCs88RJGAF5oGHs@UT<czDym2>%C8l{{>UMz5dMt|<b>y?nNtWNA
zQnr)sKATZXZGzcTzdAr>=hK1}E%Xcqr^Q^}G4_^pLSt$@FCVp7LiiGE*pztBVR!k7
zorx@~NV&Vqp#4g>@i;#z%NpoPjq)vb{*7)cBUEPt%M>^kjF?TJG3kc~=#WbxRz%)~
z;&u(}_`tADT|JI8+N!~-70T~I7q_eV>Sz>;adx><ec@#iqDpeO=bX}9$r#X5G@Zb{
z;~;kOS#^uqK&d<oKO7ek#m{$IV>BoDK;yhB`bkr$jOE<K<qw$TY!ymO{9Q9`$e+N=
zjlR63g18?-q|L3GVIq;OSj=7qsF|U3XrgoyqO?JZTwU(H43;4<qwqdC*=^y!OEOud
zv2=9&c_tEitOrln(l`ptZ$2KpCd6C}jbxm+B?xMnHz^`A%3U;otW^@;Z_x)D4*9G2
zX_f|S6;&;)Yyma-C0=PUUxBuobzk;u7H>Gaasz7{Uz(F_#oJ5-yKY<s4r|unD%F<2
zH=BwDhXv8ed>pvl5+*aanLB^->0#z-6s7UIn$b!Bn03pcJc}{#%4o~%(-s;-(7*eC
zo(_0=Ckzs~M#T%~m(en$HzW`X8(T=M87_KDMkdl!-zA%I%b0kffNlm>Vro<%vw@$m
zs$oeok0)$sv`)sKrv$Q%LJS7WC|FUa)%KoJmkO8je~fG=1^@ZT(dL!)@LITjF<a)0
z&VWw`(C}bcQIEq`m<6#kx-{6I@V2`GR#<`MM7TWPUU1(9V9e%SR_9cL%CTAY+Fw6R
zW~dS(poGe~14STO5ST!FN!#BEJ*-U{Lhu*$I*glneC>B?INbq)Yq_lWE0+G){Pv;y
zie5O7ZbNDs5WJT{zm8mXE(>F8I5ly{X|V(Xb_V8c+GLOH+n0j5zvXfUW*ZxHQcK6K
z`={N8hj_hiEZ6E>$%g-qPc+JKW^Vnsv<b0;)F7fmvQr^zGpp&KGQJd7!(p}mY(}5^
zg2TARkb70)A^OV~^jKyEI>?i!;D7Qr%jyo1F=}CPzl!gw&Fq3|+BeUE!iRLCAfKDh
z{oZ?Yz-Ej8$KdxyWXL?n0XN?gs)z4DnUw)LMz=Fwft=V#HfXs{P;X<_o}~41A$*pi
zRr4S2d}7W6hAz_~wq)1!L!sMIn}~el8Hk4nm3Y;RNdi4{i=T>h8e#1JR7cYfh2CeR
z**wLFwzV~;|GcRUnGKrb-UkutMI)cs8;hoI)kMy^es*D9U*Q_xYR2Tz1bUj(<dG;n
zw?`N@{kk4b^^L_L^J+*yTyltX1>65ExaVGv!Y2WaDq|Xo6+qMZH>C^sr!ya`L^Jtc
zBhLs%iKD!{x@;0dd+)Cs0+dA03}T4?Ft-qQ>;w7WYY};_pBcfLh!~gVjAn{dkV|JM
zG<n6;(o1&3;ZlWRB=A~fw4BiPZFKW({Rra0ZXvD%TUUBHmc$ruK2xJ%{S2aY04#YJ
zevUS>{qi@-$9@}R{+u>fP@kZ4wmhDxJ&E)-7ROqpNBm8=maM7a4jFz+U<6X-ZCB`A
zVh#7Nt6zNO7)Vh%Pk7r$dXA$%IPj;-dZ+o({6qBt1|+@c9_9jX90fH*wACqZ)KamJ
z;p6CLO!Zw6ZnqHdjOYqoXpJhY@!Gq$-k0SfbM8(lkvCIMo-V2LtQyi+wTep0OLlc;
zoaf~y_{ON4g~v^=)zysRC1tf@V*u5?L6EMaN`yhPk)hRXQySnw$|~@{Y-RJM3#q1Y
z1jA3`V$j)vfk;>@#gDk49E&(6|IgnQ1lNl@bvdiA?~0;G?}Mpp1tMfI@LiYuEJlU7
zgZzhSdZxXR!?CRRGarndkDBA<H7H>R?Q%K%an*COgLH7^Z+i@)sX|0}F&w@_8YJCe
zUi>J_ADjtXMgxA1Q2&k;ForKtx*3rf8*ohh#Zt-+B*V!?2>SSUhSLp32Q)S%hMtnY
z8iWz#dGmd2fcp_N9-w4h_;!AQwLRa^vQ4d%=&OX{QM!#OankSt0+*4Q%8+?B0yEqd
z0UeOI?d%Hw{@XwW-}}cwW8TAUi%6{f(FeEn`9tf5i#&Zda^WN^f$lT80gMv)i=vf+
zoZZ&2)Tr)vV2C(B#8jrH$-d7_|De8ngAE<u4@FQn-~{SWGqg5v?75K~<>MQQ=<(s#
zmjsN{0uIMIDc@}30KdjH$y01~<amDay^U_$2CMq#U|5f+!PwHWhMapf`)3k6bE^5I
zxOuTnN;Zj5JLv(B;Jy_GE}JN7it=!OX_wr>se;~?`N{z|rO;ckOq{9ecEA5}D27)G
zF(fH|H5N_?iLr^jg%OQ&L&!;suIBb&CzpBG;Blm`zr81HI|f2#clkNR^V@W-s$)_$
z&K3-!VnhZfPAKO{k2+5tO9erNSg!jyT8f|NXp638tps_wc2Cb(F-Y~xfI}?$2Yuxd
zRM<hq-d;lXLYLW({f_;&n$gZL*qY&3bShGx#wztwD)GZI!4ZpA{4F}vma*H|n5*^Q
ze%ZqookIE6?7z04<?+x0#*p_|l4+H`)1va&hTTOZ;W|XyoHd<i(5pzfM5!oY*^&)N
zV3`v<7bogz8fn+M+B%`|Rz^Yqw?)AFS|l1$53mObxf61VvuN)X&WV`lvc5={SWPgb
z6v9Tb4|Y@$BzaA*ee4|tN-xm5Is&3;sM-7e6cjo`Q=P5Y6&MQJ`q<qm)MZ?hB=xGQ
z?<7o_7(@e*`jT7Ys5rd?+u?YgazqZ1d{<5yyxEwK)XWWQoN+;vAnpYuc&7ofmn_Q!
zV&UQPag{kxWIqUbBz<XK@6^E<`yWhjXb=ggXEwD<fS;2F0@-v#f#%D17}-!e9BH09
z`9h<g-1cZ)?U`Vei$m`ucqm)7X>Dd4EeL_-iQ>-XNt1+?i`Q7ldFBMY$Wm_34TBFs
z7)@o|O}sQ%7L3lcS?yJx>BPv=t^9WyVEn_i%BDI*eB%wC`!Rc-1<=?+XP<ZV0Q~mg
zi-4WZX2A2t_vOy7$Avf`p!QoK37xN-E|R(jJu%fsi1&MYC*)gp$krQ#{wZprV=$x*
zEILADfRx0Om3mTv)X!R)ixTY!6kt;=fO4x~wfl8v-=-s#xoWLr$HR&(X|t;gQ_n3e
zzbDcXC3qM7b2CR*mX}NhhMa>DTu*j*NoovUDN4t~3y!Ukzo%HIj5<SZksH_9{bXPB
zb^^133dR%ac-Qr>$F3hk#}jY|7yw^q^!l~ay-(_VtJ*MpJMD!Bkf)5C5!VWrJEGmE
z#Bl0AbYhKx1s#!8FT*rl9$W<Lh~hq%!8T4$-cQ4eQ7ddNkEC-aNz?v#P$P|jmZu&G
z#;{B@X-w;bH~jXodY{VUbA-}se{K`p-1d=jaK$keFDB;<k{l)VuAq!d#_u$D!swIa
z3KU&-hxkau(dG4Jj5&tPIPkXPG4~I+IhY+n_Q^2X?sTQ>@Hl+;%~<|<C$MBW^;c>q
z%5-W1J7xnU0>ddFwq2~_x%BW#cz<;0&%{_~!JuUT4vYUX19CZ7MDtZr5GHe8#Kuy(
zX8*i~AdDnV<_DY3JpsRa)*ZjQlS@H$pNrYNe}9uY?qr)pIVAM>_=-NiU9MkvS7zhy
z;V0}H?g?U-d0*B_9z!L1fdJOB;(Z-2OB31v`LSy`A>VWLDE{VLvUgY<>!eFRU%Thp
zdY*GJH|G=dc&^~usey}vn=PI`&wczr+;tD)k@8Xvkj<WULj4}V`|62!sOO**a^K9u
zCtxb!=Lg_<Ey45qi)8)F_4Z2B>85?)vp$5YzL2uMa_Obfi339^Hl8wb?;y}La0FQj
zbv@v@gaM>};4{H9z%z$Vvs@GIaIfn3`A%Y7xztHLTSA{?FtpL7Apr6X!}L#e=@1LY
zhTn~%(<kdupRXP=F`)VJwB*Y8T^H|ETde`AgdfF1Se$LOLuMFVLt2EC)=V-hsnpW}
zUSFit0`ScAxYG%EzOCH@Ap3GL@pm~Pt~Nd2_7NNOV_t0sW9A{Z?ObhlNc@;9RZ%|;
z8)LpLr<zOmBj+J38kRZbbaf0#lr@zP6BSyu3I-%Tr_LzaFNNS7q_SG|ccFefq8p{i
zoR&jtS2tHun%R0Gg7}Em#N_tCJp=~HH&v*;My?gwW(+$T3K*vWgX~A?plQr!?!m-*
z!D#i(lcS9d31lC(6ZWl>b|-{GjmZaUy9b{1x{drG|FpZF!1uC&j&8p12@xA~Bs;nU
z)4xNx8Zvir7P5z}`QM(9fFf05h0%>)<sH)jSmS0n{sx`xaJaLPDN*|E^;*=~E0-ls
zBV~ol3UGg~4tvrSfKj9+ZVUPT!6})@lP+*sC2})lD)vfzX&#!QCbpY-VUBYNI0l;h
z$#jTr%v*zYgu-r<>feh<uOw2;(k`tINts)qNbi%TPtJ<q3a$<~fKfYMMQ(_l?w0nh
z0%;9jx~t{<q8Haan0(R4%P+++J(>E~G~jVo<{Sr*T)`9dJbeN@H^@Bu-l5oPyW5`y
zy9j>WZ(cugl{VjaP1KCfx^@x{BlTFh3eL@z$cKUY{<-qJHrfF6jhswy4THixm?t?z
zo<Q85?Fqc)m4e>pIh*N}kw6SI^qr44+zjIVc9kerHWf-LP0k_#Ol7=e4vcVl^!X|{
zBtkDgXgFyL!IcWp<SNBzPUbF4b_RY4iumK5UmEzJ=zjo99x{CTzNcoN@O|3_kB2~d
zj-Y>*G+8xyYEx*YNl-vmlDjNLeSc|2LcGlVp<s<FRfU-tv5wkx4IMu83cnnL+zp#f
z%7Ldy-~-+~{`3d>eplBJ;f~$t^d~5g6zL_|lUd~@k&3Sj{QV-m#>Aj-QF^SUXUU+q
zeIJ58Yq|ZAfx!b)Lo(+BBQ8x{$K7+OoQ%hZ)HarNeg51y<_(iJj)N`<wED$AGB&2`
zl{#T{jY#BE>&#p&2_%e_I@_7)2GHppz^8A`xHKebTf40BH7>dVE%+xx<!%17GJ%w#
zMcV{Rr>v|q*ggSru!G?pDc14;<@`=Gn2iV%6Dd779MoEwiK>kXh-W-%*Qwl|dp+Ls
zqq_N-JfeU2GKH%S$_a5cm8d?3i?@T=4)pgK{Sp3kSNr)rX^O;VU7uBmW~N=AB^-!h
zFN5G~O_lB%3qrzsQpq=@9EA|t>LIZktt-Qm4@<$KNQi#+xKX~GPVGIez52t+*}k_R
z)n>z2y*~m6wgWD=Hsby4623XZ^^>+=ub1+mkuAjiVcd6?uTTutj#mtN9f&^mSwj!=
z_yQnDLnT*lSkUGN6?I#VCo{5K!1%}rq_97*BLkcybg9NJ>LJ&@#yxS(v1k{2zoBb-
z48BUH4d;jUHW~IBTxf0zN_OaeaV+J&8$sUqMRq6E2W<MjJA(Z2==)-PxqUvX>39*j
z-MC4@w0~yp@I+_#4Gq5gUem(%+<jBKIdEq@+*#@Sw`F1;wf|cY`G8uUeXrad&uykZ
zqLm+v>%sFv@zptaUq;|*H~U55-u+V5cLaPMWt|@5cFG?9ElXiroztHRPR~!*T<yWT
z+ehapYN6DhVG{F4)w{tQ(9gS}+$2xyLSIcx+HG;;WqI5$-ogmcEqh;JQIaVhj6r&_
ze?_%VhAZrrXESAC3?lN;5n_1k&)NksCqKWi7m8KncfZVT_QPU>opCtK<M&Zve{Om@
z-fRItLFGMseuO2-&d7g@(&2i&(NR<rr4x6?QV??-5fQIq=*cEiixx3`lstdEZr?Cc
zapc%Mi}cc2`<%A0KT==HO?E*;zL}HKvW}$6&JEZ}9Eg*{tGt8Gol5S!_IN%v?+iTv
z9KY&&nl~KqU=CgcluTZaXrjWzR!XJQd=9`IQ8bb;tHx!)HMd;C!O_rceZfLcgUEiP
zK>O?~t56u0d#tn@|I>x+BsiM}r+^<>PVB^5c>?;jH$LZ9icw3}6EX`Dlo3cjlLfvr
z{dotqQN%f47_*PTL<zby2@n@|KqB$hI98#T&lc>yFv+}p^zK`H!W}i35r+;h1NDh)
z`5395gyB(CoUr;Uz)#NVvD>Y*5Qe}rexf@HZRHD!X<HEw<vi~;AI`nJtwxDr0<$~?
zcp)@<m6rU~uDV?x5J8$USyIz`uxi@p)W6WV>j9`xUtes6MzAevR%{HOOturjWK4>6
zOns)pKPsf{=b!?F-6NIMecs4@J)QktqgRlh$R{qnBxS1^_Y_g<RT|Y$&8VLe$v}&A
zromQ7R(GwRP_NX7DQj<?e}Xq5b85!8O(JW~`t`=@&I3w?>^3fC54oQQjGOe_yW`(}
zGQfHoPB1)2Plh`AEHNF-Nq9Li7kT#7W2B!SQTWB!0;Ee3GCS_Xcr?7|vF=0lQURNm
z?N$Gx+nncctuIU~+$&j`p$45Y;>dm%2Hr94FI^tvhjv{~>AXD*gV2oc*rhCxtfwWh
z0O%$t*!cYzi%q@FD4yM_D94ZWO8%XWRF=ac+ca!WR&_ZZcl|lL3u<mS{R@gbL&?)4
zSFsOh_nfyX$z+ZGHmTwEh!qopnQ!F7zSdN=6Co5MCOtgcyM4e90XuuHhD_qB!%Z~z
zH2Rw*?NWiX;DN>j-5iC4Ane`vGew6zy!XWGH5QYyRz9k}+|gDZ-Pu6ulR+QdZQ4kn
zI#_u@PtQPesCK3sKUKPu6mX5Z*3=tO!)B~IMC%A`{MDZp7I_qcY~nnKnY9x>K@lw?
zkA_soAD4M`=g)o61aqmw1tl=0Z_Da@Eat=N)Zz{V<sUYbv7*Rm|7{alGB&Nd&?6p^
zx4D@Jop8iW%{JdOMwVY}%XwrRD0?$NW!=;MCL-*O>Zf&}_V}I3uP}9zxL5WzG?Z!e
zNf;LKPg7epyj2j{hd$!fm2c7AK`*YeFDM$@ws6KBSGRlh>WLS(t1Jf!Mg;!f<#hhj
zHT};E1>{uZKOX+y)J{l{|C!qPzvXKFuWJ8YX6OH^2Kui~VUXy5s`lR}c>dp^{8w-9
z|BMp&Z}9)T^FNgTuhaMc9l(Fpc>QO94Z;5c@PANz{of({SJw1DBb;gd7li*Khx*@P
VKtlb;M}z!#1OLr>x$A$7{s*Do^HTr-

literal 0
HcmV?d00001


From a2a957ef5b29aa5c0a1e22a097f095fa2edc5b3b Mon Sep 17 00:00:00 2001
From: HaoHaoXueXiHxy <20211005@bjtu.edu.cn>
Date: Thu, 4 Dec 2025 13:56:13 +0800
Subject: [PATCH 2/2] =?UTF-8?q?=E4=BF=AE=E6=94=B9=E4=BA=86=E4=B8=80?=
 =?UTF-8?q?=E7=82=B9readme?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../README.md"                                | 239 ++++++++----------
 1 file changed, 103 insertions(+), 136 deletions(-)

diff --git "a/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md" "b/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md"
index 0b0ad724..bd83306d 100644
--- "a/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md"
+++ "b/2025-Ascend-Innovation-Contest/S1/MoE/\346\210\221\346\203\263\345\245\275\345\245\275\347\235\241\344\270\200\350\247\211/README.md"
@@ -1,78 +1,61 @@
-#### 1. MoE模块前向优化
-在moe模块可以拆分 prefill（长序列预填充）和 decode（单 token 解码）阶段逻辑，能够解决解决 “两阶段结果不一致” 和 “推理效率低” 问题。
+#### 1. MoE模块拆分 prefill 和 decode 阶段逻辑
 
 DeepseekMoE模块
-显式拆分prefill/decode阶段
 ```python
 def forward(self, hidden_states):
-    if orig_shape[1] == 1:  # decode阶段（单token）
+    if orig_shape[1] == 1:
         y = self.moe_infer_decode(...).view(*orig_shape)
-    else:  # prefill阶段（长序列）
+    ...
+    else:
+	    # 复用原代码 
         y = self.moe_infer_prefill(...).view(*orig_shape)
 ```
-通过`sequence_length == 1`显式区分两阶段，分别调用专属推理函数，避免共享逻辑导致的结果偏差。
 
-decode阶段（单token推理）：极简逐专家计算
+decode阶段（单token推理）
 ```python
 @no_grad()
 def moe_infer_decode(self, x, flat_expert_indices, flat_expert_weights):
-    expert_cache = ops.zeros_like(x)
-    for i in range(self.num_experts_per_tok):
-        expert_id = flat_expert_indices[i].item()  # 取第i个选中的专家ID
-        weight = flat_expert_weights[i].item()    # 取对应权重
-        expert = self.experts[expert_id]          # 选中专家网络
-        expert_out = expert(x)                    # 专家处理单token输入
-        expert_cache += expert_out * weight       # 加权累加输出
-    return expert_cache
+	# Decode时单token直接遍历激活专家
+	expert_cache = ops.zeros_like(x)
+	for i in range(self.num_experts_per_tok):
+		expert_id = flat_expert_indices[i].item()
+		weight = flat_expert_weights[i].item()
+		expert = self.experts[expert_id]
+		expert_out = expert(x)
+		expert_cache += expert_out * weight
+	return expert_cache
 ```
-单token场景下，直接遍历门控网络选中的`num_experts_per_tok`个专家，逐个计算并加权累加；
-无冗余操作：单token无需批量索引/掩码，直接遍历激活专家，计算效率最大化；
-  `@no_grad()`：关闭梯度计算，避免推理阶段显存占用；
-权重直接乘：跳过复杂的掩码/索引，保证计算语义简单可追溯。
+直接遍历门控网络选中的`num_experts_per_tok`个专家，逐个计算并加权累加，计算效率最大化。
 
-prefill阶段（长序列推理）：批量高效计算
+prefill阶段（长序列推理）
 ```python
 @no_grad()
 def moe_infer_prefill(self, x, flat_expert_indices, flat_expert_weights):
-    expert_cache = ops.zeros_like(x)
-    idxs = flat_expert_indices.argsort()  # 按专家ID排序，便于批量处理
-    tokens_per_expert = flat_expert_indices.bincount().cumsum(0)  # 每个专家的token数量
-    token_idxs = idxs // self.num_experts_per_tok  # 还原token在原序列中的索引
-    
-    for i, end_idx in enumerate(tokens_per_expert):
-        start_idx = 0 if i == 0 else tokens_per_expert[i-1]
-        if start_idx == end_idx: continue  # 无token的专家跳过
-        
-        expert = self.experts[i]
-        length = (end_idx - start_idx).item()
-        # 1. 切片获取当前专家处理的token索引（MindSpore原生操作）
-        exp_token_idx = ops.narrow(token_idxs, 0, start_idx, length)
-        # 2. 批量提取token的输入特征
-        expert_tokens = F.embedding(exp_token_idx, x)
-        # 3. 专家计算 + 权重加权
-        expert_out = expert(expert_tokens)
-        expert_out = expert_out.mul(F.embedding(ops.narrow(idxs, 0, start_idx, length), flat_expert_weights))
-        # 4. 批量累加结果到缓存（替代循环赋值）
-        expert_cache = mindspore.mint.scatter_add(
-            expert_cache, 0, 
-            exp_token_idx.view(-1, 1).tile((1, x.shape[-1])),  # 扩展索引到特征维度
-            expert_out
-        )
-    return expert_cache
+	expert_cache = ops.zeros_like(x)
+	idxs = flat_expert_indices.argsort()
+	tokens_per_expert = flat_expert_indices.bincount().cumsum(0)
+	token_idxs = idxs // self.num_experts_per_tok
+	for i, end_idx in enumerate(tokens_per_expert):
+		start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+		if start_idx == end_idx:
+			continue
+		expert = self.experts[i]
+		length = (end_idx - start_idx).item()
+		exp_token_idx = ops.narrow(token_idxs, 0, start_idx, length)
+		expert_tokens = F.embedding(exp_token_idx, x)
+		expert_out = expert(expert_tokens)
+		expert_out = expert_out.mul(F.embedding(ops.narrow(idxs, 0, start_idx, length), flat_expert_weights))
+		expert_cache = mindspore.mint.scatter_add(expert_cache, 0, exp_token_idx.view(-1, 1).tile((1, x.shape[-1])), expert_out)
+	return expert_cache
 ```
-长序列场景下，先按专家ID排序token，再批量处理每个专家的所有token，避免逐token循环；
-
-排序+批量处理：将同专家的token归拢，一次调用专家网络处理所有token，计算效率提升数倍；
-`ops.narrow`替代Python切片：适配MindSpore张量操作，避免维度不匹配；
-`scatter_add`批量累加：替代逐token赋值，减少显存交互开销；
-跳过空专家：无token的专家直接跳过，减少无效循环。
+这里直接复用原代码中逻辑。
 
 ---
 
 Qwen2MoeSparseMoeBlock模块
 （deepseek模块的实现参考了培训给出的一些示例，基于这个经验，在qwen2_moe模块也进行了拆分获取了收益）
 
-与DeepseekMoE的“完全拆分函数”不同，Qwen2-MoE在同一个`forward`函数内通过分支区分阶段（代码写的没那么漂亮，主打一个效率）：
+Qwen2-MoE在同一个`forward`函数内通过分支区分阶段（代码写的没那么漂亮，主打一个效率）：
 ```python
 def forward(self, hidden_states):
     if sequence_length == 1:  # decode阶段
@@ -86,12 +69,11 @@ def forward(self, hidden_states):
 decode阶段（单token）：活跃专家筛选+复用prefill核心逻辑
 ```python
 if sequence_length == 1:
-    # 1. 生成专家掩码（与prefill格式完全一致）
     expert_mask = nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
-    # 2. 筛选活跃专家（仅处理被选中的专家，减少循环次数）
+    # 筛选活跃专家（仅处理被选中的专家，减少循环次数）
     expert_usage = ops.sum(expert_mask, dim=(1, 2))  # 统计每个专家被选中的次数
     active_experts = ops.nonzero(expert_usage > 0, as_tuple=False).squeeze(-1)
-    # 3. 遍历活跃专家（复用prefill的index_add逻辑）
+    # 遍历活跃专家（复用prefill的index_add逻辑）
     for expert_idx_tensor in active_experts:
         expert_idx = int(expert_idx_tensor.asnumpy().item())
         expert_layer = self.experts[expert_idx]
@@ -99,50 +81,35 @@ if sequence_length == 1:
         if 0 not in idx.shape:
             current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
             current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
-            # 关键：与prefill使用相同的index_add累加，保证结果一致
+            
             final_hidden_states = final_hidden_states.index_add(
                 0, top_x.int(), current_hidden_states.to(hidden_states.dtype)
             )
 ```
-- **核心优化**：
-  - 掩码格式对齐：decode阶段生成与prefill完全一致的`expert_mask`，消除掩码维度/排列差异导致的结果偏差；
-  - 活跃专家筛选：仅循环被选中的专家（而非所有专家），单token推理循环次数从`num_experts`降至`top_k`（通常2），效率提升；
-  - 累加逻辑一致：复用`index_add`而非直接赋值，保证两阶段输出的数学语义完全一致。
+
+  decode阶段生成与prefill完全一致的`expert_mask`，消除掩码维度/排列差异导致的结果偏差；
+  仅循环被选中的专家，单token推理循环次数从`num_experts`降至`top_k`。
 
 prefill阶段（长序列）：标准掩码+全专家遍历
 ```python
 else:
-    # 1. 生成专家掩码：(num_experts, top_k, batch*seq_len)
     expert_mask = nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
-    # 2. 遍历所有专家（保证长序列处理兼容性）
     for expert_idx in range(self.num_experts):
         expert_layer = self.experts[expert_idx]
         idx, top_x = ops.nonzero(expert_mask[expert_idx], as_tuple=True)
         if 0 not in idx.shape:
             current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
             current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
-            # 核心：index_add累加，保证批量token结果正确
             final_hidden_states = final_hidden_states.index_add(
                 0, top_x.int(), current_hidden_states.to(hidden_states.dtype)
             )
 ```
-- **核心保留**：
-  - 全专家遍历：长序列场景下，确保所有可能被选中的专家都被处理，避免遗漏；
-  - 掩码索引：通过`nonzero`提取每个专家处理的token索引，批量计算后用`index_add`累加，保证结果与原始逻辑一致。
-
-共享专家（Shared Expert）：两阶段统一逻辑
-```python
-shared_expert_output = self.shared_expert(hidden_states)
-shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output
-final_hidden_states = final_hidden_states + shared_expert_output
-```
-- 共享专家是所有输入都会经过的“基础专家”，用于兜底计算；
-- 门控权重：通过`sigmoid`激活的门控网络（`shared_expert_gate`）控制共享专家的输出强度，两阶段逻辑完全一致，避免额外偏差。
+这部分复用原本的MoE部分代码，遍历全部专家。
 
 这样把prefill和decode分开来写带来了在decode阶段有达到秒级的收益。这样做完总分提升到了160+。
 
 
-#### 2. 培训中提到的使用ops算子替换索引操作
+#### 2. 使用ops算子替换索引操作
 
 ```python
 def rotate_half(x):
@@ -166,15 +133,15 @@ def rotate_half(x):
     hidden_states = hidden_states_expand.broadcast_to((batch, num_key_value_heads, n_rep, slen, head_dim))
 
 ...
-            # self.cos_cached[:seq_len].to(dtype=x.dtype),
-            # self.sin_cached[:seq_len].to(dtype=x.dtype),
-            ops.narrow(self.cos_cached, 0, 0, seq_len).to(dtype=x.dtype),
-            ops.narrow(self.sin_cached, 0, 0, seq_len).to(dtype=x.dtype),
+    # self.cos_cached[:seq_len].to(dtype=x.dtype),
+    # self.sin_cached[:seq_len].to(dtype=x.dtype),
+    ops.narrow(self.cos_cached, 0, 0, seq_len).to(dtype=x.dtype),
+    ops.narrow(self.sin_cached, 0, 0, seq_len).to(dtype=x.dtype),
             
 ...
 
-            attention_mask_expanded = ops.unsqueeze(ops.unsqueeze(attention_mask, dim=1), dim=2)
-            padding_mask = ops.narrow(causal_mask, -1, 0, mask_length) + attention_mask_expanded
+    attention_mask_expanded = ops.unsqueeze(ops.unsqueeze(attention_mask, dim=1), dim=2)
+    padding_mask = ops.narrow(causal_mask, -1, 0, mask_length) + attention_mask_expanded
 
 ```
 使用ops替换掉tensor索引的一些操作能够带来一些细微的收益（几百ms），有的替换甚至没有收益（也可能是那块代码没执行）。
@@ -183,70 +150,70 @@ def rotate_half(x):
 
 使用StaticCache需要先在cache_utils.py文件中修改一下StaticCache类的update函数中的某一部分（被注释掉的是原来的）：
 ```python
-            else:
-                # use index_add for mindspore since tensor slice is too slow and no implementation of index_copy
-                # k_out = ops.index_add(k_out, 2, cache_position.int(), key_states)
-                # v_out = ops.index_add(v_out, 2, cache_position.int(), value_states)
-                k_out.index_add_(2, cache_position.int(), key_states)
-                v_out.index_add_(2, cache_position.int(), value_states)
+    else:
+        # use index_add for mindspore since tensor slice is too slow and no implementation of index_copy
+        # k_out = ops.index_add(k_out, 2, cache_position.int(), key_states)
+        # v_out = ops.index_add(v_out, 2, cache_position.int(), value_states)
+        k_out.index_add_(2, cache_position.int(), key_states)
+        v_out.index_add_(2, cache_position.int(), value_states)
 ```
 
 由于默认是DynamicCache，因此需要在utils.py中generate接口处修改一些地方，支持StaticCache的创建使用
 ```python
+else:
+    model_type = getattr(self.config, 'model_type', '')
+    supports_cache_position = model_type in ['qwen2_moe']
+    # print('StaticCache')
+    if (
+        hasattr(self, '_supports_static_cache')
+        and self._supports_static_cache
+        and not requires_cross_attention_cache
+        and supports_cache_position  
+    ):
+        if hasattr(self.config, "_pre_quantization_dtype"):
+            cache_dtype = self.config._pre_quantization_dtype
         else:
-            model_type = getattr(self.config, 'model_type', '')
-            supports_cache_position = model_type in ['qwen2_moe']
-            # print('StaticCache')
-            if (
-                hasattr(self, '_supports_static_cache')
-                and self._supports_static_cache
-                and not requires_cross_attention_cache
-                and supports_cache_position  
-            ):
-                if hasattr(self.config, "_pre_quantization_dtype"):
-                    cache_dtype = self.config._pre_quantization_dtype
-                else:
-                    cache_dtype = self.dtype
-                # print('StaticCache')
-                model_kwargs[cache_name] = self._get_cache(
-                    cache_implementation="static",
-                    max_batch_size=batch_size,
-                    max_cache_len=max_cache_length,
-                    model_kwargs=model_kwargs,
-                )
-            else:
-                num_hidden_layers = self.config.get_text_config().num_hidden_layers
-                model_kwargs[cache_name] = (
-                    DynamicCache(num_hidden_layers)
-                    if not requires_cross_attention_cache
-                    else EncoderDecoderCache(DynamicCache(num_hidden_layers), DynamicCache(num_hidden_layers))
-                )
+            cache_dtype = self.dtype
+        # print('StaticCache')
+        model_kwargs[cache_name] = self._get_cache(
+            cache_implementation="static",
+            max_batch_size=batch_size,
+            max_cache_len=max_cache_length,
+            model_kwargs=model_kwargs,
+        )
+    else:
+        num_hidden_layers = self.config.get_text_config().num_hidden_layers
+        model_kwargs[cache_name] = (
+            DynamicCache(num_hidden_layers)
+            if not requires_cross_attention_cache
+            else EncoderDecoderCache(DynamicCache(num_hidden_layers), DynamicCache(num_hidden_layers))
+        )
 ```
 这里StaticCache只限制了在qwen2_moe中使用，因为我尝试在deepseek中使用最后好像出现了mismatch，可能精度上有一些损失。
 
-最后JIT优化使用在了RMSNorm部分，还在Qwen2MoeModel的_update_causal_mask部分加了JIT，都带来了收益
+最后JIT优化使用在了RMSNorm部分，还在Qwen2MoeModel的_update_causal_mask部分加了JIT，都带来了收益。但是不知道为什么使用rms_norm的融合算子会导致mismatch，可能是精度有误差，所以放弃了融合算子，同样flash attention也会有这个问题。
 ```python
-    @mindspore.jit
-    def forward(self, hidden_states):
-        # if use_pyboost():
-        #     return F.rms_norm(hidden_states, self.weight, self.variance_epsilon)
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(mindspore.float32)
-        variance = ops.mean(hidden_states.pow(2), -1, keepdim=True)
-        hidden_states = hidden_states * ops.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-        
-    ...  
+@mindspore.jit
+def forward(self, hidden_states):
+    # if use_pyboost():
+    #     return F.rms_norm(hidden_states, self.weight, self.variance_epsilon)
+    input_dtype = hidden_states.dtype
+    hidden_states = hidden_states.to(mindspore.float32)
+    variance = ops.mean(hidden_states.pow(2), -1, keepdim=True)
+    hidden_states = hidden_states * ops.rsqrt(variance + self.variance_epsilon)
+    return self.weight * hidden_states.to(input_dtype)
     
-    @mindspore.jit
-    def _update_causal_mask(
-        self,
-        attention_mask: mindspore.Tensor,
-        input_tensor: mindspore.Tensor,
-        cache_position: mindspore.Tensor,
-        past_key_values: Cache,
-        output_attentions: bool,
-    ):
+...  
+
+@mindspore.jit
+def _update_causal_mask(
+    self,
+    attention_mask: mindspore.Tensor,
+    input_tensor: mindspore.Tensor,
+    cache_position: mindspore.Tensor,
+    past_key_values: Cache,
+    output_attentions: bool,
+):
 ```
 
 最后JIT优化牺牲了一些prefill的时延，但是对decode的时延带来了很大的提升，所以最后总分也提升到了280+。