2025年昇腾AI创新大赛-昇思模型开发挑战赛（S1赛季)--MoE赛题--HUST_ASCEND队提交

2025-Ascend-Innovation-Contest/S1/MoE/HUST_ASCEND/MoE模型昇腾优化技术报告.md

@@ -0,0 +1,144 @@
+# MoE模型昇腾迁移优化技术报告
+
+## 评测结果
+
+| 评测指标 | 平均得分 |
+|---------|---------|
+| 峰值显存得分 | 100 |
+| Prefill时延得分 | 109.3774     |
+| Decode时延得分 | 360.0786     |
+| **总分** | **189.8187** |
+
+## 优化模型
+
+本项目针对以下两个MoE（Mixture of Experts）模型进行了昇腾NPU适配与性能优化：
+
+1. **DeepSeek-MoE-16B-Chat** - 深度求索开源的MoE大模型
+2. **Qwen1.5-MoE-A2.7B-Chat** - 通义千问开源的MoE大模型
+
+---
+
+## 核心优化技术
+
+### 1. MindSpore算子适配优化
+
+#### 1.1 使用mint算子替代ops算子
+
+将原始的`ops`算子替换为Ascend硬件亲和的`mint`算子
+
+```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)
+```
+
+
+#### 1.2 使用mint.narrow替代切片操作
+
+```python
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, dtype=x.dtype)
+
+        return (
+            # 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),
+        )
+```
+
+**收益**：mint.narrow避免切片操作的额外内存拷贝。
+
+---
+
+### 2. FlashAttention优化
+
+```python
+else: # prefill开启
+    # 融合后，采用mindspore的融合算子，flash_attention_score
+    sparse_mode = 0
+    if attention_mask is not None:
+        attention_mask = ~attention_mask
+
+    if self.is_causal:
+        sparse_mode = 3
+    global_attn_mask = ops.ones(2048, 2048, dtype=mindspore.bool_).triu(diagonal=1)
+    attn_output = mindspore.ops.flash_attention_score(query_states, key_states, value_states,
+head_num=self.num_heads, input_layout='BNSD', real_shift = None,padding_mask = None, attn_mask=global_attn_mask,scalar_value=1/math.sqrt(self.head_dim), keep_prob=1-self.attention_dropout,pre_tokens = 2147483647, next_tokens = 2147483647, inner_precise = 0,drop_mask = None, prefix = None, actual_seq_qlen = None, actual_seq_kvlen = None,sparse_mode=sparse_mode)
+```
+
+
+
+---
+
+### 3. MoE路由与专家计算优化
+
+#### 3.1 Qwen2-MoE: decode优化
+
+```python
+if routing_weights.shape[0] == 1:
+            # 遍历激活的 top-k 专家
+            final_hidden_states = ops.zeros((batch_size * sequence_length, hidden_dim), dtype=mindspore.float32)
+            flat_topk_idx = selected_experts.view(-1)
+            # idt = ops.zeros(1,dtype = mindspore.int64)
+            for i in range(self.top_k):
+                expert_idx = flat_topk_idx[i].item()
+                weight = routing_weights[0, i].to(mindspore.float32) # no item, no precision loss
+                expert_layer = self.experts[expert_idx]
+                final_hidden_states += expert_layer(hidden_states).to(mindspore.float32).mul(weight)
+            final_hidden_states = final_hidden_states.to(hidden_states.dtype)
+```
+
+#### 3.2 DeepSeek-MoE: decode优化
+
+**Decode阶段**
+
+```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()
+            weight = flat_expert_weights[i].item()
+            expert = self.experts[expert_id]
+            expert_out = expert(x)
+            expert_cache += expert_out * weight
+        return expert_cache
+```
+
+
+
+---
+
+## 优化效果分析
+
+### Prefill阶段优化
+
+| 优化项 | 技术手段 | 预估收益 |
+|-------|---------|---------|
+| FlashAttention | 硬件加速 | 40-60% |
+| mint算子替换 | 底层优化 | 10-20% |
+
+
+### Decode阶段优化
+
+| 优化项 | 技术手段 | 预估收益 |
+|-------|---------|---------|
+| 分场景MoE策略 | 减少padding | 40-50% |
+
+
+
+
+---
+
+## 关键技术总结
+
+1. **算子层优化**：替换mint算子，充分使能昇腾NPU加速计算。
+2. **注意力优化**：集成FlashAttention，加速prefill阶段推理能力。
+3. **MoE优化**：针对decode场景进行优化。
+