Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
262 changes: 262 additions & 0 deletions 2025-Ascend-Innovation-Contest/S1/MultiModal/肆拾贰/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,262 @@
# MindNLP 模型优化 (Qwen2-VL & janus_pro)

本文档详细记录了针对 Qwen2-VL 和 janus_pro 模型的关键性能优化点,并附带了相应的核心代码实现。



## 一、Qwen2-VL 模型优化

### 1、使用融合算子

#### ① RoPE:mindspore.ops.rotary_position_embedding

修改前:

```python
mrope_section = mrope_section * 2
cos = ops.cat([m[i % 3] for i, m in enumerate(ops.split(cos, mrope_section, dim=-1))], dim=-1).unsqueeze(unsqueeze_dim)
sin = ops.cat([m[i % 3] for i, m in enumerate(ops.split(sin, mrope_section, dim=-1))], dim=-1).unsqueeze(unsqueeze_dim)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
```

修改后:

```python
q_embed = mindspore.ops.rotary_position_embedding(q, cos, sin)
k_embed = mindspore.ops.rotary_position_embedding(k, cos, sin)
```



#### ② RMSNorm:mindnlp.core.nn.rms_norm

修改前:

```python
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

不用贴这么多的代码, 列出核心修改即可, 最好有前后对比, 下面的大段代码也是一样

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

已修改

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)
```

修改后:

```python
return F.rms_norm(hidden_states, self.weight, self.variance_epsilon)
```



#### ③ FlashAttention

- 在 VisionAttention 中,使用 mindspore.ops.flash_attention_score,需要对 qk 先进行 scale,$\frac{q}{\sqrt{\sqrt{d}}}$, $\frac{k}{\sqrt{\sqrt{d}}}$,然后计算 flash_attention 时 scale 设为默认 1.0,否则精度不对齐(感觉可能跟大算子底层的计算顺序有关系,但这个方法只在这里有用,迁到 janus_pro 模型还是 mismatch)

```python
self.scalar_value = 1 / math.sqrt(math.sqrt(self.head_dim))
seq_length = hidden_states.shape[0]
q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb) * self.scalar_value
k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb) * self.scalar_value
attn_output = mindspore.ops.flash_attention_score(q, k, v.unsqueeze(0), self.num_heads, input_layout='BSND')
attn_output = attn_output.reshape(seq_length, -1)
attn_output = self.proj(attn_output)
```



- 在 Qwen2VLAttention 中,prefill 阶段,使用 mindspore.ops.fused_infer_attention_score,decoder阶段保持原来的计算,全部使用 flash_attention 会导致精度不对齐

```python
if query_states.shape[-2] != 1: # 判定 prefill 阶段还是 decoder 阶段
attn_mask = (attention_mask != 0).to(dtype=mindspore.uint8)
attn_output = mindspore.ops.fused_infer_attention_score(query_states*self.scalar_value, key_states*self.scalar_value, value_states, num_key_value_heads=self.num_key_value_heads, num_heads=self.num_heads, input_layout='BNSD', atten_mask=attn_mask)[0]

else:
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
attn_weights = ops.matmul(query_states, mint.permute(key_states, (0, 1, 3, 2))) / self.head_dim_sqrt
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=mindspore.bfloat16)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_output = ops.matmul(attn_weights, value_states)
```



### 2、mint 算子替换

#### ① nn.Conv3d 改用 mindspore.mint.Conv3D,需要进行权重转换

修改前:
```python
self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)
```

修改后:

```python
self.proj = mint.nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False, dtype=mindspore.bfloat16)
```



#### ② .swapaxes 改用 mindspore.mint.permute



### 3、旋转位置编码优化

预计算 sin / cos 表,避免在前向传播中重复计算



### 4、其它改进

#### ① Qwen2VLAttention 的 q_proj、k_proj、v_proj 合成一个 w_qkv

修改前:

```python
def __intit__():
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)

def forward():
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
```

修改后:

```python
def __intit__():
self.w_qkv = nn.Linear(self.hidden_size, self.num_heads * self.head_dim + self.num_key_value_heads * self.head_dim * 2, bias=True)

def forward():
qkv = self.w_qkv(hidden_states)
query_states, key_states, value_states = ops.split(qkv, [self.hidden_size, self.num_key_value_heads * self.head_dim, self.num_key_value_heads * self.head_dim], dim=2)
```



#### ② repeat_kv 优化

修改前:

```python
def repeat_kv(hidden_states: mindspore.Tensor, n_rep: int) -> mindspore.Tensor:
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
if n_rep == 1:
return hidden_states
hidden_states = hidden_states[:, :, None, :, :].broadcast_to((batch, num_key_value_heads, n_rep, slen, head_dim))
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
```

修改后:

```python
def repeat_kv(hidden_states: mindspore.Tensor, n_rep: int) -> mindspore.Tensor:
return ops.repeat_interleave(hidden_states, repeats=n_rep, dim=1)
```



## 二、janus_pro 模型优化

### 1、数据预处理(主要的性能瓶颈所在)

#### ① 重写 VLChatProcessor 的处理逻辑

原始的方法中存在 `image_token_mask = input_ids == self.image_id` 以及 `batched_images_seq_mask[i, -seq_len:] = input_ids == self.image_id` 等使用 `==` 逐元素比较的方法,很慢,参考 qwen2-vl 的方法去生成 input_ids,以及重写 images_seq_mask 的生成逻辑,避免使用 `==`

```python
class VLChatProcessor(ProcessorMixin):
def process_one():
# 此处只给出核心改进代码
tmp_sft_format = sft_format
tmp_sft_format = tmp_sft_format.split(self.image_tag)[0]
tmp_input_ids = self.tokenizer.encode(tmp_sft_format)
tmp_mask_before_len = len(tmp_input_ids)
mask = [0] * tmp_mask_before_len

index = 0
while self.image_tag in sft_format:
mask += [0]
sft_format = sft_format.replace(
self.image_tag, self.image_start_tag+"<|placeholder|>"*self.num_image_tokens+self.image_end_tag, 1
)
mask += [1] * self.num_image_tokens
index += 1
sft_format = sft_format.replace("<|placeholder|>", self.image_tag)
num_image_tokens = mindspore.Tensor([self.num_image_tokens] * index, mindspore.int32)

# tokenize
input_ids = self.tokenizer.encode(sft_format)
tmp_mask_last_len = len(input_ids) - len(mask)
mask += [0] * tmp_mask_last_len
images_seq_mask = mindspore.Tensor(mask, dtype=mindspore.bool_)
input_ids = mindspore.Tensor(input_ids, dtype=mindspore.int64)

# ...
return prepare, images_seq_mask
```



#### ② 使用 opencv 代替 PIL 加载图像

opencv 读取图像的速度大概是 PIL 的10倍左右,但这块对整体的提升不大,主要瓶颈在 resize、rescale 等操作上。

前期尝试过使用 opencv 加载图像后,用 numpy 重写数据预处理过程,但是遇到 ms.dataset.vision.Resize 的 BICUBIC 插值对针对相同数据但不同格式(PIL 和 numpy)存在精度误差,导致最终 mismatch,没找到好的解决方法。



### 2、其它改进(与 Qwen2-VL 模型类似)

#### ① 使用融合算子 F.rms_norm

#### ② 旋转位置编码优化——预计算 sin / cos 表,避免在前向传播中重复计算

#### ③ repeat_kv 优化

#### ④ rotate_half 优化

修改前:

```python
def rotate_half(x):
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return ops.cat((-x2, x1), dim=-1)
```

修改后:

```python
def rotate_half(x):
x1, x2 = ops.split(x, x.shape[-1] // 2, dim=-1)
return ops.cat((-x2, x1), dim=-1)
```

####

## 三、最终收益

| model_name | memory_reserved | memory_allocated | avg_prefill_latency | avg_decode_latency |
| ---------- | --------------- | ---------------- | ------------------- | ------------------- |
| Qwen2-VL | 6.442450944 | 5.672920576 | 0.2023613452911377 | 0.04043297529220581 |
| janus_pro | 17.179869184 | 15.238398464 | 0.13930201530456543 | 0.04886315107345581 |



## 四、评测结果

| 评测指标 | 平均得分 |
| -------------- | ------------ |
| 峰值显存得分 | 116.6667 |
| Prefill时延 | 425.6324 |
| Decode时延得分 | 208.4923 |
| **总分** | **250.2638** |
Binary file not shown.