2025年昇腾AI创新大赛-昇思模型开发挑战赛（S1赛季)--MultiModal赛题--好树花生队提交

2025-Ascend-Innovation-Contest/S1/MultiModal/好树花生/README.md

@@ -0,0 +1,44 @@
+
+# 2025年昇腾AI创新大赛-昇思模型开发挑战赛（S1赛季)--MultiModal赛题--好树花生队提交说明
+
+## 运行时长如下：
+
+<img src="./imgs/results.png"></img>
+
+## 优化点分为以下方面：
+
+1.qwenvl模型中有大量api替换，分别有mindspore.ops.rms_norm替换原来的实现，然后大量直接从tensor调用的方法，比如tensor.broadcast替换成mint中的方法，还有transpose、unsqueeze等方法
+
+2.qwenvl模型中的conv3d操作，原本调用的方法速度慢，改成mindspore.mint.nn.functional.conv3d后能够明显提升
+
+<img src="./imgs/2.png"></img>
+
+3.两个模型中softmax的操作，默认使用了fp32数据类型，改成默认的bf16，能够有轻微提升
+
+<img src="./imgs/3.png"></img>
+
+4.qwenvl模型的预处理阶段，在processing_vlm.py文件中，其使用tokenizer获得tag的4个方法操作很耗时，每次都重复执行，且每次操作都是一样的，所以改成在类初始化时就获取，后面直接使用现成的值，这样能够大大降低预处理的时间
+
+<img src="./imgs/4_1.png"></img>
+<img src="./imgs/4_2.png"></img>
+
+5.janus模型中的modeling_vlm.py文件里有大量打印，发现删除那些print后，速度偶尔有轻微提升
+
+6.janus的siglip_vit.py文件中，对forward_features方法中的blocks运算加jit，速度能明显提升，forward_head方法也加了jit，但速度提升不明显，有时感觉有轻微提升，有时却一点都没有提升
+
+<img src="./imgs/6.png"></img>
+
+7.qwenvl模型中vision模块的attention_mask重复计算了，将其提取到layer的for循环之外，保证就执行一次
+
+<img src="./imgs/7_1.png"></img>
+<img src="./imgs/7_2.png"></img>
+
+以上就是主要的几个修改点
+
+# 最终优化结果：
+| 评测指标 | 平均得分 |
+|---------|---------|
+| 峰值显存得分 | 116.6667 |
+| Prefill时延得分 | 382.3937    |
+| Decode时延得分 | 154.2476     |
+| **总分** | **217.7693** |

2025-Ascend-Innovation-Contest/S1/MultiModal/好树花生/patches/0001-.patch

@@ -0,0 +1,141 @@
+From 4fffb31ea6073e689089caa02e0a17cf4b6f73af Mon Sep 17 00:00:00 2001
+From: Alice <[email protected]>
+Date: Tue, 4 Nov 2025 22:41:54 +0800
+Subject: [PATCH 1/3] =?UTF-8?q?=E5=A4=9A=E6=A8=A1=E6=80=81=E8=B0=83?=
+ =?UTF-8?q?=E4=BC=98=E6=8F=90=E4=BA=A4?=
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+---
+ .../janus_pro/janus/models/processing_vlm.py  | 28 ++++++++++++++++---
+ .../models/qwen2_vl/modeling_qwen2_vl.py      | 14 ++++++----
+ 2 files changed, 33 insertions(+), 9 deletions(-)
+
+diff --git a/llm/inference/janus_pro/janus/models/processing_vlm.py b/llm/inference/janus_pro/janus/models/processing_vlm.py
+index 7f881e39..c93bfd92 100644
+--- a/llm/inference/janus_pro/janus/models/processing_vlm.py
++++ b/llm/inference/janus_pro/janus/models/processing_vlm.py
+@@ -110,9 +110,13 @@ class VLChatProcessor(ProcessorMixin):
+             print(f"Add image tag = {image_tag} to the tokenizer")
+
+         self.image_tag = image_tag
++        self.image_tag_id = None
+         self.image_start_tag = image_start_tag
++        self.image_start_tag_id = None
+         self.image_end_tag = image_end_tag
++        self.image_end_tag_id = None
+         self.pad_tag = pad_tag
++        self.pad_tag_id = None
+
+         self.num_image_tokens = num_image_tokens
+         self.add_special_token = add_special_token
+@@ -185,17 +189,29 @@ class VLChatProcessor(ProcessorMixin):
+
+     @property
+     def image_id(self):
+-        image_id = self.tokenizer.vocab.get(self.image_tag)
++        if self.image_tag_id is None:
++            image_id = self.tokenizer.vocab.get(self.image_tag)
++            self.image_tag_id = image_id
++        else:
++            image_id = self.image_tag_id
+         return image_id
+
+     @property
+     def image_start_id(self):
+-        image_start_id = self.tokenizer.vocab.get(self.image_start_tag)
++        if self.image_start_tag_id is None:
++            image_start_id = self.tokenizer.vocab.get(self.image_start_tag)
++            self.image_start_tag_id = image_start_id
++        else:
++            image_start_id = self.image_start_tag_id
+         return image_start_id
+
+     @property
+     def image_end_id(self):
+-        image_end_id = self.tokenizer.vocab.get(self.image_end_tag)
++        if self.image_end_tag_id is None:
++            image_end_id = self.tokenizer.vocab.get(self.image_end_tag)
++            self.image_end_tag_id = image_end_id
++        else:
++            image_end_id = self.image_end_tag_id
+         return image_end_id
+
+     @property
+@@ -208,7 +224,11 @@ class VLChatProcessor(ProcessorMixin):
+
+     @property
+     def pad_id(self):
+-        pad_id = self.tokenizer.vocab.get(self.pad_tag)
++        if self.pad_tag_id is None:
++            pad_id = self.tokenizer.vocab.get(self.pad_tag)
++            self.pad_tag_id = pad_id
++        else:
++            pad_id = self.pad_tag_id
+         # pad_id = self.tokenizer.pad_token_id
+         # if pad_id is None:
+         #     pad_id = self.tokenizer.eos_token_id
+diff --git a/mindnlp/transformers/models/qwen2_vl/modeling_qwen2_vl.py b/mindnlp/transformers/models/qwen2_vl/modeling_qwen2_vl.py
+index d059dcbe..ffb100cf 100644
+--- a/mindnlp/transformers/models/qwen2_vl/modeling_qwen2_vl.py
++++ b/mindnlp/transformers/models/qwen2_vl/modeling_qwen2_vl.py
+@@ -265,6 +265,7 @@ class PatchEmbed(nn.Module):
+         self.embed_dim = embed_dim
+
+         kernel_size = (temporal_patch_size, patch_size, patch_size)
++        self.kernel_size = kernel_size
+         self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)
+
+     def forward(self, hidden_states: mindspore.Tensor) -> mindspore.Tensor:
+@@ -272,7 +273,10 @@ class PatchEmbed(nn.Module):
+         hidden_states = hidden_states.view(
+             -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
+         )
+-        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
++        
++        hidden_states = mindspore.mint.nn.functional.conv3d(hidden_states.to(dtype=target_dtype), self.proj.weight, 
++                                                            stride=self.kernel_size).view(-1, self.embed_dim)
++        
+         return hidden_states
+
+
+@@ -330,7 +334,7 @@ class VisionAttention(nn.Module):
+         v = v.swapaxes(0, 1)
+         attn_weights = ops.matmul(q, k.swapaxes(1, 2)) / math.sqrt(self.head_dim)
+         attn_weights = attn_weights + attention_mask
+-        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=mindspore.float32).to(q.dtype)
++        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+         attn_output = ops.matmul(attn_weights, v)
+         attn_output = attn_output.swapaxes(0, 1)
+         attn_output = attn_output.reshape(seq_length, -1)
+@@ -559,7 +563,7 @@ class Qwen2VLAttention(nn.Module):
+             attn_weights = attn_weights + causal_mask
+
+         # upcast attention to fp32
+-        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=mindspore.float32).to(query_states.dtype)
++        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+         attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+         attn_output = ops.matmul(attn_weights, value_states)
+
+@@ -637,7 +641,7 @@ class Qwen2VLDecoderLayer(nn.Module):
+
+         residual = hidden_states
+
+-        hidden_states = self.input_layernorm(hidden_states)
++        hidden_states, _ = mindspore.ops.rms_norm(hidden_states, self.input_layernorm.weight, self.input_layernorm.variance_epsilon)
+
+         # Self Attention
+         hidden_states, self_attn_weights, present_key_value = self.self_attn(
+@@ -654,7 +658,7 @@ class Qwen2VLDecoderLayer(nn.Module):
+
+         # Fully Connected
+         residual = hidden_states
+-        hidden_states = self.post_attention_layernorm(hidden_states)
++        hidden_states, _ = mindspore.ops.rms_norm(hidden_states, self.post_attention_layernorm.weight, self.post_attention_layernorm.variance_epsilon)
+         hidden_states = self.mlp(hidden_states)
+         hidden_states = residual + hidden_states
+
+-- 
+2.47.1.windows.2
+

2025-Ascend-Innovation-Contest/S1/MultiModal/好树花生/patches/0002-rmsnorm-jit.patch

@@ -0,0 +1,183 @@
+From a9b4c9b85f237fcae52ea396cf7be606cdf16410 Mon Sep 17 00:00:00 2001
+From: Alice <[email protected]>
+Date: Thu, 6 Nov 2025 16:39:55 +0800
+Subject: [PATCH 2/3] =?UTF-8?q?=E7=BB=A7=E7=BB=AD=E4=BC=98=E5=8C=96?=
+ =?UTF-8?q?=EF=BC=8C=E5=8A=A0=E5=85=A5rmsnorm,jit?=
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+---
+ .../janus_pro/janus/models/clip_encoder.py    |  6 +++--
+ .../janus_pro/janus/models/modeling_vlm.py    | 27 ++++++++-----------
+ .../janus_pro/janus/models/siglip_vit.py      | 11 +++++---
+ .../janus_pro/janus/models/timm_layers.py     |  2 +-
+ .../models/llama/modeling_llama.py            |  6 ++---
+ 5 files changed, 27 insertions(+), 25 deletions(-)
+
+diff --git a/llm/inference/janus_pro/janus/models/clip_encoder.py b/llm/inference/janus_pro/janus/models/clip_encoder.py
+index a0620cfe..acb6ea3f 100644
+--- a/llm/inference/janus_pro/janus/models/clip_encoder.py
++++ b/llm/inference/janus_pro/janus/models/clip_encoder.py
+@@ -56,6 +56,7 @@ class CLIPVisionTower(nn.Module):
+         self.vision_tower, self.forward_kwargs = self.build_vision_tower(
+             vision_tower_params
+         )
++        self.vision_tower.jit()
+
+         if pixel_mean is not None and pixel_std is not None:
+             image_norm = Normalize(
+@@ -112,10 +113,11 @@ class CLIPVisionTower(nn.Module):
+         Returns:
+             image_features (torch.Tensor): [b, n_patch, d]
+         """
+-
++        
+         if self.image_norm is not None:
+             images = self.image_norm(images)
+-
++        
+         image_forward_outs = self.vision_tower(images, **self.forward_kwargs)
++        
+         image_features = self.feature_select(image_forward_outs)
+         return image_features
+diff --git a/llm/inference/janus_pro/janus/models/modeling_vlm.py b/llm/inference/janus_pro/janus/models/modeling_vlm.py
+index 7178c398..3d2d2d74 100644
+--- a/llm/inference/janus_pro/janus/models/modeling_vlm.py
++++ b/llm/inference/janus_pro/janus/models/modeling_vlm.py
+@@ -241,12 +241,16 @@ class MultiModalityCausalLM(MultiModalityPreTrainedModel):
+         Returns:
+             input_embeds (torch.Tensor): [b, T, D]
+         """
+-
++        
++        
+         bs, n = pixel_values.shape[0:2]
+         # "b n c h w -> (b n) c h w"
+         images = ops.reshape(
+             pixel_values, (bs * n, pixel_values.shape[2], pixel_values.shape[3], pixel_values.shape[4]))
+-        images_embeds = self.aligner(self.vision_model(images))
++        
++        vr = self.vision_model(images)
++        
++        images_embeds = self.aligner(vr)
+
+         # "(b n) t d -> b (n t) d"
+         images_embeds = ops.reshape(
+@@ -259,33 +263,24 @@ class MultiModalityCausalLM(MultiModalityPreTrainedModel):
+         condition = input_ids < 0
+         input_ids = (1-condition) * input_ids + condition * \
+             0  # ignore the image embeddings
++        
+         inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
++        
+
+         # replace with the image embeddings
+         # 627                               576
+         # inputs_embeds[images_seq_mask] = images_embeds[images_emb_mask]
+-        print("inputs_embeds:", inputs_embeds.shape)
+-        print("images_embeds[images_emb_mask].dtype", images_embeds[images_emb_mask].dtype)
+-        print("inputs_embeds.dtype", inputs_embeds.dtype)
++        
++        
+         padding_size = images_seq_mask.shape[1] - images_emb_mask.shape[1]
+         padding = Tensor(np.full((images_seq_mask.shape[0], padding_size), False), dtype=images_emb_mask.dtype)
+         padded_images_emb_mask = ops.concat((images_emb_mask, padding), dim=1)
+-        print("padded_images_emb_mask.shape:",padded_images_emb_mask.shape)
+-        print("images_embeds.shape:",images_embeds.shape)
+-        print("images_seq_mask.shape:",images_seq_mask.shape)
+         first_true = images_seq_mask.nonzero().squeeze()[0][1] # 42
+         last_true = images_seq_mask.nonzero().squeeze()[-1][1] # 42
+-        print("first_true:",first_true)
+-        print("last_true:",last_true)
+         left = inputs_embeds[:,:first_true]
+-        print(left.shape)
+         right = inputs_embeds[:, last_true+1:]
+-        print(right.shape)
+         inputs_embeds = ops.cat((left, images_embeds, right),1)
+-        print("inputs_embeds.shape:",inputs_embeds.shape)
+-        print("inputs_embeds.dtype:",inputs_embeds.dtype)
+-
+-
++        
+
+         # inputs_embeds = images_embeds[padded_images_emb_mask] * images_seq_mask + inputs_embeds * (1 - images_seq_mask)
+         return inputs_embeds
+diff --git a/llm/inference/janus_pro/janus/models/siglip_vit.py b/llm/inference/janus_pro/janus/models/siglip_vit.py
+index 56a6f299..d896eeb2 100644
+--- a/llm/inference/janus_pro/janus/models/siglip_vit.py
++++ b/llm/inference/janus_pro/janus/models/siglip_vit.py
+@@ -580,7 +580,11 @@ class VisionTransformer(nn.Module):
+         if return_prefix_tokens:
+             return tuple(zip(outputs, prefix_tokens))
+         return tuple(outputs)
+-
++    
++    @mindspore.jit(backend='GE')
++    def run_blocks_jit(self, x: mindspore.Tensor) -> mindspore.Tensor:
++        return self.blocks(x)
++    
+     def forward_features(self, x: mindspore.Tensor) -> mindspore.Tensor:
+         x = self.patch_embed(x)
+         x = self._pos_embed(x)
+@@ -590,10 +594,11 @@ class VisionTransformer(nn.Module):
+         #     x = checkpoint_seq(self.blocks, x)
+         # else:
+         #     x = self.blocks(x)
+-        x = self.blocks(x)
++        x = self.run_blocks_jit(x)
+         x = self.norm(x)
+         return x
+-
++    
++    @mindspore.jit(backend='GE')
+     def forward_head(self, x: mindspore.Tensor, pre_logits: bool = False) -> mindspore.Tensor:
+         if self.attn_pool is not None:
+             x = self.attn_pool(x)
+diff --git a/llm/inference/janus_pro/janus/models/timm_layers.py b/llm/inference/janus_pro/janus/models/timm_layers.py
+index 8960d256..61e68c54 100644
+--- a/llm/inference/janus_pro/janus/models/timm_layers.py
++++ b/llm/inference/janus_pro/janus/models/timm_layers.py
+@@ -46,7 +46,7 @@ class Mlp(nn.Module):
+
+     def forward(self, x):
+         x = self.fc1(x)
+-        x = self.act(x)
++        x = mindspore.ops.gelu(x)
+         x = self.drop1(x)
+         x = self.norm(x)
+         x = self.fc2(x)
+diff --git a/mindnlp/transformers/models/llama/modeling_llama.py b/mindnlp/transformers/models/llama/modeling_llama.py
+index 9c5cb555..c8c55492 100644
+--- a/mindnlp/transformers/models/llama/modeling_llama.py
++++ b/mindnlp/transformers/models/llama/modeling_llama.py
+@@ -429,7 +429,7 @@ class LlamaAttention(nn.Module):
+             attn_weights = attn_weights + causal_mask
+
+         # upcast attention to fp32
+-        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=mindspore.float32).to(query_states.dtype)
++        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+         attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+         attn_output = ops.matmul(attn_weights, value_states)
+
+@@ -508,7 +508,7 @@ class LlamaDecoderLayer(nn.Module):
+         """
+         residual = hidden_states
+
+-        hidden_states = self.input_layernorm(hidden_states)
++        hidden_states, _ = mindspore.ops.rms_norm(hidden_states, self.input_layernorm.weight, self.input_layernorm.variance_epsilon)
+
+         # Self Attention
+         hidden_states, self_attn_weights, present_key_value = self.self_attn(
+@@ -526,7 +526,7 @@ class LlamaDecoderLayer(nn.Module):
+
+         # Fully Connected
+         residual = hidden_states
+-        hidden_states = self.post_attention_layernorm(hidden_states)
++        hidden_states, _ = mindspore.ops.rms_norm(hidden_states, self.post_attention_layernorm.weight, self.post_attention_layernorm.variance_epsilon)
+         hidden_states = self.mlp(hidden_states)
+         hidden_states = residual + hidden_states
+
+-- 
+2.47.1.windows.2
+