[llama3.2] refine docs

models/Llama3_2-Vision/README.md

@@ -62,9 +62,9 @@ source ./envsetup.sh
 
 ``` shell
 pip install -r requirements.txt
-cp ./compile/files/Llama-3.2-11B-Vision-Instruct/modeling_llama.py /usr/local/lib/python3.10/dist-packages/transformers/models/mllama/modeling_llama.py
+cp ./compile/files/Llama-3.2-11B-Vision-Instruct/modeling_mllama.py /usr/local/lib/python3.10/dist-packages/transformers/models/mllama/modeling_mllama.py
 ```
-* PS：不一定是/usr/local/lib/python3.10/dist-packages/transformers/models/llama/modeling_llama.py这个路径，建议替换前先pip show transformers查看一下
+* PS：不一定是/usr/local/lib/python3.10/dist-packages/transformers/models/llama/modeling_mllama.py这个路径，建议替换前先pip show transformers查看一下
 
 ### 步骤五：生成onnx文件
 
@@ -127,7 +127,7 @@ python3 pipeline.py --help
 主要对.shape等操作转换为静态shape，避免模型过大导致常量折叠失败，进而导致动态子网
 还有一些scatterND操作，通过改为concat实现，理由同上
 
-- 在`modeling_llama.py`文件中如下代码：
+- 在`modeling_mllama.py`文件中如下代码：
 
 ```python
     def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
@@ -151,7 +151,7 @@ python3 pipeline.py --help
         return q_embed, k_embed
 ```
 
-- 在`modeling_llama.py`文件中如下代码：
+- 在`modeling_mllama.py`文件中如下代码：
 
 ```python
     def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
@@ -171,7 +171,7 @@ python3 pipeline.py --help
         return hidden_states.reshape(batch, slen, num_key_value_heads * n_rep, head_dim).transpose(1, 2)
 ```
 
-- 在`modeling_llama.py`文件中如下代码：
+- 在`modeling_mllama.py`文件中如下代码：
 
 ```python
     class LlamaAttention(nn.Module):

models/Llama3_2-Vision/compile/export_onnx.py

@@ -10,7 +10,6 @@
 
 import os
 import torch
-import numpy
 import argparse
 import transformers
 from tqdm import tqdm
@@ -438,37 +437,23 @@ def test_net_with_mask(image_path):
             block_kvs.append(CrossBlockCache(i))
 
     # inference
-    numpy.savez('verify/vit_in.npz',input_0=inputs['pixel_values'].squeeze(0).squeeze(0), input_1=inputs['aspect_ratio_ids'], input_2=inputs['aspect_ratio_mask'].squeeze(0))
     vit_out = vit(inputs['pixel_values'].squeeze(0).squeeze(0),
                   inputs['aspect_ratio_ids'],
                   inputs['aspect_ratio_mask'].squeeze(0))
-    numpy.savez('verify/vit_out.npz',output_0=vit_out)
-    numpy.savez('verify/embed_in.npz',input_0=input_ids)
     out = embed(input_ids).view(1, SEQ_LENGTH, HIDDEN_SIZE)
-    numpy.savez('verify/embed_ref.npz',output_0=out)
     k_cache = []
     v_cache = []
     for i in range(NUM_LAYERS):
         if i not in CROSS_ATTN_LAYERS:
-            if i == 2:
-                numpy.savez('verify/block_in.npz',input_0=out,input_1=position_ids,input_2=attention_mask)
             out, k, v = blocks[i](out, position_ids, attention_mask)
-            if i == 2:
-                numpy.savez('verify/block_ref.npz',output_0=out,output_1=k,output_2=v)
             k_cache.append(k)
             v_cache.append(v)
         else:
-            if i == 3:
-                numpy.savez('verify/cblock_in.npz',input_0=out,input_1=vit_out,input_2=text_row_mask,input_3=cross_attn_mask)
             out, k, v = blocks[i](out, vit_out, text_row_mask, cross_attn_mask)
-            if i == 3:
-                numpy.savez('verify/cblock_ref.npz',output_0=out,output_1=k,output_2=v)
             k_cache.append(k)
             v_cache.append(v)
     out = out[:, token_len - 1: token_len].view(1, 1, HIDDEN_SIZE)
-    numpy.savez('verify/head_in.npz',input_0=out)
     token = greedy_head(lm_head(out)).view(1)
-    numpy.savez('verify/head_ref.npz',output_0=token)
     out_ids = [int(token)]
     word = processor.decode([int(token)])
     print(word, end="")