Refine model of qwen for long sequence in eval.

.gitignore

@@ -1,2 +1,3 @@
 __pycache__
-.vscode
+.vscode
+*.txt

qwen/demo.py

@@ -52,11 +52,10 @@ print(model)
 
 
 tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
-model = model.from_pretrained(
-    model_dir, config=config, device_map="auto", trust_remote_code=True
-).train()
-# model.train()
-# model.zero_grad()
+model = model.from_pretrained(model_dir, config=config, device_map="auto", trust_remote_code=True)
+
+# model = model.eval()
+model = model.train() # control by @torch.no_grad()
 
 # 可指定不同的生成长度、top_p等相关超参
 # model.generation_config = GenerationConfig.from_pretrained(
@@ -74,16 +73,14 @@ print(decode_tokens)
 # 日本的首都东京。<|im_end|><|endoftext|>
 
 
-
-
 # # 第一轮对话
 # response, history, decode_tokens = model.chat(tokenizer, "你好", "", history=None)
 # print(decode_tokens)
 # # 你好！很高兴为你提供帮助。
 
 # 第二轮对话
-# response, history = model.chat(tokenizer, "给我讲一个年轻人奋斗创业最终取得成功的故事。", history=None)
-# print(response)
+response, history, decode_tokens = model.chat(tokenizer, "给我讲一个年轻人奋斗创业最终取得成功的故事。", "", history=None)
+print(response)
 
 
 # <|im_start|>system
@@ -93,4 +90,4 @@ print(decode_tokens)
 # <|im_start|>assistant
 # 莎士比亚是头一个使用“你好”这个词的文学家，他在《哈姆雷特》中写道：“你是谁？你在哪儿？
 # ”他的这一段话，通常被认为是最早的使用“你好”这个词的文学记载。这句话在英国语中非常常见，
-# 特别是在正式或礼貌的情况下。<|im_end|><|endoftext|>
+# 特别是在正式或礼貌的情况下。<|im_end|><|endoftext|>

qwen/modeling_qwen.py

@@ -41,8 +41,10 @@ import sys
 
 sys.path.append("..")
 from tools import show
+from tools import mem_tracker
 
-logger = logging.get_logger(__name__)
+# tracker = mem_tracker.MemTracker()
+# tracker.track()
 
 
 class QWenAttention(nn.Module):
@@ -110,8 +112,6 @@ class QWenAttention(nn.Module):
         query = apply_rotary_pos_emb(query, q_pos_emb)
         key = apply_rotary_pos_emb(key, k_pos_emb)
 
-        present = (key, value)
-
         key_size = key.size(1)
         if key_size > self.seq_length and not self.training:
             seq_start = key.size(1) - query.size(1)
@@ -148,8 +148,8 @@ class QWenAttention(nn.Module):
         attn_output = F.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask).transpose(1, 2)
         context_layer = self._merge_heads(attn_output, self.num_heads, self.head_dim)
         attn_output = self.c_proj(context_layer)
-        outputs = (attn_output, present)
-        return outputs
+
+        return attn_output
 
 
 class QWenMLP(nn.Module):
@@ -199,7 +199,6 @@ class QWenBlock(nn.Module):
             attention_mask=attention_mask,
         )
         attn_output = attn_outputs[0]
-        outputs = attn_outputs[1:]
         residual = hidden_states
         layernorm_input = attn_output + residual
 
@@ -207,8 +206,7 @@ class QWenBlock(nn.Module):
         residual = layernorm_input
         mlp_output = self.mlp(layernorm_output)
         hidden_states = residual + mlp_output
-        outputs = (hidden_states,) + outputs
-        return outputs
+        return hidden_states
 
 
 class QWenPreTrainedModel(PreTrainedModel):
@@ -312,16 +310,13 @@ class QWenModel(QWenPreTrainedModel):
         hidden_states = self.drop(hidden_states)
         output_shape = input_shape + (hidden_states.size(-1),)
 
-        presents = ()
         all_hidden_states = None
-        for i, block in enumerate(self.h):
-            outputs = block(
+        for block in self.h:
+            hidden_states = block(
                 hidden_states,
                 rotary_pos_emb_list=rotary_pos_emb_list,
                 attention_mask=attention_mask,
             )
-            hidden_states = outputs[0]
-            presents = presents + (outputs[1],)
 
         hidden_states = self.ln_f(hidden_states)
         hidden_states = hidden_states.view(output_shape)
@@ -392,6 +387,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             attentions=transformer_outputs.attentions,
         )
 
+    @torch.no_grad()
     def chat(
         self,
         tokenizer: PreTrainedTokenizer,
@@ -454,15 +450,9 @@ class QWenLMHeadModel(QWenPreTrainedModel):
         # 2. Set generation parameters if not already defined
 
         if generation_config.pad_token_id is None and generation_config.eos_token_id is not None:
-            if model_kwargs.get("attention_mask", None) is None:
-                logger.warning(
-                    "The attention mask and the pad token id were not set. As a consequence, you may observe "
-                    "unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results."
-                )
             eos_token_id = generation_config.eos_token_id
             if isinstance(eos_token_id, list):
                 eos_token_id = eos_token_id[0]
-            logger.warning(f"Setting `pad_token_id` to `eos_token_id`:{eos_token_id} for open-end generation.")
             generation_config.pad_token_id = eos_token_id
 
         # 3. Define model inputs
@@ -571,7 +561,6 @@ class QWenLMHeadModel(QWenPreTrainedModel):
 
             if this_peer_finished:
                 break
-
         return input_ids
 
 

tools/__init__.py

@@ -1 +1,2 @@
-from tools import show
+from tools import show
+from tools import mem_tracker

tools/mem_tracker.py

@@ -0,0 +1,171 @@
+import gc
+import datetime
+import inspect
+
+import torch
+import numpy as np
+import torch.nn as nn
+
+dtype_memory_size_dict = {
+    torch.float64: 64 / 8,
+    torch.double: 64 / 8,
+    torch.float32: 32 / 8,
+    torch.float: 32 / 8,
+    torch.float16: 16 / 8,
+    torch.half: 16 / 8,
+    torch.int64: 64 / 8,
+    torch.long: 64 / 8,
+    torch.int32: 32 / 8,
+    torch.int: 32 / 8,
+    torch.int16: 16 / 8,
+    torch.short: 16 / 6,
+    torch.uint8: 8 / 8,
+    torch.int8: 8 / 8,
+}
+# compatibility of torch1.0
+if getattr(torch, "bfloat16", None) is not None:
+    dtype_memory_size_dict[torch.bfloat16] = 16 / 8
+if getattr(torch, "bool", None) is not None:
+    dtype_memory_size_dict[torch.bool] = (
+        8 / 8
+    )  # pytorch use 1 byte for a bool, see https://github.com/pytorch/pytorch/issues/41571
+
+
+def get_mem_space(x):
+    try:
+        ret = dtype_memory_size_dict[x]
+    except KeyError:
+        print(f"dtype {x} is not supported!")
+    return ret
+
+
+class MemTracker(object):
+    """
+    Class used to track pytorch memory usage
+    Arguments:
+        detail(bool, default True): whether the function shows the detail gpu memory usage
+        path(str): where to save log file
+        verbose(bool, default False): whether show the trivial exception
+        device(int): GPU number, default is 0
+    """
+
+    def __init__(self, detail=True, path="", verbose=False, device=0):
+        self.print_detail = detail
+        self.last_tensor_sizes = set()
+        self.gpu_profile_fn = path + f"{datetime.datetime.now():%d-%b-%y-%H:%M:%S}-gpu_mem_track.txt"
+        self.verbose = verbose
+        self.begin = True
+        self.device = device
+
+    def get_tensors(self):
+        for obj in gc.get_objects():
+            try:
+                if torch.is_tensor(obj) or (hasattr(obj, "data") and torch.is_tensor(obj.data)):
+                    tensor = obj
+                else:
+                    continue
+                if tensor.is_cuda:
+                    yield tensor
+            except Exception as e:
+                if self.verbose:
+                    print("A trivial exception occured: {}".format(e))
+
+    def get_tensor_usage(self):
+        sizes = [np.prod(np.array(tensor.size())) * get_mem_space(tensor.dtype) for tensor in self.get_tensors()]
+        return np.sum(sizes) / 1024**2
+
+    def get_allocate_usage(self):
+        return torch.cuda.memory_allocated() / 1024**2
+
+    def clear_cache(self):
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def print_all_gpu_tensor(self, file=None):
+        for x in self.get_tensors():
+            print(x.size(), x.dtype, np.prod(np.array(x.size())) * get_mem_space(x.dtype) / 1024**2, file=file)
+
+    def track(self):
+        """
+        Track the GPU memory usage
+        """
+        frameinfo = inspect.stack()[1]
+        where_str = frameinfo.filename + " line " + str(frameinfo.lineno) + ": " + frameinfo.function
+
+        with open(self.gpu_profile_fn, "a+") as f:
+            if self.begin:
+                f.write(
+                    f"GPU Memory Track | {datetime.datetime.now():%d-%b-%y-%H:%M:%S} |"
+                    f" Total Tensor Used Memory:{self.get_tensor_usage():<7.1f}Mb"
+                    f" Total Allocated Memory:{self.get_allocate_usage():<7.1f}Mb\n\n"
+                )
+                self.begin = False
+
+            if self.print_detail is True:
+                ts_list = [(tensor.size(), tensor.dtype) for tensor in self.get_tensors()]
+                new_tensor_sizes = {
+                    (
+                        type(x),
+                        tuple(x.size()),
+                        ts_list.count((x.size(), x.dtype)),
+                        np.prod(np.array(x.size())) * get_mem_space(x.dtype) / 1024**2,
+                        x.dtype,
+                    )
+                    for x in self.get_tensors()
+                }
+                for t, s, n, m, data_type in new_tensor_sizes - self.last_tensor_sizes:
+                    f.write(
+                        f"+ | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} | {data_type}\n"
+                    )
+                for t, s, n, m, data_type in self.last_tensor_sizes - new_tensor_sizes:
+                    f.write(
+                        f"- | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} | {data_type}\n"
+                    )
+
+                self.last_tensor_sizes = new_tensor_sizes
+
+            f.write(
+                f"\nAt {where_str:<50}"
+                f" Total Tensor Used Memory:{self.get_tensor_usage():<7.1f}Mb"
+                f" Total Allocated Memory:{self.get_allocate_usage():<7.1f}Mb\n\n"
+            )
+
+
+def ModelSize(model, input, type_size=4):
+    para = sum([np.prod(list(p.size())) for p in model.parameters()])
+    # print('Model {} : Number of params: {}'.format(model._get_name(), para))
+    print("Model {} : params: {:4f}M".format(model._get_name(), para * type_size / 1000 / 1000))
+
+    input_ = input.clone()
+    input_.requires_grad_(requires_grad=False)
+
+    mods = list(model.modules())
+    out_sizes = []
+
+    for i in range(1, len(mods)):
+        m = mods[i]
+        if isinstance(m, nn.ReLU):
+            if m.inplace:
+                continue
+        out = m(input_)
+        out_sizes.append(np.array(out.size()))
+        input_ = out
+
+    total_nums = 0
+    for i in range(len(out_sizes)):
+        s = out_sizes[i]
+        nums = np.prod(np.array(s))
+        total_nums += nums
+
+    # print('Model {} : Number of intermedite variables without backward: {}'.format(model._get_name(), total_nums))
+    # print('Model {} : Number of intermedite variables with backward: {}'.format(model._get_name(), total_nums*2))
+    print(
+        "Model {} : intermedite variables: {:3f} M (without backward)".format(
+            model._get_name(), total_nums * type_size / 1000 / 1000
+        )
+    )
+    print(
+        "Model {} : intermedite variables: {:3f} M (with backward)".format(
+            model._get_name(), total_nums * type_size * 2 / 1000 / 1000
+        )
+    )

tools/test.py

@@ -1,5 +1,6 @@
 import show
 import torch
+import mem_tracker
 
 
 # radata = torch.randn(8192, 128)
@@ -10,11 +11,10 @@ radata = torch.randn(127)
 show.DumpTensorToImage(radata, "test.png")
 
 
-radata = torch.randn(3,127,127)
+radata = torch.randn(3, 127, 127)
 show.DumpTensorToImage(radata, "test.png")
 
 
-
 radata = torch.randn(127, 127)
 show.DumpTensorToLog(radata, "test.log")
 
@@ -22,3 +22,7 @@ show.DumpTensorToLog(radata, "test.log")
 radata = torch.randn(127, 127) - 0.5
 show.ProbGE0(radata)
 show.DumpProb()
+
+radata = torch.randn(127, 127).cuda()
+tracker = mem_tracker.MemTracker()
+tracker.track()