Merge pull request #523 from Anhforth/fit_torch

Fit torch

examples/Aquila/Aquila-chat/Aquila-chat-lora.yaml

@@ -1,19 +1,18 @@
-batch_size: 4
+batch_size: 1
+epochs: 5
 gradient_accumulation_steps: 1
-lr: 3.0e-4
+lr: 4.0e-5
 warm_up: 0.01
-warm_up_iters: 200
-lora_r: 16
+lora_r: 8
 lora_alpha: 32
-epochs: 30000000
-save_interval: 300
-log_interval: 10
+save_interval: 200
+log_interval: 1
 bmt_cpu_offload: False
 bmt_pre_load: True
-
+bmt_lr_decay_style: 'cosine'
 save_optim: True
 save_rng: True
-
+enable_flash_attn_models: False
 eps: 1.0e-8
 lora: True
 

examples/Aquila/Aquila-chat/aquila_chat.py

@@ -61,7 +61,7 @@
     import sys
     sys.exit(0)
 
-print(f"Trainer effective env_args={env_args} local_rank={trainer.local_rank}", flush=True)
+print(f"Trainer effective env_args={env_args} local_rank={os.environ['LOCAL_RANK']}", flush=True)
 
 checkpoints = env_args.pre_load_dir
 
@@ -78,7 +78,7 @@
 # avoid sync loading models in case of Mem OOM
 if env_args.bmt_async_load:
     import time
-    time.sleep(10*60*(trainer.local_rank%4))
+    time.sleep(10*60*(os.environ['LOCAL_RANK']%4))
 
 
 config_file = os.path.join(cache_dir, 'config.json')
@@ -98,7 +98,7 @@
 
 trainer.pre_train(model)
 
-# print('*'*20, "model", model, flush=True)
+print('*'*20, "model", model, flush=True)
 
 assert env_args.enable_sft_dataset_dir is not None and \
         env_args.enable_sft_dataset_file is not None

examples/Aquila/Aquila-chat/bmtrain_mgpu.sh

@@ -80,7 +80,7 @@ OPTS=" --batch_size $BATCH_SIZE \
 ## Trigger job on Each Node when bmt or ddp.
 
 mkdir -p $PRE_LOAD_DIR
-python -m torch.distributed.launch \
+torchrun \
        --nproc_per_node $GPU_NUM_PER_NODE \
        --nnodes $NODES_NUM \
        --node_rank $RANK \

examples/Aquila/Aquila-chat/generate_chat_lora.py

@@ -20,7 +20,7 @@
                     use_cache=True,
                     fp16=True,
                     device='cuda',
-                    adapter_dir='/data2/yzd/FlagAI/examples/Aquila/Aquila-chat/checkpoints_out/aquila_experiment75new/2023070515/') # eg: /mnt/yzd/git/FlagAI/examples/Aquila/Aquila-chat/checkpoints_out/aquila_experiment/2023062909
+                    adapter_dir='/data2/yzd/FlagAI/examples/Aquila/Aquila-chat/checkpoints_out/aquila_experiment/2023080216/') # Directory to adapter_model.bin and adapter_config.json
 model = loader.get_model()
 
 tokenizer = loader.get_tokenizer()
@@ -32,6 +32,11 @@
 
 texts = [
     "Find the product of the numbers: 5 and 8",
+    "Create a list of potential topics for a company newsletter",
+    "Explain the theory of relativity in simple terms.",
+    "Write a short story about a dragon and a knight.",
+    "翻译成英文: '我饿了想吃饭'",
+    "write a fairy tale for me",
 ]
 
 for text in texts:

examples/Aquila/Aquila-chat/hostfile

@@ -1 +1 @@
-192.168.21.7 slots=2
+192.168.20.3 slots=1

examples/Aquila/Aquila-code/aquila_code_pretrain.py

@@ -77,7 +77,7 @@
 # avoid sync loading models in case of Mem OOM
 if env_args.bmt_async_load:
     import time
-    time.sleep(10 * 60 * (trainer.local_rank % 4))
+    time.sleep(10 * 60 * (os.environ['LOCAL_RANK'] % 4))
 
 config_file = os.path.join(cache_dir, 'config.json')
 model = AQUILAModel.init_from_json(config_file=config_file)

examples/Aquila/Aquila-code/bmtrain_mgpu.sh

@@ -80,7 +80,7 @@ OPTS=" --batch_size $BATCH_SIZE \
 ## Trigger job on Each Node when bmt or ddp.
 
 mkdir -p $PRE_LOAD_DIR
-python -m torch.distributed.launch \
+torchrun \
        --nproc_per_node $GPU_NUM_PER_NODE \
        --nnodes $NODES_NUM \
        --node_rank $RANK \

examples/Aquila/Aquila-pretrain/aquila_pretrain.py

@@ -5,7 +5,6 @@
 import torch
 from torch.utils.data import Dataset
 import gc
-
 gc.collect()
 torch.cuda.empty_cache()
 from flagai.auto_model.auto_loader import AutoLoader
@@ -62,7 +61,7 @@
     import sys
     sys.exit(0)
 
-print(f"Trainer effective env_args={env_args} local_rank={trainer.local_rank}",
+print(f"Trainer effective env_args={env_args} local_rank={os.environ['LOCAL_RANK']}",
       flush=True)
 checkpoints = env_args.pre_load_dir
 model_name = env_args.model_name
@@ -77,7 +76,7 @@
 # avoid sync loading models in case of Mem OOM
 if env_args.bmt_async_load:
     import time
-    time.sleep(10 * 60 * (trainer.local_rank % 4))
+    time.sleep(10 * 60 * (os.environ['LOCAL_RANK'] % 4))
 
 config_file = os.path.join(cache_dir, 'config.json')
 model = AQUILAModel.init_from_json(config_file=config_file)

examples/Aquila/Aquila-pretrain/bmtrain_mgpu.sh

@@ -80,7 +80,7 @@ OPTS=" --batch_size $BATCH_SIZE \
 ## Trigger job on Each Node when bmt or ddp.
 
 mkdir -p $PRE_LOAD_DIR
-python -m torch.distributed.launch \
+torchrun \
        --nproc_per_node $GPU_NUM_PER_NODE \
        --nnodes $NODES_NUM \
        --node_rank $RANK \

examples/Aquila/aquila_pretrain.py

@@ -62,7 +62,7 @@
     import sys
     sys.exit(0)
 
-print(f"Trainer effective env_args={env_args} local_rank={trainer.local_rank}",
+print(f"Trainer effective env_args={env_args} local_rank={os.environ['LOCAL_RANK']}",
       flush=True)
 checkpoints = env_args.pre_load_dir
 model_name = env_args.model_name
@@ -77,7 +77,7 @@
 # avoid sync loading models in case of Mem OOM
 if env_args.bmt_async_load:
     import time
-    time.sleep(10 * 60 * (trainer.local_rank % 4))
+    time.sleep(10 * 60 * (os.environ['LOCAL_RANK'] % 4))
 
 config_file = os.path.join(cache_dir, 'config.json')
 model = AQUILAModel.init_from_json(config_file=config_file)

examples/Aquila/bmtrain_mgpu.sh

@@ -80,7 +80,7 @@ OPTS=" --batch_size $BATCH_SIZE \
 ## Trigger job on Each Node when bmt or ddp.
 
 mkdir -p $PRE_LOAD_DIR
-python -m torch.distributed.launch \
+torchrun \
        --nproc_per_node $GPU_NUM_PER_NODE \
        --nnodes $NODES_NUM \
        --node_rank $RANK \

examples/Aquila/hostfile

@@ -1 +1 @@
-192.168.21.2 slots=4
+192.168.20.3 slots=2

examples/swinv2/inference_swinv2.py

@@ -6,7 +6,7 @@
 from tqdm import tqdm
 from flagai.auto_model.auto_loader import AutoLoader
 
-data_path = "./imagenet2012/"
+data_path = "/data2/yzd/FlagAI/examples/swinv2/imagenet2012/"
 
 # swinv2 model_name support:
 # 1. swinv2-base-patch4-window16-256,

flagai/env_args.py

@@ -146,7 +146,7 @@ def __init__(self,
         self.parser.add_argument('--num_nodes', default=num_nodes, type=int, help='start training from saved checkpoint')
         self.parser.add_argument('--num_gpus', default=num_gpus, type=int, help='start training from saved checkpoint')
         self.parser.add_argument('--not_call_launch', action="store_true", help='start training from saved checkpoint')
-        self.parser.add_argument('--local_rank', default=0, type=int, help='start training from saved checkpoint')
+        self.parser.add_argument('--local-rank', default=0, type=int, help='start training from saved checkpoint')
 
         self.parser.add_argument('--wandb', default=wandb, type=str2bool, help='whether to use wandb')
         self.parser.add_argument('--wandb_dir', default=wandb_dir, type=str, help='wandb directory')
@@ -204,6 +204,7 @@ def parse_args(self):
         if args.env_type == "pytorch":
             # not need the "not_call_launch" parameter
             args.not_call_launch = True
+        print(args)
         for arg in vars(args):
             # change string format list to back to python list object
             value = getattr(args, arg)

flagai/env_trainer_v1.py

@@ -156,7 +156,10 @@ def __init__(self,
             training_paras = get_args_list(env_args)
             self.rank = int(os.environ.get('RANK', 0))
             self.world_size = int(os.environ.get('WORLD_SIZE', 1))
-            self.local_rank = env_args.local_rank
+            # self.local_rank = env_args.local_rank
+            print(os.environ['LOCAL_RANK'])
+            self.local_rank = int(os.environ['LOCAL_RANK'])
+
             log_dist("not_call_launch: {}".format(self.not_call_launch))
             # Implement for AutoLaunch
             # >>> python train.py # will call get_dist_args()
@@ -196,7 +199,7 @@ def initialize_distributed(self):
             device = self.rank % torch.cuda.device_count()
             if self.local_rank is not None:
                 device = self.local_rank
-            torch.cuda.set_device(device)
+            torch.cuda.set_device(int(device))
             # Call the init process
             init_method = 'tcp://'
             self.master_ip = os.getenv('MASTER_ADDR', 'localhost')

flagai/model/predictor/aquila.py

@@ -30,7 +30,7 @@ def aquila_generate(
         total_len = min(2048, max_gen_len + max_prompt_size)
 
         tokens = torch.full((bsz, total_len), 0).cuda().long()
-        #tokens = torch.full((bsz, total_len), 0).to("cuda:5").long()
+
         for k, t in enumerate(prompt_tokens):
             tokens[k, : len(t)] = t.clone().detach().long()
         input_text_mask = tokens != 0

flagai/model/vision/swinv2.py

@@ -159,7 +159,7 @@ def forward(self, x, mask=None):
 
         # cosine attention
         attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
-        logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01))).exp()
+        logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01)).to(self.logit_scale.device)).exp()
         attn = attn * logit_scale
 
         relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
@@ -518,7 +518,8 @@ def __init__(self, img_size=224,
                  patch_norm=True,
                  pretrained_window_sizes=[0, 0, 0, 0],
                  checkpoint_activations=False,
-                 num_classes=1000):
+                 num_classes=1000,
+                 **kwargs):
         self.num_classes = num_classes
         self.img_size = img_size
         self.patch_size = patch_size
@@ -566,7 +567,7 @@ class SwinTransformerV2(BaseModel):
 
     def __init__(self, config, num_classes=1000,  **kwargs):
         super().__init__(config, **kwargs)
-        swin_config = SwinTransformerConfig(**config)
+        swin_config = SwinTransformerConfig(**config.json_config)
 
         embed_dim = swin_config.embed_dim
         img_size = swin_config.img_size

setup.py

@@ -5,7 +5,7 @@
 
 setup(
     name="flagai",
-    version="v1.7.3",
+    version="v1.7.5",
     description="FlagAI aims to help researchers and developers to freely train and test large-scale models for NLP/CV/VL tasks.",
     long_description=open("README.md", encoding="utf-8").read(),
     long_description_content_type="text/markdown",
@@ -21,8 +21,8 @@
         'sentencepiece>=0.1.96',
         'boto3==1.17.32',
         'pandas>=1.3.5',
-        'jieba==0.42.1',
-        'scikit-learn==1.0.2',
+        'jieba>=0.42.1',
+        'scikit-learn>=1.0.2',
         'tensorboard>=2.9.0',
         'transformers>=4.20.1',
         'datasets>=2.0.0',
@@ -32,13 +32,14 @@
         'Pillow>=9.3.0',
         'einops>=0.3.0',
         'diffusers==0.7.2',
-        'pytorch-lightning==1.6.5',
+        'pytorch-lightning>=1.6.5',
         'taming-transformers-rom1504==0.0.6',
         'rouge-score',
         'sacrebleu>=2.3.1',
         'jsonlines',
         'accelerate',
         'PyYAML==5.4.1',
         'safetensors',
+        'timm',
     ]
 )