add training code

.gitignore

@@ -0,0 +1 @@
+__pycache__/

configs/base.yaml

@@ -0,0 +1,31 @@
+seed: 5934875
+# Model settings
+seq_length: 4096  # Maximum sequence length
+
+# Data settings
+per_device_batch_size: 1
+n_device: 8  # Number of devices
+
+# Training settings
+optim: adamw_torch_fused
+steps: 500  # Number of training steps
+learning_rate: 0.00001  # Learning rate
+weight_decay: 0.1  # Weight decay for optimizer
+warmup_steps: 0  # Number of warmup steps for learning rate scheduler
+logging_steps: 10  # Log every X steps
+adam_beta1: 0.9
+adam_beta2: 0.95
+random_concat_ratio: 0.2  # Ratio of random concatenation
+
+
+# Evaluation settings
+eval_steps: 100  # Evaluate every X steps
+save_steps: 100  # Save model every X steps
+
+# Tokenizer settings
+
+# Additional settings (if needed)
+gradient_checkpointing: true
+gradient_accumulation_steps: 16  # Number of updates steps to accumulate before performing a backward/update pass
+max_grad_norm: 1.0  # Max gradient norm for gradient clipping
+ep_size: 2

deepseek/configuration_deepseek.py

@@ -0,0 +1,206 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+class DeepseekV2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeepseekV2Model`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V2.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 102400):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DeepseekV2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1407):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        n_shared_experts (`int`, *optional*, defaults to None):
+            Number of shared experts, None means dense model.
+        n_routed_experts (`int`, *optional*, defaults to None):
+            Number of routed experts, None means dense model.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
+            Scaling factor or routed experts.
+        topk_method (`str`, *optional*, defaults to `gready`):
+            Topk method used in routed gate.
+        n_group (`int`, *optional*, defaults to None):
+            Number of groups for routed experts.
+        topk_group (`int`, *optional*, defaults to None):
+            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
+        num_experts_per_tok (`int`, *optional*, defaults to None):
+            Number of selected experts, None means dense model.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
+        first_k_dense_replace (`int`, *optional*, defaults to 0):
+            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                            \--k dense layers--/
+        norm_topk_prob (`bool`, *optional*, defaults to False):
+            Whether to normalize the weights of the routed experts.
+        scoring_func (`str`, *optional*, defaults to 'softmax'):
+            Method of computing expert weights.
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss weight coefficient.
+        seq_aux = (`bool`, *optional*, defaults to True):
+            Whether to compute the auxiliary loss for each individual sample.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import DeepseekV2Model, DeepseekV2Config
+
+    >>> # Initializing a Deepseek-V2 style configuration
+    >>> configuration = DeepseekV2Config()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        hidden_size=4096,
+        intermediate_size=11008,
+        moe_intermediate_size = 1407,
+        num_hidden_layers=30,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        n_shared_experts = None,
+        n_routed_experts = None,
+        ep_size = 1,
+        routed_scaling_factor = 1.0,
+        kv_lora_rank = 512,
+        q_lora_rank = 1536,
+        qk_rope_head_dim = 64,
+        v_head_dim = 128,
+        qk_nope_head_dim = 128,
+        topk_method = 'gready',
+        n_group = None,
+        topk_group = None,
+        num_experts_per_tok = None,
+        moe_layer_freq = 1,
+        first_k_dense_replace = 0,
+        norm_topk_prob = False,
+        scoring_func = 'softmax',
+        aux_loss_alpha = 0.001,
+        seq_aux = True,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=100000,
+        eos_token_id=100001,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        self.aux_loss_alpha = aux_loss_alpha
+        self.seq_aux = seq_aux
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

esft.py

@@ -7,6 +7,8 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 def to_buffer(module, mark_param=True):
     """Turns all parameters of a module into buffers."""
+    if module is None:
+        return
     modules = module.modules()
     module = next(modules)
     delattrs = []
@@ -25,6 +27,8 @@ def to_buffer(module, mark_param=True):
 
 def to_param(module):
     """Turns all buffers of a module into parameterss."""
+    if module is None:
+        return
     modules = module.modules()
     module = next(modules)
     param_list = getattr(module, 'param_list', [])
@@ -57,7 +61,7 @@ def to_esft(model, adapter_config):
         to_buffer(model)
     else:
         to_param(model)
-    for idx, layer in enumerate(model.layers):
+    for idx, layer in enumerate(model.model.layers):
         if type(layer.mlp).__name__ != "DeepseekV2MoE":
             continue
         if adapter_config.get('shared_experts', False):
@@ -72,15 +76,25 @@ def to_esft(model, adapter_config):
                 to_buffer(layer.mlp.experts[expert_id])
     return model
 
+
 def load_state_dict(folder_path):
+    # 初始化空的 state_dict
     combined_state_dict = {}
 
+    # 遍历文件夹中的所有文件
     for file_name in os.listdir(folder_path):
         if file_name.endswith('.safetensors'):
             file_path = os.path.join(folder_path, file_name)
             state_dict = load_file(file_path)
             combined_state_dict.update(state_dict)
-    
+
+    # legacy for loading v1 checkpoints: add prefix "model." for parameters
+    for k in list(combined_state_dict.keys()):
+        if k.startswith("layers"):
+            k_new = "model." + k
+            combined_state_dict[k_new] = combined_state_dict[k]
+            del combined_state_dict[k]
+
     return combined_state_dict
     
 
@@ -89,21 +103,24 @@ def load_esft_model(base_model_path, adapter_dir):
     adapter_state_dict = load_state_dict(adapter_dir)
 
     # load pretrained model:
-    model, tokenizer = AutoModelForCausalLM.from_pretrained(base_model_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="auto"), AutoTokenizer.from_pretrained(base_model_path)
+    model, tokenizer = AutoModelForCausalLM.from_pretrained(base_model_path, trust_remote_code=True, torch_dtype=torch.bfloat16), AutoTokenizer.from_pretrained(base_model_path)
 
-    to_esft(model.model, adapter_config)
-    model.model.load_state_dict(adapter_state_dict)
+    to_esft(model, adapter_config)
+    model.load_state_dict(adapter_state_dict)
 
     return model, tokenizer
 
 def load_base_model(base_model_path):
     # load pretrained model:
-    model, tokenizer = AutoModelForCausalLM.from_pretrained(base_model_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="auto"), AutoTokenizer.from_pretrained(base_model_path)
+    model, tokenizer = AutoModelForCausalLM.from_pretrained(base_model_path, trust_remote_code=True, torch_dtype=torch.bfloat16), AutoTokenizer.from_pretrained(base_model_path)
 
     return model, tokenizer
 
-def add_adapter(base_model, adapter_dir, return_original_states=False):
-    adapter_config = json.load(open(adapter_dir + "/expert_cfg.json"))
+def add_adapter(base_model, adapter_dir, return_original_states=False, expert_config=None):
+    if expert_config is not None:
+        adapter_config = json.load(open(expert_config))
+    else:
+        adapter_config = json.load(open(adapter_dir + "/expert_cfg.json"))
     adapter_state_dict = load_state_dict(adapter_dir)
     
     to_esft(base_model, adapter_config)

scripts/train.sh

@@ -0,0 +1,12 @@
+
+export TOKENIZERS_PARALLELISM=false
+
+exp_name="test/eval_translation"
+base_model_path="/hf3fs-jd/prod/deepseek/shared/wangzihan/models/huggingface/vanilla_model"
+# turn above to for loop
+python train.py \
+    --base_model_path=${base_model_path} \
+    --expert_config=results/expert_configs/translation.json \
+    --train_dataset=translation \
+    --train_config=configs/base.yaml \
+    --output_dir=results/checkpoints/${exp_name}

scripts/train_ep.sh

@@ -0,0 +1,11 @@
+
+export TOKENIZERS_PARALLELISM=false
+
+exp_name="test/eval_translation"
+base_model_path="/hf3fs-jd/prod/deepseek/shared/wangzihan/models/huggingface/vanilla_model"
+torchrun --nproc-per-node=8 train_ep.py \
+    --base_model_path=${base_model_path} \
+    --expert_config=results/expert_configs/translation.json \
+    --train_dataset=translation \
+    --train_config=configs/base.yaml \
+    --output_dir=results/checkpoints/${exp_name}

train.py

@@ -0,0 +1,117 @@
+import argparse
+import json
+import yaml
+import os
+import random
+import torch
+from torch.utils.data import TensorDataset
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, logging
+
+from benchmarks import *
+from utils import get_formatted_input_and_target, get_examples_from_buffer_pad
+from esft import to_esft
+from deepseek.modeling_deepseek import DeepseekV2ForCausalLM
+
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--base_model_path", type=str, required=True)
+    parser.add_argument("--expert_config", type=str, required=True)
+    parser.add_argument("--train_dataset", type=str, required=True)
+    parser.add_argument("--output_dir", type=str, required=True)
+    parser.add_argument("--train_config", type=str, required=True)
+    parser.add_argument("--wandb_api_key", type=str, required=False)
+    args = parser.parse_args()
+
+    expert_config = json.load(open(args.expert_config))
+    output_dir = args.output_dir
+    base_model_path = args.base_model_path
+    config = yaml.safe_load(open(args.train_config))
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    seed = config['seed']
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    random.seed(seed)
+
+    if args.wandb_api_key is not None:
+        import wandb
+        wandb.login(key=args.wandb_api_key)
+
+    # Prepare data
+    tokenizer =  AutoTokenizer.from_pretrained(base_model_path)
+    samples = [json.loads(i) for i in open(f"datasets/train/{args.train_dataset}.jsonl").readlines()]
+    buffer = []
+    for instance in samples:
+        input_ids, target_ids = get_formatted_input_and_target(instance['messages'], tokenizer, -100)
+        buffer.append((input_ids, target_ids))
+    seq_length = config['seq_length']
+    random_concat_ratio = config['random_concat_ratio']
+    concated_examples = get_examples_from_buffer_pad(buffer, seq_length, tokenizer, random_concat_ratio)
+
+    dataset = TensorDataset(concated_examples['input_ids'], concated_examples['labels'])
+    train_dataset, valid_dataset = torch.utils.data.random_split(dataset, [int(len(dataset) * 0.98), len(dataset) - int(len(dataset) * 0.98)])
+
+    # Training arguments
+    training_args = TrainingArguments(
+        output_dir=output_dir,
+        max_steps=config['steps'],
+        per_device_train_batch_size=config['per_device_batch_size'],
+        per_device_eval_batch_size=config['per_device_batch_size'],
+        warmup_steps=config['warmup_steps'],
+        weight_decay=config['weight_decay'],
+        logging_dir=f"{output_dir}/logs",
+        logging_steps=config['logging_steps'],
+        save_steps=config['save_steps'],
+        eval_strategy="steps",
+        eval_steps=config['eval_steps'],
+        gradient_accumulation_steps=config['gradient_accumulation_steps'],
+        load_best_model_at_end=True,
+        metric_for_best_model="loss",
+        greater_is_better=False,
+        bf16=True,
+        lr_scheduler_type='constant',
+        save_total_limit=5,
+        learning_rate=config['learning_rate'],
+        optim=config['optim'],
+        adam_beta1=config['adam_beta1'],
+        adam_beta2=config['adam_beta2'],
+        gradient_checkpointing=config['gradient_checkpointing'],
+        gradient_checkpointing_kwargs={"use_reentrant": False} if config['gradient_checkpointing'] else {}, # if set to True, backward will raise bug
+    )
+
+    def data_collator(data):
+        input_ids = torch.stack([item[0] for item in data])
+        labels = torch.stack([item[1] for item in data])
+        return {"input_ids": input_ids, "labels": labels}
+
+
+    model = DeepseekV2ForCausalLM.from_pretrained(base_model_path, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2")
+    to_esft(model, expert_config)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset,
+        eval_dataset=valid_dataset,
+        data_collator=data_collator,
+    )
+    # Training
+    if os.path.exists(output_dir) and len(os.listdir(output_dir)) > 1: # has checkpoints already
+        trainer.train(resume_from_checkpoint=True)
+    else:
+        trainer.train()
+
+    # Save the model and tokenizer
+    trainer.save_model(output_dir + "/last_checkpoint")
+    tokenizer.save_pretrained(output_dir + "/last_checkpoint")
+
+    print("Training complete")
+
+if __name__ == "__main__":
+    main()

train_ep.py

@@ -0,0 +1,154 @@
+import argparse
+import json
+import yaml
+import os
+import random
+import torch
+import torch.distributed as dist
+from types import MethodType
+from torch.utils.data import TensorDataset
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, logging
+
+from benchmarks import *
+from utils import get_formatted_input_and_target, get_examples_from_buffer_pad, init_parallel_groups
+from esft import to_esft
+from deepseek.modeling_deepseek import DeepseekV2ForCausalLM
+
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+os.environ["NCCL_AVOID_RECORD_STREAMS"] = "1"
+logging.set_verbosity_error()
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--base_model_path", type=str, required=True)
+    parser.add_argument("--expert_config", type=str, required=True)
+    parser.add_argument("--train_dataset", type=str, required=True)
+    parser.add_argument("--output_dir", type=str, required=True)
+    parser.add_argument("--train_config", type=str, required=True)
+    parser.add_argument("--wandb_api_key", type=str, required=False)
+    args = parser.parse_args()
+
+    expert_config = json.load(open(args.expert_config))
+    output_dir = args.output_dir
+    base_model_path = args.base_model_path
+    config = yaml.safe_load(open(args.train_config))
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    seed = config['seed']
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    random.seed(seed)
+
+    if args.wandb_api_key is not None:
+        import wandb
+        wandb.login(key=args.wandb_api_key)
+
+    ep_size = config.get("ep_size", 1)
+    world_size, local_rank, ep_group, edp_group = init_parallel_groups(ep_size)
+    edp_size = world_size // ep_size
+
+    # Prepare data
+    tokenizer = AutoTokenizer.from_pretrained(base_model_path)
+    samples = [json.loads(i) for i in open(f"datasets/train/{args.train_dataset}.jsonl").readlines()]
+    buffer = []
+    for instance in samples:
+        input_ids, target_ids = get_formatted_input_and_target(instance['messages'], tokenizer, -100)
+        buffer.append((input_ids, target_ids))
+    seq_length = config['seq_length']
+    random_concat_ratio = config['random_concat_ratio']
+    concated_examples = get_examples_from_buffer_pad(buffer, seq_length, tokenizer, random_concat_ratio)
+
+    dataset = TensorDataset(concated_examples['input_ids'], concated_examples['labels'])
+    train_dataset, valid_dataset = torch.utils.data.random_split(dataset, [int(len(dataset) * 0.98), len(dataset) - int(len(dataset) * 0.98)])
+
+    # Training arguments
+    training_args = TrainingArguments(
+        output_dir=output_dir,
+        max_steps=config['steps'],
+        per_device_train_batch_size=config['per_device_batch_size'],
+        per_device_eval_batch_size=config['per_device_batch_size'],
+        warmup_steps=config['warmup_steps'],
+        weight_decay=config['weight_decay'],
+        logging_dir=f"{output_dir}/logs",
+        logging_steps=config['logging_steps'],
+        save_steps=config['save_steps'],
+        eval_strategy="steps",
+        eval_steps=config['eval_steps'],
+        gradient_accumulation_steps=config['gradient_accumulation_steps'],
+        load_best_model_at_end=True,
+        metric_for_best_model="loss",
+        greater_is_better=False,
+        bf16=True,
+        lr_scheduler_type='constant',
+        save_total_limit=5,
+        learning_rate=config['learning_rate'],
+        optim=config['optim'],
+        adam_beta1=config['adam_beta1'],
+        adam_beta2=config['adam_beta2'],
+        disable_tqdm=False,
+        gradient_checkpointing=config['gradient_checkpointing'],
+        gradient_checkpointing_kwargs={"use_reentrant": False} if config['gradient_checkpointing'] else {}, # if set to True, backward will raise bug
+    )
+
+    def data_collator(data):
+        input_ids = torch.stack([item[0] for item in data])
+        labels = torch.stack([item[1] for item in data])
+        return {"input_ids": input_ids, "labels": labels}
+
+
+    model = DeepseekV2ForCausalLM.from_pretrained(base_model_path, trust_remote_code=True, torch_dtype=torch.bfloat16,
+                                                  ep_size=ep_size, attn_implementation="flash_attention_2")
+    model._ddp_params_and_buffers_to_ignore = [n for n, _ in model.named_parameters() if ".expert" in n]    # we manage grad synchronization of expert parameters
+    to_esft(model, expert_config)
+    model.dummy = torch.nn.Parameter(torch.zeros(1, dtype=model.dtype))    # prevent DDP from having no trainable parameters
+    model._keys_to_ignore_on_save = ["dummy"]
+    expert_params = [p for n, p in model.named_parameters() if p.requires_grad and ".expert" in n]
+    for layer in model.model.layers:
+        if type(layer.mlp).__name__ != "DeepseekV2MoE":
+            continue
+        layer.mlp.ep_group = ep_group
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset,
+        eval_dataset=valid_dataset,
+        data_collator=data_collator,
+    )
+
+    accelerator = trainer.accelerator
+    backward = accelerator.backward
+    def custom_backward(self, loss, **kwargs):
+        backward(loss, **kwargs)
+        if not self.sync_gradients or edp_size == 1:
+            return
+        return
+        for p in expert_params:
+            g = p.grad if p.grad is not None else torch.zeros_like(p)
+            dist.all_reduce(g, op=dist.ReduceOp.AVG, group=edp_group)
+            if p.grad is not g:
+                p.grad = g
+    accelerator.backward = MethodType(custom_backward, accelerator)
+
+    # Training
+    ckpt_path = f"{output_dir}/last_checkpoint_ep{local_rank}"
+    if os.path.exists(output_dir) and len(os.listdir(output_dir)) > 1: # has checkpoints already
+        trainer.train(resume_from_checkpoint=ckpt_path)
+    else:
+        trainer.train()
+
+    # Save the model and tokenizer
+    if local_rank == 0:
+        trainer.save_model(ckpt_path)
+        tokenizer.save_pretrained(ckpt_path)
+    elif 0 < local_rank < ep_size:
+        model.save_pretrained(ckpt_path)
+
+    print("Training complete")
+
+if __name__ == "__main__":
+    main()

utils.py

@@ -1,3 +1,7 @@
+import os
+import random
+import torch
+import torch.distributed as dist
 # given a message object, convert to prompt and response
 
 PROMPT_USER: str = 'User: {input}\n\n'
@@ -38,3 +42,57 @@ def get_formatted_input_and_target(messages, tokenizer, IGNORE_TOKEN_ID=-100, ma
             assert False, f"Unknown role: {message['role']}"
 
     return [input_ids, target_ids]
+
+
+def get_examples_from_buffer_pad(buffer, seq_length, tokenizer, random_concat_ratio, IGNORE_TOKEN_ID=-100):
+    all_input_ids_list, all_target_ids_list = [], []
+    all_input_ids, all_target_ids = [], []
+
+    for input_ids, target_ids in buffer:
+        if len(input_ids) > seq_length - len(all_input_ids):
+            input_ids = input_ids[-(seq_length - len(all_input_ids)):]
+            target_ids = target_ids[-(seq_length - len(all_target_ids)):]
+        if len(all_input_ids) > 0 and random.random() < random_concat_ratio:
+            input_ids = input_ids[1:]
+            target_ids = target_ids[1:]
+        all_input_ids.extend(input_ids)
+        all_target_ids.extend(target_ids)
+        if len(all_input_ids) >= seq_length:
+            assert len(all_input_ids) == seq_length, f"{len(all_input_ids)=}, {seq_length=}, {len(buffer)=}"
+            all_input_ids_list.append(all_input_ids)
+            all_target_ids_list.append(all_target_ids)
+            all_input_ids, all_target_ids = [], []
+
+    all_input_ids = all_input_ids + [tokenizer.pad_token_id for i in range(seq_length - len(all_input_ids))]
+    all_target_ids = all_target_ids + [IGNORE_TOKEN_ID for i in range(seq_length - len(all_target_ids))]
+    all_input_ids_list.append(all_input_ids)
+    all_target_ids_list.append(all_target_ids)
+
+    if len(all_input_ids) <= 0:
+        return None
+    return {
+        "input_ids": torch.tensor(all_input_ids_list, dtype=torch.long),
+        "labels": torch.tensor(all_target_ids_list, dtype=torch.long)
+    }
+
+
+def init_parallel_groups(ep_size=1):
+    dist.init_process_group("nccl")
+    world_size = int(os.getenv("WORLD_SIZE", "0"))
+    local_rank = int(os.getenv("LOCAL_RANK", "0"))
+    torch.cuda.set_device(local_rank)
+    ep_group = edp_group = None
+    for i in range(0, world_size, ep_size):
+        ranks = list(range(i, i + ep_size))
+        group = dist.new_group(ranks)
+        if local_rank in ranks:
+            ep_group = group
+    edp_group = None
+    for i in range(ep_size):
+        ranks = list(range(i, world_size, ep_size))
+        group = dist.new_group(ranks)
+        if local_rank in ranks:
+            edp_group = group
+    dist.all_reduce(torch.zeros(1, device="cuda"), group=ep_group)
+    dist.all_reduce(torch.zeros(1, device="cuda"), group=edp_group)
+    return world_size, local_rank, ep_group, edp_group