Add LoRA and Prefix-Tuning as Modeling Options for Improved Memory Ef…

…ficiency + performance (potentially) (#2840)

This PR adds LoRA and prefix-tuning as modelling options (training and
sampling code).

Both have shown strong performance and can outperform fine-tuning. They
also can protect against the catastrophic forgetting problem which is
important for chatbots. They keep the whole language model frozen so
they can be distributed freely independent of the base language model.

They also allow much more memory-efficient training as there is no need
for the optimizer states of the base model.

Benefits of LoRA
-  Can run only DS 2 with 30B.
- Run 65B without intense CPU usage
- Less overfitting, maintains performance on datasets. 
- Only need to share/push to hub the OS component (small file in Mbs)

— See [Andrej Karpathy
(OpenAI)](https://twitter.com/karpathy/status/1649127655122550784?s=20)
comment
— See purported [google
leak](https://www.semianalysis.com/i/119223672/lora-is-an-incredibly-powerful-technique-we-should-probably-be-paying-more-attention-to)


Implementation Details:
- Explicitly set input_ids as a keyword argument in the sampling code to
ensure proper functionality with PEFT `generate.`
- Manually enable gradients for input to leverage gradient checkpointing
effectively, as frozen embeddings would otherwise prevent gradient
attachment.
- Include saving code from `pytorch_model.bin` for special tokens.
Although these tokens are randomly initialized, they must be stored and
saved as an external module since PEFT parameters learn to utilize them.
Making them trainable is an option, but it is unlikely to make a
significant difference.
- Integrate prefix-tuning, a powerful technique that involves stacking
keys and values, by incorporating custom LLama modelling code. (Training
speed lower than LoRA in my initial tests and performance
similar/worse).

---------

Co-authored-by: jordiclive <jordiclive19@imperial.ac.uk>

model/model_eval/manual/sampling_report.py

@@ -10,6 +10,7 @@
 
 import pydantic
 import torch
+from model_training.models.peft_modeling import load_peft_model
 from tqdm import tqdm
 from transformers import AutoTokenizer, PreTrainedTokenizer
 
@@ -132,9 +133,9 @@ def sample(
     ).to(device)
     input_ids = inputs.input_ids
     outputs = model.generate(
-        input_ids,
-        **sampling_params,
+        input_ids=input_ids,
         pad_token_id=tokenizer.eos_token_id,
+        **sampling_params,
     )
     if skip_input_tokens:
         output_tokens = outputs[0, input_ids.size(1) :]
@@ -255,6 +256,7 @@ def parse_args():
     parser.add_argument("--max-input-len", type=int, help="max token counts for input")
     parser.add_argument("--auth-token", type=str)
     parser.add_argument("--num-threads", type=int, default=8)
+    parser.add_argument("--peft_model", type=str, default=None)
 
     return parser.parse_args()
 
@@ -314,6 +316,10 @@ def main():
     else:
         raise RuntimeError("Invalid model_type specified")
 
+    if args.peft_model is not None:
+        tokenizer = AutoTokenizer.from_pretrained(args.peft_model)
+        model = load_peft_model(model, args.peft_model, tokenizer)
+
     print("special_tokens_map:", tokenizer.special_tokens_map)
     print(f"eos_token='{tokenizer.eos_token}', eos_token_id={tokenizer.eos_token_id}")
 

model/model_training/configs/config.yaml

@@ -84,6 +84,8 @@ defaults:
   per_digit_tokens: false
   is_reward_model: false
   deepspeed_config: configs/zero_config.json
+  peft_model: false
+  peft_type: "lora"
 
 use_system_tag:
   use_system_tag: True
@@ -442,6 +444,78 @@ llama-30b-pretrain:
   num_train_epochs: 1
   save_total_limit: 2
 
+lora-llama-13b:
+  dtype: fp16
+  log_dir: "llama_lora_log_13b"
+  learning_rate: 5e-5
+  model_name: /home/ubuntu/llama_hf/13B
+  output_dir: llama_model_13b_lora
+  weight_decay: 0.0
+  max_length: 2048
+  warmup_steps: 300
+  gradient_checkpointing: true
+  gradient_accumulation_steps: 1
+  per_device_train_batch_size: 24
+  per_device_eval_batch_size: 5
+  eval_steps: 500
+  num_train_epochs: 12
+  save_total_limit: 2
+  save_strategy: epoch
+  use_flash_attention: True
+  residual_dropout: 0.0
+  deepspeed_config: configs/zero_config.json
+  peft_model: true
+  peft_type: "lora"
+  use_custom_sampler: true
+
+lora-llama-30b:
+  dtype: fp16
+  log_dir: "llama_lora_log_30b"
+  learning_rate: 5e-5
+  model_name: /home/ubuntu/llama_hf/30B
+  output_dir: llama_model_30b_lora
+  weight_decay: 0.0
+  max_length: 2048
+  warmup_steps: 300
+  gradient_checkpointing: true
+  gradient_accumulation_steps: 1
+  per_device_train_batch_size: 4
+  per_device_eval_batch_size: 2
+  eval_steps: 500
+  num_train_epochs: 12
+  save_total_limit: 2
+  save_strategy: epoch
+  use_flash_attention: True
+  residual_dropout: 0.0
+  deepspeed_config: configs/zero_config.json
+  peft_model: true
+  peft_type: "lora"
+  use_custom_sampler: true
+
+lora-llama-65b:
+  dtype: fp16
+  log_dir: "llama_lora_log_65b"
+  learning_rate: 5e-5
+  model_name: /home/ubuntu/llama_hf/65B
+  output_dir: llama_model_65b_lora
+  weight_decay: 0.0
+  max_length: 2048
+  warmup_steps: 300
+  gradient_checkpointing: true
+  gradient_accumulation_steps: 1
+  per_device_train_batch_size: 12
+  per_device_eval_batch_size: 5
+  eval_steps: 250
+  num_train_epochs: 12
+  save_total_limit: 2
+  save_strategy: epoch
+  use_flash_attention: True
+  residual_dropout: 0.0
+  deepspeed_config: configs/zero_config_sft_65b.json
+  peft_model: true
+  peft_type: "lora"
+  use_custom_sampler: true
+
 pythia-70m-deduped:
   learning_rate: 8e-6
   # model_name: EleutherAI/pythia-1b-deduped

model/model_training/configs/zero_config_sft_65b.json

@@ -0,0 +1,58 @@
+{
+  "fp16": {
+    "enabled": "auto",
+    "loss_scale": 0,
+    "loss_scale_window": 1000,
+    "initial_scale_power": 16,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
+  "optimizer": {
+    "type": "AdamW",
+    "params": {
+      "lr": "auto",
+      "betas": "auto",
+      "eps": "auto",
+      "weight_decay": "auto"
+    }
+  },
+  "scheduler": {
+    "type": "WarmupDecayLR",
+    "params": {
+      "warmup_min_lr": "auto",
+      "warmup_max_lr": "auto",
+      "warmup_num_steps": "auto",
+      "warmup_type": "linear",
+      "total_num_steps": "auto"
+    }
+  },
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 2e9,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 2e9,
+    "stage3_max_reuse_distance": 2e9,
+    "stage3_gather_16bit_weights_on_model_save": true,
+    "offload_param": {
+      "device": "cpu",
+      "pin_memory": true
+    },
+    "offload_optimizer": {
+      "device": "cpu",
+      "pin_memory": true
+    }
+  },
+  "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
+  "steps_per_print": 2000,
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "wall_clock_breakdown": false
+}

model/model_training/models/peft_modeling.py

@@ -0,0 +1,101 @@
+from dataclasses import dataclass
+from pathlib import Path
+
+import torch
+from huggingface_hub import hf_hub_download
+from model_training.utils.utils import get_model, get_tokenizer
+from peft import LoraConfig, PeftModel, PrefixTuningConfig, get_peft_model, prepare_model_for_int8_training
+
+
+def load_peft_model(model, peft_model_path, tokenizer):
+    model.resize_token_embeddings(len(tokenizer))
+    model.config.eos_token_id = tokenizer.eos_token_id
+    model.config.bos_token_id = tokenizer.bos_token_id
+    model.config.pad_token_id = tokenizer.pad_token_id
+    model = PeftModel.from_pretrained(
+        model,
+        peft_model_path,
+        torch_dtype=model.dtype,
+    )
+    model.eos_token_id = tokenizer.eos_token_id
+    extra_embeds = hf_hub_download(peft_model_path, "extra_embeddings.pt")
+    embed_weights = torch.load(extra_embeds, map_location=model.device)
+    model.base_model.model.model.embed_tokens.weight[len(tokenizer) - embed_weights.shape[0] :, :] = embed_weights.to(
+        model.base_model.model.model.embed_tokens.weight.dtype
+    )
+    return model
+
+
+def prepare_model_for_gradient_checkpointing(model):
+    r"""
+    Prepares the model for gradient checkpointing if necessary
+    """
+    if not getattr(model, "is_loaded_in_8bit", False):
+        if hasattr(model, "enable_input_require_grads"):
+            model.enable_input_require_grads()
+        else:
+
+            def make_inputs_require_grad(module, input, output):
+                output.requires_grad_(True)
+
+            model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+    return model
+
+
+def peft_model(model, peft_type="lora", int8_training=False, gradient_checkpointing=False):
+    if peft_type == "lora":
+        config = LoraConfig(
+            r=16,
+            lora_alpha=32,
+            target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
+            lora_dropout=0.05,
+            bias="none",
+            task_type="CAUSAL_LM",
+        )
+    elif peft_type == "prefix-tuning":
+        config = PrefixTuningConfig(
+            num_virtual_tokens=30, prefix_projection=True, encoder_hidden_size=1024, task_type="CAUSAL_LM"
+        )
+    else:
+        raise ValueError("peft_method config is lora or prefix-tuning")
+    model = get_peft_model(model, config)
+    if int8_training:
+        model = prepare_model_for_int8_training(model)
+
+    if gradient_checkpointing:
+        model = prepare_model_for_gradient_checkpointing(model)
+    model.print_trainable_parameters()
+    return model
+
+
+@dataclass
+class SaveLoraConfig:
+    dtype: torch.dtype = torch.float16
+    is_reward_model: bool = False
+    quantization: bool = False
+    seq2seqmodel: bool = False
+    freeze_layer: bool = False
+    residual_dropout: float = 0
+    use_flash_attention: bool = False
+    adapter_save_path: str = "adapter"
+    cache_dir: str = ""
+    model_name: str = ""
+    torch_ckpt_path: str = ""
+    peft_type: str = "lora"
+
+
+def save_adapter_model_from_ckpt(save_config: SaveLoraConfig):
+    tokenizer = get_tokenizer(save_config)
+    model = get_model(save_config, tokenizer)
+    model = peft_model(model)
+    model.load_state_dict(torch.load(save_config.torch_ckpt_path))
+    vocab_size = tokenizer.vocab_size
+    num_special_tokens = len(tokenizer.additional_special_tokens)
+
+    new_embs = model.state_dict()["base_model.model.model.embed_tokens.weight"][
+        vocab_size : vocab_size + num_special_tokens, :
+    ].clone()
+    new_embs = new_embs.to(save_config.dtype)
+    model.save_pretrained(save_config.adapter_save_path, torch_dtype=save_config.dtype)
+    tokenizer.save_pretrained(save_config.adapter_save_path)
+    torch.save(new_embs, Path(save_config.adapter_save_path).joinpath("extra_embeddings.pt"))