add sft stage

finetune.py

@@ -0,0 +1,324 @@
+import os
+import torch
+import torch.nn as nn
+import transformers
+from transformers import LlamaForCausalLM, LlamaTokenizer
+from datasets import load_dataset
+import json
+from peft import (
+    prepare_model_for_int8_training,
+    LoraConfig,
+    get_peft_model,
+    get_peft_model_state_dict,
+)
+
+from utils import (
+    QuantLinear,
+    get_peft_qmodel,
+    generate_alpaca_prompt,
+    prepare_model_for_fp16_training,
+    Prompter,
+)
+
+from evaluate_mmlu import eval_mmlu
+
+def replace_linear2qlinear(module, prefix="", **kwargs):
+    for n, m in module.named_children():
+        full_name = prefix + "." + n
+        if isinstance(m, nn.Linear) and "lm_head" not in n:  # hack for lm_head
+            qlinear = QuantLinear(
+                m.in_features,
+                m.out_features,
+                bit_width=kwargs.get("bit_width", None),
+                groupsize=kwargs.get("groupsize", -1),
+                enable_gscales_decompose=kwargs.get("enable_gscales_decompose", False),
+                enable_svd_deltaW=kwargs.get("enable_svd_deltaW", False),
+                svd_deltaW_rank=kwargs.get("svd_deltaW_rank", 8),
+            )
+            setattr(module, n, qlinear)
+        new_prefix = prefix + "." + n if prefix else n
+        replace_linear2qlinear(m, new_prefix, **kwargs)
+
+
+def unpack_model_weight(model):
+    for n, m in model.named_modules():
+        if isinstance(m, QuantLinear):
+            print("unpack module weight:", n)
+            m.unpack()
+
+
+def build_model(args, device_map="auto"):
+    if args.weight_dtype == "llm_int8":
+        model = LlamaForCausalLM.from_pretrained(
+            args.model_name,
+            load_in_8bit=True,
+            device_map=device_map,
+            low_cpu_mem_usage=True,
+        )
+    elif args.weight_dtype == "fp16":
+        model = LlamaForCausalLM.from_pretrained(
+            args.model_name,
+            low_cpu_mem_usage=True,
+            torch_dtype=torch.float16,
+            device_map=device_map,
+        )
+    elif args.weight_dtype == "uniform_qlora":
+        assert os.path.exists(
+            args.checkpoint
+        ), "loading uniform_qlora model requires checkpoint"
+        ckpt = torch.load(args.checkpoint)
+        cvt_hyparams = ckpt["hyparams"]
+        config = transformers.AutoConfig.from_pretrained(args.model_name)
+
+        def get_non_init_llama(config):
+            def skip(*args, **kwargs):
+                pass
+
+            torch.nn.init.kaiming_uniform_ = skip
+            torch.nn.init.uniform_ = skip
+            torch.nn.init.normal_ = skip
+            transformers.modeling_utils._init_weights = False
+            model = LlamaForCausalLM(config)
+            return model
+
+        model = get_non_init_llama(config)
+        replace_linear2qlinear(model, **cvt_hyparams)
+        model.load_state_dict(ckpt["model"])
+        device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+        model.to(device)
+        unpack_model_weight(model)
+        model = model.half()
+    else:
+        raise RuntimeError("unavaliable weight_dtype {}".format(args.weight_dtype))
+    model.eval()
+    return model
+
+
+def build_peft_model(args, model):
+    config = LoraConfig(
+        r=args.lora_rank,
+        lora_alpha=args.lora_alpha,
+        target_modules=args.lora_target_modules,
+        lora_dropout=args.lora_dropout,
+        bias="none",
+        task_type="QLORA_CAUSAL_LM"
+        if args.weight_dtype == "uniform_qlora"
+        else "CAUSAL_LM",
+    )
+    if args.weight_dtype == "uniform_qlora":
+        model.is_loaded_in_8bit = True  # hack for gradient-checkpoint
+        model = prepare_model_for_int8_training(model)
+        model.is_loaded_in_8bit = False
+        model = get_peft_qmodel(model, config)
+    else:
+        if args.weight_dtype == "llm_int8":
+            model = prepare_model_for_int8_training(model)
+        else:
+            model = prepare_model_for_fp16_training(model)
+        model = get_peft_model(model, config)
+    return model
+
+
+def build_tokenizer(args, model):
+    tokenizer = LlamaTokenizer.from_pretrained(args.model_name)
+    # unk. we want this to be different from the eos token
+    tokenizer.pad_token_id = 0
+    tokenizer.padding_side = "left"  # Allow batched inference
+    return tokenizer
+
+
+def build_dataset(args, tokenizer):
+    if args.dataset_path.endswith(".json") or args.dataset_path.endswith(".jsonl"):
+        dataset = load_dataset("json", data_files=args.dataset_path)
+    else:
+        dataset = load_dataset(args.dataset_path)
+
+    prompter = Prompter("alpaca")
+
+    def tokenize(prompt, add_eos_token=True):
+        # there's probably a way to do this with the tokenizer settings
+        # but again, gotta move fast
+        result = tokenizer(
+            prompt,
+            truncation=True,
+            max_length=args.cutoff_len,
+            padding=False,
+            return_tensors=None,
+        )
+        if (
+            result["input_ids"][-1] != tokenizer.eos_token_id
+            and len(result["input_ids"]) < args.cutoff_len
+            and add_eos_token
+        ):
+            result["input_ids"].append(tokenizer.eos_token_id)
+            result["attention_mask"].append(1)
+
+        result["labels"] = result["input_ids"].copy()
+
+        return result
+
+    def generate_and_tokenize_prompt(data_point):
+        full_prompt = prompter.generate_prompt(
+            data_point["instruction"],
+            data_point["input"],
+            data_point["output"],
+        )
+        tokenized_full_prompt = tokenize(full_prompt)
+        if False: #not train_on_inputs:
+            user_prompt = prompter.generate_prompt(
+                data_point["instruction"], data_point["input"]
+            )
+            tokenized_user_prompt = tokenize(
+                user_prompt, add_eos_token=args.add_eos_token
+            )
+            user_prompt_len = len(tokenized_user_prompt["input_ids"])
+
+            if args.add_eos_token:
+                user_prompt_len -= 1
+
+            tokenized_full_prompt["labels"] = [
+                -100
+            ] * user_prompt_len + tokenized_full_prompt["labels"][
+                user_prompt_len:
+            ]  # could be sped up, probably
+        return tokenized_full_prompt
+
+    if False:  # debug
+        TRAIN_SIZE = 500  # None
+        VAL_SIZE = 100  # 2000
+    else:
+        TRAIN_SIZE = None
+        VAL_SIZE = 2000
+
+    train_val = dataset["train"].train_test_split(
+        train_size=TRAIN_SIZE, test_size=VAL_SIZE, shuffle=True, seed=42
+    )
+    train_data = (
+        train_val["train"].shuffle().map(generate_and_tokenize_prompt)
+        #train_val["train"].map(lambda x: tokenize(generate_alpaca_prompt(x)))
+    )
+    val_data = train_val["test"].shuffle().map(generate_and_tokenize_prompt)
+    return train_data, val_data
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name", default="decapoda-research/llama-7b-hf")
+    parser.add_argument("--dataset_path", default="yahma/alpaca-cleaned")
+    parser.add_argument(
+        "--checkpoint", type=str, default="", help="the path of converted checkpoint"
+    )
+    # training params
+    parser.add_argument("--epochs", type=int, default=3)
+    parser.add_argument("--batch_size", type=int, default=128)
+    parser.add_argument("--micro_batch_size", type=int, default=4)
+    parser.add_argument("--lr", type=float, default=3e-4)
+    parser.add_argument("--cutoff_len", type=int, default=256)
+    parser.add_argument("--max_grad_norm", type=float, default=1.0)
+    # lora params
+    parser.add_argument("--lora_rank", type=int, required=True)
+    parser.add_argument("--lora_alpha", type=float, required=True)
+    parser.add_argument("--lora_dropout", type=float, required=True)
+    parser.add_argument(
+        "--lora_target_modules", type=str, nargs="+", default=["q_proj", "v_proj"]
+    )
+    # others
+    parser.add_argument(
+        "--weight_dtype", required=True, choices=["uniform_qlora", "llm_int8", "fp16"]
+    )
+    parser.add_argument(
+        "--group_by_length",
+        action="store_true",
+        help="a flag enable group data by length",
+    )
+    parser.add_argument(
+        "--add_eos_token",
+        action="store_true",
+        help="a flag enable group data by length",
+    )
+    parser.add_argument(
+        "--enable_mmlu_evaluation",
+        action="store_true",
+        help="a flag enable mmlu evaluation after finetuning",
+    )
+
+    parser.add_argument("--output_dir", type=str, required=True)
+
+    args = parser.parse_args()
+
+    gradient_accumulation_steps = args.batch_size // args.micro_batch_size
+
+    device_map = "auto"
+
+    world_size = int(os.environ.get("WORLD_SIZE", 1))
+    ddp = world_size != 1
+    if ddp:
+        device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
+        gradient_accumulation_steps = gradient_accumulation_steps // world_size
+
+    if int(os.environ.get("LOCAL_RANK", 0)) == 0:
+        os.makedirs(args.output_dir, exist_ok=True)
+        # show&save convert configure
+        for k, v in args.__dict__.items():
+            print(k + " : ", v)
+        with open(os.path.join(args.output_dir, "train_config.json"), "w") as f:
+            json.dump(args.__dict__, f)
+
+    # load model
+    model = build_model(args, device_map=device_map)
+    tokenizer = build_tokenizer(args, model)
+    model = build_peft_model(args, model)
+    tokenizer.save_pretrained(args.output_dir)
+
+    train_dataset, val_dataset = build_dataset(args, tokenizer)
+
+    model.save_pretrained(args.output_dir)
+
+    trainer = transformers.Trainer(
+        model=model,
+        train_dataset=train_dataset,
+        eval_dataset=val_dataset,
+        args=transformers.TrainingArguments(
+            per_device_train_batch_size=args.micro_batch_size,
+            gradient_accumulation_steps=gradient_accumulation_steps,
+            warmup_steps=100,
+            num_train_epochs=args.epochs,
+            learning_rate=args.lr,
+            fp16=True,
+            logging_steps=1,
+            optim="adamw_torch",
+            evaluation_strategy="steps",
+            save_strategy="steps",
+            eval_steps=200,
+            save_steps=200,
+            output_dir=args.output_dir,
+            save_total_limit=3,
+            load_best_model_at_end=True,
+            ddp_find_unused_parameters=False if ddp else None,
+            group_by_length=args.group_by_length,
+            max_grad_norm=args.max_grad_norm,
+            report_to="tensorboard",
+        ),
+        data_collator=transformers.DataCollatorForSeq2Seq(
+            tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
+        ),
+    )
+
+    model.config.use_cache = False
+
+    old_state_dict = model.state_dict
+    model.state_dict = (
+        lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
+    ).__get__(model, type(model))
+
+    trainer.train()
+
+    model.save_pretrained(args.output_dir)
+
+    if args.enable_mmlu_evaluation:
+        if args.weight_dtype == "fp16":
+            model = model.half()
+        print("Run mmlu evaluation")
+        eval_mmlu(model, args.model_name, tokenizer, save_dir=args.output_dir)