4bit quantized support (wip)

README.md

@@ -29,8 +29,8 @@ shuf -n2000 data/vicuna_cleaned.jsonl > data/vicuna_cleaned.subset0.jsonl
 ```
 
 - Create a new or update the existing YAML config (config/pythia_1_2B_alpaca.yml)[config/pythia_1_2B_alpaca.yml]
-- Install python dependencies `pip3 install -r requirements.txt`
-- Configure accelerate `accelerate launch` or update `~/.cache/huggingface/accelerate/default_config.yaml`
+- Install python dependencies `pip3 install -e .[triton]` or `pip3 install -e .[cuda]`
+- Configure accelerate `accelerate config` or update `~/.cache/huggingface/accelerate/default_config.yaml`
 
 ```yaml
 compute_environment: LOCAL_MACHINE

configs/cerebras_1_3B_alpaca.yml

@@ -11,7 +11,7 @@ datasets:
     type: gpteacher
   - path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
     type: gpteacher
-dataset_prepared_path: data/last_run
+dataset_prepared_path: last_run_prepared
 val_set_size: 0.05
 adapter: lora
 sequence_len: 2048
@@ -34,6 +34,7 @@ train_on_inputs: false
 group_by_length: false
 bf16: True
 tf32: True
+gradient_checkpointing:
 early_stopping_patience:
 resume_from_checkpoint:
 local_rank:

configs/llama_65B_alpaca.yml

@@ -11,7 +11,7 @@ datasets:
     type: gpteacher
   - path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
     type: gpteacher
-dataset_prepared_path: data/last_run
+dataset_prepared_path: last_run_prepared
 val_set_size: 0.04
 adapter: lora
 lora_model_dir:

configs/llama_7B_alpaca.yml

@@ -11,7 +11,7 @@ datasets:
     type: gpteacher
   - path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
     type: gpteacher
-dataset_prepared_path: data/last_run
+dataset_prepared_path: last_run_prepared
 val_set_size: 0.04
 adapter: lora
 lora_model_dir:

configs/pythia_1_2B_alpaca.yml

@@ -11,7 +11,7 @@ datasets:
     type: gpteacher
   - path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
     type: gpteacher
-dataset_prepared_path: data/last_run
+dataset_prepared_path: last_run_prepared
 val_set_size: 0.05
 adapter: lora
 lora_model_dir:

pyproject.toml

@@ -1,3 +0,0 @@
-[build-system]
-requires = ["setuptools", "wheel"]
-build-backend = "setuptools.build_meta"

requirements.txt

@@ -1,5 +1,4 @@
-git+https://github.com/huggingface/peft.git
-git+https://github.com/huggingface/transformers.git
+transformers @ git+https://github.com/huggingface/transformers.git
 attrdict
 fire
 PyYAML==6.0

scripts/finetune.py

@@ -13,12 +13,6 @@ import transformers
 import yaml
 from attrdict import AttrDefault
 from datasets import load_dataset, IterableDataset, Dataset, load_from_disk
-from peft import (
-    LoraConfig,
-    get_peft_model,
-    prepare_model_for_int8_training,
-    PeftModel,
-)
 from torch import nn
 from transformers import (
     AutoModelForCausalLM,
@@ -45,7 +39,7 @@ from axolotl.prompt_tokenizers import (
 from axolotl.prompters import AlpacaPrompter, GPTeacherPrompter, ShareGPTPrompter
 
 logging.basicConfig(level=os.getenv("LOG_LEVEL", "INFO"))
-DEFAULT_DATASET_PREPARED_PATH = "data/last_run"
+DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
 
 
 def setup_wandb_env_vars(cfg):
@@ -60,7 +54,11 @@ def setup_wandb_env_vars(cfg):
             os.environ["WANDB_RUN_ID"] = cfg.wandb_run_id
 
 
-def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora", inference: bool=False):
+def load_model(base_model, base_model_config, model_type, tokenizer_type, cfg, adapter="lora", inference: bool=False):
+    # TODO refactor as a kwarg
+    load_in_8bit = cfg.load_in_8bit
+    tokenizer = None
+
     if adapter != "lora":
         raise NotImplementedError(f"{adapter} peft adapter not available")
     if "llama" in base_model:
@@ -70,7 +68,43 @@ def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora", infe
 
     torch_dtype = torch.float16 if cfg.load_in_8bit or cfg.fp16 else torch.float32,
     try:
-        if "llama" in base_model:
+        if cfg.load_4bit:
+            from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import replace_peft_model_with_int4_lora_model
+            replace_peft_model_with_int4_lora_model()
+
+        from peft import (
+            LoraConfig,
+            get_peft_model,
+            prepare_model_for_int8_training,
+            PeftModel,
+        )
+    except Exception as e:
+        logging.exception(e)
+        raise e
+
+    try:
+        if cfg.load_4bit and "llama" in base_model:
+            from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
+            from huggingface_hub import snapshot_download
+
+            cache_model_path = Path(snapshot_download(base_model))
+            # TODO search .glob for a .pt, .safetensor, or .bin
+            cache_model_path.glob("*.pt")
+            files = list(cache_model_path.glob('*.pt')) + list(cache_model_path.glob('*.safetensor')) + list(cache_model_path.glob('*.bin'))
+            if len(files) > 0:
+                model_path = str(files[0])
+            else:
+                logging.warning("unable to find a cached model file, this will likely fail...")
+                model_path = str(cache_model_path)
+            model, tokenizer = load_llama_model_4bit_low_ram(
+                base_model_config if base_model_config else base_model,
+                model_path,
+                device_map=cfg.device_map,
+                groupsize=-1,
+                is_v1_model=True,
+            )
+            load_in_8bit = False
+        elif "llama" in base_model:
             model = LlamaForCausalLM.from_pretrained(
                 base_model,
                 load_in_8bit=cfg.load_in_8bit,
@@ -92,13 +126,14 @@ def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora", infe
             device_map=cfg.device_map,
         )
 
-    try:
-        if "llama" in base_model:
-            tokenizer = LlamaTokenizer.from_pretrained(model)
-        else:
-            tokenizer = getattr(transformers, tokenizer_type).from_pretrained(model)
-    except:
-        tokenizer = AutoTokenizer.from_pretrained(base_model)
+    if not tokenizer:
+        try:
+            if "llama" in base_model:
+                tokenizer = LlamaTokenizer.from_pretrained(model)
+            else:
+                tokenizer = getattr(transformers, tokenizer_type).from_pretrained(model)
+        except:
+            tokenizer = AutoTokenizer.from_pretrained(base_model)
 
     if tokenizer.__class__.__name__ in ["LlamaTokenizer", "LlamaTokenizerFast"]:
         tokenizer.pad_token = LLAMA_DEFAULT_PAD_TOKEN
@@ -107,7 +142,7 @@ def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora", infe
         tokenizer.add_special_tokens({"pad_token": "[PAD]"})
         os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
-    if cfg.load_in_8bit:
+    if load_in_8bit:
         model = prepare_model_for_int8_training(model)
 
     lora_config = LoraConfig(
@@ -128,6 +163,16 @@ def load_model(base_model, model_type, tokenizer_type, cfg, adapter="lora", infe
     if cfg.ddp:
         model.to(f"cuda:{cfg.local_rank}")
 
+    if cfg.load_4bit:
+        # Scales to half
+        print('Fitting 4bit scales and zeros to half')
+        for n, m in model.named_modules():
+            if 'Autograd4bitQuantLinear' in str(type(m)) or 'Linear4bitLt' in str(type(m)):
+                if hasattr(m, "is_v1_model") and m.is_v1_model:
+                    m.zeros = m.zeros.half()
+                m.scales = m.scales.half()
+                m.bias = m.bias.half()
+
     # TODO resume_from_checkpoint handling
     model.print_trainable_parameters()
     return model, tokenizer, lora_config
@@ -243,6 +288,8 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     training_arguments_kwargs["tf32"] = cfg.tf32
     training_arguments_kwargs["warmup_steps"] = warmup_steps
     training_arguments_kwargs["logging_steps"] = logging_steps
+    if cfg.gradient_checkpointing is not None:
+        training_arguments_kwargs["gradient_checkpointing"] = cfg.gradient_checkpointing
 
     training_args = transformers.TrainingArguments(
         per_device_train_batch_size=cfg.micro_batch_size,
@@ -260,7 +307,6 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         group_by_length=cfg.group_by_length,
         report_to="wandb" if cfg.use_wandb else None,
         run_name=cfg.wandb_run_id if cfg.use_wandb else None,
-        gradient_checkpointing=cfg.gradient_checkpointing,
         **training_arguments_kwargs,
     )
 
@@ -356,11 +402,13 @@ def train(
         cfg.bf16 = False
 
     # Load the model and tokenizer
+    logging.info("loading model, tokenizer, and lora_config...")
     model, tokenizer, lora_config = load_model(
-        cfg.base_model, cfg.model_type, cfg.tokenizer_type, cfg, adapter=cfg.adapter, inference=("inference" in kwargs)
+        cfg.base_model, cfg.base_model_config, cfg.model_type, cfg.tokenizer_type, cfg, adapter=cfg.adapter, inference=("inference" in kwargs)
     )
 
     if "inference" in kwargs:
+        logging.info("calling do_inference function")
         do_inference(cfg, model, tokenizer)
         return
 
@@ -369,6 +417,7 @@ def train(
         dataset = load_from_disk(cfg.dataset_prepared_path)
         logging.info("Prepared dataset loaded from disk...")
     else:
+        logging.info("Loading raw datasets...")
         datasets = []
         for d in cfg.datasets:
             if Path(d.path).exists():
@@ -402,6 +451,7 @@ def train(
         constant_len_dataset = ConstantLengthDataset(
             tokenizer, datasets, seq_length=cfg.sequence_len
         )
+        logging.info("merging, packing, shuffling, and splitting master dataset")
         dataset = Dataset.from_list(
             [_ for _ in constant_len_dataset]
         ).train_test_split(test_size=cfg.val_set_size, shuffle=True, seed=42)

setup.cfg

@@ -1,33 +0,0 @@
-[metadata]
-name = axolotl
-version = 0.1.0
-description = You know you're going to axolotl questions
-author = Wing Lian
-author_email = [email protected]
-license = MIT
-
-[options]
-package_dir =
-    =src
-packages = find:
-install_requires =
-    transformers @ git+https://github.com/huggingface/transformers.git@main
-    peft @ git+https://github.com/huggingface/peft.git@main
-    attrdict
-    fire
-    PyYAML == 6.0
-    black
-    bitsandbytes
-    datasets
-    accelerate
-    sentencepiece
-    wandb
-    flash-attn
-    einops
-
-[options.packages.find]
-where = src
-
-[options.extras_require]
-gptq_cuda = alpaca_lora_4bit[cuda] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip#egg=alpaca_lora_4bit[cuda]
-gptq_triton = alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip#egg=alpaca_lora_4bit[triton]

setup.py

@@ -0,0 +1,30 @@
+import sys
+from setuptools import setup, find_packages
+
+install_requires = []
+with open("./requirements.txt", "r") as requirements_file:
+    # don't include peft yet until we check the int4
+    reqs = [r.strip() for r in requirements_file.readlines() if "peft" not in r]
+    reqs = [r for r in reqs if r[0] != "#"]
+    for r in reqs:
+        install_requires.append(r)
+
+setup(
+    name='axolotl',
+    version='0.1',
+    description="You know you're going to axolotl questions",
+    package_dir={'': 'src'},
+    packages=find_packages(),
+    install_requires=install_requires,
+    extras_require={
+        None: [
+            "peft @ git+https://github.com/huggingface/peft.git",
+        ],
+        'int4_cuda': [
+            "alpaca_lora_4bit[cuda] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip#egg=alpaca_lora_4bit[cuda]",
+        ],
+        'int4_triton': [
+            "alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip#egg=alpaca_lora_4bit[triton]",
+        ],
+    },
+)

src/axolotl/datasets.py

@@ -31,6 +31,7 @@ class TokenizedPromptDataset(IterableDataset):
                 pass
 
 
+# TODO this isn't the best since it can't interleave datasets
 class ConstantLengthDataset(IterableDataset):
     """
     Iterable dataset that returns constant length chunks of tokens from stream of text files.