refactor(param): rename load_4bit config param by gptq

README.md

@@ -176,7 +176,7 @@ tokenizer_type: AutoTokenizer
 trust_remote_code:
 
 # whether you are training a 4-bit GPTQ quantized model
-load_4bit: true
+gptq: true
 gptq_groupsize: 128 # group size
 gptq_model_v1: false # v1 or v2
 

configs/quickstart.yml

@@ -40,6 +40,6 @@ early_stopping_patience: 3
 resume_from_checkpoint:
 auto_resume_from_checkpoints: true
 local_rank:
-load_4bit: true
+gptq: true
 xformers_attention: true
 flash_attention:

examples/4bit-lora-7b/config.yml

@@ -4,7 +4,7 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 trust_remote_code:
 load_in_8bit: true
-load_4bit: true
+gptq: true
 datasets:
   - path: vicgalle/alpaca-gpt4
     type: alpaca

src/axolotl/utils/models.py

@@ -73,7 +73,7 @@ def load_model(
     else:
         torch_dtype = torch.float32
     try:
-        if cfg.load_4bit:
+        if cfg.gptq:
             from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
                 replace_peft_model_with_int4_lora_model,
             )
@@ -95,7 +95,7 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     try:
-        if cfg.load_4bit and is_llama_derived_model:
+        if cfg.gptq and is_llama_derived_model:
             from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
             from huggingface_hub import snapshot_download
 
@@ -248,7 +248,7 @@ def load_model(
 
     if (
         ((cfg.adapter == "lora" and load_in_8bit) or cfg.adapter == "qlora")
-        and not cfg.load_4bit
+        and not cfg.gptq
         and (load_in_8bit or cfg.load_in_4bit)
     ):
         logging.info("converting PEFT model w/ prepare_model_for_int8_training")
@@ -259,7 +259,7 @@ def load_model(
     if cfg.ddp and not load_in_8bit:
         model.to(f"cuda:{cfg.local_rank}")
 
-    if cfg.load_4bit:
+    if cfg.gptq:
         # Scales to half
         logging.info("Fitting 4bit scales and zeros to half")
         for n, m in model.named_modules():
@@ -274,7 +274,7 @@ def load_model(
     if (
         torch.cuda.device_count() > 1
         and int(os.getenv("WORLD_SIZE", "1")) > 1
-        and cfg.load_4bit
+        and cfg.gptq
     ):
         # llama is PROBABLY model parallelizable, but the default isn't that it is
         # so let's only set it for the 4bit, see

src/axolotl/utils/trainer.py

@@ -63,7 +63,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     training_arguments_kwargs["warmup_steps"] = warmup_steps
     training_arguments_kwargs["logging_steps"] = logging_steps
     if cfg.gradient_checkpointing is not None:
-        if cfg.load_4bit:
+        if cfg.gptq:
             from alpaca_lora_4bit.gradient_checkpointing import (
                 apply_gradient_checkpointing,
             )
@@ -138,7 +138,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
             importlib.import_module("torchdistx")
     if (
         cfg.optimizer == "adamw_bnb_8bit"
-        and not cfg.load_4bit
+        and not cfg.gptq
         and not "deepspeed" in training_arguments_kwargs
         and not cfg.fsdp
     ):

src/axolotl/utils/validation.py

@@ -2,16 +2,20 @@ import logging
 
 
 def validate_config(cfg):
+    if cfg.load_4bit:
+        raise ValueError("cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq")
+
     if cfg.adapter == "qlora":
         if cfg.merge_lora:
             # can't merge qlora if loaded in 8bit or 4bit
             assert cfg.load_in_8bit is False
-            assert cfg.load_4bit is False
+            assert cfg.gptq is False
             assert cfg.load_in_4bit is False
         else:
             assert cfg.load_in_8bit is False
-            assert cfg.load_4bit is False
+            assert cfg.gptq is False
             assert cfg.load_in_4bit is True
+
     if not cfg.load_in_8bit and cfg.adapter == "lora":
         logging.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")