Upload qat_sft.py with huggingface_hub

qat_sft.py

@@ -7,7 +7,7 @@ import re
 import torch
 import torch.distributed as dist
 
-from datasets import load_dataset
+from datasets import Dataset, load_dataset
 from torch.distributed.elastic.multiprocessing.api import record
 from torch.distributed.fsdp import (
     FullOptimStateDictConfig,
@@ -23,7 +23,7 @@ from torchao.prototype.parq.api import QuantConfig, create_optimizer
 from torchao.quantization.qat import QATConfig, IntxFakeQuantizeConfig
 from torchao.quantization.quant_api import quantize_
 from transformers import AutoTokenizer, Trainer
-from transformers.trainer import OPTIMIZER_NAME_BIN
+from transformers.trainer import OPTIMIZER_NAME, OPTIMIZER_NAME_BIN
 from transformers.trainer_utils import get_last_checkpoint
 from trl import SFTConfig, SFTTrainer, ScriptArguments, TrlParser
 from dataclasses import field
@@ -116,6 +116,8 @@ class PARQTrainer(Trainer):
             "Only single GPU is supported for torchao_convert from FSDP checkpoint"
         )
         optim_path = os.path.join(resume_from_checkpoint, OPTIMIZER_NAME_BIN)
+        if not os.path.isfile(optim_path):
+            optim_path = os.path.join(resume_from_checkpoint, OPTIMIZER_NAME)
         assert os.path.isfile(optim_path), f"Optimizer file not found at {optim_path}"
 
         if isinstance(resume_from_checkpoint, bool) and resume_from_checkpoint:
@@ -137,7 +139,7 @@ class PARQTrainer(Trainer):
         optimizer.load_state_dict(optimizer_state_dict)
 
         self.accelerator.print(
-            f"Loaded FSDP model and optimizer state from {resume_from_checkpoint}"
+            f"Loaded model and optimizer state from {resume_from_checkpoint}"
         )
         return True
 
@@ -292,6 +294,11 @@ def tokenize_fn(examples, tokenizer):
     }
 
 
+def disable_thinking(examples):
+    examples.setdefault("chat_template_kwargs", {})["enable_thinking"] = False
+    return examples
+
+
 @record
 def train(sft_config, custom_args, parq_config, world_size):
     tokenizer = AutoTokenizer.from_pretrained(custom_args.model_name_or_path)
@@ -337,14 +344,15 @@ def train(sft_config, custom_args, parq_config, world_size):
                 num_shards=sft_config.dataloader_num_workers * world_size,
             )
 
+    map_kwargs = (
+        {"num_proc": sft_config.dataset_num_proc}
+        if isinstance(dataset, Dataset)
+        else {}
+    )
+
     if not custom_args.enable_thinking:
         with sft_config.main_process_first(desc="disable thinking mode"):
-            dataset = dataset.map(
-                lambda x: x.update(
-                    {"chat_template_kwargs": {"enable_thinking": False}}
-                ),
-                num_proc=sft_config.dataset_num_proc,
-            )
+            dataset = dataset.map(disable_thinking, **map_kwargs)
 
     if prompt_format_fn is not None:
         sft_config.dataset_text_field = "messages"
@@ -361,12 +369,12 @@ def train(sft_config, custom_args, parq_config, world_size):
             dataset = dataset.map(
                 prompt_format_fn,
                 batched=True,
-                num_proc=sft_config.dataset_num_proc,
                 remove_columns=[
                     c
                     for c in dataset.column_names
                     if c != sft_config.dataset_text_field
                 ],
+                **map_kwargs,
             )
 
     if parq_config.weight_bits is not None: