fix sharegpt tokenization, refactor tokenization debugging

scripts/finetune.py

@@ -11,6 +11,8 @@ import yaml
 from attrdict import AttrDefault
 
 # add src to the pythonpath so we don't need to pip install this
+from axolotl.utils.tokenization import check_dataset_labels
+
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 src_dir = os.path.join(project_root, "src")
 sys.path.insert(0, src_dir)
@@ -42,36 +44,6 @@ def choose_device(cfg):
         cfg.device_map = {"": cfg.device}
 
 
-def check_dataset_labels(dataset, tokenizer):
-    from termcolor import colored
-
-    # the dataset is already shuffled, so let's just check the first 5 elements
-    for idx in range(5):
-        # Get the input_ids, labels, and attention_mask from the dataset
-        input_ids = dataset[idx]["input_ids"]
-        labels = dataset[idx]["labels"]
-        attention_mask = dataset[idx]["attention_mask"]
-
-        # You can compare the input_ids and labels element-wise
-        # Remember to ignore positions with IGNORE_TOKEN_ID (if you use it) or attention_mask equal to 0
-        colored_tokens = []
-        for i, (input_id, label_id, mask) in enumerate(
-            zip(input_ids, labels, attention_mask)
-        ):
-            decoded_input_token = tokenizer.decode(input_id)
-            # Choose the color based on whether the label has the ignore value or not
-            color = (
-                "red" if label_id == -100 else ("yellow" if label_id == 0 else "green")
-            )
-            colored_token = colored(decoded_input_token, color) + colored(
-                f"({label_id}, {mask})", "white"
-            )
-            colored_tokens.append(colored_token)
-
-        logging.info(" ".join(colored_tokens))
-        logging.info("\n\n\n")
-
-
 def do_inference(cfg, model, tokenizer):
     tokenizer.add_special_tokens({"unk_token": "<unk>"})
     tokenizer.add_special_tokens({"bos_token": "<s>"})
@@ -199,8 +171,9 @@ def train(
         return
 
     if cfg.debug:
+        logging.info("check_dataset_labels...")
         check_dataset_labels(
-            train_dataset.select([random.randrange(0, len(train_dataset) - 1)]),
+            train_dataset.select([random.randrange(0, len(train_dataset) - 1) for i in range(5)]),
             tokenizer,
         )
 

src/axolotl/prompters.py

@@ -127,7 +127,7 @@ conv_vicuna_v1_1 = Conversation(
 
 
 class ShareGPTPrompter:
-    def build_prompt(self, source, tokenizer):
+    def build_prompt(self, source, tokenizer, sequence_len=2048):
         # ignore the system prompt if provided
         if source[0]["from"] == "system":
             source.pop(0)
@@ -157,13 +157,14 @@ class ShareGPTPrompter:
             role = roles[sentence["from"]]
             assert role == conv.roles[j % 2]
             conv.append_message(role, sentence["value"])
+        # TODO, this concatenates everything, but doesn't seem to properly add the eos_token_id, as the eos_token gets split up
         conversation = conv.get_prompt()
 
         # Tokenize conversations
         tokenized_result = tokenizer(
             conversation,
             truncation=True,
-            max_length=2048,  # FIXME
+            max_length=sequence_len,  # FIXME
             padding=False,
             return_tensors=None,
         )
@@ -173,7 +174,9 @@ class ShareGPTPrompter:
         sep = conv.sep + conv.roles[1] + ": "
 
         rounds = conversation.split(conv.sep2)
+        rounds = [r + conv.sep2 for r in rounds]
         cur_len = 1
+        target[0] = IGNORE_TOKEN_ID  # mask out the bos
         for i, rou in enumerate(rounds):
             if rou == "":
                 break
@@ -182,19 +185,27 @@ class ShareGPTPrompter:
             if len(parts) != 2:
                 break
             parts[0] += sep
-            round_len = len(tokenizer(rou)["input_ids"])
+            round_len = len(tokenizer(rou)["input_ids"]) - 1  # -1 ignores the bos_token generated for this
-            instruction_len = len(tokenizer(parts[0])["input_ids"]) - 2
+            # we have to strip the initial part, any dangling whitespace creates an additional ghost token
+            instruction_len = len(tokenizer(parts[0].strip())["input_ids"]) - 1  # -1 ignores the bos_token generated for this
             target[cur_len : cur_len + instruction_len] = [
                 IGNORE_TOKEN_ID
             ] * instruction_len
 
             cur_len += round_len
-        target[cur_len:] = [IGNORE_TOKEN_ID] * (len(target) - cur_len)
+            if cur_len >= sequence_len:
+                break
+
+        # Fix: Truncate the target to have the same length as input_ids
+        target = target[:len(tokenized_result["input_ids"])]
+        # target[cur_len:] = [IGNORE_TOKEN_ID] * (len(target) - cur_len)
+
         attention_mask = [
             1 if x != tokenizer.pad_token_id else 0
             for x in tokenized_result["input_ids"]
         ]
 
+        # TODO truncate len to sequence_len
         return dict(
             input_ids=tokenized_result["input_ids"],
             labels=target,

src/axolotl/utils/models.py

@@ -53,7 +53,7 @@ def load_model(
         logging.info("patching with xformers attention")
         hijack_llama_attention()
 
-    torch_dtype = (torch.float16 if cfg.load_in_8bit or cfg.fp16 else torch.float32,)
+    torch_dtype = torch.float16 if cfg.load_in_8bit or cfg.fp16 or cfg.bf16 else torch.float32
     try:
         if cfg.load_4bit:
             from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
@@ -161,11 +161,11 @@ def load_model(
         tokenizer.add_special_tokens({"pad_token": "[PAD]"})
         os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
-    if cfg.special_tokens:
+    if cfg.tokens:
-        for k, v in cfg.special_tokens.items():
+        for k, v in cfg.tokens.items():
-            setattr(tokenizer, k, v)
+            tokenizer.add_special_tokens({k: v})
 
-    if load_in_8bit and not cfg.load_4bit:
+    if load_in_8bit and cfg.load_4bit:
         logging.info("converting model w/ prepare_model_for_int8_training")
         model = prepare_model_for_int8_training(model)
 

src/axolotl/utils/tokenization.py

@@ -0,0 +1,33 @@
+from termcolor import colored
+import logging
+
+def check_dataset_labels(dataset, tokenizer):
+    # the dataset is already shuffled, so let's just check the first 5 elements
+    for idx in range(5):
+        check_example_labels(dataset[idx], tokenizer)
+
+
+def check_example_labels(example, tokenizer):
+    # Get the input_ids, labels, and attention_mask from the dataset
+    input_ids = example["input_ids"]
+    labels = example["labels"]
+    attention_mask =example["attention_mask"]
+
+    # You can compare the input_ids and labels element-wise
+    # Remember to ignore positions with IGNORE_TOKEN_ID (if you use it) or attention_mask equal to 0
+    colored_tokens = []
+    for i, (input_id, label_id, mask) in enumerate(
+        zip(input_ids, labels, attention_mask)
+    ):
+        decoded_input_token = tokenizer.decode(input_id)
+        # Choose the color based on whether the label has the ignore value or not
+        color = (
+            "red" if label_id == -100 else ("yellow" if label_id == 0 else "green")
+        )
+        colored_token = colored(decoded_input_token, color) + colored(
+            f"({label_id}, {mask}, {input_id})", "white"
+        )
+        colored_tokens.append(colored_token)
+
+    logging.info(" ".join(colored_tokens))
+    logging.info("\n\n\n")

src/axolotl/utils/trainer.py

@@ -61,6 +61,11 @@ 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,
+        optim=cfg.optimizer if cfg.optimizer != "adam8bit" else cfg.optimizer,
+        lr_scheduler_type=cfg.lr_scheduler if cfg.lr_scheduler else None,
+        weight_decay=cfg.weight_decay if cfg.weight_decay else 0.0,
+        fsdp=cfg.fsdp.split(" ") if cfg.fsdp else None,
+        fsdp_transformer_layer_cls_to_wrap=cfg.fsdp_transformer_layer_cls_to_wrap if cfg.fsdp_transformer_layer_cls_to_wrap else None,
         **training_arguments_kwargs,
     )