add gemma instruct chat template (#1341)

* add gemma instruct chat template

* support for chat tempalte strategy too

src/axolotl/prompt_strategies/chat_template.py

@@ -0,0 +1,78 @@
+"""
+HF Chat Templates prompt strategy
+"""
+from typing import Any, Dict, Optional
+
+from axolotl.prompt_tokenizers import PromptTokenizingStrategy
+from axolotl.prompters import Prompter
+from axolotl.utils.chat_templates import chat_templates
+
+
+class ChatTemplatePrompter(Prompter):
+    """prompter for HF chat templates"""
+
+    def __init__(self, tokenizer, chat_template=None, max_length=2048):
+        self.tokenizer = tokenizer
+        self.chat_template = chat_template
+        self.max_length = max_length
+
+    def build_prompt(self, conversation, add_generation_prompt=False):
+        return self.tokenizer.apply_chat_template(
+            conversation,
+            truncation=True,
+            max_length=self.max_length,
+            add_generation_prompt=add_generation_prompt,
+            chat_template=self.chat_template,
+        )
+
+
+class ChatTemplateStrategy(PromptTokenizingStrategy):
+    """
+    Tokenizing strategy for instruction-based prompts.
+    """
+
+    def tokenize_prompt(self, prompt):
+        turns = self.get_conversation_thread(prompt)
+        prompt_ids = self.prompter.build_prompt([turns[0]], add_generation_prompt=True)
+        input_ids = self.prompter.build_prompt(turns)
+
+        if not self.train_on_inputs:
+            user_prompt_len = len(prompt_ids)
+            labels = [-100] * user_prompt_len + input_ids[user_prompt_len:]
+        else:
+            labels = input_ids
+
+        tokenized_prompt = {
+            "input_ids": input_ids,
+            "labels": labels,
+            "attention_mask": [1] * len(input_ids),
+        }
+
+        return tokenized_prompt
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        # remap roles - allow for assistant turn
+        role_map = {
+            "human": "user",
+            "user": "user",
+            "assistant": "assistant",
+            "gpt": "assistant",
+        }
+        turns = [
+            {"role": role_map[t["from"]], "content": t["value"]} for t in conversations
+        ]
+        return turns
+
+
+def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    chat_template = (
+        ds_cfg["chat_template"] if ds_cfg and "chat_template" in ds_cfg else "chatml"
+    )
+    strategy = ChatTemplateStrategy(
+        ChatTemplatePrompter(tokenizer, chat_templates(chat_template)),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+    return strategy

src/axolotl/utils/chat_templates.py

@@ -22,6 +22,7 @@ def chat_templates(user_choice: str):
         "alpaca": "{% for message in messages %}{% if message['role'] == 'user' %}{{ '### Instruction: ' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ '### Response: ' + message['content'] + eos_token}}{% endif %}{% endfor %}",
         "inst": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",  # I don't know what this one is called. Used by Mistral/Mixtral.
         "chatml": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = 'You are a helpful assistant.' %}{% endif %}{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in loop_messages %}{% if loop.index0 == 0 %}{{'<|im_start|>system\n' + system_message + '<|im_end|>\n'}}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+        "gemma": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}",
     }
 
     if user_choice in templates:

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -76,6 +76,7 @@ class SFTDataset(BaseModel):
     type: Optional[Union[str, UserDefinedPrompterType]] = None
     shards: Optional[int] = None
     conversation: Optional[str] = None
+    chat_template: Optional[str] = None
     data_files: Optional[Union[str, List[str]]] = None
     name: Optional[str] = None
     ds_type: Optional[str] = None