Upload 4 files

kraken_model/__init__.py

@@ -0,0 +1,55 @@
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_kraken_lora": ["KrakenConfig"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_kraken_lora"] = [
+        "KrakenForCausalLM",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_kraken_lora import KrakenConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_kraken_lora import (
+            KrakenForCausalLM,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

kraken_model/configuration_kraken_lora.py

@@ -0,0 +1,8 @@
+from transformers import PretrainedConfig
+
+class KrakenConfig(PretrainedConfig):
+    model_type = "kraken"
+    
+    def __init__(self, config_dict=None, **kwargs):
+        super().__init__(**kwargs)
+        self.config_dict = config_dict or {}

kraken_model/modeling_kraken_lora.py

@@ -0,0 +1,86 @@
+import torch
+from transformers import PreTrainedModel, AutoTokenizer, AutoModelForCausalLM, AutoModelForSequenceClassification, TextClassificationPipeline
+from configuration_kraken_lora import KrakenConfig
+import tokenizer_template_switch
+from peft import PeftModel, PeftConfig  # Import necessary modules for LoRA
+
+class KrakenForCausalLM(PreTrainedModel):
+    config_class = KrakenConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.tokenizers = {key: AutoTokenizer.from_pretrained(name, device_map="auto") for key, name in config.config_dict['tokenizers'].items()}
+        self.model = self.load_base_model(config.config_dict['models']['base'], config.config_dict['quantization']['base'])  # Load only expert1 as the base model
+        self.lora_adapters = config.config_dict['lora_adapters']  # Load LoRA adapter paths
+        self.router_model = AutoModelForSequenceClassification.from_pretrained(config.config_dict['router'], trust_remote_code=True, device_map="auto")
+        self.tokenizer = AutoTokenizer.from_pretrained(config.config_dict['router'], trust_remote_code=True, device_map="auto")
+        self.router = TextClassificationPipeline(model=self.router_model, tokenizer=self.tokenizer)
+        self.models_indices = config.config_dict['class_indices']
+
+    def load_base_model(self, model_name, quantization):
+        if quantization == "8bit":
+            return AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="auto", load_in_8bit=True, torch_dtype="auto")
+        elif quantization == "4bit":
+            return AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="auto", load_in_4bit=True, torch_dtype="auto")
+        elif quantization == "awq":
+            return self.load_awq_model(model_name)
+        else:
+            return AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="auto", torch_dtype="auto")
+
+    def load_awq_model(self, name):
+        return AutoModelForCausalLM.from_pretrained(name, trust_remote_code=True, device_map="auto")
+
+    def load_lora_adapter(self, base_model, adapter_path):
+        print("Loading adapter: "+adapter_path)
+        return PeftModel.from_pretrained(base_model, adapter_path)
+
+    def tokenize_inputs(self, text, adapter_key):
+        return self.tokenizers[adapter_key](text, return_tensors="pt")
+
+    def determine_adapter(self, text):
+        prediction = self.router(text)[0]["label"]
+        model_decision_index = self.models_indices[prediction]
+        adapter_keys = ['lora_expert1', 'lora_expert2', 'lora_expert3', 'lora_expert4', 'lora_expert5']
+        return adapter_keys[model_decision_index]
+
+    def expert_tokenizer(self, text):
+        adapter_key = self.determine_adapter(text)
+        return self.tokenizers[adapter_key]
+
+
+    def generate(self, input_ids, **generate_kwargs):
+        # Tokenize the input_ids
+        text = self.tokenizer.batch_decode(input_ids, skip_special_tokens=False)[0]
+
+        msgs = tokenizer_template_switch.recover_chat_messages(text, self.tokenizer)
+        if msgs and msgs[0]['role'] == 'system' and msgs[0]['content']=='<|im_start|>system':
+            # Delete the first element
+            msgs.pop(0)  
+        # Check if the last element has the role 'assistant'
+        if msgs and msgs[-1]['role'] == 'assistant':
+            # Delete the last element
+            msgs.pop()  
+
+
+        # Determine the appropriate LoRA adapter
+        adapter_key = self.determine_adapter(text)
+        print(f"Choosing LoRA adapter for {adapter_key} ..")
+        # Load and apply the LoRA adapter to the base model (expert1)
+        lora_adapter_path = self.lora_adapters[adapter_key]
+        model_with_lora = self.load_lora_adapter(self.model, lora_adapter_path)
+
+        # Use the tokenizer for the selected expert to tokenize the inputs
+        mod_txt = self.tokenizers[adapter_key].apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+        current_device = input_ids.device if isinstance(input_ids, torch.Tensor) else 'cpu'
+        
+        # Tokenize accordingly to the best model
+
+        tok = self.tokenizers[adapter_key](mod_txt, return_tensors="pt")
+        tok_input_ids = tok.input_ids.to(current_device)
+        tok_attention_mask = tok.attention_mask.to(current_device)
+
+        # Generate text using the modified model
+        return model_with_lora.generate(tok_input_ids, attention_mask=tok_attention_mask, **generate_kwargs)
+    
+
+      

kraken_model/tokenizer_template_switch.py

@@ -0,0 +1,95 @@
+import re
+from transformers import AutoTokenizer
+
+def extract_separators(template):
+    """
+    Extracts separators used in the tokenization template.
+    """
+    # Adjust the regex to correctly match the specific pattern between '{{' and '+ message["content"] +'
+    pattern = r"\{\{\s*([^{}]+?)\s*\+ message\['content'\]"
+    matches = re.findall(pattern, template)
+    # Clean up any extra spaces and return the matches
+    separators = [match.strip() for match in matches]
+
+    if any("message['role']" in element for element in separators):
+        roles = ["system", "user", "assistant"]
+        separators_ = []
+        for role in roles:
+            separators_.append(separators[0].replace(" + message['role'] + ", role).replace("'",""))
+        return separators_
+
+    return separators
+
+def detect_eos_token(jinja_template, tokenizer):
+    if "<|im_end|>" in jinja_template:
+        return "<|im_end|>"
+    if "</s>" in jinja_template:
+        return "</s>"
+    if "eos_token" in jinja_template:
+        return tokenizer.eos_token    
+    else:
+        return "<|endoftext|>"
+
+def recover_messages(formatted_message, separators, eos_token):
+    """
+    Recovers the original messages from the formatted message string.
+    """
+    # Split the formatted message using the end-of-string token
+    split_messages = formatted_message.split(eos_token)
+    
+    # Remove the last empty string if it exists due to a trailing separator
+    if split_messages and split_messages[-1].strip() == '':
+        split_messages.pop()
+
+    # Prepare the list to hold the recovered messages
+    recovered_messages = []
+
+    # Define roles after the first message, alternating between "user" and "assistant"
+    alternate_roles = ["user", "assistant"]
+    
+    # Iterate over the split messages
+    for index, message_content in enumerate(split_messages):
+        # Determine the role, starting with "system" for the first message
+        # then alternating between "user" and "assistant" for subsequent messages
+        if index == 0:
+            role = "system"
+        else:
+            role = alternate_roles[(index - 1) % 2]
+
+        # Clean the message content by removing leading/trailing whitespace and separators
+        clean_content = message_content.strip()
+        for separator in separators:
+            clean_content = clean_content.replace(separator.strip("'"), '', 1).strip()
+
+        # Append the cleaned message with its role to the list
+        recovered_messages.append({"role": role, "content": clean_content})
+
+    return recovered_messages
+
+def recover_chat_messages(tokenized_chat, tokenizer):
+    """
+    Given a tokenized_chat string and a tokenizer, returns the list of message dictionaries.
+    """
+    jinja_template = tokenizer.chat_template
+    separators = extract_separators(jinja_template)
+    eos_token = eos_token = detect_eos_token(jinja_template, tokenizer)
+    recovered_messages = recover_messages(tokenized_chat, separators, eos_token)
+    return recovered_messages
+
+# Example usage
+if __name__ == "__main__":
+    checkpoint = "Qwen/Qwen1.5-0.5B"
+    tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+    
+    messages = [
+        {
+            "role": "system",
+            "content": "You are a friendly chatbot who always responds in the style of a pirate",
+        },
+        {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
+    ]
+    tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=False)
+    print(tokenized_chat)
+    
+    recovered_messages = recover_chat_messages(tokenized_chat, tokenizer)
+    print(recovered_messages)