Add application file

app.py

@@ -0,0 +1,220 @@
+import os
+import sys
+
+import fire
+import gradio as gr
+import torch
+import transformers
+from peft import PeftModel
+from transformers import GenerationConfig, LlamaForCausalLM, LlamaTokenizer
+
+from utils.callbacks import Iteratorize, Stream
+from utils.prompter import Prompter
+
+if torch.cuda.is_available():
+    device = "cuda"
+else:
+    device = "cpu"
+
+try:
+    if torch.backends.mps.is_available():
+        device = "mps"
+except:  # noqa: E722
+    pass
+
+
+def main(
+    load_8bit: bool = False,
+    base_model: str = "",
+    lora_weights: str = "tloen/alpaca-lora-7b",
+    prompt_template: str = "",  # The prompt template to use, will default to alpaca.
+    server_name: str = "0.0.0.0",  # Allows to listen on all interfaces by providing '0.
+    share_gradio: bool = False,
+):
+    print("lora_weights: " + str(lora_weights))
+    base_model = base_model or os.environ.get("BASE_MODEL", "")
+    assert (
+        base_model
+    ), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'"
+
+    prompter = Prompter(prompt_template)
+    tokenizer = LlamaTokenizer.from_pretrained(base_model)
+    if device == "cuda":
+        model = LlamaForCausalLM.from_pretrained(
+            base_model,
+            load_in_8bit=load_8bit,
+            torch_dtype=torch.float16,
+            device_map="auto",
+        )
+        model = PeftModel.from_pretrained(
+            model,
+            lora_weights,
+            torch_dtype=torch.float16,
+        )
+    elif device == "mps":
+        model = LlamaForCausalLM.from_pretrained(
+            base_model,
+            device_map={"": device},
+            torch_dtype=torch.float16,
+        )
+        model = PeftModel.from_pretrained(
+            model,
+            lora_weights,
+            device_map={"": device},
+            torch_dtype=torch.float16,
+        )
+    else:
+        model = LlamaForCausalLM.from_pretrained(
+            base_model, device_map={"": device}, low_cpu_mem_usage=True
+        )
+        model = PeftModel.from_pretrained(
+            model,
+            lora_weights,
+            device_map={"": device},
+        )
+
+    # unwind broken decapoda-research config
+    model.config.pad_token_id = tokenizer.pad_token_id = 0  # unk
+    model.config.bos_token_id = 1
+    model.config.eos_token_id = 2
+
+    if not load_8bit:
+        model.half()  # seems to fix bugs for some users.
+
+    model.eval()
+    if torch.__version__ >= "2" and sys.platform != "win32":
+        model = torch.compile(model)
+
+    def evaluate(
+        instruction,
+        input=None,
+        temperature=0.1,
+        top_p=0.75,
+        top_k=40,
+        num_beams=4,
+        max_new_tokens=128,
+        stream_output=False,
+        **kwargs,
+    ):
+        prompt = prompter.generate_prompt(instruction, input)
+        inputs = tokenizer(prompt, return_tensors="pt")
+        input_ids = inputs["input_ids"].to(device)
+        generation_config = GenerationConfig(
+            temperature=temperature,
+            top_p=top_p,
+            top_k=top_k,
+            num_beams=num_beams,
+            **kwargs,
+        )
+
+        generate_params = {
+            "input_ids": input_ids,
+            "generation_config": generation_config,
+            "return_dict_in_generate": True,
+            "output_scores": True,
+            "max_new_tokens": max_new_tokens,
+        }
+
+        if stream_output:
+            # Stream the reply 1 token at a time.
+            # This is based on the trick of using 'stopping_criteria' to create an iterator,
+            # from https://github.com/oobabooga/text-generation-webui/blob/ad37f396fc8bcbab90e11ecf17c56c97bfbd4a9c/modules/text_generation.py#L216-L243.
+
+            def generate_with_callback(callback=None, **kwargs):
+                kwargs.setdefault(
+                    "stopping_criteria", transformers.StoppingCriteriaList()
+                )
+                kwargs["stopping_criteria"].append(
+                    Stream(callback_func=callback)
+                )
+                with torch.no_grad():
+                    model.generate(**kwargs)
+
+            def generate_with_streaming(**kwargs):
+                return Iteratorize(
+                    generate_with_callback, kwargs, callback=None
+                )
+
+            with generate_with_streaming(**generate_params) as generator:
+                for output in generator:
+                    # new_tokens = len(output) - len(input_ids[0])
+                    decoded_output = tokenizer.decode(output)
+
+                    if output[-1] in [tokenizer.eos_token_id]:
+                        break
+
+                    yield prompter.get_response(decoded_output)
+            return  # early return for stream_output
+
+        # Without streaming
+        with torch.no_grad():
+            generation_output = model.generate(
+                input_ids=input_ids,
+                generation_config=generation_config,
+                return_dict_in_generate=True,
+                output_scores=True,
+                max_new_tokens=max_new_tokens,
+            )
+        s = generation_output.sequences[0]
+        output = tokenizer.decode(s)
+        yield prompter.get_response(output)
+
+    gr.Interface(
+        fn=evaluate,
+        inputs=[
+            gr.components.Textbox(
+                lines=2,
+                label="Instruction",
+                placeholder="Generate an Ad for the iPhone 14.",
+            ),
+            gr.components.Textbox(lines=2, label="Input", placeholder="none"),
+            gr.components.Slider(
+                minimum=0, maximum=1, value=0.1, label="Temperature"
+            ),
+            gr.components.Slider(
+                minimum=0, maximum=1, value=0.75, label="Top p"
+            ),
+            gr.components.Slider(
+                minimum=0, maximum=100, step=1, value=40, label="Top k"
+            ),
+            gr.components.Slider(
+                minimum=1, maximum=4, step=1, value=4, label="Beams"
+            ),
+            gr.components.Slider(
+                minimum=1, maximum=2000, step=1, value=128, label="Max tokens"
+            ),
+            gr.components.Checkbox(label="Stream output"),
+        ],
+        outputs=[
+            gr.inputs.Textbox(
+                lines=5,
+                label="Output",
+            )
+        ],
+        title="🦙🛍️ LLaMA-E-LoRA",
+        description="LLaMA-E-LoRA is a series of fine-tuned LLaMA model following the E-commerce instructions. It is developed by DSMI (http://dsmi.tech/) @ University of Technology Sydney, and trained on the 120k instruction set. This model is for academic research use only. For more details please contact: Kaize.Shi@uts.edu.au",
+        # noqa: E501
+    ).queue().launch(server_name="0.0.0.0", share=share_gradio)
+    # Old testing code follows.
+
+    """
+    # testing code for readme
+    for instruction in [
+        "Tell me about alpacas.",
+        "Tell me about the president of Mexico in 2019.",
+        "Tell me about the king of France in 2019.",
+        "List all Canadian provinces in alphabetical order.",
+        "Write a Python program that prints the first 10 Fibonacci numbers.",
+        "Write a program that prints the numbers from 1 to 100. But for multiples of three print 'Fizz' instead of the number and for the multiples of five print 'Buzz'. For numbers which are multiples of both three and five print 'FizzBuzz'.",  # noqa: E501
+        "Tell me five words that rhyme with 'shock'.",
+        "Translate the sentence 'I have no mouth but I must scream' into Spanish.",
+        "Count up from 1 to 500.",
+    ]:
+        print("Instruction:", instruction)
+        print("Response:", evaluate(instruction))
+        print()
+    """
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

requirements.txt

@@ -0,0 +1,12 @@
+accelerate
+appdirs
+loralib
+bitsandbytes
+black
+black[jupyter]
+datasets
+fire
+git+https://github.com/huggingface/peft.git
+transformers>=4.28.0
+sentencepiece
+gradio

templates/README.md

@@ -0,0 +1,46 @@
+# Prompt templates
+
+This directory contains template styles for the prompts used to finetune LoRA models.
+
+## Format
+
+A template is described via a JSON file with the following keys:
+
+- `prompt_input`: The template to use when input is not None. Uses `{instruction}` and `{input}` placeholders.
+- `prompt_no_input`: The template to use when input is None. Uses `{instruction}` placeholders.
+- `description`: A short description of the template, with possible use cases.
+- `response_split`: The text to use as separator when cutting real response from the model output.
+
+No `{response}` placeholder was used, since the response is always the last element of the template and is just to be concatenated to the rest.
+
+## Example template
+
+The default template, used unless otherwise specified, is `alpaca.json`
+
+```json
+{
+    "description": "Template used by Alpaca-LoRA.",
+    "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+    "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n",
+    "response_split": "### Response:"    
+}
+
+```
+
+## Current templates
+
+### alpaca
+
+Default template used for generic LoRA fine tunes so far.
+
+### alpaca_legacy
+
+Legacy template used by the original alpaca repo, with no `\n` after the response field. Kept for reference and experiments.
+
+### alpaca_short
+
+A trimmed down alpaca template which seems to perform just as well and spare some tokens. Models created with the default template seem to be queryable by the short tempalte as well. More experiments are welcome.
+
+### vigogne
+
+The default alpaca template, translated to french. This template was used to train the "Vigogne" LoRA and is to be used to query it, or for extra fine tuning.

templates/alpaca.json

@@ -0,0 +1,6 @@
+{
+    "description": "Template used by Alpaca-LoRA.",
+    "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+    "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n",
+    "response_split": "### Response:"    
+}

templates/alpaca_legacy.json

@@ -0,0 +1,6 @@
+{
+    "description": "Legacy template, used by Original Alpaca repository.",
+    "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:",
+    "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:",
+    "response_split": "### Response:"    
+}

templates/alpaca_short.json

@@ -0,0 +1,6 @@
+{
+    "description": "A shorter template to experiment with.",
+    "prompt_input": "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+    "prompt_no_input": "### Instruction:\n{instruction}\n\n### Response:\n",
+    "response_split": "### Response:"    
+}

templates/vigogne.json

@@ -0,0 +1,6 @@
+{
+    "description": "French template, used by Vigogne for finetuning.",
+    "prompt_input": "Ci-dessous se trouve une instruction qui décrit une tâche, associée à une entrée qui fournit un contexte supplémentaire. Écrivez une réponse qui complète correctement la demande.\n\n### Instruction:\n{instruction}\n\n### Entrée:\n{input}\n\n### Réponse:\n",
+    "prompt_no_input": "Ci-dessous se trouve une instruction qui décrit une tâche. Écrivez une réponse qui complète correctement la demande.\n\n### Instruction:\n{instruction}\n\n### Réponse:\n",
+    "response_split": "### Réponse:"
+}

utils/README.md

@@ -0,0 +1,13 @@
+# Directory for helpers modules
+
+## prompter.py
+
+Prompter class, a template manager.
+
+`from utils.prompter import Prompter`
+
+## callbacks.py
+
+Helpers to support streaming generate output.
+
+`from utils.callbacks import Iteratorize, Stream`

utils/callbacks.py

@@ -0,0 +1,75 @@
+"""
+Helpers to support streaming generate output.
+Borrowed from https://github.com/oobabooga/text-generation-webui/blob/ad37f396fc8bcbab90e11ecf17c56c97bfbd4a9c/modules/callbacks.py
+"""
+
+import gc
+import traceback
+from queue import Queue
+from threading import Thread
+
+import torch
+import transformers
+
+
+class Stream(transformers.StoppingCriteria):
+    def __init__(self, callback_func=None):
+        self.callback_func = callback_func
+
+    def __call__(self, input_ids, scores) -> bool:
+        if self.callback_func is not None:
+            self.callback_func(input_ids[0])
+        return False
+
+
+class Iteratorize:
+
+    """
+    Transforms a function that takes a callback
+    into a lazy iterator (generator).
+    """
+
+    def __init__(self, func, kwargs={}, callback=None):
+        self.mfunc = func
+        self.c_callback = callback
+        self.q = Queue()
+        self.sentinel = object()
+        self.kwargs = kwargs
+        self.stop_now = False
+
+        def _callback(val):
+            if self.stop_now:
+                raise ValueError
+            self.q.put(val)
+
+        def gentask():
+            try:
+                ret = self.mfunc(callback=_callback, **self.kwargs)
+            except ValueError:
+                pass
+            except:
+                traceback.print_exc()
+                pass
+
+            self.q.put(self.sentinel)
+            if self.c_callback:
+                self.c_callback(ret)
+
+        self.thread = Thread(target=gentask)
+        self.thread.start()
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        obj = self.q.get(True, None)
+        if obj is self.sentinel:
+            raise StopIteration
+        else:
+            return obj
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.stop_now = True

utils/prompter.py

@@ -0,0 +1,51 @@
+"""
+A dedicated helper to manage templates and prompt building.
+"""
+
+import json
+import os.path as osp
+from typing import Union
+
+
+class Prompter(object):
+    __slots__ = ("template", "_verbose")
+
+    def __init__(self, template_name: str = "", verbose: bool = False):
+        self._verbose = verbose
+        if not template_name:
+            # Enforce the default here, so the constructor can be called with '' and will not break.
+            template_name = "alpaca"
+        file_name = osp.join("templates", f"{template_name}.json")
+        if not osp.exists(file_name):
+            raise ValueError(f"Can't read {file_name}")
+        with open(file_name) as fp:
+            self.template = json.load(fp)
+        if self._verbose:
+            print(
+                f"Using prompt template {template_name}: {self.template['description']}"
+            )
+
+    def generate_prompt(
+        self,
+        instruction: str,
+        input: Union[None, str] = None,
+        label: Union[None, str] = None,
+    ) -> str:
+        # returns the full prompt from instruction and optional input
+        # if a label (=response, =output) is provided, it's also appended.
+        if input:
+            res = self.template["prompt_input"].format(
+                instruction=instruction, input=input
+            )
+        else:
+            res = self.template["prompt_no_input"].format(
+                instruction=instruction
+            )
+        if label:
+            res = f"{res}{label}"
+        if self._verbose:
+            print(res)
+        return res
+
+    def get_response(self, output: str) -> str:
+        return output.split(self.template["response_split"])[1].strip()